Method: Direct token optimization (DTO) identifies target tokens for unlearning and non-target tokens for utility preservation, then optimizes token-level objectives directly without external resources.

Result: DTO achieves up to 16.8× improvement in forget quality on benchmark datasets compared to latest baselines while maintaining comparable model utility.

Conclusion: DTO provides an effective, self-contained solution for machine unlearning in LLMs that eliminates dependency on external resources and associated privacy risks.

Abstract: Machine unlearning is an emerging technique that removes the influence of a subset of training data (forget set) from a model without full retraining, with applications including privacy protection, content moderation, and model correction. The key challenge lies in ensuring that the model completely forgets the knowledge of the forget set without compromising its overall utility. Existing unlearning methods for large language models (LLMs) often utilize auxiliary language models, retain datasets, or even commercial AI services for effective unlearning and maintaining the model utility. However, dependence on these external resources is often impractical and could potentially introduce additional privacy risks. In this work, we propose direct token optimization (DTO), a novel self-contained unlearning approach for LLMs that directly optimizes the token level objectives and eliminates the need for external resources. Given a sequence to unlearn, we identify two categories of tokens: target tokens, which capture critical knowledge for unlearning, and the remaining non-target tokens, which are crucial for maintaining the model utility. The former are used to optimize the unlearning objective, while the latter serve to preserve the model’s performance. The experimental results show that the proposed DTO achieves up to 16.8$\times$ improvement in forget quality on several benchmark datasets than the latest baselines while maintaining a comparable level of model utility.

Method: Proposed TAMA framework enables interleaved multimodal reasoning using multimedia-returning tools without requiring training, featuring agentic flexible tool selection.

Result: Experimental results on ProMQA-Assembly dataset show improved performance for vision-language models (GPT-5 and MiMo-VL), with ablation studies confirming effectiveness of multimedia-returning tools and flexible tool selection.

Conclusion: The framework facilitates ’thinking with images’ paradigm for video and multimodal tasks and advances development of procedural activity assistants.

Abstract: Procedural activity assistants potentially support humans in a variety of settings, from our daily lives, e.g., cooking or assembling flat-pack furniture, to professional situations, e.g., manufacturing or biological experiments. Despite its potential use cases, the system development tailored for such an assistant is still underexplored. In this paper, we propose a novel framework, called TAMA, a Tool-Augmented Multimodal Agent, for procedural activity understanding. TAMA enables interleaved multimodal reasoning by making use of multimedia-returning tools in a training-free setting. Our experimental result on the multimodal procedural QA dataset, ProMQA-Assembly, shows that our approach can improve the performance of vision-language models, especially GPT-5 and MiMo-VL. Furthermore, our ablation studies provide empirical support for the effectiveness of two features that characterize our framework, multimedia-returning tools and agentic flexible tool selection. We believe our proposed framework and experimental results facilitate the thinking with images paradigm for video and multimodal tasks, let alone the development of procedural activity assistants.

Method: Created 15 deep research tasks across 10 domains using a synthesis pipeline with human verification. Tasks are grounded in realistic personas and enterprise context, spanning productivity software, cloud files, emails, chats, and web.

Result: Evaluated diverse DR agents across open- and closed-source models (GPT, Llama, Qwen), revealing their strengths, weaknesses, and critical advancement paths for enterprise deep research.

Conclusion: DRBench provides an effective benchmark for assessing AI agents on complex enterprise research tasks, highlighting the need for improved multi-source integration and reasoning capabilities.

Abstract: We introduce DRBench, a benchmark for evaluating AI agents on complex, open-ended deep research tasks in enterprise settings. Unlike prior benchmarks that focus on simple questions or web-only queries, DRBench evaluates agents on multi-step queries (for example, ``What changes should we make to our product roadmap to ensure compliance with this standard?") that require identifying supporting facts from both the public web and private company knowledge base. Each task is grounded in realistic user personas and enterprise context, spanning a heterogeneous search space that includes productivity software, cloud file systems, emails, chat conversations, and the open web. Tasks are generated through a carefully designed synthesis pipeline with human-in-the-loop verification, and agents are evaluated on their ability to recall relevant insights, maintain factual accuracy, and produce coherent, well-structured reports. We release 15 deep research tasks across 10 domains, such as Sales, Cybersecurity, and Compliance. We demonstrate the effectiveness of DRBench by evaluating diverse DR agents across open- and closed-source models (such as GPT, Llama, and Qwen) and DR strategies, highlighting their strengths, weaknesses, and the critical path for advancing enterprise deep research. Code is available at https://github.com/ServiceNow/drbench.

Method: Developed the first open-source RAG pipeline for ELMs with baselines and ablation studies for natural language generation, tested on three public datasets.

Result: ELMs with RAG consistently improved performance over non-RAG baselines and highlighted key ELM design considerations.

Conclusion: The presented RAG pipeline successfully enhances ELM performance and provides important design insights for ECG-language models.

Abstract: Interest in generative Electrocardiogram-Language Models (ELMs) is growing, as they can produce textual responses conditioned on ECG signals and textual queries. Unlike traditional classifiers that output label probabilities, ELMs are more versatile, supporting domain-specific tasks (e.g., waveform analysis, diagnosis, prognosis) as well as general tasks (e.g., open-ended questions, dialogue). Retrieval-Augmented Generation (RAG), widely used in Large Language Models (LLMs) to ground LLM outputs in retrieved knowledge, helps reduce hallucinations and improve natural language generation (NLG). However, despite its promise, no open-source implementation or systematic study of RAG pipeline design for ELMs currently exists. To address this gap, we present the first open-source RAG pipeline for ELMs, along with baselines and ablation studies for NLG. Experiments on three public datasets show that ELMs with RAG consistently improves performance over non-RAG baselines and highlights key ELM design considerations. Our code is available at: https://github.com/willxxy/ECG-Bench.

Method: Developed PrimeX dataset containing public opinion survey data with belief explanations and Primal World Belief survey assessments, then analyzed how this belief information personalizes language models.

Result: Showed that belief explanations and worldview data provide valuable information for personalizing language models and improving opinion prediction.

Conclusion: PrimeX dataset opens new research avenues for both NLP and psychology communities by demonstrating the value of incorporating belief systems into language model personalization.

Abstract: As the adoption of language models advances, so does the need to better represent individual users to the model. Are there aspects of an individual’s belief system that a language model can utilize for improved alignment? Following prior research, we investigate this question in the domain of opinion prediction by developing PrimeX, a dataset of public opinion survey data from 858 US residents with two additional sources of belief information: written explanations from the respondents for why they hold specific opinions, and the Primal World Belief survey for assessing respondent worldview. We provide an extensive initial analysis of our data and show the value of belief explanations and worldview for personalizing language models. Our results demonstrate how the additional belief information in PrimeX can benefit both the NLP and psychological research communities, opening up avenues for further study.

Method: PREFDISCO transforms static benchmarks into interactive personalization tasks using psychologically-grounded personas with sparse preferences, creating scenarios where identical questions require different reasoning chains depending on user context.

Result: Evaluation of 21 frontier models across 10 tasks shows 29.0% of naive personalization attempts produce worse preference alignment than generic responses, while generic responses also fail to serve individual user needs effectively.

Conclusion: Personalized reasoning requires dedicated development rather than emerging naturally in current LLMs, establishing it as a measurable research frontier with implications for education, healthcare, and technical domains where personalization is critical.

Abstract: Current large language model (LLM) development treats task-solving and preference alignment as separate challenges, optimizing first for objective correctness, then for alignment to aggregated human preferences. This paradigm fails in human-facing applications where solving a problem correctly is insufficient if the response mismatches the user’s needs. This challenge intensifies in just-in-time scenarios where no prior user interaction history exists due to cold-start conditions or privacy constraints. LLMs need to identify what they don’t know about user preferences, strategically elicit preference values through questioning, then adapt their reasoning processes and responses accordingly – a complicated chain of cognitive processes which we term personalized reasoning. We introduce PREFDISCO, an evaluation methodology that transforms static benchmarks into interactive personalization tasks using psychologically-grounded personas with sparse preferences. Our framework creates scenarios where identical questions require different reasoning chains depending on user context, as optimal explanation approaches vary by individual expertise and preferences while maintaining factual accuracy. Evaluation of 21 frontier models across 10 tasks reveals 29.0% of naive personalization attempts produce worse preference alignment than generic responses, yet generic responses also fail to serve individual user needs effectively. These findings suggest personalized reasoning requires dedicated development rather than emerging naturally. PREFDISCO establishes personalized reasoning as a measurable research frontier and reveals fundamental limitations in current LLMs’ interactive capabilities, providing a foundation for developing systems that can adapt to individual users in education, healthcare, and technical domains where personalization is critical.

Method: Created BiasFreeBench benchmark with 8 mainstream bias mitigation techniques (4 prompting-based, 4 training-based) tested on multi-choice QA and open-ended multi-turn QA scenarios. Reorganized existing datasets into unified query-response format and introduced Bias-Free Score metric.

Result: Systematically compared debiasing performances across key dimensions: prompting vs. training paradigm, model size, and generalization of training strategies to unseen bias types. Established comprehensive evaluation framework.

Conclusion: BiasFreeBench provides a unified testbed for bias mitigation research, enabling consistent evaluation across methods and bridging the gap between technical metrics and real-world fairness requirements.

Abstract: Existing studies on bias mitigation methods for large language models (LLMs) use diverse baselines and metrics to evaluate debiasing performance, leading to inconsistent comparisons among them. Moreover, their evaluations are mostly based on the comparison between LLMs’ probabilities of biased and unbiased contexts, which ignores the gap between such evaluations and real-world use cases where users interact with LLMs by reading model responses and expect fair and safe outputs rather than LLMs’ probabilities. To enable consistent evaluation across debiasing methods and bridge this gap, we introduce BiasFreeBench, an empirical benchmark that comprehensively compares eight mainstream bias mitigation techniques (covering four prompting-based and four training-based methods) on two test scenarios (multi-choice QA and open-ended multi-turn QA) by reorganizing existing datasets into a unified query-response setting. We further introduce a response-level metric, Bias-Free Score, to measure the extent to which LLM responses are fair, safe, and anti-stereotypical. Debiasing performances are systematically compared and analyzed across key dimensions: the prompting vs. training paradigm, model size, and generalization of different training strategies to unseen bias types. We will publicly release our benchmark, aiming to establish a unified testbed for bias mitigation research.

Method: Uses Large Reasoning Models (LRMs) with structured prompting templates for systematic step-by-step evaluation of translation quality, tested with varying reasoning efforts on OpenAI’s o3 model.

Result: Achieved highest soft pairwise accuracy in system-level evaluation for both reference-based and reference-free settings, with reference-free variant ranking as top-performing metric among all reference-free approaches.

Conclusion: Large Reasoning Models represent a measurable advancement in translation quality assessment, combining improved accuracy with transparent evaluation across diverse language pairs.

Abstract: We introduce TASER (Translation Assessment via Systematic Evaluation and Reasoning), a metric that uses Large Reasoning Models (LRMs) for automated translation quality assessment. TASER harnesses the explicit reasoning capabilities of LRMs to conduct systematic, step-by-step evaluation of translation quality. We evaluate TASER on the WMT24 Metrics Shared Task across both reference-based and reference-free scenarios, demonstrating state-of-the-art performance. In system-level evaluation, TASER achieves the highest soft pairwise accuracy in both reference-based and reference-free settings, outperforming all existing metrics. At the segment level, TASER maintains competitive performance with our reference-free variant ranking as the top-performing metric among all reference-free approaches. Our experiments reveal that structured prompting templates yield superior results with LRMs compared to the open-ended approaches that proved optimal for traditional LLMs. We evaluate o3, a large reasoning model from OpenAI, with varying reasoning efforts, providing insights into the relationship between reasoning depth and evaluation quality. The explicit reasoning process in LRMs offers interpretability and visibility, addressing a key limitation of existing automated metrics. Our results demonstrate that Large Reasoning Models show a measurable advancement in translation quality assessment, combining improved accuracy with transparent evaluation across diverse language pairs.

Method: Two learning methods: 1) supervised fine-tuning for dense probabilistic labels, and 2) reinforcement learning for sparse binary labels, both using distribution-matching objectives.

Result: Fine-tuned autoraters show better alignment with target preference distributions, improved calibration, significantly lower positional bias, while maintaining performance on objective tasks.

Conclusion: Modeling full preference distributions rather than discrete labels enables more reliable probabilistic autoraters for LLM alignment.

Abstract: The alignment of large language models (LLMs) with human values increasingly relies on using other LLMs as automated judges, or ``autoraters’’. However, their reliability is limited by a foundational issue: they are trained on discrete preference labels, forcing a single ground truth onto tasks that are often subjective, ambiguous, or nuanced. We argue that a reliable autorater must learn to model the full distribution of preferences defined by a target population. In this paper, we propose a general framework for calibrating probabilistic autoraters to any given preference distribution. We formalize the problem and present two learning methods tailored to different data conditions:

1. a direct supervised fine-tuning for dense, probabilistic labels, and 2) a reinforcement learning approach for sparse, binary labels. Our empirical results show that finetuning autoraters with a distribution-matching objective leads to verbalized probability predictions that are better aligned with the target preference distribution, with improved calibration and significantly lower positional bias, all while preserving performance on objective tasks.

Method: A plug-and-play layer-wise PEFT framework that fine-tunes only important LLM layers selected based on gradient norm distribution while freezing redundant layers.

Result: IR-Tuning outperforms several layer-wise PEFT baselines across diverse text revisions, with fast convergence, low GPU memory consumption, and effectiveness on small corpora.

Conclusion: The proposed IR-Tuning framework successfully addresses the challenge of fine-tuning LLMs for nuanced text classification tasks with limited data, providing an efficient and effective solution.

Abstract: Large Language Models (LLMs) have shown extraordinary success across various text generation tasks; however, their potential for simple yet essential text classification remains underexplored, as LLM pre-training tends to emphasize generation over classification. While LLMs with instruction tuning can transform classification into a generation task, they often struggle to categorize nuanced texts. One such example is text revision, which involves nuanced edits between pairs of texts. Although simply fine-tuning LLMs for revision classification seems plausible, it requires a large amount of revision annotations, which are exceptionally expensive and scarce in the community. To address this issue, we introduce a plug-and-play layer-wise parameter-efficient fine-tuning (PEFT) framework, i.e., IR-Tuning, which fine-tunes a subset of important LLM layers that are dynamically selected based on their gradient norm distribution, while freezing those of redundant layers. Extensive experiments suggest that IR-Tuning surpasses several layer-wise PEFT baselines over diverse text revisions, while achieving fast convergence, low GPU memory consumption, and effectiveness on small revision corpora.

Method: Three-step pipeline: (1) LLM extractor generates structured entities and their contexts, (2) string-based global aligner, and (3) scoring model to verify consistency between extracted information and grounded context.

Result: Reduces hallucinations by up to 85% on IE tasks with minimal false positives. Fine-tuned transformer encoder outperforms LLM scoring model. High agreement with human judgments enables dual use for LLM evaluation.

Conclusion: SafePassage effectively mitigates LLM hallucinations in information extraction through grounded context verification, and surprisingly, task-specific fine-tuned models can outperform LLMs for safety scoring with minimal annotation effort.

Abstract: Black box large language models (LLMs) make information extraction (IE) easy to configure, but hard to trust. Unlike traditional information extraction pipelines, the information “extracted” is not guaranteed to be grounded in the document. To prevent this, this paper introduces the notion of a “safe passage”: context generated by the LLM that is both grounded in the document and consistent with the extracted information. This is operationalized via a three-step pipeline, SafePassage, which consists of: (1) an LLM extractor that generates structured entities and their contexts from a document, (2) a string-based global aligner, and (3) a scoring model. Results show that using these three parts in conjunction reduces hallucinations by up to 85% on information extraction tasks with minimal risk of flagging non-hallucinations. High agreement between the SafePassage pipeline and human judgments of extraction quality mean that the pipeline can be dually used to evaluate LLMs. Surprisingly, results also show that using a transformer encoder fine-tuned on a small number of task-specific examples can outperform an LLM scoring model at flagging unsafe passages. These annotations can be collected in as little as 1-2 hours.

Method: Proposed a clinically grounded Meta-Evaluation framework with criteria spanning clinical alignment, discrimination, robustness, and monotonicity. Used a fine-grained dataset with annotated report pairs containing error types, clinical significance labels, and explanations.

Result: Systematic evaluation revealed limitations in existing metrics: failure to distinguish clinically significant errors, over-penalizing harmless variations, and lacking consistency across error severity levels.

Conclusion: The framework provides guidance for developing more clinically reliable evaluation methods that better align with clinical needs and improve trust in automated radiology reporting systems.

Abstract: Automatically generated radiology reports often receive high scores from existing evaluation metrics but fail to earn clinicians’ trust. This gap reveals fundamental flaws in how current metrics assess the quality of generated reports. We rethink the design and evaluation of these metrics and propose a clinically grounded Meta-Evaluation framework. We define clinically grounded criteria spanning clinical alignment and key metric capabilities, including discrimination, robustness, and monotonicity. Using a fine-grained dataset of ground truth and rewritten report pairs annotated with error types, clinical significance labels, and explanations, we systematically evaluate existing metrics and reveal their limitations in interpreting clinical semantics, such as failing to distinguish clinically significant errors, over-penalizing harmless variations, and lacking consistency across error severity levels. Our framework offers guidance for building more clinically reliable evaluation methods.

Method: The paper presents o-mega, a hyperparameter optimization tool that systematically explores different explainable AI methods and their configurations within the semantic matching domain, evaluated on a post-claim matching pipeline using social media posts paired with refuting claims.

Result: The tool demonstrates improved transparency in automated fact-checking systems by automatically identifying optimal explainability approaches and configurations.

Conclusion: Automated optimization of explanation methods can significantly enhance interpretability of claim-matching models in critical applications like misinformation detection, contributing to more trustworthy and transparent AI systems.

Abstract: The proliferation of transformer-based language models has revolutionized NLP domain while simultaneously introduced significant challenges regarding model transparency and trustworthiness. The complexity of achieving explainable systems in this domain is evidenced by the extensive array of explanation methods and evaluation metrics developed by researchers. To address the challenge of selecting optimal explainability approaches, we present \textbf{\texttt{o-mega}}, a hyperparameter optimization tool designed to automatically identify the most effective explainable AI methods and their configurations within the semantic matching domain. We evaluate o-mega on a post-claim matching pipeline using a curated dataset of social media posts paired with refuting claims. Our tool systematically explores different explainable methods and their hyperparameters, demonstrating improved transparency in automated fact-checking systems. As a result, such automated optimization of explanation methods can significantly enhance the interpretability of claim-matching models in critical applications such as misinformation detection, contributing to more trustworthy and transparent AI systems.

Method: Multi-agent LLM architecture with specialized agents: behavior-analysis agent inspects activity sequences, evidence-gathering agents query external systems, and reasoning agent synthesizes findings into auditable decisions.

Result: CORTEX substantially reduces false positives and improves investigation quality over state-of-the-art single-agent LLMs across diverse enterprise scenarios. A dataset of fine-grained SOC investigations from production environments is released.

Conclusion: The multi-agent approach provides better performance, transparency, and auditability for high-stakes alert triage compared to single-agent LLM systems.

Abstract: Security Operations Centers (SOCs) are overwhelmed by tens of thousands of daily alerts, with only a small fraction corresponding to genuine attacks. This overload creates alert fatigue, leading to overlooked threats and analyst burnout. Classical detection pipelines are brittle and context-poor, while recent LLM-based approaches typically rely on a single model to interpret logs, retrieve context, and adjudicate alerts end-to-end – an approach that struggles with noisy enterprise data and offers limited transparency. We propose CORTEX, a multi-agent LLM architecture for high-stakes alert triage in which specialized agents collaborate over real evidence: a behavior-analysis agent inspects activity sequences, evidence-gathering agents query external systems, and a reasoning agent synthesizes findings into an auditable decision. To support training and evaluation, we release a dataset of fine-grained SOC investigations from production environments, capturing step-by-step analyst actions and linked tool outputs. Across diverse enterprise scenarios, CORTEX substantially reduces false positives and improves investigation quality over state-of-the-art single-agent LLMs.

Method: TokMem stores recurring procedures as trainable embeddings where each memory token encodes both an address to a procedure and a control signal that steers generation, with backbone model kept frozen for continual adaptation.

Result: TokMem consistently outperforms retrieval-augmented generation on 1,000 atomic recall tasks and compositional function-calling tasks while avoiding repeated context overhead and using far fewer parameters than fine-tuning.

Conclusion: TokMem provides a scalable and modular alternative to prompt engineering and fine-tuning by offering explicit procedural memory for LLMs.

Abstract: Large language models rely heavily on prompts to specify tasks, recall knowledge and guide reasoning. However, this reliance is inefficient as prompts must be re-read at each step, scale poorly across tasks, and lack mechanisms for modular reuse. We introduce TokMem, a tokenized procedural memory that stores recurring procedures as compact, trainable embeddings. Each memory token encodes both an address to a procedure and a control signal that steers generation, enabling targeted behavior with constant-size overhead. To support continual adaptation, TokMem keeps the backbone model frozen, allowing new procedures to be added without interfering with existing ones. We evaluate TokMem on 1,000 tasks for atomic recall, and on function-calling tasks for compositional recall, where it consistently outperforms retrieval-augmented generation while avoiding repeated context overhead, and fine-tuning with far fewer parameters. These results establish TokMem as a scalable and modular alternative to prompt engineering and fine-tuning, offering an explicit procedural memory for LLMs.

Method: Dual-stage strategy: (1) coarse-grained compression identifies and ranks function-level chunks using conditional perplexity, (2) fine-grained compression segments retained functions into blocks and selects optimal subset under adaptive token budget.

Result: Achieves up to 5.6x compression ratio without degrading task performance across code completion, summarization, and question answering tasks.

Conclusion: LongCodeZip enables LLMs to better scale to real-world, large-scale code scenarios by effectively reducing context size while preserving essential information.

Abstract: Code generation under long contexts is becoming increasingly critical as Large Language Models (LLMs) are required to reason over extensive information in the codebase. While recent advances enable code LLMs to process long inputs, high API costs and generation latency remain substantial bottlenecks. Existing context pruning techniques, such as LLMLingua, achieve promising results for general text but overlook code-specific structures and dependencies, leading to suboptimal performance in programming tasks. In this paper, we propose LongCodeZip, a novel plug-and-play code compression framework designed specifically for code LLMs. LongCodeZip employs a dual-stage strategy: (1) coarse-grained compression, which identifies and ranks function-level chunks using conditional perplexity with respect to the instruction, retaining only the most relevant functions; and (2) fine-grained compression, which segments retained functions into blocks based on perplexity and selects an optimal subset under an adaptive token budget to maximize relevance. Evaluations across multiple tasks, including code completion, summarization, and question answering, show that LongCodeZip consistently outperforms baseline methods, achieving up to a 5.6x compression ratio without degrading task performance. By effectively reducing context size while preserving essential information, LongCodeZip enables LLMs to better scale to real-world, large-scale code scenarios, advancing the efficiency and capability of code intelligence applications.

Method: Comprehensive experiments with eight off-the-shelf LLM models across three datasets, testing different in-context information including user-written reviews and general preference descriptions. Also tested prompting LLMs to generate hypothetical reviews first.

Result: User-written reviews significantly improve LLM rating prediction performance, comparable to traditional methods like matrix factorization. Reviews for concrete items are more effective than general preference descriptions. Prompting LLMs to generate hypothetical reviews first further enhances performance.

Conclusion: LLMs show promise as a solution for the cold-start problem in rating prediction. Concrete user reviews are more valuable than abstract preference descriptions, and generating hypothetical reviews first can boost prediction accuracy.

Abstract: Personalizing the outputs of large language models (LLMs) to align with individual user preferences is an active research area. However, previous studies have mainly focused on classification or ranking tasks and have not considered Likert-scale rating prediction, a regression task that requires both language and mathematical reasoning to be solved effectively. This task has significant industrial applications, but the utilization of LLMs remains underexplored, particularly regarding the capabilities of off-the-shelf LLMs. This study investigates the performance of off-the-shelf LLMs on rating prediction, providing different in-context information. Through comprehensive experiments with eight models across three datasets, we demonstrate that user-written reviews significantly improve the rating prediction performance of LLMs. This result is comparable to traditional methods like matrix factorization, highlighting the potential of LLMs as a promising solution for the cold-start problem. We also find that the reviews for concrete items are more effective than general preference descriptions that are not based on any specific item. Furthermore, we discover that prompting LLMs to first generate a hypothetical review enhances the rating prediction performance. Our code is available at https://github.com/ynklab/rating-prediction-with-reviews.

Method: Integrates learnable query-based architecture with multilingual awareness and large-scale pretraining on simulated code-switching data.

Result: Achieves state-of-the-art performance on several language diarization benchmarks with 23% to 52% relative performance improvement over previous methods.

Conclusion: This work advances language diarization research and establishes a foundational framework for code-switching speech technologies.

Abstract: In this paper, we present a neural spoken language diarization model that supports an unconstrained span of languages within a single framework. Our approach integrates a learnable query-based architecture grounded in multilingual awareness, with large-scale pretraining on simulated code-switching data. By jointly leveraging these two components, our method overcomes the limitations of conventional approaches in data scarcity and architecture optimization, and generalizes effectively to real-world multilingual settings across diverse environments. Experimental results demonstrate that our approach achieves state-of-the-art performance on several language diarization benchmarks, with a relative performance improvement of 23% to 52% over previous methods. We believe that this work not only advances research in language diarization but also establishes a foundational framework for code-switching speech technologies.

Method: Fine-tuned LLMs using JP1-specific datasets from domain manuals and distilled reasoning trajectories generated by LLMs. Used agentic prompts with retrieval-augmented generation and a context-answer extractor during inference.

Result: Achieved 14% performance improvement over base model on JP1 certification exam questions, enhancing decision making accuracy and search efficiency.

Conclusion: Agent fine-tuning shows significant potential for domain-specific reasoning in complex technical microdomains, enabling LLMs to internalize procedural knowledge and improve performance in specialized contexts.

Abstract: Agentic large language models (LLMs) have become prominent for autonomously interacting with external environments and performing multi-step reasoning tasks. Most approaches leverage these capabilities via in-context learning with few-shot prompts, but this often results in lengthy inputs and higher computational costs. Agent fine-tuning offers an alternative by enabling LLMs to internalize procedural reasoning and domain-specific knowledge through training on relevant data and demonstration trajectories. While prior studies have focused on general domains, their effectiveness in specialized technical microdomains remains unclear. This paper explores agent fine-tuning for domain adaptation within Hitachi’s JP1 middleware, a microdomain for specialized IT operations. We fine-tuned LLMs using JP1-specific datasets derived from domain manuals and distilled reasoning trajectories generated by LLMs themselves, enhancing decision making accuracy and search efficiency. During inference, we used an agentic prompt with retrieval-augmented generation and introduced a context-answer extractor to improve information relevance. On JP1 certification exam questions, our method achieved a 14% performance improvement over the base model, demonstrating the potential of agent fine-tuning for domain-specific reasoning in complex microdomains.

Method: Cascading domain-specific encoders with LLMs, testing backdoor attacks across four speech encoders and three datasets covering four tasks (ASR, emotion recognition, gender/age prediction).

Result: Attack consistently achieves high success rates from 90.76% to 99.41%. Component-wise analysis identifies most vulnerable pipeline stages.

Conclusion: Proposes fine-tuning-based defense to mitigate threat of poisoned pretrained encoders, highlighting the security risks in multimodal LLM architectures.

Abstract: Large Language Models (LLMs) and their multimodal extensions are becoming increasingly popular. One common approach to enable multimodality is to cascade domain-specific encoders with an LLM, making the resulting model inherit vulnerabilities from all of its components. In this work, we present the first systematic study of audio backdoor attacks against speech language models. We demonstrate its effectiveness across four speech encoders and three datasets, covering four tasks: automatic speech recognition (ASR), speech emotion recognition, and gender and age prediction. The attack consistently achieves high success rates, ranging from 90.76% to 99.41%. To better understand how backdoors propagate, we conduct a component-wise analysis to identify the most vulnerable stages of the pipeline. Finally, we propose a fine-tuning-based defense that mitigates the threat of poisoned pretrained encoders.

Method: Proposed Agent-ScanKit framework with three orthogonal probing paradigms (visual-guided, text-guided, structure-guided) to quantify memorization vs reasoning contributions without accessing model internals.

Result: Evaluation on 5 GUI benchmarks with 18 multimodal agents showed mechanical memorization often outweighs systematic reasoning, with models functioning mainly as retrievers of training-aligned knowledge with limited generalization.

Conclusion: Findings highlight the necessity for robust reasoning modeling in multimodal agents for real-world scenarios and provide insights for developing more reliable agents.

Abstract: Although numerous strategies have recently been proposed to enhance the autonomous interaction capabilities of multimodal agents in graphical user interface (GUI), their reliability remains limited when faced with complex or out-of-domain tasks. This raises a fundamental question: Are existing multimodal agents reasoning spuriously? In this paper, we propose \textbf{Agent-ScanKit}, a systematic probing framework to unravel the memory and reasoning capabilities of multimodal agents under controlled perturbations. Specifically, we introduce three orthogonal probing paradigms: visual-guided, text-guided, and structure-guided, each designed to quantify the contributions of memorization and reasoning without requiring access to model internals. In five publicly available GUI benchmarks involving 18 multimodal agents, the results demonstrate that mechanical memorization often outweighs systematic reasoning. Most of the models function predominantly as retrievers of training-aligned knowledge, exhibiting limited generalization. Our findings underscore the necessity of robust reasoning modeling for multimodal agents in real-world scenarios, offering valuable insights toward the development of reliable multimodal agents.

Method: Combines modality-based layer-splitting architecture with frozen pre-training strategy to preserve reasoning and knowledge of pretrained text LLMs while adding native speech capabilities.

Result: Achieves state-of-the-art results in spoken question answering and comparable speech-to-speech performance to text-guided systems while maintaining competitive text performance.

Conclusion: Establishes a new paradigm for expressive and efficient end-to-end speech interaction by narrowing the gap between text-guided and direct speech generation.

Abstract: Spoken dialogue systems often rely on cascaded pipelines that transcribe, process, and resynthesize speech. While effective, this design discards paralinguistic cues and limits expressivity. Recent end-to-end methods reduce latency and better preserve these cues, yet still rely on text intermediates, creating a fundamental bottleneck. We present MOSS-Speech, a true speech-to-speech large language model that directly understands and generates speech without relying on text guidance. Our approach combines a modality-based layer-splitting architecture with a frozen pre-training strategy, preserving the reasoning and knowledge of pretrained text LLMs while adding native speech capabilities. Experiments show that our model achieves state-of-the-art results in spoken question answering and delivers comparable speech-to-speech performance relative to existing text-guided systems, while still maintaining competitive text performance. By narrowing the gap between text-guided and direct speech generation, our work establishes a new paradigm for expressive and efficient end-to-end speech interaction.

Method: Uses knowledge graphs constructed from multi-source external data as task space, translating semantic relations into structured multimodal tasks via subgraph sampling, task templates, and meta-paths. Implements multi-stage filtering pipeline with node reachability, LLM scoring, and similarity analysis to ensure task quality and executability.

Result: Created Graph2Eval-Bench with 1,319 tasks spanning document comprehension and web interaction scenarios. Experiments show the framework efficiently generates tasks that differentiate agent and model performance, revealing gaps in reasoning, collaboration, and web interaction across different settings.

Conclusion: Graph2Eval offers a comprehensive evaluation framework for multimodal LLM-driven agents, enabling systematic assessment of reasoning, collaboration, and interactive capabilities in dynamic web environments, providing new perspective for agent evaluation.

Abstract: As multimodal LLM-driven agents continue to advance in autonomy and generalization, evaluation based on static datasets can no longer adequately assess their true capabilities in dynamic environments and diverse tasks. Existing LLM-based synthetic data methods are largely designed for LLM training and evaluation, and thus cannot be directly applied to agent tasks that require tool use and interactive capabilities. While recent studies have explored automatic agent task generation with LLMs, most efforts remain limited to text or image analysis, without systematically modeling multi-step interactions in web environments. To address these challenges, we propose Graph2Eval, a knowledge graph-based framework that automatically generates both multimodal document comprehension tasks and web interaction tasks, enabling comprehensive evaluation of agents’ reasoning, collaboration, and interactive capabilities. In our approach, knowledge graphs constructed from multi-source external data serve as the task space, where we translate semantic relations into structured multimodal tasks using subgraph sampling, task templates, and meta-paths. A multi-stage filtering pipeline based on node reachability, LLM scoring, and similarity analysis is applied to guarantee the quality and executability of the generated tasks. Furthermore, Graph2Eval supports end-to-end evaluation of multiple agent types (Single-Agent, Multi-Agent, Web Agent) and measures reasoning, collaboration, and interaction capabilities. We instantiate the framework with Graph2Eval-Bench, a curated dataset of 1,319 tasks spanning document comprehension and web interaction scenarios. Experiments show that Graph2Eval efficiently generates tasks that differentiate agent and model performance, revealing gaps in reasoning, collaboration, and web interaction across different settings and offering a new perspective for agent evaluation.

Method: Two-stage high-copying response preference training with three prompting methods to enhance copying degree, creating an automated pipeline that transforms generated responses into high-copying preference data.

Result: Achieves best performance on FaithEval, ConFiQA and PubMedQA with 12.2% to 24.5% accuracy improvements on FaithEval over best baseline, requiring only 365 training samples (1/50th of baseline data).

Conclusion: CopyPasteLLM recalibrates LLMs to rely more on internal parametric knowledge rather than external knowledge during generation, effectively reducing context-unfaithful hallucinations through high-copying response generation.

Abstract: While Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to generate contextually grounded responses, contextual faithfulness remains challenging as LLMs may not consistently trust provided context, leading to hallucinations that undermine reliability. We observe an inverse correlation between response copying degree and context-unfaithful hallucinations on RAGTruth, suggesting that higher copying degrees reduce hallucinations by fostering genuine contextual belief. We propose CopyPasteLLM, obtained through two-stage high-copying response preference training. We design three prompting methods to enhance copying degree, demonstrating that high-copying responses achieve superior contextual faithfulness and hallucination control. These approaches enable a fully automated pipeline that transforms generated responses into high-copying preference data for training CopyPasteLLM. On FaithEval, ConFiQA and PubMedQA, CopyPasteLLM achieves best performance in both counterfactual and original contexts, remarkably with 12.2% to 24.5% accuracy improvements on FaithEval over the best baseline, while requiring only 365 training samples – 1/50th of baseline data. To elucidate CopyPasteLLM’s effectiveness, we propose the Context-Parameter Copying Capturing algorithm. Interestingly, this reveals that CopyPasteLLM recalibrates reliance on internal parametric knowledge rather than external knowledge during generation. All codes are available at https://github.com/longyongchao/CopyPasteLLM

Method: Integrates three core components: collective multi-agent framework (planning/execution agents with critic voting), hierarchical memory system (working/semantic/procedural layers), and refined tool suite (search, code execution, multimodal parsing).

Result: Consistently outperforms open-source baselines and approaches performance of proprietary systems on comprehensive benchmark.

Conclusion: Demonstrates importance of system-level integration for scalable, resilient, and adaptive AI assistants capable of operating across diverse domains and tasks.

Abstract: Large Language Models are increasingly deployed as autonomous agents for complex real-world tasks, yet existing systems often focus on isolated improvements without a unifying design for robustness and adaptability. We propose a generalist agent architecture that integrates three core components: a collective multi-agent framework combining planning and execution agents with critic model voting, a hierarchical memory system spanning working, semantic, and procedural layers, and a refined tool suite for search, code execution, and multimodal parsing. Evaluated on a comprehensive benchmark, our framework consistently outperforms open-source baselines and approaches the performance of proprietary systems. These results demonstrate the importance of system-level integration and highlight a path toward scalable, resilient, and adaptive AI assistants capable of operating across diverse domains and tasks.

Method: A scalable pipeline with robust media retrieval and a two-stage alignment algorithm designed to handle non-verbatim transcripts and long-form audio from 22 European parliaments.

Result: Extracted over 61k hours of aligned speech segments with 19 languages exceeding 1k hours and 22 languages exceeding 500 hours. Achieved 41.8% average reduction in word error rates when finetuning ASR models.

Conclusion: The proposed pipeline effectively addresses data scarcity in multilingual speech processing by leveraging parliamentary recordings to create large-scale, high-quality speech datasets.

Abstract: Recent progress in speech processing has highlighted that high-quality performance across languages requires substantial training data for each individual language. While existing multilingual datasets cover many languages, they often contain insufficient data for most languages. Thus, trained models perform poorly on the majority of the supported languages. Our work addresses this challenge by introducing a scalable pipeline for constructing speech datasets from parliamentary recordings. The proposed pipeline includes robust components for media retrieval and a two-stage alignment algorithm designed to handle non-verbatim transcripts and long-form audio. Applying this pipeline to recordings from 22 European parliaments, we extract over 61k hours of aligned speech segments, achieving substantial per-language coverage with 19 languages exceeding 1k hours and 22 languages exceeding 500 hours of high-quality speech data. We obtain an average 41.8% reduction in word error rates over baselines when finetuning an existing ASR model on our dataset, demonstrating the usefulness of our approach.

Method: Study a family of probability-based objectives and characterize their effectiveness across different conditions, conducting comprehensive experiments across 7 model backbones, 14 benchmarks, and 3 domains.

Result: Found that near the model-strong end, prior-leaning objectives (e.g., $-p$, $-p^{10}$, thresholded variants) consistently outperform NLL; toward the model-weak end, NLL dominates; in between, no single objective prevails.

Conclusion: Objective effectiveness depends on the model-capability continuum, providing a principled foundation for adapting objectives to model capability rather than using NLL by default.

Abstract: Supervised fine-tuning (SFT) is the standard approach for post-training large language models (LLMs), yet it often shows limited generalization. We trace this limitation to its default training objective: negative log likelihood (NLL). While NLL is classically optimal when training from scratch, post-training operates in a different paradigm and could violate its optimality assumptions, where models already encode task-relevant priors and supervision can be long and noisy. To this end, we study a general family of probability-based objectives and characterize their effectiveness under different conditions. Through comprehensive experiments and extensive ablation studies across 7 model backbones, 14 benchmarks, and 3 domains, we uncover a critical dimension that governs objective behavior: the model-capability continuum. Near the model-strong end, prior-leaning objectives that downweight low-probability tokens (e.g., $-p$, $-p^{10}$, thresholded variants) consistently outperform NLL; toward the model-weak end, NLL dominates; in between, no single objective prevails. Our theoretical analysis further elucidates how objectives trade places across the continuum, providing a principled foundation for adapting objectives to model capability. Our code is available at https://github.com/GaotangLi/Beyond-Log-Likelihood.

Method: GUI-KV combines two techniques: spatial saliency guidance (augmenting attention scores with hidden state L2 norms) and temporal redundancy scoring (projecting previous frames’ keys onto current frame’s key subspace to prune redundant history). Uses uniform budget allocation based on analysis showing uniform attention sparsity across transformer layers.

Result: Outperforms competitive KV compression baselines, closely matching full-cache accuracy at modest budgets. In 5-screenshot setting on AgentNetBench, reduces decoding FLOPs by 38.9% while increasing step accuracy by 4.1% over full-cache baseline.

Conclusion: Exploiting GUI-specific redundancies enables efficient and reliable agent performance without requiring retraining, demonstrating that simple uniform budget allocation combined with spatial-temporal awareness effectively addresses GUI agent efficiency challenges.

Abstract: Graphical user interface (GUI) agents built on vision-language models have emerged as a promising approach to automate human-computer workflows. However, they also face the inefficiency challenge as they process long sequences of high-resolution screenshots and solving long-horizon tasks, making inference slow, costly and memory-bound. While key-value (KV) caching can mitigate this, storing the full cache is prohibitive for image-heavy contexts. Existing cache-compression methods are sub-optimal as they do not account for the spatial and temporal redundancy of GUIs. In this work, we first analyze attention patterns in GUI agent workloads and find that, unlike in natural images, attention sparsity is uniformly high across all transformer layers. This insight motivates a simple uniform budget allocation strategy, which we show empirically outperforms more complex layer-varying schemes. Building on this, we introduce GUI-KV, a plug-and-play KV cache compression method for GUI agents that requires no retraining. GUI-KV combines two novel techniques: (i) spatial saliency guidance, which augments attention scores with the L2 norm of hidden states to better preserve semantically important visual tokens, and (ii) temporal redundancy scoring, which projects previous frames’ keys onto the current frame’s key subspace to preferentially prune redundant history. Across standard GUI agent benchmarks and models, GUI-KV outperforms competitive KV compression baselines, closely matching full-cache accuracy at modest budgets. Notably, in a 5-screenshot setting on the AgentNetBench benchmark, GUI-KV reduces decoding FLOPs by 38.9% while increasing step accuracy by 4.1% over the full-cache baseline. These results demonstrate that exploiting GUI-specific redundancies enables efficient and reliable agent performance.

Method: Diagnosed overthinking via oracle rollouts that inject at sentence boundaries. Introduced ThinkBrake, a training-free decoding heuristic that monitors the log-probability margin between and the current top token at sentence boundaries, triggering termination when this margin becomes small.

Result: Oracle termination lifted average accuracy from 85.8% to 94.2% while reducing tokens by 80-94%. ThinkBrake preserved or improved accuracy while reducing tokens up to 25% across BFCL’s single turn, non-live and live splits, outperforming various baselines.

Conclusion: ThinkBrake effectively addresses overthinking in small reasoning models during tool use, demonstrating that early termination based on log-probability monitoring can significantly improve efficiency while maintaining or improving accuracy.

Abstract: Small reasoning models (SRMs) often overthink during tool use: they reach a correct tool-argument configuration, then continue reasoning and overwrite it with an incorrect final call. We diagnose overthinking via oracle rollouts that inject at sentence boundaries. On the Berkeley Function Calling Leaderboard (BFCL), this oracle termination lifts average accuracy from 85.8% to 94.2% while reducing tokens by 80-94%, revealing substantial recoverable headroom and potential redundant reasoning. While prior work on concise reasoning has largely targeted mathematics, tool reasoning remains underexplored. We adapt various early-termination baselines to tool use and introduce ThinkBrake, a training-free decoding heuristic. ThinkBrake monitors the log-probability margin between and the current top token at sentence boundaries and triggers termination when this margin becomes small. Across BFCL’s single turn, non-live and live splits, ThinkBrake preserves or improves accuracy while reducing tokens up to 25%, outperforming various baselines.

Method: Created CHIME dataset with popular Chinese phrase-based memes annotated with meanings, origins, examples, and types. Designed two evaluation tasks: 1) meme explanation including meaning, origin, and example generation; 2) multiple-choice questions for selecting appropriate memes in contextual sentences.

Result: LLMs can explain meanings of some memes but performance significantly declines for culturally nuanced types. Models consistently struggle to provide accurate origins. In multiple-choice tasks, models perform below human levels despite providing some correct answers.

Conclusion: Current LLMs have limited understanding of culturally specific internet memes, particularly struggling with origins and nuanced cultural aspects. The CHIME dataset is made public to facilitate future research on computational meme understanding.

Abstract: Large language models (LLMs) are trained on vast amounts of text from the Internet, but do they truly understand the viral content that rapidly spreads online – commonly known as memes? In this paper, we introduce CHIME, a dataset for CHinese Internet Meme Explanation. The dataset comprises popular phrase-based memes from the Chinese Internet, annotated with detailed information on their meaning, origin, example sentences, types, etc. To evaluate whether LLMs understand these memes, we designed two tasks. In the first task, we assessed the models’ ability to explain a given meme, identify its origin, and generate appropriate example sentences. The results show that while LLMs can explain the meanings of some memes, their performance declines significantly for culturally and linguistically nuanced meme types. Additionally, they consistently struggle to provide accurate origins for the memes. In the second task, we created a set of multiple-choice questions (MCQs) requiring LLMs to select the most appropriate meme to fill in a blank within a contextual sentence. While the evaluated models were able to provide correct answers, their performance remains noticeably below human levels. We have made CHIME public and hope it will facilitate future research on computational meme understanding.

Method: Introduces a self-correction mechanism with JUDGE action for dynamic error identification and recovery, plus a dense process reward function decomposing into correctness and utility rewards.

Result: Agents trained with ReSeek significantly outperform state-of-the-art baselines in task success rate and path faithfulness on the new FictionalHot benchmark.

Conclusion: ReSeek provides an effective framework for training search agents with self-correction capabilities, addressing limitations of previous RL approaches.

Abstract: Search agents powered by Large Language Models (LLMs) have demonstrated significant potential in tackling knowledge-intensive tasks. Reinforcement learning (RL) has emerged as a powerful paradigm for training these agents to perform complex, multi-step reasoning. However, prior RL-based methods often rely on sparse or rule-based rewards, which can lead agents to commit to suboptimal or erroneous reasoning paths without the ability to recover. To address these limitations, we propose ReSeek, a novel self-correcting framework for training search agents. Our framework introduces a self-correction mechanism that empowers the agent to dynamically identify and recover from erroneous search paths during an episode. By invoking a special JUDGE action, the agent can judge the information and re-plan its search strategy. To guide this process, we design a dense, instructive process reward function, which decomposes into a correctness reward for retrieving factual information and a utility reward for finding information genuinely useful for the query. Furthermore, to mitigate the risk of data contamination in existing datasets, we introduce FictionalHot, a new and challenging benchmark with recently curated questions requiring complex reasoning. Being intuitively reasonable and practically simple, extensive experiments show that agents trained with ReSeek significantly outperform SOTA baselines in task success rate and path faithfulness.

Method: Proposed CoT Vectors - compact representations encoding task-general reasoning knowledge, optimized under teacher-student framework to address layer-wise instability observed in extracted versions.

Result: CoT Vectors outperform existing baselines, achieve performance comparable to parameter-efficient fine-tuning methods with fewer trainable parameters, and reveal insights about LLM reasoning organization through latent space analysis.

Conclusion: CoT Vectors provide an efficient alternative to costly CoT implementations while offering new insights into multi-step reasoning functional organization in LLMs.

Abstract: Chain-of-Thought (CoT) prompting has emerged as a powerful approach to enhancing the reasoning capabilities of Large Language Models (LLMs). However, existing implementations, such as in-context learning and fine-tuning, remain costly and inefficient. To improve CoT reasoning at a lower cost, and inspired by the task vector paradigm, we introduce CoT Vectors, compact representations that encode task-general, multi-step reasoning knowledge. Through experiments with Extracted CoT Vectors, we observe pronounced layer-wise instability, manifesting as a U-shaped performance curve that reflects a systematic three-stage reasoning process in LLMs. To address this limitation, we propose Learnable CoT Vectors, optimized under a teacher-student framework to provide more stable and robust guidance. Extensive evaluations across diverse benchmarks and models demonstrate that CoT Vectors not only outperform existing baselines but also achieve performance comparable to parameter-efficient fine-tuning methods, while requiring fewer trainable parameters. Moreover, by treating CoT Vectors as a probe, we uncover how their effectiveness varies due to latent space structure, information density, acquisition mechanisms, and pre-training differences, offering new insights into the functional organization of multi-step reasoning in LLMs. The source code will be released.

Method: Created a corpus of 10,120 syllabified Tenyidie words and applied deep learning models (LSTM, BLSTM, BLSTM+CRF, Encoder-decoder) using an 80:10:10 train:validation:test split.

Result: Achieved highest accuracy of 99.21% with BLSTM model on the test set.

Conclusion: This work enables numerous NLP applications for Tenyidie including morphological analysis, POS tagging, and machine translation.

Abstract: The Tenyidie language is a low-resource language of the Tibeto-Burman family spoken by the Tenyimia Community of Nagaland in the north-eastern part of India and is considered a major language in Nagaland. It is tonal, Subject-Object-Verb, and highly agglutinative in nature. Being a low-resource language, very limited research on Natural Language Processing (NLP) has been conducted. To the best of our knowledge, no work on syllabification has been reported for this language. Among the many NLP tasks, syllabification or syllabication is an important task in which the given word syllables are identified. The contribution of this work is the creation of 10,120 syllabified Tenyidie words and the application of the Deep Learning techniques on the created corpus. In this paper, we have applied LSTM, BLSTM, BLSTM+CRF, and Encoder-decoder deep learning architectures on our created dataset. In our dataset split of 80:10:10 (train:validation:test) set, we achieved the highest accuracy of 99.21% with BLSTM model on the test set. This work will find its application in numerous other NLP applications, such as morphological analysis, part-of-speech tagging, machine translation, etc, for the Tenyidie Language. Keywords: Tenyidie; NLP; syllabification; deep learning; LSTM; BLSTM; CRF; Encoder-decoder

Method: Multi-faceted Cross-modal Direct Preference Optimization (MCM-DPO) that optimizes preferences across single, paired, and multi-preference dimensions covering textual, visual, and cross-modal factors. Also constructed two large-scale datasets TAlt and PAlt with 202k annotated samples and 18k preference pairs.

Result: MCM-DPO consistently outperforms both DPO and SFT, establishing a new state of the art in alt-text generation.

Conclusion: The proposed MCM-DPO method effectively addresses limitations of existing approaches by learning from preference pairs without requiring precise annotations, and the released datasets support further research in alt-text generation.

Abstract: The alt-text generation task produces concise, context-relevant descriptions of images, enabling blind and low-vision users to access online images. Despite the capabilities of large vision-language models, alt-text generation performance remains limited due to noisy user annotations, inconsistent standards, and MLLMs’ insensitivity to contextual information. Previous efforts to fine-tune MLLMs using supervised fine-tuning (SFT) have struggled, as SFT relies on accurate target annotations, which are often flawed in user-generated alt-text. To address this, we propose Multi-faceted Cross-modal Direct Preference Optimization (MCM-DPO), which improves alt-text generation by learning to identify better options in preference pairs without requiring precise annotations. MCM-DPO optimizes preferences across single, paired, and multi-preference dimensions, covering textual, visual, and cross-modal factors. In light of the scarcity of high-quality annotated and preference-labeled datasets for alt-text, we constructed two large-scale, high-quality datasets named TAlt and PAlt, sourced from Twitter and Pinterest. These datasets include 202k annotated alt-text samples and 18k preference pairs that cover diverse preference dimensions, aiming to support further research in this domain. Experimental results show that our proposed MCM-DPO method consistently outperforms both DPO and SFT, establishing a new state of the art in alt-text generation. We release the code and data here: https://github.com/LVUGAI/MCM-DPO

Method: Proposed multi-task learning approach combining text summarization, text simplification, and ETR generation. Tested two strategies: multi-task RAG for in-context learning and MTL-LoRA for parameter-efficient fine-tuning using Mistral-7B and LLaMA-3-8B models on the new ETR-fr dataset.

Result: Multi-task setups outperformed single-task baselines across all configurations. RAG-based strategy enabled generalization in out-of-domain settings, while MTL-LoRA achieved best performance in in-domain configurations.

Conclusion: Multi-task learning approaches effectively automate ETR content generation, with different strategies offering complementary benefits for in-domain and out-of-domain performance.

Abstract: Simplifying complex texts is essential for ensuring equitable access to information, especially for individuals with cognitive impairments. The Easy-to-Read (ETR) initiative offers a framework for making content accessible to the neurodivergent population, but the manual creation of such texts remains time-consuming and resource-intensive. In this work, we investigate the potential of large language models (LLMs) to automate the generation of ETR content. To address the scarcity of aligned corpora and the specificity of ETR constraints, we propose a multi-task learning (MTL) approach that trains models jointly on text summarization, text simplification, and ETR generation. We explore two different strategies: multi-task retrieval-augmented generation (RAG) for in-context learning, and MTL-LoRA for parameter-efficient fine-tuning. Our experiments with Mistral-7B and LLaMA-3-8B, based on ETR-fr, a new high-quality dataset, demonstrate the benefits of multi-task setups over single-task baselines across all configurations. Moreover, results show that the RAG-based strategy enables generalization in out-of-domain settings, while MTL-LoRA outperforms all learning strategies within in-domain configurations.

Method: Introduced ETR-fr dataset compliant with European ETR guidelines, implemented parameter-efficient fine-tuning on PLMs and LLMs, and created an evaluation framework with automatic metrics and human assessments using a 36-question evaluation form aligned with guidelines.

Result: PLMs perform comparably to LLMs and adapt effectively to out-of-domain texts, demonstrating that lightweight models can achieve similar performance to larger models in ETR text generation tasks.

Conclusion: The proposed approach provides a scalable solution for ETR text generation, with PLMs showing competitive performance to LLMs while being more efficient, enabling broader accessibility for individuals with cognitive impairments.

Abstract: Ensuring accessibility for individuals with cognitive impairments is essential for autonomy, self-determination, and full citizenship. However, manual Easy-to-Read (ETR) text adaptations are slow, costly, and difficult to scale, limiting access to crucial information in healthcare, education, and civic life. AI-driven ETR generation offers a scalable solution but faces key challenges, including dataset scarcity, domain adaptation, and balancing lightweight learning of Large Language Models (LLMs). In this paper, we introduce ETR-fr, the first dataset for ETR text generation fully compliant with European ETR guidelines. We implement parameter-efficient fine-tuning on PLMs and LLMs to establish generative baselines. To ensure high-quality and accessible outputs, we introduce an evaluation framework based on automatic metrics supplemented by human assessments. The latter is conducted using a 36-question evaluation form that is aligned with the guidelines. Overall results show that PLMs perform comparably to LLMs and adapt effectively to out-of-domain texts.

Method: Created dataset with court cases including facts, reasoning, verdicts, and cited clauses; defined challenging legal reasoning tasks; benchmarked Arabic LLMs and used dataset for instruction tuning.

Result: Instruction-tuning a 12B parameter model with ALARB significantly improved verdict prediction and Arabic verdict generation, achieving performance comparable to GPT-4o.

Conclusion: ALARB effectively evaluates and enhances Arabic LLMs’ legal reasoning capabilities, bridging a gap in Arabic NLP benchmarks.

Abstract: We introduce ALARB, a dataset and suite of tasks designed to evaluate the reasoning capabilities of large language models (LLMs) within the Arabic legal domain. While existing Arabic benchmarks cover some knowledge-intensive tasks such as retrieval and understanding, substantial datasets focusing specifically on multistep reasoning for Arabic LLMs, especially in open-ended contexts, are lacking. The dataset comprises over 13K commercial court cases from Saudi Arabia, with each case including the facts presented, the reasoning of the court, the verdict, as well as the cited clauses extracted from the regulatory documents. We define a set of challenging tasks leveraging this dataset and reflecting the complexity of real-world legal reasoning, including verdict prediction, completion of reasoning chains in multistep legal arguments, and identification of relevant regulations based on case facts. We benchmark a representative selection of current open and closed Arabic LLMs on these tasks and demonstrate the dataset’s utility for instruction tuning. Notably, we show that instruction-tuning a modest 12B parameter model using ALARB significantly enhances its performance in verdict prediction and Arabic verdict generation, reaching a level comparable to that of GPT-4o.

Method: Start from English models, continue pre-training on Scandinavian languages (individually or merged), then fine-tune on Faroese using full fine-tuning or LoRA. Create new Faroese evaluation benchmarks and conduct human evaluations.

Result: Transfer from related languages is crucial but task-dependent: Icelandic improves linguistic accuracy while Danish boosts comprehension. LoRA enhances linguistic acceptability and human scores, while full fine-tuning yields better comprehension and preserves model capabilities.

Conclusion: Successful adaptation of LLMs to low-resource languages requires strategic transfer learning from related languages and task-dependent tuning methods, with different source languages and tuning approaches optimal for different objectives.

Abstract: We investigate how to adapt small, efficient LLMs to Faroese, a low-resource North Germanic language. Starting from English models, we continue pre-training on related Scandinavian languages, either individually or combined via merging, before fine-tuning on Faroese. We compare full fine-tuning with parameter-efficient tuning using LoRA, evaluating their impact on both linguistic accuracy and text comprehension. Due to the lack of existing Faroese evaluation data, we construct two new minimal-pair benchmarks from adapted and newly collected datasets and complement them with human evaluations by Faroese linguists. Our results demonstrate that transfer from related languages is crucial, though the optimal source language depends on the task: Icelandic enhances linguistic accuracy, whereas Danish boosts comprehension. Similarly, the choice between full fine-tuning and LoRA is task-dependent: LoRA improves linguistic acceptability and slightly increases human evaluation scores on the base model, while full fine-tuning yields stronger comprehension performance and better preserves model capabilities during downstream fine-tuning.

Method: Proposed a noise synthesis framework and new metrics to systematically evaluate REAL-MT robustness. Instantiated REAL-MT with Qwen-series models (standard LLMs and large reasoning models) and evaluated on idiomatic translation across different resource language pairs under synthesized noise.

Result: Low-resource language pairs degrade more severely under noise than high-resource ones, often producing nonsensical translations. Large reasoning models show no improvement in error correction and are more susceptible to noise, rationalizing incorrect contexts. Attention shifts away from source idioms to noisy content while confidence increases despite declining accuracy.

Conclusion: Current approaches have limitations, revealing a fundamental trade-off between robustness and clean context performance. Training-free and fine-tuning strategies improve robustness but at performance cost. Highlights need for self-verifying integration mechanisms.

Abstract: \textbf{RE}trieval-\textbf{A}ugmented \textbf{L}LM-based \textbf{M}achine \textbf{T}ranslation (REAL-MT) shows promise for knowledge-intensive tasks like idiomatic translation, but its reliability under noisy retrieval contexts remains poorly understood despite this being a common challenge in real-world deployment. To address this gap, we propose a noise synthesis framework and new metrics to evaluate the robustness of REAL-MT systematically. Using this framework, we instantiate REAL-MT with Qwen-series models, including standard LLMs and large reasoning models (LRMs) with enhanced reasoning, and evaluate their performance on idiomatic translation across high-, medium-, and low-resource language pairs under synthesized noise. Our results show that low-resource language pairs, which rely more heavily on retrieved context, degrade more severely under noise than high-resource ones and often produce nonsensical translations. Although LRMs possess enhanced reasoning capabilities, they show no improvement in error correction and are even more susceptible to noise, tending to rationalize incorrect contexts. We find that this stems from an attention shift away from the source idiom to noisy content, while confidence increases despite declining accuracy, indicating poor calibration. To mitigate these issues, we investigate training-free and fine-tuning strategies, which improve robustness at the cost of performance in clean contexts, revealing a fundamental trade-off. Our findings highlight the limitations of current approaches, underscoring the need for self-verifying integration mechanisms.

Method: Created ManagerBench with human-validated managerial scenarios that force choices between pragmatic but harmful actions and safe but less effective actions, plus a parallel control set with harm directed only at inanimate objects to measure pragmatism.

Result: Frontier LLMs perform poorly in navigating the safety-pragmatism trade-off - many consistently choose harmful options for operational goals, while others become overly safe and ineffective. The misalignment stems from flawed prioritization rather than inability to perceive harm.

Conclusion: ManagerBench addresses a critical gap in evaluating agentic behavior where operational goals and alignment values conflict, revealing significant challenges in LLM decision-making for autonomous agents.

Abstract: As large language models (LLMs) evolve from conversational assistants into autonomous agents, evaluating the safety of their actions becomes critical. Prior safety benchmarks have primarily focused on preventing generation of harmful content, such as toxic text. However, they overlook the challenge of agents taking harmful actions when the most effective path to an operational goal conflicts with human safety. To address this gap, we introduce ManagerBench, a benchmark that evaluates LLM decision-making in realistic, human-validated managerial scenarios. Each scenario forces a choice between a pragmatic but harmful action that achieves an operational goal, and a safe action that leads to worse operational performance. A parallel control set, where potential harm is directed only at inanimate objects, measures a model’s pragmatism and identifies its tendency to be overly safe. Our findings indicate that the frontier LLMs perform poorly when navigating this safety-pragmatism trade-off. Many consistently choose harmful options to advance their operational goals, while others avoid harm only to become overly safe and ineffective. Critically, we find this misalignment does not stem from an inability to perceive harm, as models’ harm assessments align with human judgments, but from flawed prioritization. ManagerBench is a challenging benchmark for a core component of agentic behavior: making safe choices when operational goals and alignment values incentivize conflicting actions. Benchmark & code available at https://github.com/technion-cs-nlp/ManagerBench.

Method: Proposed Erasable Reinforcement Learning (ERL) - a framework that explicitly identifies faulty reasoning steps, erases them, and regenerates reasoning in place to prevent defective logic from propagating through the chain.

Result: ESearch models trained with ERL achieved substantial improvements: 3B model +8.48% EM and +11.56% F1, 7B model +5.38% EM and +7.22% F1 over previous SOTA on HotpotQA, MuSiQue, 2Wiki, and Bamboogle benchmarks.

Conclusion: Erasable reinforcement learning provides a powerful paradigm shift for robust multi-step reasoning in LLMs, transforming fragile reasoning into a more resilient process.

Abstract: While search-augmented large language models (LLMs) exhibit impressive capabilities, their reliability in complex multi-hop reasoning remains limited. This limitation arises from three fundamental challenges: decomposition errors, where tasks are incorrectly broken down; retrieval missing, where key evidence fails to be retrieved; and reasoning errors, where flawed logic propagates through the reasoning chain. A single failure in any of these stages can derail the final answer. We propose Erasable Reinforcement Learning (ERL), a novel framework that transforms fragile reasoning into a robust process. ERL explicitly identifies faulty steps, erases them, and regenerates reasoning in place, preventing defective logic from propagating through the reasoning chain. This targeted correction mechanism turns brittle reasoning into a more resilient process. Models trained with ERL, termed ESearch, achieve substantial improvements on HotpotQA, MuSiQue, 2Wiki, and Bamboogle, with the 3B model achieving +8.48% EM and +11.56% F1, and the 7B model achieving +5.38% EM and +7.22% F1 over previous state-of-the-art(SOTA) results. These findings suggest that erasable reinforcement learning provides a powerful paradigm shift for robust multi-step reasoning in LLMs.

Method: Combines domain-agnostic synthetic dataset from FineWeb with multi-stage curation, synthetic grounded/hallucinated claims, and preference-based fine-tuning using Odds Ratio Preference Optimization to distill large-model reasoning into a smaller model.

Result: Achieves 84.0% balanced accuracy on RAGTruth subset (matching specialized models) and 75.7% BAcc on full benchmark (matching GPT-4o), while using roughly half the parameters of comparable models.

Conclusion: HalluGuard demonstrates that smaller specialized models can effectively detect hallucinations and compete with larger general-purpose LLMs, providing a practical solution for trustworthy RAG applications.

Abstract: Large Language Models (LLMs) excel in many NLP tasks but remain prone to hallucinations, limiting trust in real-world applications. We present HalluGuard, a 4B-parameter Small Reasoning Model (SRM) for mitigating hallucinations in Retrieval-Augmented Generation (RAG). HalluGuard classifies document-claim pairs as grounded or hallucinated and produces evidence-grounded justifications for transparency. Our approach combines (i) a domain-agnostic synthetic dataset derived from FineWeb and refined through multi-stage curation and data reformation, (ii) synthetic grounded and hallucinated claims, and (iii) preference-based fine-tuning with Odds Ratio Preference Optimization to distill large-model reasoning into a smaller backbone. On the RAGTruth subset of the LLM-AggreFact benchmark, HalluGuard achieves 84.0% balanced accuracy (BAcc), rivaling specialized models, MiniCheck (7B; 84.0%) and Granite Guardian 3.3 (8B; 82.2%) while using roughly half their parameters. Over the full benchmark it reaches 75.7% BAcc, matching larger general-purpose LLMs such as GPT-4o (75.9%). We will release HalluGuard and datasets under Apache 2.0 upon acceptance.

Method: Combines section-conditioned stylistic modeling with multi-level contrastive learning to capture human-AI differences while reducing topic dependence. Integrates BIO-CRF sequence labeling with pointer-based boundary decoding and confidence calibration for precise span-level detection.

Result: Achieves state-of-the-art performance with F1(AI) of 80.17, AUROC of 92.63, and Span-F1 of 74.36. Shows strong resilience under adversarial rewriting and maintains balanced accuracy across IMRaD sections and diverse disciplines, substantially surpassing existing baselines.

Conclusion: Sci-SpanDet provides an effective solution for AI-generated text detection in scholarly documents with superior performance and cross-domain robustness, and the dataset and source code will be publicly released to foster further research.

Abstract: The rapid adoption of large language models (LLMs) in scientific writing raises serious concerns regarding authorship integrity and the reliability of scholarly publications. Existing detection approaches mainly rely on document-level classification or surface-level statistical cues; however, they neglect fine-grained span localization, exhibit weak calibration, and often fail to generalize across disciplines and generators. To address these limitations, we present Sci-SpanDet, a structure-aware framework for detecting AI-generated scholarly texts. The proposed method combines section-conditioned stylistic modeling with multi-level contrastive learning to capture nuanced human-AI differences while mitigating topic dependence, thereby enhancing cross-domain robustness. In addition, it integrates BIO-CRF sequence labeling with pointer-based boundary decoding and confidence calibration to enable precise span-level detection and reliable probability estimates. Extensive experiments on a newly constructed cross-disciplinary dataset of 100,000 annotated samples generated by multiple LLM families (GPT, Qwen, DeepSeek, LLaMA) demonstrate that Sci-SpanDet achieves state-of-the-art performance, with F1(AI) of 80.17, AUROC of 92.63, and Span-F1 of 74.36. Furthermore, it shows strong resilience under adversarial rewriting and maintains balanced accuracy across IMRaD sections and diverse disciplines, substantially surpassing existing baselines. To ensure reproducibility and to foster further research on AI-generated text detection in scholarly documents, the curated dataset and source code will be publicly released upon publication.

Method: Created PhoPile multimodal dataset for Olympiad physics, benchmarked RAG-augmented LLMs and LMMs with multiple retrievers.

Result: Integration of retrieval with physics corpora improves model performance, but challenges remain for further research.

Conclusion: RAG shows promise for enhancing physics reasoning in foundation models, though more work is needed to address remaining challenges.

Abstract: Retrieval-augmented generation (RAG) with foundation models has achieved strong performance across diverse tasks, but their capacity for expert-level reasoning-such as solving Olympiad-level physics problems-remains largely unexplored. Inspired by the way students prepare for competitions by reviewing past problems, we investigate the potential of RAG to enhance physics reasoning in foundation models. We introduce PhoPile, a high-quality multimodal dataset specifically designed for Olympiad-level physics, enabling systematic study of retrieval-based reasoning. PhoPile includes diagrams, graphs, and equations, capturing the inherently multimodal nature of physics problem solving. Using PhoPile, we benchmark RAG-augmented foundation models, covering both large language models (LLMs) and large multimodal models (LMMs) with multiple retrievers. Our results demonstrate that integrating retrieval with physics corpora can improve model performance, while also highlighting challenges that motivate further research in retrieval-augmented physics reasoning.

Method: FusioN uses a general LLM judge to synthesize the most informative elements from each sample in a pool of N generations into a single final answer. It’s applied in test-time scaling (aggregating from single model) and synthetic data generation (fusing samples from diverse teachers).

Result: FusioN consistently outperforms BoN across 11 languages, 3 diverse tasks and varying model scales. It shows versatility and robustness in both test-time scaling and downstream gains from synthetic data generation, with surprising strengths under challenging settings.

Conclusion: We should shift from monolithic quality measurement to embracing the polylithic nature of LLM generations, integrating diverse strengths to unlock latent potential and achieve improvements inaccessible through selection alone.

Abstract: Obtaining high-quality generations in modern LLMs has largely been framed as a selection problem: identifying a single winning generation from a diverse pool of N samples, the Best-of-N (BoN). Yet, this approach is inherently zero-sum, discarding diverse and potentially useful information from the pool. Instead, we explore a collaborative setup, where all candidates can potentially contribute to the final winning generation. To this end, we propose Fusion-of-N (FusioN): a method that uses a general LLM judge to synthesize the most informative elements of each sample into a single final answer. We compare FusioN to BoN in two settings, (i) test-time scaling, where we sample and aggregate from a single model at test-time (ii) synthetic data generation, where we fuse samples from a pool of diverse teachers to improve a student model. We extensively benchmark both setups across 11 languages, 3 diverse tasks and varying model scales. Across the bench, FusioN consistently outperforms BoN showing versatility and robustness both in test-time scaling and in downstream gains from synthetic data generation. We also perform extensive analysis on FusioN, where it shows surprising strengths and robustness under challenging settings. These results show that we should shift how we think about evaluating and utilizing LLM generations from a monolithic measure of quality, to embracing their polylithic nature. This shift allows us to integrate diverse strengths, unlock latent potential, and achieve improvements that were previously inaccessible through selection alone.

Method: Analyzed effects of converting “standard” American English questions into non-standard dialectal variants on multiple choice question answering tasks, and investigated the grammatical basis of performance differences.

Result: Found up to 20% reduction in accuracy for non-standard dialect questions, with three specific grammar rules (existential “it”, zero copula, and y’all) explaining the majority of performance degradation across multiple dialects.

Conclusion: Calls for future work to develop bias mitigation methods focused on individual, high-impact grammatical structures rather than treating dialects as monolithic entities.

Abstract: Large language models (LLMs) are ubiquitous in modern day natural language processing. However, previous work has shown degraded LLM performance for under-represented English dialects. We analyze the effects of typifying “standard” American English language questions as non-“standard” dialectal variants on multiple choice question answering tasks and find up to a 20% reduction in accuracy. Additionally, we investigate the grammatical basis of under-performance in non-“standard” English questions. We find that individual grammatical rules have varied effects on performance, but some are more consequential than others: three specific grammar rules (existential “it”, zero copula, and y’all) can explain the majority of performance degradation observed in multiple dialects. We call for future work to investigate bias mitigation methods focused on individual, high-impact grammatical structures.

Method: Proposes cascaded and noncascaded architectures that generate syntactic guidance before text generation, incorporating syntactic information and shared representation for personalization.

Result: Extensive experiments show superiority in fluency, diversity, and stylistic fidelity. Qualitative analyses highlight interpretability and flexibility in learning personalized patterns.

Conclusion: The proposed model effectively enhances text quality, diversity, and controllability by integrating structural supervision and personalized conditioning through diffusion-based approaches.

Abstract: Large language models have made revolutionary progress in generating human-like text, yet their outputs often tend to be generic, exhibiting insufficient structural diversity, which limits personalized expression. Recent advances in diffusion models have opened new opportunities for improving language generation beyond the limitations of autoregressive paradigms. In this work, we propose a syntax-guided diffusion language model that integrates structural supervision and personalized conditioning to enhance text quality, diversity, and controllability. We introduce a cascaded framework that generates syntactic guidance before conditional text generation, and further generalize it to a novel noncascaded architecture for better alignment between structure and content. By incorporating syntactic information in the generating process, the proposed model better captures the lexical and structural characteristics of stylistic sentence construction. To enable fine-grained personalization, we develop a shared representation mechanism that facilitates information integration across users, supporting both faithful stylistic generation and generalizable zero-shot inference. Extensive experiments on multiple tasks demonstrate the superiority of our approach in fluency, diversity, and stylistic fidelity. Further qualitative analyses highlight its interpretability and flexibility in learning personalized patterns.

Method: Survey methodology - synthesizing existing literature on concept description generation methods, evaluation metrics (automated and human), and supporting datasets in this emerging field.

Result: The survey reveals key trends in concept description research, including the evolution of evaluation approaches and identifies that current methods lack rigorous causal evaluation frameworks.

Conclusion: The paper provides a roadmap for future research to improve model transparency through better concept description methods, emphasizing the need for more causal evaluation approaches.

Abstract: Understanding the decision-making processes of neural networks is a central goal of mechanistic interpretability. In the context of Large Language Models (LLMs), this involves uncovering the underlying mechanisms and identifying the roles of individual model components such as neurons and attention heads, as well as model abstractions such as the learned sparse features extracted by Sparse Autoencoders (SAEs). A rapidly growing line of work tackles this challenge by using powerful generator models to produce open-vocabulary, natural language concept descriptions for these components. In this paper, we provide the first survey of the emerging field of concept descriptions for model components and abstractions. We chart the key methods for generating these descriptions, the evolving landscape of automated and human metrics for evaluating them, and the datasets that underpin this research. Our synthesis reveals a growing demand for more rigorous, causal evaluation. By outlining the state of the art and identifying key challenges, this survey provides a roadmap for future research toward making models more transparent.

Method: Hybrid DST model using rule-based methods with BERT for slot filling/intent detection, XGBoost for intent validation, GPT for DST, and online agents for real-time answer generation.

Result: Significantly improved accuracy and coherence over existing methods in Persian-based chatbots when evaluated on comprehensive Persian multi-turn dialogue dataset.

Conclusion: Hybrid approach effectively improves DST capabilities, enabling more customized, adaptable, and human-like conversational AI systems.

Abstract: Dialogue State Tracking (DST) is an essential element of conversational AI with the objective of deeply understanding the conversation context and leading it toward answering user requests. Due to high demands for open-domain and multi-turn chatbots, the traditional rule-based DST is not efficient enough, since it cannot provide the required adaptability and coherence for human-like experiences in complex conversations. This study proposes a hybrid DST model that utilizes rule-based methods along with language models, including BERT for slot filling and intent detection, XGBoost for intent validation, GPT for DST, and online agents for real-time answer generation. This model is uniquely designed to be evaluated on a comprehensive Persian multi-turn dialogue dataset and demonstrated significantly improved accuracy and coherence over existing methods in Persian-based chatbots. The results demonstrate how effectively a hybrid approach may improve DST capabilities, paving the way for conversational AI systems that are more customized, adaptable, and human-like.

Method: Developed a novel integration model through theoretical explanation and experimental exploration

Result: The model demonstrated high effectiveness across various text processing tasks

Conclusion: The integration approach shows strong applicative potential for intelligent text processing

Abstract: This article addresses embodied intelligence and reinforcement learning integration in the field of text processing, aiming to enhance text handling with more intelligence on the basis of embodied intelligence’s perception and action superiority and reinforcement learning’s decision optimization capability. Through detailed theoretical explanation and experimental exploration, a novel integration model is introduced. This model has been demonstrated to be very effective in a wide range oftext processing tasks, validating its applicative potential

Method: Conducted a systematic literature review using PRISMA 2020 procedures, screening 71 out of 2,062 records from major databases (DBLP, ACM, Google Scholar, Semantic Scholar, arXiv) published between 2020-2025.

Result: Identified 74 datasets across 111 African languages (~11,206 hours of speech), but fewer than 15% provided reproducible materials. Self-supervised learning shows promise but faces data limitations. Evaluation relies heavily on WER with minimal use of linguistically-informed metrics.

Conclusion: Sustainable ASR development for African languages requires stakeholder partnerships, ethically balanced datasets, lightweight modeling, and active benchmarking to address current limitations in dataset availability, reproducibility, and evaluation.

Abstract: ASR has achieved remarkable global progress, yet African low-resource languages remain rigorously underrepresented, producing barriers to digital inclusion across the continent with more than +2000 languages. This systematic literature review (SLR) explores research on ASR for African languages with a focus on datasets, models and training methods, evaluation techniques, challenges, and recommends future directions. We employ the PRISMA 2020 procedures and search DBLP, ACM Digital Library, Google Scholar, Semantic Scholar, and arXiv for studies published between January 2020 and July 2025. We include studies related to ASR datasets, models or metrics for African languages, while excluding non-African, duplicates, and low-quality studies (score <3/5). We screen 71 out of 2,062 records and we record a total of 74 datasets across 111 languages, encompassing approximately 11,206 hours of speech. Fewer than 15% of research provided reproducible materials, and dataset licensing is not clear. Self-supervised and transfer learning techniques are promising, but are hindered by limited pre-training data, inadequate coverage of dialects, and the availability of resources. Most of the researchers use Word Error Rate (WER), with very minimal use of linguistically informed scores such as Character Error Rate (CER) or Diacritic Error Rate (DER), and thus with limited application in tonal and morphologically rich languages. The existing evidence on ASR systems is inconsistent, hindered by issues like dataset availability, poor annotations, licensing uncertainties, and limited benchmarking. Nevertheless, the rise of community-driven initiatives and methodological advancements indicates a pathway for improvement. Sustainable development for this area will also include stakeholder partnership, creation of ethically well-balanced datasets, use of lightweight modelling techniques, and active benchmarking.

Method: Developed mR3 through comprehensive study of data and curriculum selection strategies, including integration of target-language reasoning datasets, trained on 72 languages using rubric-agnostic reward modeling approach.

Result: Achieved state-of-the-art performance on multilingual reward model benchmarks, surpassing much larger models (including GPT-OSS-120B) while being up to 9x smaller, with effectiveness confirmed through extensive ablation studies.

Conclusion: The mR3 model demonstrates that carefully designed multilingual training with appropriate data selection can create highly effective reward models that outperform much larger alternatives across diverse languages.

Abstract: Evaluation using Large Language Model (LLM) judges has been widely adopted in English and shown to be effective for automatic evaluation. However, their performance does not generalize well to non-English settings, and it remains unclear what constitutes effective multilingual training for such judges. In this paper, we introduce mR3, a massively multilingual, rubric-agnostic reward reasoning model trained on 72 languages, achieving the broadest language coverage in reward modeling to date. We present a comprehensive study of data and curriculum selection for training to identify effective strategies and data sources for building high-quality reward models, including the integration of target-language reasoning datasets. Our approach attains state-of-the-art performance on multilingual reward model benchmarks, surpassing much larger models (i.e., GPT-OSS-120B) while being up to 9x smaller, and its effectiveness is further confirmed through extensive ablation studies. Our models, data, and code are available as open source at https://github.com/rubricreward/mr3.

Method: Uses reinforcement learning with a pay-per-search reward that penalizes external help-seeking while rewarding answer accuracy. This treats search tool use as a proxy for abstention without requiring pre-determined knowledge boundaries.

Result: MASH substantially improves selective help-seeking performance over prior approaches, achieving 7.6% higher answer accuracy on multi-hop datasets. It can distinguish between answerable/unanswerable questions and selectively generate responses.

Conclusion: MASH effectively aligns search tool use with parametric knowledge boundaries, enabling LLMs to make abstention decisions as a by-product of training for selective help-seeking, without needing pre-defined knowledge boundaries.

Abstract: LLMs cannot reliably recognize their parametric knowledge boundaries and often hallucinate answers to outside-of-boundary questions. In contrast, humans recognize their limitations and can either seek external help for such questions or abstain. In this paper, we introduce MASH (Modeling Abstention via Selective Help-seeking), a training framework that readily extracts abstentions from LLMs. Our key idea is that any external help-seeking by an LLM, i.e. search tool use, can serve as a proxy for abstention if the external help (search) is appropriately penalized while simultaneously rewarding answer accuracy. MASH operationalizes this idea using reinforcement learning with a pay-per-search reward. We run experiments on three knowledge-intensive QA datasets. Our results show that MASH substantially improves upon the selective help-seeking performance of prior efficient search approaches; on multi-hop datasets, MASH improves answer accuracy by 7.6%. Furthermore, MASH demonstrates strong off-the-shelf abstention – it can distinguish between unanswerable/answerable questions and selectively generate responses for answerable questions – showcasing behavior analogous to specialized abstention approaches. We emphasize that contrary to prior abstention methods, MASH does not require pre-determining knowledge boundaries to construct training data. Instead, MASH’s abstentions are a by-product of training for the auxiliary selective help-seeking task. Overall, we show that MASH training effectively aligns search tool use with parametric knowledge, which can be successfully leveraged for making abstention decisions.

Method: Created a dynamic simulation environment (SWF Benchmark) where LLMs act as sovereign allocators distributing tasks to heterogeneous communities, measuring trade-offs between efficiency (ROI) and fairness (Gini coefficient).

Result: Evaluated 20 state-of-the-art LLMs and found: (i) conversational ability doesn’t predict allocation skill, (ii) most models default to utilitarian approaches favoring productivity over equality, (iii) allocation strategies are vulnerable to output constraints and social framing.

Conclusion: Current LLMs pose risks as societal decision-makers and require specialized benchmarks and targeted alignment for responsible AI governance.

Abstract: Large language models (LLMs) are increasingly entrusted with high-stakes decisions that affect human welfare. However, the principles and values that guide these models when distributing scarce societal resources remain largely unexamined. To address this, we introduce the Social Welfare Function (SWF) Benchmark, a dynamic simulation environment where an LLM acts as a sovereign allocator, distributing tasks to a heterogeneous community of recipients. The benchmark is designed to create a persistent trade-off between maximizing collective efficiency (measured by Return on Investment) and ensuring distributive fairness (measured by the Gini coefficient). We evaluate 20 state-of-the-art LLMs and present the first leaderboard for social welfare allocation. Our findings reveal three key insights: (i) A model’s general conversational ability, as measured by popular leaderboards, is a poor predictor of its allocation skill. (ii) Most LLMs exhibit a strong default utilitarian orientation, prioritizing group productivity at the expense of severe inequality. (iii) Allocation strategies are highly vulnerable, easily perturbed by output-length constraints and social-influence framing. These results highlight the risks of deploying current LLMs as societal decision-makers and underscore the need for specialized benchmarks and targeted alignment for AI governance.

Method: Train an LLM model to generate input-specific concise demonstrations, using token budget constraints for both demonstrations and final output. The method is trained solely on math data but tested across multiple domains.

Result: GRAD consistently outperforms strong baselines on Qwen2.5-14B across mathematical reasoning and advanced STEM questions, showing robust generalization to physics, chemistry, and computer science. Smaller trained models can effectively guide larger target models.

Conclusion: GRAD introduces a scalable demonstration generator model as the first step toward dynamic few-shot learning in resource-constrained settings, with released code for the project.

Abstract: Large Language Models (LLMs) achieve strong performance across diverse tasks, but their effectiveness often depends on the quality of the provided context. Retrieval-Augmented Generation (RAG) enriches prompts with external information, but its reliance on static databases constrains adaptability and can result in irrelevant demonstrations. In this work, we propose a Generative Retrieval-Aligned Demonstrator (GRAD), a dynamic demonstration-based approach where an LLM model is trained to generate input-specific concise demonstrations. By tailoring demonstrations to each input, our method offers better contextual support than traditional RAG approaches. We demonstrate the superiority of GRAD under budget constraints, where we limit both the number of tokens used per demonstration and the number of tokens used for the final output. Trained solely on a math dataset, GRAD consistently outperforms strong baselines on Qwen2.5-14B across mathematical reasoning and advanced STEM questions, highlighting GRAD’s robust generalization to out-of-distribution (OOD) domains such as physics, chemistry, and computer science. Furthermore, we show that demonstrations generated by trained smaller models can effectively guide larger target models, reducing training costs while maintaining competitive accuracy. Overall, this work introduces a scalable demonstration generator model presenting the first step toward a dynamic few-shot learning paradigm in resource-constrained settings. We release the code used for the project.

Method: Verbalized Sampling (VS) - a training-free prompting strategy that asks the model to verbalize a probability distribution over multiple responses (e.g., “Generate 5 jokes about coffee and their corresponding probabilities”).

Result: VS significantly improves performance across creative writing (1.6-2.1x diversity increase), dialogue simulation, open-ended QA, and synthetic data generation without sacrificing factual accuracy and safety. More capable models benefit more from VS.

Conclusion: The work provides a data-centric perspective on mode collapse and a practical inference-time solution (VS) that helps unlock pre-trained generative diversity.

Abstract: Post-training alignment often reduces LLM diversity, leading to a phenomenon known as mode collapse. Unlike prior work that attributes this effect to algorithmic limitations, we identify a fundamental, pervasive data-level driver: typicality bias in preference data, whereby annotators systematically favor familiar text as a result of well-established findings in cognitive psychology. We formalize this bias theoretically, verify it on preference datasets empirically, and show that it plays a central role in mode collapse. Motivated by this analysis, we introduce Verbalized Sampling, a simple, training-free prompting strategy to circumvent mode collapse. VS prompts the model to verbalize a probability distribution over a set of responses (e.g., ``Generate 5 jokes about coffee and their corresponding probabilities’’). Comprehensive experiments show that VS significantly improves performance across creative writing (poems, stories, jokes), dialogue simulation, open-ended QA, and synthetic data generation, without sacrificing factual accuracy and safety. For instance, in creative writing, VS increases diversity by 1.6-2.1x over direct prompting. We further observe an emergent trend that more capable models benefit more from VS. In sum, our work provides a new data-centric perspective on mode collapse and a practical inference-time remedy that helps unlock pre-trained generative diversity.

Method: The authors use Hyperspherical Energy (HE) to quantify neuron uniformity and propose SPHERE - a regularization strategy that identifies sparse space complementary to principal hyperspherical directions and projects new knowledge onto it, minimizing perturbations to existing knowledge.

Result: SPHERE outperforms the best baseline in editing capability by an average of 16.41% on LLaMA3 (8B) and Qwen2.5 (7B), while better preserving general model performance.

Conclusion: SPHERE offers a principled path toward reliable large-scale knowledge editing by stabilizing neuron weight distributions through hyperspherical energy regularization, enabling sequential updates without catastrophic forgetting.

Abstract: Large language models (LLMs) require constant updates to remain aligned with evolving real-world knowledge. Model editing offers a lightweight alternative to retraining, but sequential editing often destabilizes representations and induces catastrophic forgetting. In this work, we seek to better understand and mitigate performance degradation caused by sequential editing. We hypothesize that hyperspherical uniformity, a property that maintains uniform distribution of neuron weights on a hypersphere, helps the model remain stable, retain prior knowledge, while still accommodate new updates. We use Hyperspherical Energy (HE) to quantify neuron uniformity during editing, and examine its correlation with editing performance. Empirical studies across widely used editing methods reveals a strong correlation between HE dynamics and editing performance, with editing failures consistently coinciding with high HE fluctuations. We further theoretically prove that HE dynamics impose a lower bound on the degradation of pretrained knowledge, highlighting why HE stability is crucial for knowledge retention. Motivated by these insights, we propose SPHERE (Sparse Projection for Hyperspherical Energy-Regularized Editing), an HE-driven regularization strategy that stabilizes neuron weight distributions, ultimately preserving prior knowledge while enabling reliable sequential updates. Specifically, SPHERE identifies a sparse space complementary to the principal hyperspherical directions of the pretrained weight matrices and projects new knowledge onto it, attenuating perturbations on the principal directions. Extensive experiments on LLaMA3 (8B) and Qwen2.5 (7B) show that SPHERE outperforms the best baseline in editing capability by an average of 16.41%, while most faithfully preserving general model performance, thereby offering a principled path toward reliable large-scale knowledge editing.

Method: Calculates phylogenetic distance metric based on LLM output similarity, constructs dendrograms from these distances to capture model relationships.

Result: Successfully captured known relationships across 111 open-source and 45 closed models, and phylogenetic distance predicted performance in standard benchmarks.

Conclusion: PhyloLM provides a validated tool for evaluating LLM development, relationships, and capabilities without requiring transparent training information, bridging population genetics concepts to machine learning.

Abstract: This paper introduces PhyloLM, a method adapting phylogenetic algorithms to Large Language Models (LLMs) to explore whether and how they relate to each other and to predict their performance characteristics. Our method calculates a phylogenetic distance metric based on the similarity of LLMs’ output. The resulting metric is then used to construct dendrograms, which satisfactorily capture known relationships across a set of 111 open-source and 45 closed models. Furthermore, our phylogenetic distance predicts performance in standard benchmarks, thus demonstrating its functional validity and paving the way for a time and cost-effective estimation of LLM capabilities. To sum up, by translating population genetic concepts to machine learning, we propose and validate a tool to evaluate LLM development, relationships and capabilities, even in the absence of transparent training information.

Method: Built a legislative testbed where LLM lobbyists propose amendments benefiting specific companies while avoiding identification of the benefactor, using real-world bills and companies. Employed LLM-based re-planning and re-sampling for optimization.

Result: LLM lobbyists successfully drafted subtle phrasing to evade detection by strong LLM-based detectors. Optimization increased deception rates by up to 40 percentage points. Human evaluations confirmed generation quality and intent retention.

Conclusion: LLMs pose risks for strategic phrasing through seemingly neutral language to achieve self-serving goals, calling for future research to detect and protect against such subtle deception.

Abstract: We explore the ability of large language models (LLMs) to engage in subtle deception through strategically phrasing and intentionally manipulating information. This harmful behavior can be hard to detect, unlike blatant lying or unintentional hallucination. We build a simple testbed mimicking a legislative environment where a corporate \textit{lobbyist} module is proposing amendments to bills that benefit a specific company while evading identification of this benefactor. We use real-world legislative bills matched with potentially affected companies to ground these interactions. Our results show that LLM lobbyists can draft subtle phrasing to avoid such identification by strong LLM-based detectors. Further optimization of the phrasing using LLM-based re-planning and re-sampling increases deception rates by up to 40 percentage points. Our human evaluations to verify the quality of deceptive generations and their retention of self-serving intent show significant coherence with our automated metrics and also help in identifying certain strategies of deceptive phrasing. This study highlights the risk of LLMs' capabilities for strategic phrasing through seemingly neutral language to attain self-serving goals. This calls for future research to uncover and protect against such subtle deception.

Method: Progressive vocabulary expansion using a modified BPE algorithm that gradually extends Arabic subwords in the dynamic vocabulary during training to balance OOV ratio at every stage.

Result: Ablation study showed effectiveness of progressive vocabulary expansion. AraLLaMA achieves performance comparable to best Arabic LLMs across various Arabic benchmarks.

Conclusion: Progressive vocabulary expansion effectively addresses Arabic LLM development challenges. All models, training data, benchmarks, and code will be open-sourced to support Arabic NLP development.

Abstract: This paper addresses the critical need for democratizing large language models (LLM) in the Arab world, a region that has seen slower progress in developing models comparable to state-of-the-art offerings like GPT-4 or ChatGPT 3.5, due to a predominant focus on mainstream languages (e.g., English and Chinese). One practical objective for an Arabic LLM is to utilize an Arabic-specific vocabulary for the tokenizer that could speed up decoding. However, using a different vocabulary often leads to a degradation of learned knowledge since many words are initially out-of-vocabulary (OOV) when training starts. Inspired by the vocabulary learning during Second Language (Arabic) Acquisition for humans, the released AraLLaMA employs progressive vocabulary expansion, which is implemented by a modified BPE algorithm that progressively extends the Arabic subwords in its dynamic vocabulary during training, thereby balancing the OOV ratio at every stage. The ablation study demonstrated the effectiveness of Progressive Vocabulary Expansion. Moreover, AraLLaMA achieves decent performance comparable to the best Arabic LLMs across a variety of Arabic benchmarks. Models, training data, benchmarks, and codes will be all open-sourced.

Method: APP modularizes utterance generation prompts and dynamically prunes prompt segments based on attention scores, allowing users to control diversity through a single lambda parameter while maintaining compatibility with existing diversity control methods.

Result: APP effectively modulates diversity in experiments, with analysis showing all prompt components constrain diversity (Memory being most influential) and high-attention contents consistently suppress output diversity.

Conclusion: The proposed APP method successfully addresses diversity degradation in long-term simulations and provides a practical approach to balance diversity-stability trade-offs through adaptive prompt pruning.

Abstract: Controlling diversity in LLM-agent simulations is essential for balancing stability in structured tasks with variability in open-ended interactions. However, we observe that dialogue diversity tends to degrade over long-term simulations. To explore the role of prompt design in this phenomenon, we modularized the utterance generation prompt and found that reducing contextual information leads to more diverse outputs. Based on this insight, we propose Adaptive Prompt Pruning (APP), a novel method that allows users to control diversity via a single parameter, lambda. APP dynamically prunes prompt segments based on attention scores and is compatible with existing diversity control methods. We demonstrate that APP effectively modulates diversity through extensive experiments and propose a method to balance the control trade-offs. Our analysis reveals that all prompt components impose constraints on diversity, with the Memory being the most influential. Additionally, high-attention contents consistently suppress output diversity.

Method: OmniThink emulates human-like iterative expansion and reflection processes, simulating how learners slowly deepen their knowledge of topics through cognitive behaviors.

Result: Experimental results show OmniThink improves knowledge density of generated articles without compromising coherence and depth. Human evaluations and expert feedback confirm its effectiveness for long-form article generation.

Conclusion: OmniThink demonstrates potential to address real-world challenges in long-form article generation by incorporating slow-thinking cognitive processes.

Abstract: Machine writing with large language models often relies on retrieval-augmented generation. However, these approaches remain confined within the boundaries of the model’s predefined scope, limiting the generation of content with rich information. Specifically, vanilla-retrieved information tends to lack depth, novelty, and suffers from redundancy, which negatively impacts the quality of generated articles, leading to shallow, unoriginal, and repetitive outputs. To address these issues, we propose OmniThink, a slow-thinking machine writing framework that emulates the human-like process of iterative expansion and reflection. The core idea behind OmniThink is to simulate the cognitive behavior of learners as they slowly deepen their knowledge of the topics. Experimental results demonstrate that OmniThink improves the knowledge density of generated articles without compromising metrics such as coherence and depth. Human evaluations and expert feedback further highlight the potential of OmniThink to address real-world challenges in the generation of long-form articles. Code is available at https://github.com/zjunlp/OmniThink.

Method: ATLAS uses a concept repository, expert iteration with knowledge distillation, and novel augmentation strategies exploiting formal language characteristics. It runs for 10 iterations to generate datasets.

Result: Created an undergraduate-level dataset of 117k theorem statements and developed ATLAS Translator (fine-tuned Llama3.1-8B-Instruct) that outperforms existing models across all benchmarks with statistical significance.

Conclusion: The framework successfully addresses the data scarcity problem in autoformalization and demonstrates that fine-tuning stronger base models on ATLAS data leads to superior performance.

Abstract: Autoformalization, the automatic translation of mathematical content from natural language into machine-verifiable formal languages, has seen significant progress driven by advances in large language models (LLMs). Nonetheless, a primary barrier to further improvements is the limited availability of parallel corpora that map informal mathematical text to its formal counterpart. To address this limitation, we propose ATLAS (Autoformalizing Theorems through Lifting, Augmentation, and Synthesis of Data), a novel data generation framework designed to produce large-scale, high-quality parallel corpora of theorem statements. Distinct from prior approaches, ATLAS begins with a concept repository, accelerates the improvement of the student model through expert iteration combined with knowledge distillation, and introduces two novel augmentation strategies that exploit the structural characteristics of formal languages. Running the proposed ATLAS framework for 10 iterations, we construct an undergraduate-level dataset of 117k theorem statements and develop the ATLAS Translator by fine-tuning Llama3.1-8B-Instruct with LoRA. This model establishes a new state of the art, demonstrating statistically significant improvements over both the Herald Translator and the Kimina-Autoformalizer across all benchmarks (p<0.05, two-sided t-test). Furthermore, we demonstrate that the full-parameter fine-tuning of a stronger base model on the ATLAS dataset leads to superior performance. The datasets, model, and code are available at https://github.com/XiaoyangLiu-sjtu/ATLAS.

Method: Two complementary approaches: iterative model editing applies successive edits to mitigate UnderEdit, and neighbor-assisted model editing incorporates neighboring knowledge to reduce OverEdit.

Result: Experiments show improved editing performance across multiple LLMs, reducing UnderEdit by up to 38 percentage points and OverEdit by up to 6 percentage points.

Conclusion: The proposed methods effectively address key limitations in model editing and are broadly applicable to any locate-and-edit approach, enhancing LLM knowledge updating efficiency.

Abstract: Large Language Models (LLMs) are widely deployed in downstream tasks, but keeping their knowledge up-to-date via retraining or fine-tuning is often computationally expensive. Model editing provides a more efficient alternative by updating a targeted subset of parameters, which often follows the locate-and-edit paradigm. Despite this efficiency, existing methods are limited: edits may fail to inject knowledge (UnderEdit) or unintentionally disrupt unrelated neighboring knowledge (OverEdit). To address these challenges, we propose two complementary methods: iterative model editing, which applies successive edits to mitigate UnderEdit, and neighbor-assisted model editing, which incorporates neighboring knowledge during editing to reduce OverEdit. Our extensive experiments show that these techniques improve editing performance across multiple LLMs, algorithms, and benchmarks, reducing UnderEdit by up to 38 percentage points and OverEdit by up to 6, while remaining broadly applicable to any locate-and-edit method. We release our code at https://github.com/bhimanbaghel/ResolveUnderOverEdit.

Method: Uses doctor, patient, and measurement agents in multi-turn conversations; incorporates self-improvement mechanisms, multi-agent discussions, chain-of-thought reasoning, and experience-based knowledge retrieval; supports both automated and user-controlled modes.

Result: Comprehensive evaluations demonstrate the effectiveness of the approach in various simulated diagnostic scenarios, showing enhanced LLM performance through progressive learning.

Conclusion: MedAgentSim provides an effective framework for evaluating and improving LLM performance in clinical diagnostics, with the code, simulation tool, and benchmark publicly available.

Abstract: In this work, we introduce MedAgentSim, an open-source simulated clinical environment with doctor, patient, and measurement agents designed to evaluate and enhance LLM performance in dynamic diagnostic settings. Unlike prior approaches, our framework requires doctor agents to actively engage with patients through multi-turn conversations, requesting relevant medical examinations (e.g., temperature, blood pressure, ECG) and imaging results (e.g., MRI, X-ray) from a measurement agent to mimic the real-world diagnostic process. Additionally, we incorporate self improvement mechanisms that allow models to iteratively refine their diagnostic strategies. We enhance LLM performance in our simulated setting by integrating multi-agent discussions, chain-of-thought reasoning, and experience-based knowledge retrieval, facilitating progressive learning as doctor agents interact with more patients. We also introduce an evaluation benchmark for assessing the LLM’s ability to engage in dynamic, context-aware diagnostic interactions. While MedAgentSim is fully automated, it also supports a user-controlled mode, enabling human interaction with either the doctor or patient agent. Comprehensive evaluations in various simulated diagnostic scenarios demonstrate the effectiveness of our approach. Our code, simulation tool, and benchmark are available at \href{https://medagentsim.netlify.app/}.

Method: Designed improved cross-lingual retrieval systems trained with both sentence-level and word-level matching objectives, tested with three different RANMT architectures in controlled and real-world settings.

Result: Achieved performance matching standard TM-based models in controlled settings, and showed strong improvements over baseline and general-purpose cross-lingual retrievers in real-world settings with larger monolingual corpora.

Conclusion: Monolingual target language corpora can effectively replace bilingual resources in RANMT systems when combined with properly designed cross-lingual retrieval methods, offering a practical alternative when bilingual data is scarce.

Abstract: Conventional retrieval-augmented neural machine translation (RANMT) systems leverage bilingual corpora, e.g., translation memories (TMs). Yet, in many settings, monolingual corpora in the target language are often available. This work explores ways to take advantage of such resources by directly retrieving relevant target language segments, based on a source-side query. For this, we design improved cross-lingual retrieval systems, trained with both sentence level and word-level matching objectives. In our experiments with three RANMT architectures, we assess such cross-lingual objectives in a controlled setting, reaching performances that match those of standard TM-based models. We also showcase our method on a real-world settings, using much larger monolingual and observe strong improvements over both the baseline setting and general-purpose cross-lingual retrievers.

Method: Characterizes representation differences of ambiguous text in latent space, introduces path kernel-based distance measure over concepts to detect sentence-level ambiguity, and proposes missing concept prediction for improving ambiguous tool calling.

Result: Achieves state-of-the-art results in both ambiguity detection and improving LLM performance on ambiguous agentic tool calling.

Conclusion: The proposed approach effectively handles ambiguity in text-to-structured data mapping through novel distance measurement and concept-based methods, outperforming existing techniques.

Abstract: Ambiguity in natural language is a significant obstacle for achieving accurate text to structured data mapping through large language models (LLMs), which affects the performance of tasks such as mapping text to agentic tool calling and text-to-SQL queries. Existing methods to ambiguity handling either rely on the ReACT framework to obtain correct mappings through trial and error, or on supervised fine-tuning to bias models toward specific tasks. In this paper, we adopt a different approach that characterizes representation differences of ambiguous text in the latent space and leverages these differences to identify ambiguity before mapping them to structured data. To detect sentence-level ambiguity, we focus on the relationship between ambiguous questions and their interpretations. Unlike distances calculated by dense embeddings, we introduce a new distance measure based on a path kernel over concepts. With this measurement, we identify patterns to distinguish ambiguous from unambiguous questions. Furthermore, we propose a method for improving LLM performance on ambiguous agentic tool calling through missing concept prediction. Both achieve state-of-the-art results.

Method: Propose Generative query REwriter (GuRE) that trains Large Language Models to rewrite queries, helping retrievers find relevant legal passages more effectively.

Result: GuRE significantly improves retrieval performance in a retriever-agnostic manner, outperforming all baseline methods.

Conclusion: Query rewriting with LLMs is more suitable than direct retriever fine-tuning for real-world legal applications, with different training objectives leading to distinct retrieval behaviors.

Abstract: Legal Passage Retrieval (LPR) systems are crucial as they help practitioners save time when drafting legal arguments. However, it remains an underexplored avenue. One primary reason is the significant vocabulary mismatch between the query and the target passage. To address this, we propose a simple yet effective method, the Generative query REwriter (GuRE). We leverage the generative capabilities of Large Language Models (LLMs) by training the LLM for query rewriting. “Rewritten queries” help retrievers to retrieve target passages by mitigating vocabulary mismatch. Experimental results show that GuRE significantly improves performance in a retriever-agnostic manner, outperforming all baseline methods. Further analysis reveals that different training objectives lead to distinct retrieval behaviors, making GuRE more suitable than direct retriever fine-tuning for real-world applications. Codes are avaiable at github.com/daehuikim/GuRE.

Method: Proposes Gradient Interaction Modifications (GIM) technique that accounts for self-repair during backpropagation by addressing softmax redistribution in attention mechanisms.

Result: Extensive experiments across multiple LLMs (Gemma 2B/9B, LLAMA 1B/3B/8B, Qwen 1.5B/3B) show GIM significantly improves faithfulness over existing circuit identification and feature attribution methods.

Conclusion: GIM represents a significant step toward better understanding LLM inner mechanisms, which is crucial for improving model safety and reliability.

Abstract: Ensuring faithful interpretability in large language models is imperative for trustworthy and reliable AI. A key obstacle is self-repair, a phenomenon where networks compensate for reduced signal in one component by amplifying others, masking the true importance of the ablated component. While prior work attributes self-repair to layer normalization and back-up components that compensate for ablated components, we identify a novel form occurring within the attention mechanism, where softmax redistribution conceals the influence of important attention scores. This leads traditional ablation and gradient-based methods to underestimate the significance of all components contributing to these attention scores. We introduce Gradient Interaction Modifications (GIM), a technique that accounts for self-repair during backpropagation. Extensive experiments across multiple large language models (Gemma 2B/9B, LLAMA 1B/3B/8B, Qwen 1.5B/3B) and diverse tasks demonstrate that GIM significantly improves faithfulness over existing circuit identification and feature attribution methods. Our work is a significant step toward better understanding the inner mechanisms of LLMs, which is crucial for improving them and ensuring their safety. Our code is available at https://github.com/JoakimEdin/gim.

Method: Introduces a point-and-copy approach where the model identifies relevant image patches and copies their embeddings back into the reasoning stream, using semantic representations as keys to select patches directly.

Result: v1 consistently outperforms comparable baselines across various multimodal mathematical reasoning benchmarks.

Conclusion: Point-and-copy is established as a practical mechanism for grounded reasoning, with the model checkpoint and dataset made publicly available.

Abstract: When thinking with images, humans rarely rely on a single glance: they revisit visual information repeatedly during reasoning. However, existing models typically process images only once and thereafter generate reasoning entirely in text, lacking mechanisms to re-access or ground inference in visual representations. We empirically confirm this: as reasoning chains lengthen, models progressively lose focus on relevant regions. In response, we introduce v1, a lightweight extension that enables active visual referencing through a simple point-and-copy approach. This allows the model to identify relevant image patches and copy their embeddings back into the reasoning stream, ensuring that evolving hypotheses remain grounded in perceptual evidence. Crucially, our pointing strategy lets the MLLM directly select image patches using their semantic representations as keys, keeping perceptual evidence embedded in the same space as the model’s reasoning. To train this capability, we construct v1g, a dataset of 300K multimodal reasoning traces with interleaved visual grounding annotations. Across various multimodal mathematical reasoning benchmarks, v1 consistently outperforms comparable baselines, establishing point-and-copy as a practical mechanism for grounded reasoning. The model checkpoint and dataset are available at github.com/jun297/v1.

Method: Testing human-scale transformer language models on a bespoke synthetic benchmark that targets syntactic and semantic properties of the LET-ALONE construction, filtering out related constructions from the dataset.

Result: Human-scale LMs are sensitive to the form of the LET-ALONE construction but do not make correct generalizations about its meaning, even when related constructions are excluded.

Conclusion: Current language model architectures show an asymmetry in sample efficiency between learning language form versus meaning, which differs from human language learners who efficiently acquire both.

Abstract: Humans have a remarkable ability to acquire and understand grammatical phenomena that are seen rarely, if ever, during childhood. Recent evidence suggests that language models with human-scale pretraining data may possess a similar ability by generalizing from frequent to rare constructions. However, it remains an open question how widespread this generalization ability is, and to what extent this knowledge extends to meanings of rare constructions, as opposed to just their forms. We fill this gap by testing human-scale transformer language models on their knowledge of both the form and meaning of the (rare and quirky) English LET-ALONE construction. To evaluate our LMs we construct a bespoke synthetic benchmark that targets syntactic and semantic properties of the construction. We find that human-scale LMs are sensitive to form, even when related constructions are filtered from the dataset. However, human-scale LMs do not make correct generalizations about LET-ALONE’s meaning. These results point to an asymmetry in the current architectures’ sample efficiency between language form and meaning, something which is not present in human language learners.

Method: Proposed approach uses parameter isolation and an MLLM-based routing mechanism to prevent catastrophic interference during continual learning.

Result: Extensive experiments show the approach can integrate domain-specific knowledge and functional abilities with minimal forgetting, significantly outperforming existing methods.

Conclusion: MLLM-CL provides a comprehensive benchmark for continual learning in multimodal models, and the proposed method effectively addresses catastrophic forgetting while enabling continuous knowledge integration.

Abstract: Recent Multimodal Large Language Models (MLLMs) excel in vision-language understanding but face challenges in adapting to dynamic real-world scenarios that require continuous integration of new knowledge and skills. While continual learning (CL) offers a potential solution, existing benchmarks and methods suffer from critical limitations. In this paper, we introduce MLLM-CL, a novel benchmark encompassing domain and ability continual learning, where the former focuses on independently and identically distributed (IID) evaluation across evolving mainstream domains, whereas the latter evaluates on non-IID scenarios with new model abilities. Methodologically, we propose preventing catastrophic interference through parameter isolation and an MLLM-based routing mechanism. Extensive experiments demonstrate that our approach can integrate domain-specific knowledge and functional abilities with minimal forgetting, significantly outperforming existing methods. Our benchmark and code are available at https://github.com/bjzhb666/MLLM-CL.

Method: Proposes Precise Information Control (PIC) task with full and partial settings, creates PIC-Bench benchmark with 8 long-form generation tasks, and introduces a post-training framework using weakly supervised preference data to train PIC-LM.

Result: State-of-the-art LMs hallucinate against user-provided input in over 70% of generations. The trained 8B PIC-LM improves from 69.1% to 91.0% F1 in full PIC setting, and improves exact match recall by 17.1% on ambiguous QA and factual precision by 30.5% on birthplace fact-checking.

Conclusion: The PIC framework and PIC-LM demonstrate significant improvements in faithful generation, showing the potential of precisely grounded generation for reducing hallucinations in language models.

Abstract: A central challenge in language models (LMs) is faithfulness hallucination: the generation of information unsubstantiated by input context. To study this problem, we propose Precise Information Control (PIC), a new task formulation that requires models to generate long-form outputs grounded in a provided set of short self-contained statements, without adding any unsupported ones. PIC includes a full setting that tests a model’s ability to include exactly all input claims, and a partial setting that requires the model to selectively incorporate only relevant claims. We present PIC-Bench, a benchmark of eight long-form generation tasks (e.g., summarization, biography generation) adapted to the PIC setting, where LMs are supplied with well-formed, verifiable input claims. Our evaluation of a range of open and proprietary LMs on PIC-Bench reveals that, surprisingly, state-of-the-art LMs still hallucinate against user-provided input in over 70% of generations. To alleviate this lack of faithfulness, we introduce a post-training framework that uses a weakly supervised preference data construction method to train an 8B PIC-LM with stronger PIC ability–improving from 69.1% to 91.0% F1 in the full PIC setting. When integrated into end-to-end factual generation pipelines, PIC-LM improves exact match recall by 17.1% on ambiguous QA with retrieval, and factual precision by 30.5% on a birthplace fact-checking task, underscoring the potential of precisely grounded generation.

Method: Extensive experiments on Qwen2.5, LLaMA3 and Gemma3 model families, analyzing performance degradation across different model sizes and training data scales (8k-220k examples).

Result: Widespread Long CoT Degradation observed in SLMs, with some models losing 75% performance after training on only 8k long CoT examples. Some small models fail to recover original performance even with 220k examples. Error accumulation identified as root cause.

Conclusion: Long CoT training can be detrimental for small language models due to error accumulation, challenging common assumptions about its benefits. Practical guidance provided for building effective small-scale reasoning models.

Abstract: Long chain-of-thought (CoT) supervision has become a common strategy to enhance reasoning in language models. While effective for large models, we identify a phenomenon we call Long CoT Degradation, in which small language models (SLMs; <=3B parameters) trained on limited long CoT data experience significant performance deterioration. Through extensive experiments on the Qwen2.5, LLaMA3 and Gemma3 families, we demonstrate that this degradation is widespread across SLMs. In some settings, models trained on only 8k long CoT examples lose up to 75% of their original performance before fine-tuning. Strikingly, we further observe that for some particularly small models, even training on 220k long CoT examples fails to recover or surpass their original performance prior to fine-tuning. Our analysis attributes this effect to error accumulation: while longer responses increase the capacity for multi-step reasoning, they also amplify the risk of compounding mistakes. Furthermore, we find that Long CoT Degradation may negatively impacts downstream reinforcement learning (RL), although this can be alleviated by sufficiently scaled supervised fine-tuning (SFT). Our findings challenge common assumptions about the benefits of long CoT training for SLMs and offer practical guidance for building more effective small-scale reasoning models.

Method: REAL trains VQ-AEs on each module’s hidden activations with a shared learnable codebook to partition latent space into behavior-relevant and irrelevant subspaces. It quantifies behavioral relevance by how well VQ-AE encodings discriminate behavior-aligned vs behavior-violating responses via binary classification.

Result: REAL achieved 20% average relative improvement (up to 81.5%) over ITI method on truthfulness steering across 8 LLMs and 9 datasets. Selected modules also showed strong zero-shot generalization in cross-domain truthfulness steering.

Conclusion: REAL provides a systematic framework for identifying behavior-relevant modules in LLMs, enabling more effective inference-time interventions with strong generalization capabilities across different domains and models.

Abstract: Inference-time steering aims to alter a large language model’s (LLM’s) responses without changing its parameters, but a central challenge is identifying the internal modules that most strongly govern the target behavior. Existing approaches often rely on simplistic cues or ad hoc heuristics, leading to suboptimal or unintended effects. We introduce REAL, a framework for identifying behavior-relevant modules (attention heads or layers) in Transformer models. For each module, REAL trains a vector-quantized autoencoder (VQ-AE) on its hidden activations and uses a shared, learnable codebook to partition the latent space into behavior-relevant and behavior-irrelevant subspaces. REAL quantifies a module’s behavioral relevance by how well its VQ-AE encodings discriminate behavior-aligned from behavior-violating responses via a binary classification metric; this score guides both module selection and steering strength. We evaluate REAL across eight LLMs from the Llama and Qwen families and nine datasets spanning truthfulness enhancement, open-domain QA under knowledge conflicts, and general alignment tasks. REAL enables more effective inference-time interventions, achieving an average relative improvement of 20% (up to 81.5%) over the ITI method on truthfulness steering. In addition, the modules selected by REAL exhibit strong zero-shot generalization in cross-domain truthfulness-steering scenarios.

Method: Created CounselBench with two components: CounselBench-EVAL (2,000 expert evaluations of LLM and human therapist answers across six clinical dimensions) and CounselBench-Adv (120 adversarial questions to trigger specific model failures).

Result: LLMs achieve high scores on some dimensions but exhibit recurring issues including unconstructive feedback, overgeneralization, limited personalization, and safety risks like unauthorized medical advice. LLM judges systematically overrate responses and miss human-identified safety concerns.

Conclusion: CounselBench establishes a clinically grounded framework for benchmarking LLMs in mental health QA, revealing consistent failure patterns and highlighting the need for human expert evaluation to identify safety risks.

Abstract: Medical question answering (QA) benchmarks often focus on multiple-choice or fact-based tasks, leaving open-ended answers to real patient questions underexplored. This gap is particularly critical in mental health, where patient questions often mix symptoms, treatment concerns, and emotional needs, requiring answers that balance clinical caution with contextual sensitivity. We present CounselBench, a large-scale benchmark developed with 100 mental health professionals to evaluate and stress-test large language models (LLMs) in realistic help-seeking scenarios. The first component, CounselBench-EVAL, contains 2,000 expert evaluations of answers from GPT-4, LLaMA 3, Gemini, and human therapists on patient questions from the public forum CounselChat. Each answer is rated across six clinically grounded dimensions, with span-level annotations and written rationales. Expert evaluations show that while LLMs achieve high scores on several dimensions, they also exhibit recurring issues, including unconstructive feedback, overgeneralization, and limited personalization or relevance. Responses were frequently flagged for safety risks, most notably unauthorized medical advice. Follow-up experiments show that LLM judges systematically overrate model responses and overlook safety concerns identified by human experts. To probe failure modes more directly, we construct CounselBench-Adv, an adversarial dataset of 120 expert-authored mental health questions designed to trigger specific model issues. Evaluation of 3,240 responses from nine LLMs reveals consistent, model-specific failure patterns. Together, CounselBench establishes a clinically grounded framework for benchmarking LLMs in mental health QA.

Method: Used probing and agentic evaluations to study privacy leakage in reasoning traces, particularly examining how test-time compute approaches (especially increased reasoning steps) affect data leakage.

Result: Found that reasoning traces frequently contain sensitive user data, and increased reasoning steps amplify such leakage. While more reasoning makes models more cautious in final answers, it also leads to more verbose reasoning that leaks more information.

Conclusion: Safety efforts must extend to the model’s internal thinking processes, not just its outputs, as there is a fundamental tension between reasoning for utility and the enlarged privacy attack surface.

Abstract: We study privacy leakage in the reasoning traces of large reasoning models used as personal agents. Unlike final outputs, reasoning traces are often assumed to be internal and safe. We challenge this assumption by showing that reasoning traces frequently contain sensitive user data, which can be extracted via prompt injections or accidentally leak into outputs. Through probing and agentic evaluations, we demonstrate that test-time compute approaches, particularly increased reasoning steps, amplify such leakage. While increasing the budget of those test-time compute approaches makes models more cautious in their final answers, it also leads them to reason more verbosely and leak more in their own thinking. This reveals a core tension: reasoning improves utility but enlarges the privacy attack surface. We argue that safety efforts must extend to the model’s internal thinking, not just its outputs.

Method: Refined frame labels, developed framework to reconstruct dominant frames from sentence-level predictions, and aligned articles with topically relevant comments across eleven topics and two news outlets.

Result: Frame reuse in comments correlates highly across outlets, while topic-specific patterns vary. Released frame classifier, manually labeled dataset, and large-scale dataset with predicted frame labels.

Conclusion: Successfully created the first computational framework for analyzing framing across source content and audience responses, revealing systematic patterns in frame reuse across different contexts.

Abstract: When a news article describes immigration as an “economic burden” or a “humanitarian crisis,” it selectively emphasizes certain aspects of the issue. Although \textit{framing} shapes how the public interprets such issues, audiences do not absorb frames passively but actively reorganize the presented information. While this relationship between source content and audience response is well-documented in the social sciences, NLP approaches often ignore it, detecting frames in articles and responses in isolation. We present the first computational framework for large-scale analysis of framing across source content (news articles) and audience responses (reader comments). Methodologically, we refine frame labels and develop a framework that reconstructs dominant frames in articles and comments from sentence-level predictions, and aligns articles with topically relevant comments. Applying our framework across eleven topics and two news outlets, we find that frame reuse in comments correlates highly across outlets, while topic-specific patterns vary. We release a frame classifier that performs well on both articles and comments, a dataset of article and comment sentences manually labeled for frames, and a large-scale dataset of articles and comments with predicted frame labels.

Method: Examines pretrained and tuned models using code-mixed coreferential statements; leverages interpretability approaches; tests code-switching training and cross-lingual word alignment objectives; experiments with activation patching for test-time consistency calibration.

Result: Shows varying consistency levels influenced by language families, linguistic factors, scripts, and specific model layers; code-switching training and cross-lingual alignment yield most promising results for improving consistency and multilingual performance.

Conclusion: Cross-lingual alignment supervision and code-switching strategies are valuable for enhancing both multilingual performance and cross-lingual consistency, with activation patching showing potential for test-time consistency calibration.

Abstract: Cross-lingual consistency should be considered to assess cross-lingual transferability, maintain the factuality of the model knowledge across languages, and preserve the parity of language model performance. We are thus interested in analyzing, evaluating, and interpreting cross-lingual consistency for factual knowledge. To facilitate our study, we examine multiple pretrained models and tuned models with code-mixed coreferential statements that convey identical knowledge across languages. Interpretability approaches are leveraged to analyze the behavior of a model in cross-lingual contexts, showing different levels of consistency in multilingual models, subject to language families, linguistic factors, scripts, and a bottleneck in cross-lingual consistency on a particular layer. Code-switching training and cross-lingual word alignment objectives show the most promising results, emphasizing the worthiness of cross-lingual alignment supervision and code-switching strategies for both multilingual performance and cross-lingual consistency enhancement. In addition, experimental results suggest promising result for calibrating consistency in the test time via activation patching.

Method: Provides a library with user interface and modular backend supporting operations like searching, viewing, ingesting, exporting, inspecting, and sampling data for pretraining datasets.

Result: TokenSmith enables structured editing of pretraining data without requiring training code changes, simplifying dataset debugging, validation, and experimentation.

Conclusion: TokenSmith democratizes access to production-grade dataset tooling as a plug-and-play addition to existing LLM pretraining workflows.

Abstract: Understanding the relationship between training data and model behavior during pretraining is crucial, but existing workflows make this process cumbersome, fragmented, and often inaccessible to researchers. We present TokenSmith, an open-source library for interactive editing, inspection, and analysis of datasets used in Megatron-style pretraining frameworks such as GPT-NeoX, Megatron, and NVIDIA NeMo. TokenSmith supports a wide range of operations including searching, viewing, ingesting, exporting, inspecting, and sampling data, all accessible through a simple user interface and a modular backend. It also enables structured editing of pretraining data without requiring changes to training code, simplifying dataset debugging, validation, and experimentation. TokenSmith is designed as a plug-and-play addition to existing large language model pretraining workflows, thereby democratizing access to production-grade dataset tooling. TokenSmith is hosted on GitHub, with accompanying documentation, tutorials, and a demonstration video (available on YouTube).

Method: Created a cross-linguistic benchmark with 800 unique prompts across 5 gendered languages (French, Spanish, German, Italian, Russian) and 2 gender-neutral control languages (English, Chinese), generating 28,800 images using three state-of-the-art T2I models.

Result: Grammatical gender dramatically influences image generation: masculine grammatical markers increase male representation to 73% (vs 22% in English), feminine markers increase female representation to 38% (vs 28% in English). Effects vary by language resource availability and model architecture.

Conclusion: Language structure itself, not just content, shapes AI-generated visual outputs, introducing a new dimension for understanding bias and fairness in multilingual, multimodal systems.

Abstract: Research on bias in Text-to-Image (T2I) models has primarily focused on demographic representation and stereotypical attributes, overlooking a fundamental question: how does grammatical gender influence visual representation across languages? We introduce a cross-linguistic benchmark examining words where grammatical gender contradicts stereotypical gender associations (e.g., une sentinelle'' - grammatically feminine in French but referring to the stereotypically masculine concept guard’’). Our dataset spans five gendered languages (French, Spanish, German, Italian, Russian) and two gender-neutral control languages (English, Chinese), comprising 800 unique prompts that generated 28,800 images across three state-of-the-art T2I models. Our analysis reveals that grammatical gender dramatically influences image generation: masculine grammatical markers increase male representation to 73% on average (compared to 22% with gender-neutral English), while feminine grammatical markers increase female representation to 38% (compared to 28% in English). These effects vary systematically by language resource availability and model architecture, with high-resource languages showing stronger effects. Our findings establish that language structure itself, not just content, shapes AI-generated visual outputs, introducing a new dimension for understanding bias and fairness in multilingual, multimodal systems.

Method: Benchmarked generic multilingual ASR models and models fine-tuned for Dutch spoken by older adults, considering processing speed and using the Welzijn.AI chatbot designed for geriatric contexts.

Result: Generic multilingual models outperformed fine-tuned models, suggesting recent ASR models can generalize well to real-world datasets. Truncating generic models helps balance accuracy-speed trade-off, though some inputs still cause high word error rates.

Conclusion: Recent ASR models show good generalization capabilities for underrepresented groups like older Dutch adults, with truncation providing an effective strategy for managing accuracy-speed trade-offs in clinical applications.

Abstract: Voice-controlled interfaces can support older adults in clinical contexts – with chatbots being a prime example – but reliable Automatic Speech Recognition (ASR) for underrepresented groups remains a bottleneck. This study evaluates state-of-the-art ASR models on language use of older Dutch adults, who interacted with the Welzijn.AI chatbot designed for geriatric contexts. We benchmark generic multilingual ASR models, and models fine-tuned for Dutch spoken by older adults, while also considering processing speed. Our results show that generic multilingual models outperform fine-tuned models, which suggests recent ASR models can generalise well out of the box to real-world datasets. Moreover, our results indicate that truncating generic models is helpful in balancing the accuracy-speed trade-off. Nonetheless, we also find inputs which cause a high word error rate and place them in context.

Method: Benchmarked multiple LLMs, augmented PalmX dataset with Palm dataset and curated 22K+ culturally grounded MCQs, then performed LoRA fine-tuning of the best-performing Fanar-1-9B-Instruct model on the combined dataset.

Result: Fanar-1-9B-Instruct achieved highest performance; submitted system ranked 5th with 70.50% accuracy on blind test set and 84.1% accuracy on PalmX development set.

Conclusion: Data augmentation and LoRA fine-tuning effectively improve LLM performance on Arabic cultural knowledge tasks, with Fanar-1-9B-Instruct emerging as the most suitable model for this domain.

Abstract: In this paper, we report our participation to the PalmX cultural evaluation shared task. Our system, CultranAI, focused on data augmentation and LoRA fine-tuning of large language models (LLMs) for Arabic cultural knowledge representation. We benchmarked several LLMs to identify the best-performing model for the task. In addition to utilizing the PalmX dataset, we augmented it by incorporating the Palm dataset and curated a new dataset of over 22K culturally grounded multiple-choice questions (MCQs). Our experiments showed that the Fanar-1-9B-Instruct model achieved the highest performance. We fine-tuned this model on the combined augmented dataset of 22K+ MCQs. On the blind test set, our submitted system ranked 5th with an accuracy of 70.50%, while on the PalmX development set, it achieved an accuracy of 84.1%.

Method: Proposes FASB framework that dynamically determines intervention necessity and strength by tracking LLM internal states during generation, considering both question and generated content. Includes backtracking mechanism to correct deviated tokens.

Result: Extensive experiments on TruthfulQA and six multiple-choice datasets show FASB outperforms baselines.

Conclusion: FASB provides an effective approach for aligning LLM behaviors through dynamic activation steering with backtracking, achieving better performance than existing methods.

Abstract: Large language models (LLMs) have achieved remarkable performance across many generation tasks. Nevertheless, effectively aligning them with desired behaviors remains a significant challenge. Activation steering is an effective and cost-efficient approach that directly modifies the activations of LLMs during the inference stage, aligning their responses with the desired behaviors and avoiding the high cost of fine-tuning. Existing methods typically indiscriminately intervene to all generations or rely solely on the question to determine intervention, which limits the accurate assessment of the intervention strength. To this end, we propose the Flexible Activation Steering with Backtracking (FASB) framework, which dynamically determines both the necessity and strength of intervention by tracking the internal states of the LLMs during generation, considering both the question and the generated content. Since intervening after detecting a deviation from the desired behavior is often too late, we further propose the backtracking mechanism to correct the deviated tokens and steer the LLMs toward the desired behavior. Extensive experiments on the TruthfulQA dataset and six multiple-choice datasets demonstrate that our method outperforms baselines. Our code will be released at https://github.com/gjw185/FASB.

Method: Created DiFBench, a configurable benchmark framework with controllable parameters (document set size, distinctiveness thresholds). Evaluated 10 state-of-the-art LLMs on DFM task using this benchmark.

Result: Significant performance gap between general-purpose and reasoning-enhanced models. All models degrade substantially with increased task complexity and document count. Common failure mode is misidentifying frequent features as distinctive.

Conclusion: Contemporary LLMs have core limitations in fine-grained statistical reasoning and rarity detection, revealing gaps in their ability to perform distinctive feature mining effectively.

Abstract: Effective decision-making often relies on identifying what makes each candidate distinctive. While existing benchmarks for LLMs emphasize retrieving or summarizing information relevant to a given query, they do not evaluate a model’s ability to identify globally distinctive features across a set of documents. We introduce Distinctive Feature Mining (DFM), a new task that challenges models to analyze a small-to-medium collection (10-40 documents) and surface features that are rare in the global context (e.g., appearing in less than 10% of documents). This setting mirrors real-world scenarios such as candidate selection or product differentiation, where statistical reasoning, not retrieval, is key. To enable systematic evaluation of this capability, we present DiFBench, a configurable benchmark creation framework with controllable parameters such as document set size and distinctiveness thresholds. Using DiFBench, we perform a large-scale assessment of distinctive feature mining across ten state-of-the-art LLMs. Our findings reveal a significant performance gap between general-purpose and reasoning-enhanced models. All models, however, substantially degrade as the task complexity and document count increase. We also find that a common failure mode is misidentifying frequent features as distinctive. These insights reveal core limitations in contemporary LLMs' abilities to perform fine-grained, statistical reasoning and rarity detection.

Method: Hybrid approach combining PRISMA guidelines with fuzzy matching and regex for semi-automated deduplication and filtering. Modified splitting method for multi-arm trials. Applied to RCTs on GnRH-a efficacy for endometriosis.

Result: Workflow reduced screening workload significantly (33,444 records processed in 11 days). Pooled analysis of 7 RCTs (841 patients) showed RR=0.64 (95% CI 0.48-0.86), indicating 36% reduction in recurrence with no heterogeneity.

Conclusion: Information-retrieval-driven workflow successfully bridges clinical research and computer science, providing generalizable framework for scalable evidence synthesis with robust clinical results.

Abstract: Background: Evidence synthesis facilitates evidence-based medicine. This task becomes increasingly difficult to accomplished with applying computational solutions, since the medical literature grows at astonishing rates. Objective: This study evaluates an information retrieval-driven workflow, CASMA, to enhance the efficiency, transparency, and reproducibility of systematic reviews. Endometriosis recurrence serves as the ideal case due to its complex and ambiguous literature. Methods: The hybrid approach integrates PRISMA guidelines with fuzzy matching and regular expression (regex) to facilitate semi-automated deduplication and filtered records before manual screening. The workflow synthesised evidence from randomised controlled trials on the efficacy of a subclass of gonadotropin-releasing hormone agonists (GnRH-a). A modified splitting method addressed unit-of-analysis errors in multi-arm trials. Results: The workflow sharply reduced the screening workload, taking only 11 days to fetch and filter 33,444 records. Seven eligible RCTs were synthesized (841 patients). The pooled random-effects model yielded a Risk Ratio (RR) of $0.64$ ($95%$ CI $0.48$ to $0.86$), demonstrating a $36%$ reduction in recurrence, with non-significant heterogeneity ($I^2=0.00%$, $\tau^2=0.00$). The findings were robust and stable, as they were backed by sensitivity analyses. Conclusion: This study demonstrates an application of an information-retrieval-driven workflow for medical evidence synthesis. The approach yields valuable clinical results and a generalisable framework to scale up the evidence synthesis, bridging the gap between clinical research and computer science.

Method: SPEED employs specialized functional experts to perform descriptive analyses across multiple dimensions including hallucination detection, toxicity assessment, and lexical-contextual appropriateness, incorporating expert feedback.

Result: SPEED achieves robust and consistent evaluation performance across diverse domains and datasets while demonstrating superior resource efficiency with compact expert models compared to larger-scale evaluators.

Conclusion: SPEED significantly enhances fairness and interpretability in LLM evaluations, offering a promising alternative to existing evaluation methodologies.

Abstract: Reliable evaluation of large language models is essential to ensure their applicability in practical scenarios. Traditional benchmark-based evaluation methods often rely on fixed reference answers, limiting their ability to capture important qualitative aspects of generated responses. To address these shortcomings, we propose an integrated evaluation framework called \textit{self-refining descriptive evaluation with expert-driven diagnostics}, SPEED, which utilizes specialized functional experts to perform comprehensive, descriptive analyses of model outputs. Unlike conventional approaches, SPEED actively incorporates expert feedback across multiple dimensions, including hallucination detection, toxicity assessment, and lexical-contextual appropriateness. Experimental results demonstrate that SPEED achieves robust and consistent evaluation performance across diverse domains and datasets. Additionally, by employing relatively compact expert models, SPEED demonstrates superior resource efficiency compared to larger-scale evaluators. These findings illustrate that SPEED significantly enhances fairness and interpretability in LLM evaluations, offering a promising alternative to existing evaluation methodologies.

Method: Developed TACOS taxonomy that explicitly models varying safety thresholds and external tool dependencies, curated a TACOS-annotated dataset, and performed extensive experiments.

Result: The taxonomy demonstrates value for clinical agent settings and reveals insights about train data distribution and pretrained knowledge of base models.

Conclusion: TACOS provides a comprehensive safety framework specialized for clinical agents that integrates safety filtering and tool selection through fine-grained intent classification.

Abstract: Safety is a paramount concern in clinical chatbot applications, where inaccurate or harmful responses can lead to serious consequences. Existing methods–such as guardrails and tool calling–often fall short in addressing the nuanced demands of the clinical domain. In this paper, we introduce TACOS (TAxonomy of COmprehensive Safety for Clinical Agents), a fine-grained, 21-class taxonomy that integrates safety filtering and tool selection into a single user intent classification step. TACOS is a taxonomy that can cover a wide spectrum of clinical and non-clinical queries, explicitly modeling varying safety thresholds and external tool dependencies. To validate our taxonomy, we curate a TACOS-annotated dataset and perform extensive experiments. Our results demonstrate the value of a new taxonomy specialized for clinical agent settings, and reveal useful insights about train data distribution and pretrained knowledge of base models.

Method: The framework implements Orchestrated Test-Time Scaling with three perspectives: Internal Scaling via RL-enhanced Intrinsic Reasoning, Sequential Scaling through Iterative Refinement, and Parallel Scaling using Diverse Synthesis and Tournament Selection.

Result: Agentar-Scale-SQL achieves 81.67% execution accuracy on the BIRD test set and ranks first on the official leaderboard, demonstrating state-of-the-art performance.

Conclusion: The framework provides an effective path toward human-level performance in Text-to-SQL tasks and is designed as a general-purpose solution adaptable to new databases and more powerful language models.

Abstract: State-of-the-art (SOTA) Text-to-SQL methods still lag significantly behind human experts on challenging benchmarks like BIRD. Current approaches that explore test-time scaling lack an orchestrated strategy and neglect the model’s internal reasoning process. To bridge this gap, we introduce Agentar-Scale-SQL, a novel framework leveraging scalable computation to improve performance. Agentar-Scale-SQL implements an Orchestrated Test-Time Scaling strategy that synergistically combines three distinct perspectives: i) Internal Scaling via RL-enhanced Intrinsic Reasoning, ii) Sequential Scaling through Iterative Refinement, and iii) Parallel Scaling using Diverse Synthesis and Tournament Selection. Agentar-Scale-SQL is a general-purpose framework designed for easy adaptation to new databases and more powerful language models. Extensive experiments show that Agentar-Scale-SQL achieves SOTA performance on the BIRD benchmark, reaching 81.67% execution accuracy on the test set and ranking first on the official leaderboard, demonstrating an effective path toward human-level performance.

Method: Three approaches were compared: retrieval-augmented generation (RAG) using codebooks, prompt engineering (zero-shot, few-shot, chain-of-thought), and fine-tuning on hand-coded texts.

Result: Fine-tuned LLMs achieved median F1 score of 0.79. Discrepancies between human and LLM outputs were systematic and reflected known theoretical challenges in metaphor identification.

Conclusion: LLMs can partially automate metaphor identification and serve as testbeds for refining metaphor identification protocols and underlying theory.

Abstract: Metaphor is a pervasive feature of discourse and a powerful lens for examining cognition, emotion, and ideology. Large-scale analysis, however, has been constrained by the need for manual annotation due to the context-sensitive nature of metaphor. This study investigates the potential of large language models (LLMs) to automate metaphor identification in full texts. We compare three methods: (i) retrieval-augmented generation (RAG), where the model is provided with a codebook and instructed to annotate texts based on its rules and examples; (ii) prompt engineering, where we design task-specific verbal instructions; and (iii) fine-tuning, where the model is trained on hand-coded texts to optimize performance. Within prompt engineering, we test zero-shot, few-shot, and chain-of-thought strategies. Our results show that state-of-the-art closed-source LLMs can achieve high accuracy, with fine-tuning yielding a median F1 score of 0.79. A comparison of human and LLM outputs reveals that most discrepancies are systematic, reflecting well-known grey areas and conceptual challenges in metaphor theory. We propose that LLMs can be used to at least partly automate metaphor identification and can serve as a testbed for developing and refining metaphor identification protocols and the theory that underpins them.

Method: Carefully curate and resample open-source datasets using designed metrics to identify beneficial data, then pre-train on 4.2T tokens from these ~2T high-quality tokens followed by established post-training procedures.

Result: MobileLLM-R1-950M achieves AIME score of 15.5 (vs 0.6 for OLMo-2-1.48B and 0.3 for SmolLM-2-1.7B) and matches/surpasses Qwen3-0.6B across multiple reasoning benchmarks despite using only 11.7% of Qwen3’s training tokens.

Conclusion: High-quality data curation is more crucial than massive data scaling for reasoning emergence, enabling efficient sub-billion-parameter reasoning models with substantially reduced training data requirements.

Abstract: The paradigm shift in large language models (LLMs) from instinctive responses to chain-of-thought (CoT) reasoning has fueled two prevailing assumptions: (1) reasoning capabilities only emerge in sufficiently large models, and (2) such capabilities require training on massive datasets. While the first assumption has already been challenged by recent sub-billion-parameter reasoning models such as Qwen3-0.6B and DeepSeek distilled variants, the second remains largely unquestioned. In this work, we revisit the necessity of scaling to extremely large corpora (>10T tokens) for reasoning emergence. By carefully curating and resampling open-source datasets that we identify as beneficial under our designed metrics, we demonstrate that strong reasoning abilities can emerge with far less data. Specifically, we show that only ~2T tokens of high-quality data are sufficient, and pre-training with 4.2T tokens on the dataset resampled from these ~2T tokens, followed by a established post-training procedure, enables the development of MobileLLM-R1, a series of sub-billion-parameter reasoning models that substantially outperform prior models trained on fully open-sourced data. For example, MobileLLM-R1-950M achieves an AIME score of 15.5, compared to just 0.6 for OLMo-2-1.48B and 0.3 for SmolLM-2-1.7B. Remarkably, despite being trained on only 11.7% of the tokens compared to Qwen3’s proprietary 36T-token corpus for pretraining, MobileLLM-R1-950M matches or surpasses Qwen3-0.6B across multiple reasoning benchmarks. To facilitate further research in this direction, we have released the complete training recipe, data sources, data mixing ratio, and model checkpoints, together with the key insights obtained throughout this study.

Method: Initializes from pre-trained discrete diffusion language models and distills few-step students using integral KL-divergence minimization framework. Introduces grouped reward normalization, intermediate-state matching, and reward-guided ancestral sampler for improved training stability and inference quality.

Result: Achieves perplexity from 62.2 (8 NFEs) to 18.4 (128 NFEs) on OpenWebText, outperforming prior accelerated dLLMs and GPT-2 baseline. Reduces training time by 20x compared to competing methods with only 1% entropy loss. Validated through ablation studies, model scaling, and protein sequence generation.

Conclusion: DiDi-Instruct is an efficient and effective distillation method that enables fast language generation while maintaining quality, making it suitable for practical applications requiring rapid inference.

Abstract: Fast and high-quality language generation is the holy grail that people pursue in the age of AI. In this work, we introduce Discrete Diffusion Divergence Instruct (DiDi-Instruct), a training-based method that initializes from a pre-trained (masked) discrete diffusion language model (dLLM) and distills a few-step student for fast generation. The resulting DiDi-Instruct model achieves comparable or superior performance to its dLLM teacher and the GPT-2 baseline while enabling up to 64$\times$ acceleration. The theoretical foundation of DiDi-Instruct is a novel framework based on integral KL-divergence minimization, which yields a practical training algorithm. We further introduce grouped reward normalization, intermediate-state matching, and the reward-guided ancestral sampler that significantly improve training stability, model coverage, and inference quality. On OpenWebText, DiDi-Instruct achieves perplexity from 62.2 (8 NFEs) to 18.4 (128 NFEs), which outperforms prior accelerated dLLMs and GPT-2 baseline. These gains come with a negligible entropy loss (around $1%$) and reduce additional training wall-clock time by more than $20\times$ compared to competing dLLM distillation methods. We further validate the robustness and effectiveness of DiDi-Instruct through extensive ablation studies, model scaling, and the generation of discrete protein sequences. In conclusion, DiDi-Instruct is an efficient yet effective distillation method, enabling language generation in the blink of an eye. We will release both code and models at github.com/haoyangzheng-ai/didi-instruct.

Method: Introduces ’last but not late interaction’ - conducts causal self-attention between query and documents within the same context window, then extracts contextual embeddings from the last token of each document, unlike late interaction models that perform separate encoding followed by multi-vector matching.

Result: Achieves state-of-the-art BEIR performance with 61.94 nDCG@10 while being significantly smaller than generative listwise rerankers.

Conclusion: The proposed ’last but not late interaction’ approach enables efficient and effective document reranking with rich cross-document interactions in a compact 0.6B parameter model.

Abstract: jina-reranker-v3 is a 0.6B parameter multilingual document reranker that introduces a novel last but not late interaction. Unlike late interaction models such as ColBERT that perform separate encoding followed by multi-vector matching, our approach conducts causal self-attention between query and documents within the same context window, enabling rich cross-document interactions before extracting contextual embeddings from the last token of each document. This compact architecture achieves state-of-the-art BEIR performance with 61.94 nDCG@10 while being significant smaller than generative listwise rerankers.

Method: Uses pairwise learning within domain-year groups to reduce rating inconsistencies, introduces Review Tendency Signal (RTS) as probabilistic integration of reviewer scores and confidences, and trains on NAIDv2 dataset of 24,276 ICLR submissions with metadata and structured content.

Result: Achieves state-of-the-art performance (78.2% AUC, 0.432 Spearman) with linear-time inference efficiency. Demonstrates strong generalization on unseen NeurIPS submissions, with predicted scores increasing consistently from Rejected to Oral decision categories.

Conclusion: NAIPv2 establishes a debiased and scalable framework for automated paper quality estimation, representing progress toward future scientific intelligence systems.

Abstract: The ability to estimate the quality of scientific papers is central to how both humans and AI systems will advance scientific knowledge in the future. However, existing LLM-based estimation methods suffer from high inference cost, whereas the faster direct score regression approach is limited by scale inconsistencies. We present NAIPv2, a debiased and efficient framework for paper quality estimation. NAIPv2 employs pairwise learning within domain-year groups to reduce inconsistencies in reviewer ratings and introduces the Review Tendency Signal (RTS) as a probabilistic integration of reviewer scores and confidences. To support training and evaluation, we further construct NAIDv2, a large-scale dataset of 24,276 ICLR submissions enriched with metadata and detailed structured content. Trained on pairwise comparisons but enabling efficient pointwise prediction at deployment, NAIPv2 achieves state-of-the-art performance (78.2% AUC, 0.432 Spearman), while maintaining scalable, linear-time efficiency at inference. Notably, on unseen NeurIPS submissions, it further demonstrates strong generalization, with predicted scores increasing consistently across decision categories from Rejected to Oral. These findings establish NAIPv2 as a debiased and scalable framework for automated paper quality estimation, marking a step toward future scientific intelligence systems. Code and dataset are released at sway.cloud.microsoft/Pr42npP80MfPhvj8.

Method: Quantitative examination using 1.9K NLP paper abstracts, 4.9K author contributors, and 7.8K human-annotated contribution sentences (AfricaNLPContributions) with benchmark results.

Result: Created a dataset and continuously existing NLP progress tracking website that provides insights into AfricaNLP research trends and enables data-driven literature surveys.

Conclusion: The developed dataset and tracking platform offer a powerful tool for tracing AfricaNLP research trends and have potential for generating comprehensive, data-driven literature surveys in the field.

Abstract: Natural Language Processing (NLP) is undergoing constant transformation, as Large Language Models (LLMs) are driving daily breakthroughs in research and practice. In this regard, tracking the progress of NLP research and automatically analyzing the contributions of research papers provides key insights into the nature of the field and the researchers. This study explores the progress of African NLP (AfricaNLP) by asking (and answering) basic research questions such as: i) How has the nature of NLP evolved over the last two decades?, ii) What are the contributions of AfricaNLP papers?, and iii) Which individuals and organizations (authors, affiliated institutions, and funding bodies) have been involved in the development of AfricaNLP? We quantitatively examine the contributions of AfricaNLP research using 1.9K NLP paper abstracts, 4.9K author contributors, and 7.8K human-annotated contribution sentences (AfricaNLPContributions) along with benchmark results. Our dataset and continuously existing NLP progress tracking website provide a powerful lens for tracing AfricaNLP research trends and hold potential for generating data-driven literature surveys.

Method: Created ReFACT benchmark with 1,001 expert-annotated question-answer pairs spanning diverse scientific domains, each with correct answers and non-factual counterparts annotated with error spans and types.

Result: Benchmarked 9 state-of-the-art LLMs showing limited performance (~50% accuracy), with even top models like GPT-4o failing to distinguish factual from confabulated scientific answers.

Conclusion: Highlights need for fine-grained, human-validated benchmarks to detect and correct scientific confabulation, raising concerns about LLM-as-judge evaluation reliability.

Abstract: Large Language Models (LLMs) frequently confabulate scientific facts, severely undermining their trustworthiness. Addressing this challenge requires benchmarks that go beyond binary factuality and enable fine-grained evaluation. We introduce ReFACT (Reddit False And Correct Texts), a benchmark of 1,001 expert-annotated question-answer pairs spanning diverse scientific domains for the detection of scientific confabulation. Each instance includes both a scientifically correct answer and a non-factual counterpart annotated with precise error spans and error types. ReFACT enables multi-stage evaluation: (1) confabulation detection, (2) fine-grained error localization, and (3) correction. We benchmark 9 state-of-the-art LLMs, revealing limited performance (about 50 percent accuracy). Even top models such as GPT-4o fail to distinguish factual from confabulated scientific answers, raising concerns about the reliability of LLM-as-judge evaluation paradigms. Our findings highlight the need for fine-grained, human-validated benchmarks to detect and correct scientific confabulation in domain-specific contexts. The dataset is available at: https://github.com/ddz5431/ReFACT

Method: Hierarchical decomposition with Planner expert for macro-level strategy selection (recommend, explain, encourage) and Actor expert for detailed responses, guided by auxiliary experts for user preferences and factual grounding. Uses reinforcement learning with LLM-based reward model for automatic strategy exploration.

Result: Extensive experiments show RSO significantly improves interaction performance compared to state-of-the-art baselines.

Conclusion: Explicit hierarchical strategy optimization is effective for conversational recommender systems, enabling more tractable learning and better interaction outcomes.

Abstract: Conversational Recommender Systems (CRSs) aim to provide personalized recommendations through multi-turn natural language interactions with users. Given the strong interaction and reasoning skills of Large Language Models (LLMs), leveraging LLMs for CRSs has recently emerged as a promising direction. However, existing LLM-based methods often lack explicit optimization of interaction strategies, instead relying on unified prompts and the LLM’s internal knowledge to decide how to interact, which can lead to suboptimal outcomes. In this paper, we propose a novel Reinforced Strategy Optimization (RSO) method for CRS, which decomposes the process of generating strategy-driven response decisions into the macro-level strategy planning and micro-level strategy adaptation through a network-of-experts architecture. At the macro level, a Planner expert selects macro-level interaction strategies (e.g., recommend, explain, encourage). At the micro level, an Actor expert generates detailed responses conditioned on the selected macro-level strategy, guided by auxiliary experts that provide complementary information such as user preferences and factual grounding. This hierarchical decomposition disentangles the optimization of different sub-tasks involved in CRS response generation, enabling more tractable learning at each level. To address the scarcity of high-quality multi-turn training data, we formulate strategy learning as a reinforcement learning problem, guided by an LLM-based reward model to achieve automatic strategy exploration. Extensive experiments show that RSO significantly improves interaction performance compared to state-of-the-art baselines, demonstrating the effectiveness of explicit hierarchical strategy optimization for CRS.

Method: Fine-tuned open-weights large language models as definition generators, using datasets from AXOLOTL'24 shared task on explainable semantic change modeling for Finnish, Russian, and German languages.

Result: The fine-tuned open-source models performed higher than the best submissions from the shared task that used closed proprietary LLMs, and encoder-decoder definition generators performed on par with decoder-only models.

Conclusion: Open-weights LLMs can be effectively fine-tuned for definition generation tasks and compete with proprietary models, with encoder-decoder architectures showing comparable performance to decoder-only models.

Abstract: We apply definition generators based on open-weights large language models to the task of creating explanations of novel senses, taking target word usages as an input. To this end, we employ the datasets from the AXOLOTL'24 shared task on explainable semantic change modeling, which features Finnish, Russian and German languages. We fine-tune and provide publicly the open-source models performing higher than the best submissions of the aforementioned shared task, which employed closed proprietary LLMs. In addition, we find that encoder-decoder definition generators perform on par with their decoder-only counterparts.

Method: Uses a single agent that interacts with KGs as its environment, learning retrieval strategies through end-to-end reinforcement learning and incorporating retrieved information into reasoning and generation.

Result: KG-R1 improves answer accuracy with fewer generation tokens than prior methods, using smaller models (Qwen-2.5-3B) while outperforming larger foundation models. It maintains strong accuracy on new KGs without modification.

Conclusion: KG-R1 provides an efficient and transferable KG-RAG framework suitable for real-world deployment due to its single-agent design and plug-and-play capability across different knowledge graphs.

Abstract: Knowledge-graph retrieval-augmented generation (KG-RAG) couples large language models (LLMs) with structured, verifiable knowledge graphs (KGs) to reduce hallucinations and expose reasoning traces. However, many KG-RAG systems compose multiple LLM modules (e.g planning, reasoning, and responding), inflating inference cost and binding behavior to a specific target KG. To address this, we introduce KG-R1, an agentic KG retrieval-augmented generation (KG-RAG) framework through reinforcement learning (RL). KG-R1 utilizes a single agent that interacts with KGs as its environment, learning to retrieve at each step and incorporating the retrieved information into its reasoning and generation. The process is optimized through end-to-end RL. In controlled experiments across Knowledge-Graph Question Answering (KGQA) benchmarks, our method demonstrates both efficiency and transferability: Using Qwen-2.5-3B, KG-R1 improves answer accuracy with fewer generation tokens than prior multi-module workflow methods that use larger foundation or fine-tuned models. Furthermore, KG-R1 enables plug and play: after training, it maintains strong accuracy on new KGs without modification. These properties make KG-R1 a promising KG-RAG framework for real-world deployment. Our code is publicly available at https://github.com/Jinyeop3110/KG-R1.

Method: Proposed SSUF module integrates spectral unmixing with spectral-spatial feature extraction into ResNet-based CNN, plus custom Spatial-Spectral Gradient Loss combining MSE with spatial and spectral gradient components.

Result: Experiments on three public remote sensing hyperspectral datasets show competitive performance while reducing model complexity.

Conclusion: The hybrid deep learning model with SSUF module effectively enhances both spatial resolution and spectral integrity in hyperspectral image super-resolution.

Abstract: Hyperspectral single image super-resolution (SISR) is a challenging task due to the difficulty of restoring fine spatial details while preserving spectral fidelity across a wide range of wavelengths, which limits the performance of conventional deep learning models. To address this challenge, we introduce Spectral-Spatial Unmixing Fusion (SSUF), a novel module that can be seamlessly integrated into standard 2D convolutional architectures to enhance both spatial resolution and spectral integrity. The SSUF combines spectral unmixing with spectral–spatial feature extraction and guides a ResNet-based convolutional neural network for improved reconstruction. In addition, we propose a custom Spatial-Spectral Gradient Loss function that integrates mean squared error with spatial and spectral gradient components, encouraging accurate reconstruction of both spatial and spectral features. Experiments on three public remote sensing hyperspectral datasets demonstrate that the proposed hybrid deep learning model achieves competitive performance while reducing model complexity.

Method: Developed a unified multi-level framework for characterizing hallucinations across image and text modalities, using a task-modality interleaved approach to link hallucinations to specific mechanisms in model lifecycle.

Result: Investigations revealed that hallucinations often stem from predictable patterns in data distributions and inherited biases, providing insights into root causes rather than just surface symptoms.

Conclusion: The survey provides a foundation for developing more robust and effective solutions to hallucinations in real-world generative AI systems by offering a coherent understanding of the phenomenon.

Abstract: The widespread adoption of large language and vision models in real-world applications has made urgent the need to address hallucinations – instances where models produce incorrect or nonsensical outputs. These errors can propagate misinformation during deployment, leading to both financial and operational harm. Although much research has been devoted to mitigating hallucinations, our understanding of it is still incomplete and fragmented. Without a coherent understanding of hallucinations, proposed solutions risk mitigating surface symptoms rather than underlying causes, limiting their effectiveness and generalizability in deployment. To tackle this gap, we first present a unified, multi-level framework for characterizing both image and text hallucinations across diverse applications, aiming to reduce conceptual fragmentation. We then link these hallucinations to specific mechanisms within a model’s lifecycle, using a task-modality interleaved approach to promote a more integrated understanding. Our investigations reveal that hallucinations often stem from predictable patterns in data distributions and inherited biases. By deepening our understanding, this survey provides a foundation for developing more robust and effective solutions to hallucinations in real-world generative AI systems.

Method: Proposes RobustVLA with: 1) Output robustness via offline robust optimization against worst-case action noise, 2) Input robustness by enforcing consistent actions across semantically-preserving variations, 3) Multi-armed bandit formulation with UCB algorithm to identify most harmful noise across multiple perturbations.

Result: Achieves 12.6% absolute gain on pi0 backbone and 10.4% on OpenVLA backbone across 17 perturbations, 50.6x faster inference than visual-robust VLAs, 10.4% gain under mixed perturbations, and 65.6% absolute gain on real-world FR5 robot with limited demonstrations.

Conclusion: RobustVLA effectively addresses multi-modal robustness in VLAs, demonstrating significant performance improvements across various perturbations and real-world robotic applications with limited data.

Abstract: In Vision-Language-Action (VLA) models, robustness to real-world perturbations is critical for deployment. Existing methods target simple visual disturbances, overlooking the broader multi-modal perturbations that arise in actions, instructions, environments, and observations. Here, we first evaluate the robustness of mainstream VLAs under 17 perturbations across four modalities. We find (1) actions as the most fragile modality, (2) Existing visual-robust VLA do not gain robustness in other modality, and (3) pi0 demonstrates superior robustness with a diffusion-based action head. To build multi-modal robust VLAs, we propose RobustVLA against perturbations in VLA inputs and outputs. For output robustness, we perform offline robust optimization against worst-case action noise that maximizes mismatch in flow matching objective. This can be seen as adversarial training, label smoothing, and outlier penalization. For input robustness, we enforce consistent actions across input variations that preserve task semantics. To account for multiple perturbations, we formulate robustness as a multi-armed bandit problem and apply an upper confidence bound algorithm to automatically identify the most harmful noise. Experiments on LIBERO demonstrate our RobustVLA delivers absolute gains over baselines of 12.6% on the pi0 backbone and 10.4% on the OpenVLA backbone across all 17 perturbations, achieving 50.6x faster inference than existing visual-robust VLAs, and a 10.4% gain under mixed perturbations. Our RobustVLA is particularly effective on real-world FR5 robot with limited demonstrations, showing absolute gains by 65.6% under perturbations of four modalities.

Method: Unsupervised discovery of intrinsic capabilities from gradient-based learning trajectories, data attribution to capabilities via influence estimation, and capability-aware curriculum curation through balanced selection and staged sequencing.

Result: With only 5% of the original data, CADC surpasses full-data training performance on multimodal benchmarks.

Conclusion: Intrinsic capabilities are fundamental building blocks of model learning, and CADC establishes a principled paradigm for instruction data curation that transforms black-box tuning into a controllable, capability-driven process.

Abstract: Large vision-language models (VLMs) achieve strong benchmark performance, but controlling their behavior through instruction tuning remains difficult. Reducing the budget of instruction tuning dataset often causes regressions, as heuristic strategies treat models as black boxes and overlook the latent capabilities that govern learning. We introduce Capability-Attributed Data Curation (CADC), a framework that shifts curation from task-specific heuristics to intrinsic capability analysis. CADC discovers intrinsic capabilities in an unsupervised manner from gradient-based learning trajectories, attributes training data to these capabilities via influence estimation, and curates capability-aware curricula through balanced selection and staged sequencing. This transforms black-box instruction tuning into a controllable, capability-driven process. With as little as 5% of the original data, CADC surpasses full-data training on multimodal benchmarks. These results validate intrinsic capabilities as the fundamental building blocks of model learning and establish CADC as a principle paradigm for instruction data curation.

Method: Created C³B benchmark with over 2000 images and 18000 QA pairs across three progressively difficult tasks: basic visual recognition, cultural conflict understanding, and cultural content generation, using comics to incorporate multiple cultures per image.

Result: Evaluation of 11 open-source MLLMs revealed significant performance gaps compared to human performance, demonstrating C³B’s substantial challenges for current models.

Conclusion: C³B effectively challenges current MLLMs’ cultural awareness capabilities and encourages future research to advance this critical capability in multimodal AI systems.

Abstract: Cultural awareness capabilities has emerged as a critical capability for Multimodal Large Language Models (MLLMs). However, current benchmarks lack progressed difficulty in their task design and are deficient in cross-lingual tasks. Moreover, current benchmarks often use real-world images. Each real-world image typically contains one culture, making these benchmarks relatively easy for MLLMs. Based on this, we propose C$^3$B ($\textbf{C}$omics $\textbf{C}$ross-$\textbf{C}$ultural $\textbf{B}$enchmark), a novel multicultural, multitask and multilingual cultural awareness capabilities benchmark. C$^3$B comprises over 2000 images and over 18000 QA pairs, constructed on three tasks with progressed difficulties, from basic visual recognition to higher-level cultural conflict understanding, and finally to cultural content generation. We conducted evaluations on 11 open-source MLLMs, revealing a significant performance gap between MLLMs and human performance. The gap demonstrates that C$^3$B poses substantial challenges for current MLLMs, encouraging future research to advance the cultural awareness capabilities of MLLMs.

Method: Compress each second of video into single temporally pooled images using motion blur and weighted averaging variants (Blend Blur with Last Frame, Weighted Average, Exponential, Ramp pooling). Fine-tune QWEN-2.5-VL 7B with supervised two-turn targets including reasoning and final answer using SFT and DPO on ReasonVQA dataset.

Result: Dramatic performance improvements on ReasonVQA dataset: F1 score from 0.212 to 0.543, BLEU-4 from 0.031 to 0.291, ROUGE-L from 0.196 to 0.528. Significant rationale quality improvement. Gains persist across different pooling schemes and show strong robustness in temporal evidence summarization.

Conclusion: The POVQA pipeline effectively enables efficient long-video question answering by compressing video content and aligning LVLMs with lightweight supervision, demonstrating strong performance improvements and robustness across different temporal pooling methods.

Abstract: Video Question Answering (VQA) with Large Vision Language Models (LVLMs) has gained significant traction in research ever since the Flamingo was introduced by Deepmind. Recent advancements in large context/long video question answering have allowed VQA tasks to have context window of 1500+ frames. However, this only leads to 50 seconds of video footage without losing any significant information. We introduce POVQA, a data-efficient pipeline that compresses each second of video into a single temporally pooled image (via motion blur and weighted averaging variants) and then align LVLMs with lightweight supervision. Concretely, we build 1 fps input sources using Blend Blur with Last Frame, Weighted Average, Exponential and Ramp pooling and fine-tune QWEN-2.5-VL 7B with supervised two turn target including reasoning and final answer. We apply Supervised Fine Tuning (SFT) and Direct Preference Optimization (DPO) on our novel dataset ReasonVQA consisting of 12 movies with 239 human annotated question-answer with reasoning prompts. On our ReasonVQA dataset, this method dramatically improves performance over pooled baselines: F1 score improves from 0.212 to 0.543, BLEU-4 from 0.031 to 0.291, and ROUGE-L from 0.196 to 0.528. Rationale quality also significantly increases. Cross-evaluation of SFT + DPO on various pooling functions show that the gains persist regardless of the pooling scheme used at train or test time, indicating strong robustness on summarization of temporal evidence. Similar observations were made on zero-shot in TVQA.

Method: Generated 100 images each for 5 hospital professions using 6 open-weight TTI models with 5 portrait qualifiers, analyzing gender representation systematically.

Result: All models showed occupational stereotypes: nurses exclusively female, surgeons predominantly male. Model-specific variations: Qwen-Image/SDXL rigid male dominance, FLUX.1-dev female skew, others mixed. Prompt qualifiers (corporate=male, beautiful=female) significantly influenced outcomes.

Conclusion: Gender bias in TTI models is systematic and model-specific, with prompt wording playing critical role. Need for bias-aware design, balanced defaults, and user guidance to prevent stereotype reinforcement.

Abstract: Text-to-image (TTI) models are increasingly used in professional, educational, and creative contexts, yet their outputs often embed and amplify social biases. This paper investigates gender representation in six state-of-the-art open-weight models: HunyuanImage 2.1, HiDream-I1-dev, Qwen-Image, FLUX.1-dev, Stable-Diffusion 3.5 Large, and Stable-Diffusion-XL. Using carefully designed prompts, we generated 100 images for each combination of five hospital-related professions (cardiologist, hospital director, nurse, paramedic, surgeon) and five portrait qualifiers ("", corporate, neutral, aesthetic, beautiful). Our analysis reveals systematic occupational stereotypes: all models produced nurses exclusively as women and surgeons predominantly as men. However, differences emerge across models: Qwen-Image and SDXL enforce rigid male dominance, HiDream-I1-dev shows mixed outcomes, and FLUX.1-dev skews female in most roles. HunyuanImage 2.1 and Stable-Diffusion 3.5 Large also reproduce gender stereotypes but with varying degrees of sensitivity to prompt formulation. Portrait qualifiers further modulate gender balance, with terms like corporate reinforcing male depictions and beautiful favoring female ones. Sensitivity varies widely: Qwen-Image remains nearly unaffected, while FLUX.1-dev, SDXL, and SD3.5 show strong prompt dependence. These findings demonstrate that gender bias in TTI models is both systematic and model-specific. Beyond documenting disparities, we argue that prompt wording plays a critical role in shaping demographic outcomes. The results underscore the need for bias-aware design, balanced defaults, and user guidance to prevent the reinforcement of occupational stereotypes in generative AI.

Method: A three-agent framework: Planner structures lecture content into coherent flows with visual assets; Coder converts instructions into executable Python code with scope-guided auto-fix; Critic uses vision-language models with visual anchor prompts to refine spatial layout and ensure clarity.

Result: Code2Video achieves 40% improvement over direct code generation and produces videos comparable to human-crafted tutorials, as measured by VLM-as-a-Judge aesthetic scores, code efficiency, and the novel TeachQuiz metric that quantifies knowledge recovery from generated videos.

Conclusion: Code2Video demonstrates potential as a scalable, interpretable, and controllable approach for educational video generation, with performance comparable to human-crafted content and significant improvements over baseline methods.

Abstract: While recent generative models advance pixel-space video synthesis, they remain limited in producing professional educational videos, which demand disciplinary knowledge, precise visual structures, and coherent transitions, limiting their applicability in educational scenarios. Intuitively, such requirements are better addressed through the manipulation of a renderable environment, which can be explicitly controlled via logical commands (e.g., code). In this work, we propose Code2Video, a code-centric agent framework for generating educational videos via executable Python code. The framework comprises three collaborative agents: (i) Planner, which structures lecture content into temporally coherent flows and prepares corresponding visual assets; (ii) Coder, which converts structured instructions into executable Python codes while incorporating scope-guided auto-fix to enhance efficiency; and (iii) Critic, which leverages vision-language models (VLM) with visual anchor prompts to refine spatial layout and ensure clarity. To support systematic evaluation, we build MMMC, a benchmark of professionally produced, discipline-specific educational videos. We evaluate MMMC across diverse dimensions, including VLM-as-a-Judge aesthetic scores, code efficiency, and particularly, TeachQuiz, a novel end-to-end metric that quantifies how well a VLM, after unlearning, can recover knowledge by watching the generated videos. Our results demonstrate the potential of Code2Video as a scalable, interpretable, and controllable approach, achieving 40% improvement over direct code generation and producing videos comparable to human-crafted tutorials. The code and datasets are available at https://github.com/showlab/Code2Video.

Method: Treats template stealing as sequential decision making problem, uses reinforcement learning with multiple similarity-based feedback signals as reward functions to explore prompt space effectively.

Result: Achieves state-of-the-art performance on public benchmarks, reduces total attack cost to under 13% of existing baselines, and generalizes effectively across different image styles to steal unseen prompt templates.

Conclusion: The study exposes urgent security threats in prompt trading and provides groundwork for developing protective standards in the emerging MLLMs marketplace.

Abstract: Multimodal Large Language Models (MLLMs) have transformed text-to-image workflows, allowing designers to create novel visual concepts with unprecedented speed. This progress has given rise to a thriving prompt trading market, where curated prompts that induce trademark styles are bought and sold. Although commercially attractive, prompt trading also introduces a largely unexamined security risk: the prompts themselves can be stolen. In this paper, we expose this vulnerability and present RLStealer, a reinforcement learning based prompt inversion framework that recovers its template from only a small set of example images. RLStealer treats template stealing as a sequential decision making problem and employs multiple similarity based feedback signals as reward functions to effectively explore the prompt space. Comprehensive experiments on publicly available benchmarks demonstrate that RLStealer gets state-of-the-art performance while reducing the total attack cost to under 13% of that required by existing baselines. Our further analysis confirms that RLStealer can effectively generalize across different image styles to efficiently steal unseen prompt templates. Our study highlights an urgent security threat inherent in prompt trading and lays the groundwork for developing protective standards in the emerging MLLMs marketplace.

Method: Proposed HR-INR framework with event camera assistance, featuring: (1) regional event feature extractor using event temporal pyramid representation for regional nonlinear motion, (2) holistic event-frame feature extractor for long-term dependence and continuity motion, and (3) INR-based decoder with spatiotemporal embeddings for larger temporal perception field.

Result: Validated on four datasets (both simulated and real data), showing effectiveness, generalization, and superiority over existing methods.

Conclusion: HR-INR successfully addresses limitations of previous INR-based C-STVSR methods by capturing both holistic dependencies and regional motions, enabling better performance in dynamic scenes with fast, complex motion.

Abstract: Continuous space-time video super-resolution (C-STVSR) aims to simultaneously enhance video resolution and frame rate at an arbitrary scale. Recently, implicit neural representation (INR) has been applied to video restoration, representing videos as implicit fields that can be decoded at an arbitrary scale. However, existing INR-based C-STVSR methods typically rely on only two frames as input, leading to insufficient inter-frame motion information. Consequently, they struggle to capture fast, complex motion and long-term dependencies (spanning more than three frames), hindering their performance in dynamic scenes. In this paper, we propose a novel C-STVSR framework, named HR-INR, which captures both holistic dependencies and regional motions based on INR. It is assisted by an event camera – a novel sensor renowned for its high temporal resolution and low latency. To fully utilize the rich temporal information from events, we design a feature extraction consisting of (1) a regional event feature extractor – taking events as inputs via the proposed event temporal pyramid representation to capture the regional nonlinear motion and (2) a holistic event-frame feature extractor for long-term dependence and continuity motion. We then propose a novel INR-based decoder with spatiotemporal embeddings to capture long-term dependencies with a larger temporal perception field. We validate the effectiveness and generalization of our method on four datasets (both simulated and real data), showing the superiority of our method. The project page is available at https://github.com/yunfanLu/HR-INR

Method: EDCT extracts testable concepts from explanations, generates counterfactual image edits via inpainting, and computes consistency scores using LLM analysis.

Result: Testing on 120 OK-VQA examples revealed substantial faithfulness gaps in multiple VLMs, with regulator-aligned audit artifacts showing when concepts fail causal tests.

Conclusion: EDCT provides automated verification of VLM explanation faithfulness, uncovering significant mismatches between stated and actual reasoning.

Abstract: Vision-Language Models (VLMs) often produce fluent Natural Language Explanations (NLEs) that sound convincing but may not reflect the causal factors driving predictions. This mismatch of plausibility and faithfulness poses technical and governance risks. We introduce Explanation-Driven Counterfactual Testing (EDCT), a fully automated verification procedure for a target VLM that treats the model’s own explanation as a falsifiable hypothesis. Given an image-question pair, EDCT: (1) obtains the model’s answer and NLE, (2) parses the NLE into testable visual concepts, (3) generates targeted counterfactual edits via generative inpainting, and (4) computes a Counterfactual Consistency Score (CCS) using LLM-assisted analysis of changes in both answers and explanations. Across 120 curated OK-VQA examples and multiple VLMs, EDCT uncovers substantial faithfulness gaps and provides regulator-aligned audit artifacts indicating when cited concepts fail causal tests.

Method: Proposes Decoupled Position Detection Transformer with: 1) Explicit definition of object category, visible position, and infrared position; 2) Decoupled Position Multispectral Cross-attention module for adaptive feature fusion; 3) Query-decoupled Multispectral Decoder; 4) Decoupled Position Contrastive DeNoising Training strategy.

Result: Experiments on DroneVehicle and KAIST datasets show significant improvements over state-of-the-art methods.

Conclusion: DPDETR effectively addresses modality misalignment by learning decoupled positions and reliably fusing complementary features, achieving superior performance in infrared-visible object detection.

Abstract: Infrared-visible object detection aims to achieve robust object detection by leveraging the complementary information of infrared and visible image pairs. However, the commonly existing modality misalignment problem presents two challenges: fusing misalignment complementary features is difficult, and current methods cannot reliably locate objects in both modalities under misalignment conditions. In this paper, we propose a Decoupled Position Detection Transformer (DPDETR) to address these issues. Specifically, we explicitly define the object category, visible modality position, and infrared modality position to enable the network to learn the intrinsic relationships and output reliably positions of objects in both modalities. To fuse misaligned object features reliably, we propose a Decoupled Position Multispectral Cross-attention module that adaptively samples and aggregates multispectral complementary features with the constraint of infrared and visible reference positions. Additionally, we design a query-decoupled Multispectral Decoder structure to address the the conflict in feature focus among the three kinds of object information in our task and propose a Decoupled Position Contrastive DeNoising Training strategy to enhance the DPDETR’s ability to learn decoupled positions. Experiments on DroneVehicle and KAIST datasets demonstrate significant improvements compared to other state-of-the-art methods. The code will be released at https://github.com/gjj45/DPDETR

Method: Uses Token-wise Attention Decoupling (TAD) to identify key information tokens and align them with target visual regions, then Layout-Preserving Decoupling (LPD) to separate regions from background while preserving spatial layouts in a compact representation.

Result: Achieves new SOTA on V*Bench (92.1% for Qwen2.5-VL 7B, 91.6% for InternVL3 8B), HRBench4K, and HRBench8K, surpassing RL methods while using 75% less memory than previous training-free approaches.

Conclusion: HiDe effectively addresses background interference in high-resolution image understanding without training, demonstrating superior performance and efficiency over existing methods.

Abstract: Multimodal Large Language Models (MLLMs) have made significant strides in visual understanding tasks. However, their performance on high-resolution images remains suboptimal. While existing approaches often attribute this limitation to perceptual constraints and argue that MLLMs struggle to recognize small objects, leading them to use “zoom in” strategies for better detail, our analysis reveals a different cause: the main issue is not object size, but rather caused by complex background interference. We systematically analyze this “zoom in” operation through a series of decoupling experiments and propose the Hierarchical Decoupling Framework (HiDe), a training-free framework that uses Token-wise Attention Decoupling (TAD) to decouple the question tokens and identify the key information tokens, then leverages their attention weights to achieve precise alignment with the target visual regions. Subsequently, it employs Layout-Preserving Decoupling (LPD) to decouple these regions from the background and reconstructs a compact representation that preserves essential spatial layouts while eliminating background interference. HiDe sets a new SOTA on VBench, HRBench4K, and HRBench8K, boosting Qwen2.5-VL 7B and InternVL3 8B to SOTA (92.1% and 91.6% on VBench), even surpassing RL methods. After optimization, HiDe uses 75% less memory than the previous training-free approach. Code is provided in https://github.com/Tennine2077/HiDe.

Method: Uses spatial processing for multi-scale features, Fast Fourier Transform (FFT) for global frequency information, and Haar wavelet transform to decompose features into high/low-frequency components for boundary refinement.

Result: Achieves state-of-the-art performance on four datasets: LoveDA, Vaihingen, Potsdam, and iSAID, with significant improvements in boundary regions and grayscale transition zones.

Conclusion: The dual-domain synergy between spatial granularity and frequency-domain edge sensitivity substantially improves segmentation accuracy in challenging regions.

Abstract: To fully leverage spatial information for remote sensing image segmentation and address semantic edge ambiguities caused by grayscale variations (e.g., shadows and low-contrast regions), we propose the Frequency and Spatial Domains based Detail Enhancement Network (FSDENet). Our framework employs spatial processing methods to extract rich multi-scale spatial features and fine-grained semantic details. By effectively integrating global and frequency-domain information through the Fast Fourier Transform (FFT) in global mappings, the model’s capability to discern global representations under grayscale variations is significantly strengthened. Additionally, we utilize Haar wavelet transform to decompose features into high- and low-frequency components, leveraging their distinct sensitivity to edge information to refine boundary segmentation. The model achieves dual-domain synergy by integrating spatial granularity with frequency-domain edge sensitivity, substantially improving segmentation accuracy in boundary regions and grayscale transition zones. Comprehensive experimental results demonstrate that FSDENet achieves state-of-the-art (SOTA) performance on four widely adopted datasets: LoveDA, Vaihingen, Potsdam, and iSAID.

Method: Uses a Vision-Language Model (VLM) for end-to-end trajectory prediction directly from front-view camera input, with principled supervision strategy from statistical modeling for imitation learning from expert demonstrations.

Result: Achieves state-of-the-art performance on nuScenes dataset with over 30% improvement compared to prior baselines, and shows superior generalization on cross-domain datasets from diverse vehicles.

Conclusion: The framework enables fundamental driving behaviors and lays the foundation for more capable self-driving agents, demonstrating notable potential for cross-vehicle robustness and adaptability.

Abstract: In this work, we reconceptualize autonomous driving as a generalized language and formulate the trajectory planning task as next waypoint prediction. We introduce Max-V1, a novel framework for one-stage end-to-end autonomous driving. Our framework presents a single-pass generation paradigm that aligns with the inherent sequentiality of driving. This approach leverages the generative capacity of the VLM (Vision-Language Model) to enable end-to-end trajectory prediction directly from front-view camera input. The efficacy of this method is underpinned by a principled supervision strategy derived from statistical modeling. This provides a well-defined learning objective, which makes the framework highly amenable to master complex driving policies through imitation learning from large-scale expert demonstrations. Empirically, our method achieves the state-of-the-art performance on the nuScenes dataset, delivers an overall improvement of over 30% compared to prior baselines. Furthermore, it exhibits superior generalization performance on cross-domain datasets acquired from diverse vehicles, demonstrating notable potential for cross-vehicle robustness and adaptability. Due to these empirical strengths, this work introduces a model enabling fundamental driving behaviors, laying the foundation for the development of more capable self-driving agents. Code will be available upon publication.

Method: Uses an in-context dual-branch diffusion model that separates foreground and background processing with blob representations, plus a self-supervised disentangle-then-reconstruct training paradigm with identity-preserving loss.

Result: Achieves state-of-the-art performance in various element-level editing tasks (object addition, removal, scaling, replacement) while maintaining computational efficiency.

Conclusion: BlobCtrl provides an effective framework for controllable object-level image manipulation through blob-based representation and introduces BlobData and BlobBench for future research.

Abstract: As user expectations for image editing continue to rise, the demand for flexible, fine-grained manipulation of specific visual elements presents a challenge for current diffusion-based methods. In this work, we present BlobCtrl, a framework for element-level image editing based on a probabilistic blob-based representation. Treating blobs as visual primitives, BlobCtrl disentangles layout from appearance, affording fine-grained, controllable object-level manipulation. Our key contributions are twofold: (1) an in-context dual-branch diffusion model that separates foreground and background processing, incorporating blob representations to explicitly decouple layout and appearance, and (2) a self-supervised disentangle-then-reconstruct training paradigm with an identity-preserving loss function, along with tailored strategies to efficiently leverage blob-image pairs. To foster further research, we introduce BlobData for large-scale training and BlobBench, a benchmark for systematic evaluation. Experimental results demonstrate that BlobCtrl achieves state-of-the-art performance in a variety of element-level editing tasks, such as object addition, removal, scaling, and replacement, while maintaining computational efficiency. Project Webpage: https://liyaowei-stu.github.io/project/BlobCtrl/

Method: Uses feature map similarity for rank selection, one-shot fine-tuning process, integrates three LRF techniques for Conv layers and three for FC layers applied selectively per-layer, and combines multiple LRF methods within single models.

Result: Achieves substantial compression with minimal accuracy loss, outperforms state-of-the-art techniques across 14 CNN models and eight datasets.

Conclusion: The proposed framework effectively addresses LRF compression challenges and demonstrates that combining multiple LRF methods yields better results than uniform application of single methods.

Abstract: Low-Rank Factorization (LRF) is a widely adopted technique for compressing deep neural networks (DNNs). However, it faces several challenges, including optimal rank selection, a vast design space, long fine-tuning times, and limited compatibility with different layer types and decomposition methods. This paper presents an end-to-end Design Space Exploration (DSE) methodology and framework for compressing convolutional neural networks (CNNs) that addresses all these issues. We introduce a novel rank selection strategy based on feature map similarity, which captures non-linear interactions between layer outputs more effectively than traditional weight-based approaches. Unlike prior works, our method uses a one-shot fine-tuning process, significantly reducing the overall fine-tuning time. The proposed framework is fully compatible with all types of convolutional (Conv) and fully connected (FC) layers. To further improve compression, the framework integrates three different LRF techniques for Conv layers and three for FC layers, applying them selectively on a per-layer basis. We demonstrate that combining multiple LRF methods within a single model yields better compression results than using a single method uniformly across all layers. Finally, we provide a comprehensive evaluation and comparison of the six LRF techniques, offering practical insights into their effectiveness across different scenarios. The proposed work is integrated into TensorFlow 2.x, ensuring compatibility with widely used deep learning workflows. Experimental results on 14 CNN models across eight datasets demonstrate that the proposed methodology achieves substantial compression with minimal accuracy loss, outperforming several state-of-the-art techniques.

Method: Uses a pretrained 3D VAE to encode videos, fuses with target HR prompts using trainable spatio-temporal layers with AdaLN, and applies FiLM in decoder with fine-tuned output layer for accurate physiological modulation.

Result: Achieves PSNR of 38.96 dB and SSIM of 0.98 for visual quality, with HR modulation error of 10.00 bpm MAE and 10.09% MAPE using state-of-the-art rPPG estimator.

Conclusion: The method enables controllable HR editing for applications like anonymizing biometric signals or synthesizing realistic videos with desired vital signs.

Abstract: Camera-based physiological signal estimation provides a non-contact and convenient means to monitor Heart Rate (HR). However, the presence of vital signals in facial videos raises significant privacy concerns, as they can reveal sensitive personal information related to the health and emotional states of an individual. To address this, we propose a learned framework that edits physiological signals in videos while preserving visual fidelity. First, we encode an input video into a latent space via a pretrained 3D Variational Autoencoder (3D VAE), while a target HR prompt is embedded through a frozen text encoder. We fuse them using a set of trainable spatio-temporal layers with Adaptive Layer Normalizations (AdaLN) to capture the strong temporal coherence of remote Photoplethysmography (rPPG) signals. We apply Feature-wise Linear Modulation (FiLM) in the decoder with a fine-tuned output layer to avoid the degradation of physiological signals during reconstruction, enabling accurate physiological modulation in the reconstructed video. Empirical results show that our method preserves visual quality with an average PSNR of 38.96 dB and SSIM of 0.98 on selected datasets, while achieving an average HR modulation error of 10.00 bpm MAE and 10.09% MAPE using a state-of-the-art rPPG estimator. Our design’s controllable HR editing is useful for applications such as anonymizing biometric signals in real videos or synthesizing realistic videos with desired vital signs.

Method: Developed an automated 5S audit system based on large-scale language models (LLM) capable of assessing the five senses (Seiri, Seiton, Seiso, Seiketsu, Shitsuke) through intelligent image analysis.

Result: System reliability validated with Cohen’s concordance coefficient (kappa = 0.75), showing strong alignment with human audits. Achieved 50% faster audit process and 99.8% reduction in operating costs compared to traditional manual audits.

Conclusion: The methodology establishes a new paradigm for integrating lean systems with emerging AI technologies, offering scalability for implementation in automotive plants of different sizes and contributing significantly to continuous improvement.

Abstract: The evolution of the 5S methodology with the support of artificial intelligence techniques represents a significant opportunity to improve industrial organization audits in the automotive chain, making them more objective, efficient and aligned with Industry 4.0 standards. This work developed an automated 5S audit system based on large-scale language models (LLM), capable of assessing the five senses (Seiri, Seiton, Seiso, Seiketsu, Shitsuke) in a standardized way through intelligent image analysis. The system’s reliability was validated using Cohen’s concordance coefficient (kappa = 0.75), showing strong alignment between the automated assessments and the corresponding human audits. The results indicate that the proposed solution contributes significantly to continuous improvement in automotive manufacturing environments, speeding up the audit process by 50% of the traditional time and maintaining the consistency of the assessments, with a 99.8% reduction in operating costs compared to traditional manual audits. The methodology presented establishes a new paradigm for integrating lean systems with emerging AI technologies, offering scalability for implementation in automotive plants of different sizes.

Method: Proposed binary encoding approaches including vanilla binary encoding, ECOCs, class weighting, hard/soft decoding, class-to-codeword assignment, and label embedding trees to reduce complexity to logarithmic scale.

Result: Binary encoding methods achieved significantly lower performance (DSC 39.3-73.8) compared to one-hot encoding (DSC 82.4) in whole brain parcellation with 108 classes.

Conclusion: While binary encodings offer computational efficiency, they currently cannot match state-of-the-art segmentation quality, highlighting the need for future research on compact encoding strategies.

Abstract: This work presents novel methods to reduce computational and memory requirements for medical image segmentation with a large number of classes. We curiously observe challenges in maintaining state-of-the-art segmentation performance with all of the explored options. Standard learning-based methods typically employ one-hot encoding of class labels. The computational complexity and memory requirements thus increase linearly with the number of classes. We propose a family of binary encoding approaches instead of one-hot encoding to reduce the computational complexity and memory requirements to logarithmic in the number of classes. In addition to vanilla binary encoding, we investigate the effects of error-correcting output codes (ECOCs), class weighting, hard/soft decoding, class-to-codeword assignment, and label embedding trees. We apply the methods to the use case of whole brain parcellation with 108 classes based on 3D MRI images. While binary encodings have proven efficient in so-called extreme classification problems in computer vision, we faced challenges in reaching state-of-the-art segmentation quality with binary encodings. Compared to one-hot encoding (Dice Similarity Coefficient (DSC) = 82.4 (2.8)), we report reduced segmentation performance with the binary segmentation approaches, achieving DSCs in the range from 39.3 to 73.8. Informative negative results all too often go unpublished. We hope that this work inspires future research of compact encoding strategies for large multi-class segmentation tasks.

Method: Developed OIG-Bench with a semi-automated annotation pipeline using multiple intelligent agents to generate preliminary image descriptions, assisting humans in constructing image-text pairs. Evaluated 29 state-of-the-art MLLMs.

Result: Qwen2.5-VL-72B performed best with 77% overall accuracy. All models showed weaknesses in semantic understanding and logical reasoning. The multi-agent annotation system outperformed all MLLMs in image captioning.

Conclusion: Current MLLMs still struggle to accurately interpret complex visual-text relationships in One-Image Guides. The multi-agent annotation system shows promise as a high-quality image description generator and dataset construction tool.

Abstract: Recent advances in Multimodal Large Language Models (MLLMs) have demonstrated impressive capabilities. However, evaluating their capacity for human-like understanding in One-Image Guides remains insufficiently explored. One-Image Guides are a visual format combining text, imagery, and symbols to present reorganized and structured information for easier comprehension, which are specifically designed for human viewing and inherently embody the characteristics of human perception and understanding. Here, we present OIG-Bench, a comprehensive benchmark focused on One-Image Guide understanding across diverse domains. To reduce the cost of manual annotation, we developed a semi-automated annotation pipeline in which multiple intelligent agents collaborate to generate preliminary image descriptions, assisting humans in constructing image-text pairs. With OIG-Bench, we have conducted a comprehensive evaluation of 29 state-of-the-art MLLMs, including both proprietary and open-source models. The results show that Qwen2.5-VL-72B performs the best among the evaluated models, with an overall accuracy of 77%. Nevertheless, all models exhibit notable weaknesses in semantic understanding and logical reasoning, indicating that current MLLMs still struggle to accurately interpret complex visual-text relationships. In addition, we also demonstrate that the proposed multi-agent annotation system outperforms all MLLMs in image captioning, highlighting its potential as both a high-quality image description generator and a valuable tool for future dataset construction. Datasets are available at https://github.com/XiejcSYSU/OIG-Bench.

Method: Two-stage framework: 1) Scaffolding stage uses supervised fine-tuning on synthetic chain-of-thought exemplars to instill a geospatial thinking paradigm; 2) Elevating stage uses GRPO-based reinforcement learning on weakly-supervised cross-view pairing proxy for verifiable and scalable reward signals.

Result: Achieves state-of-the-art performance across various geospatial reasoning benchmarks, demonstrating effective geospatial reasoning capabilities in vision-language models.

Conclusion: Geo-R1 successfully extends geospatial modeling to reasoning-first post-training and provides a scalable approach for enhancing geospatial reasoning without expensive human annotations.

Abstract: We introduce Geo-R1, a reasoning-centric post-training framework that unlocks geospatial reasoning in vision-language models by combining thinking scaffolding and elevating. In the scaffolding stage, Geo-R1 instills a ``geospatial thinking paradigm" via supervised fine-tuning on synthetic chain-of-thought exemplars, enabling models to connect visual cues with geographic priors without costly human reasoning annotations. In the elevating stage, it uses GRPO-based reinforcement learning on a weakly-supervised cross-view pairing proxy. This design supplies a verifiable and scalable reward signal: teaching models to capture and reconcile features across modalities, and harnessing reasoning for accurate prediction. Geo-R1 extends geospatial modeling from domain pretraining / supervised finetuning to reasoning-first post-training, and achieves state-of-the-art performance across various geospatial reasoning benchmarks. Our model is available at https://huggingface.co/miniHui/Geo-R1.

Method: Propose USN metric to identify certifiable robustness, then prune neurons with low USN while retaining high-USN neurons. Introduce Wasserstein distance loss to concentrate pruned neurons across layers for better pruning effectiveness.

Result: Extensive experiments on robust keypoint detection with realistic brightness and contrast perturbations show superior robustness certification performance and efficiency compared to baselines.

Conclusion: The USN-based pruning approach effectively addresses over-parameterization issues in certified robustness, achieving better performance and efficiency for semantic transformation perturbations.

Abstract: Deep neural networks have been widely adopted in many vision and robotics applications with visual inputs. It is essential to verify its robustness against semantic transformation perturbations, such as brightness and contrast. However, current certified training and robustness certification methods face the challenge of over-parameterization, which hinders the tightness and scalability due to the over-complicated neural networks. To this end, we first analyze stability and variance of layers and neurons against input perturbation, showing that certifiable robustness can be indicated by a fundamental Unbiased and Smooth Neuron metric (USN). Based on USN, we introduce a novel neural network pruning method that removes neurons with low USN and retains those with high USN, thereby preserving model expressiveness without over-parameterization. To further enhance this pruning process, we propose a new Wasserstein distance loss to ensure that pruned neurons are more concentrated across layers. We validate our approach through extensive experiments on the challenging robust keypoint detection task, which involves realistic brightness and contrast perturbations, demonstrating that our method achieves superior robustness certification performance and efficiency compared to baselines.

Method: Proposes transport-based mathematical model viewing pixels as observations of template pattern undergoing unknown deformations, representing them in SCDT domain, then using unsupervised subspace modeling to construct background signal model.

Result: Comprehensive evaluations across five distinct datasets demonstrate superiority over state-of-the-art methods.

Conclusion: The proposed approach effectively detects anomalous signals as deviations from learned background models in the SCDT domain.

Abstract: Hyperspectral anomaly detection (HAD), a crucial approach for many civilian and military applications, seeks to identify pixels with spectral signatures that are anomalous relative to a preponderance of background signatures. Significant effort has been made to improve HAD techniques, but challenges arise due to complex real-world environments and, by definition, limited prior knowledge of potential signatures of interest. This paper introduces a novel HAD method by proposing a transport-based mathematical model to describe the pixels comprising a given hyperspectral image. In this approach, hyperspectral pixels are viewed as observations of a template pattern undergoing unknown deformations that enables their representation in the signed cumulative distribution transform (SCDT) domain. An unsupervised subspace modeling technique is then used to construct a model of abundant background signals in this domain, whereupon anomalous signals are detected as deviations from the learned model. Comprehensive evaluations across five distinct datasets illustrate the superiority of our approach compared to state-of-the-art methods.

Method: Systematic examination of architecture design choices, computation distribution optimization, omni-training objective for multifunctional VAE networks, and novel latent consistency loss for improved reconstruction.

Result: Achieves ultra-high compression ratios with real-time decoding on GPU and mobile devices, outperforms prior methods in reconstruction metrics by large margins, and enables fast, high-quality text-to-video generation.

Conclusion: H3AE demonstrates that optimized autoencoder design significantly enhances latent diffusion models for video generation, providing superior compression, speed, and quality compared to existing approaches.

Abstract: Autoencoder (AE) is the key to the success of latent diffusion models for image and video generation, reducing the denoising resolution and improving efficiency. However, the power of AE has long been underexplored in terms of network design, compression ratio, and training strategy. In this work, we systematically examine the architecture design choices and optimize the computation distribution to obtain a series of efficient and high-compression video AEs that can decode in real time even on mobile devices. We also propose an omni-training objective to unify the design of plain Autoencoder and image-conditioned I2V VAE, achieving multifunctionality in a single VAE network but with enhanced quality. In addition, we propose a novel latent consistency loss that provides stable improvements in reconstruction quality. Latent consistency loss outperforms prior auxiliary losses including LPIPS, GAN and DWT in terms of both quality improvements and simplicity. H3AE achieves ultra-high compression ratios and real-time decoding speed on GPU and mobile, and outperforms prior arts in terms of reconstruction metrics by a large margin. We finally validate our AE by training a DiT on its latent space and demonstrate fast, high-quality text-to-video generation capability.

Method: Formulates watermarking as key-dependent perturbation of model parameters using Mixture of LoRA Markers - binary keys activate lightweight LoRA adapters in residual and attention blocks.

Result: Preserves image quality while achieving robust key recovery against distortions, compression, regeneration, averaging attacks, and black-box adversarial attacks on the extractor.

Conclusion: MOLM provides an effective watermarking solution with desired properties of imperceptibility, fidelity, verifiability, and robustness without requiring key-specific retraining.

Abstract: Generative models can generate photorealistic images at scale. This raises urgent concerns about the ability to detect synthetically generated images and attribute these images to specific sources. While watermarking has emerged as a possible solution, existing methods remain fragile to realistic distortions, susceptible to adaptive removal, and expensive to update when the underlying watermarking key changes. We propose a general watermarking framework that formulates the encoding problem as key-dependent perturbation of the parameters of a generative model. Within this framework, we introduce Mixture of LoRA Markers (MOLM), a routing-based instantiation in which binary keys activate lightweight LoRA adapters inside residual and attention blocks. This design avoids key-specific re-training and achieves the desired properties such as imperceptibility, fidelity, verifiability, and robustness. Experiments on Stable Diffusion and FLUX show that MOLM preserves image quality while achieving robust key recovery against distortions, compression and regeneration, averaging attacks, and black-box adversarial attacks on the extractor.

Method: Proposed a fully end-to-end data-driven framework that predicts relative permittivity from measured fields using data distribution priors to compensate for lack of physical information, enabling feed-forward prediction with strong robustness to transmitter sparsity.

Result: Extensive experiments show the method outperforms state-of-the-art approaches in reconstruction accuracy and robustness, achieving high-quality results even with a single transmitter where previous methods consistently fail.

Conclusion: This work offers a fundamentally new perspective on electromagnetic inverse scattering and represents a major step toward cost-effective practical solutions for electromagnetic imaging applications like medical imaging.

Abstract: Solving Electromagnetic Inverse Scattering Problems (EISP) is fundamental in applications such as medical imaging, where the goal is to reconstruct the relative permittivity from scattered electromagnetic field. This inverse process is inherently ill-posed and highly nonlinear, making it particularly challenging, especially under sparse transmitter setups, e.g., with only one transmitter. A recent machine learning-based approach, Img-Interiors, shows promising results by leveraging continuous implicit functions. However, it requires time-consuming case-specific optimization and fails under sparse transmitter setups. To address these limitations, we revisit EISP from a data-driven perspective. The scarcity of transmitters leads to an insufficient amount of measured data, which fails to capture adequate physical information for stable inversion. Built on this insight, we propose a fully end-to-end and data-driven framework that predicts the relative permittivity of scatterers from measured fields, leveraging data distribution priors to compensate for the lack of physical information. This design enables data-driven training and feed-forward prediction of relative permittivity while maintaining strong robustness to transmitter sparsity. Extensive experiments show that our method outperforms state-of-the-art approaches in reconstruction accuracy and robustness. Notably, it achieves high-quality results even with a single transmitter, a setting where previous methods consistently fail. This work offers a fundamentally new perspective on electromagnetic inverse scattering and represents a major step toward cost-effective practical solutions for electromagnetic imaging.

Method: Proposes CROWD framework with two components: CROWD-Discover uses Submodular Conditional Gain functions to mine unknown instances distinct from known objects, and CROWD-Learn employs combinatorial objectives to disentangle known/unknown representations while maintaining discriminative coherence.

Result: Achieves improvements of 2.83% and 2.05% in known-class accuracy on M-OWODB and S-OWODB benchmarks, and nearly 2.4x unknown recall compared to leading baselines.

Conclusion: CROWD effectively mitigates semantic confusion and catastrophic forgetting in open-world object detection through combinatorial discovery and learning approaches.

Abstract: Open-World Object Detection (OWOD) enriches traditional object detectors by enabling continual discovery and integration of unknown objects via human guidance. However, existing OWOD approaches frequently suffer from semantic confusion between known and unknown classes, alongside catastrophic forgetting, leading to diminished unknown recall and degraded known-class accuracy. To overcome these challenges, we propose Combinatorial Open-World Detection (CROWD), a unified framework reformulating unknown object discovery and adaptation as an interwoven combinatorial (set-based) data-discovery (CROWD-Discover) and representation learning (CROWD-Learn) task. CROWD-Discover strategically mines unknown instances by maximizing Submodular Conditional Gain (SCG) functions, selecting representative examples distinctly dissimilar from known objects. Subsequently, CROWD-Learn employs novel combinatorial objectives that jointly disentangle known and unknown representations while maintaining discriminative coherence among known classes, thus mitigating confusion and forgetting. Extensive evaluations on OWOD benchmarks illustrate that CROWD achieves improvements of 2.83% and 2.05% in known-class accuracy on M-OWODB and S-OWODB, respectively, and nearly 2.4x unknown recall compared to leading baselines.

Method: Created a benchmark of 2,720 QA pairs using high-resolution scans of public-domain paintings with dense visual content, manually annotated across six task categories to probe thorough scene understanding.

Result: Even the best model (o3) among 37 tested achieved only 19.6% accuracy on the hardest test split and 69.5% overall, revealing multiple failure modes including counting errors, OCR failures, and logical inconsistencies.

Conclusion: VisualOverload exposes a critical gap in current vision models’ ability to handle densely populated scenes and provides a crucial resource for developing better models with improved detailed visual reasoning capabilities.

Abstract: Is basic visual understanding really solved in state-of-the-art VLMs? We present VisualOverload, a slightly different visual question answering (VQA) benchmark comprising 2,720 question-answer pairs, with privately held ground-truth responses. Unlike prior VQA datasets that typically focus on near global image understanding, VisualOverload challenges models to perform simple, knowledge-free vision tasks in densely populated (or, overloaded) scenes. Our dataset consists of high-resolution scans of public-domain paintings that are populated with multiple figures, actions, and unfolding subplots set against elaborately detailed backdrops. We manually annotated these images with questions across six task categories to probe for a thorough understanding of the scene. We hypothesize that current benchmarks overestimate the performance of VLMs, and encoding and reasoning over details is still a challenging task for them, especially if they are confronted with densely populated scenes. Indeed, we observe that even the best model (o3) out of 37 tested models only achieves 19.6% accuracy on our hardest test split and overall 69.5% accuracy on all questions. Beyond a thorough evaluation, we complement our benchmark with an error analysis that reveals multiple failure modes, including a lack of counting skills, failure in OCR, and striking logical inconsistencies under complex tasks. Altogether, VisualOverload exposes a critical gap in current vision models and offers a crucial resource for the community to develop better models. Benchmark: http://paulgavrikov.github.io/visualoverload

Method: Proposed ExpDWT-VAE uses dual branches: one processes spatial domain input with convolutional operations, while the other extracts frequency-domain features via 2D Haar wavelet decomposition, convolution, and inverse DWT reconstruction. The integrated spatial-frequency representation is refined through convolutional and diagonal Gaussian mapping.

Result: Experimental results on a new satellite imagery dataset from TerraFly mapping system demonstrate improved latent space representation across multiple performance metrics compared to existing methods.

Conclusion: The ExpDWT-VAE method effectively enhances latent space representation for satellite imagery by leveraging both spatial and frequency-domain information through wavelet transform integration, providing better foundations for LDM applications in remote sensing.

Abstract: Latent Diffusion Models (LDM), a subclass of diffusion models, mitigate the computational complexity of pixel-space diffusion by operating within a compressed latent space constructed by Variational Autoencoders (VAEs), demonstrating significant advantages in Remote Sensing (RS) applications. Though numerous studies enhancing LDMs have been conducted, investigations explicitly targeting improvements within the intrinsic latent space remain scarce. This paper proposes an innovative perspective, utilizing the Discrete Wavelet Transform (DWT) to enhance the VAE’s latent space representation, designed for satellite imagery. The proposed method, ExpDWT-VAE, introduces dual branches: one processes spatial domain input through convolutional operations, while the other extracts and processes frequency-domain features via 2D Haar wavelet decomposition, convolutional operation, and inverse DWT reconstruction. These branches merge to create an integrated spatial-frequency representation, further refined through convolutional and diagonal Gaussian mapping into a robust latent representation. We utilize a new satellite imagery dataset housed by the TerraFly mapping system to validate our method. Experimental results across several performance metrics highlight the efficacy of the proposed method at enhancing latent space representation.

Method: Proposed BiFlow - a dual-stream flow matching model that concurrently denoises historical observations and forecasts future motion using shared latent representation. Includes EgoAnchor mechanism for feature modulation to better model agent intent.

Result: BiFlow achieves state-of-the-art performance, reducing minADE and minFDE by 10-15% on average and demonstrating superior robustness compared to existing methods.

Conclusion: The EgoTraj-Bench benchmark and BiFlow model provide a critical foundation for developing trajectory forecasting systems resilient to real-world ego-centric perception challenges.

Abstract: Reliable trajectory prediction from an ego-centric perspective is crucial for robotic navigation in human-centric environments. However, existing methods typically assume idealized observation histories, failing to account for the perceptual artifacts inherent in first-person vision, such as occlusions, ID switches, and tracking drift. This discrepancy between training assumptions and deployment reality severely limits model robustness. To bridge this gap, we introduce EgoTraj-Bench, the first real-world benchmark that grounds noisy, first-person visual histories in clean, bird’s-eye-view future trajectories, enabling robust learning under realistic perceptual constraints. Building on this benchmark, we propose BiFlow, a dual-stream flow matching model that concurrently denoises historical observations and forecasts future motion by leveraging a shared latent representation. To better model agent intent, BiFlow incorporates our EgoAnchor mechanism, which conditions the prediction decoder on distilled historical features via feature modulation. Extensive experiments show that BiFlow achieves state-of-the-art performance, reducing minADE and minFDE by 10-15% on average and demonstrating superior robustness. We anticipate that our benchmark and model will provide a critical foundation for developing trajectory forecasting systems truly resilient to the challenges of real-world, ego-centric perception.

Method: Comparative analysis between supervised lightweight CNN and zero-shot BiomedCLIP VLM on pneumonia detection (PneumoniaMNIST) and tuberculosis detection (Shenzhen TB dataset), with decision threshold calibration on validation set.

Result: After calibration, BiomedCLIP achieved superior F1-score of 0.8841 vs CNN’s 0.8803 for pneumonia detection, and dramatically improved from 0.4812 to 0.7684 (close to CNN’s 0.7834) for tuberculosis detection.

Conclusion: Proper calibration is essential for unlocking the full diagnostic potential of zero-shot VLMs, enabling them to match or outperform task-specific supervised models in medical imaging tasks.

Abstract: The accurate interpretation of chest radiographs using automated methods is a critical task in medical imaging. This paper presents a comparative analysis between a supervised lightweight Convolutional Neural Network (CNN) and a state-of-the-art, zero-shot medical Vision-Language Model (VLM), BiomedCLIP, across two distinct diagnostic tasks: pneumonia detection on the PneumoniaMNIST benchmark and tuberculosis detection on the Shenzhen TB dataset. Our experiments show that supervised CNNs serve as highly competitive baselines in both cases. While the default zero-shot performance of the VLM is lower, we demonstrate that its potential can be unlocked via a simple yet crucial remedy: decision threshold calibration. By optimizing the classification threshold on a validation set, the performance of BiomedCLIP is significantly boosted across both datasets. For pneumonia detection, calibration enables the zero-shot VLM to achieve a superior F1-score of 0.8841, surpassing the supervised CNN’s 0.8803. For tuberculosis detection, calibration dramatically improves the F1-score from 0.4812 to 0.7684, bringing it close to the supervised baseline’s 0.7834. This work highlights a key insight: proper calibration is essential for leveraging the full diagnostic power of zero-shot VLMs, enabling them to match or even outperform efficient, task-specific supervised models.

Method: PAL-UI combines dual-level summarization (observation-level cues and action-level outcomes) with a dedicated retrieval tool that allows agents to recall specific historical screenshots during planning. Models are trained on 8.6K mobile GUI navigation samples based on Qwen2.5-VL.

Result: PAL-UI significantly outperforms baseline models and prior methods in mobile GUI navigation tasks, even in data-efficient settings, and shows strong cross-domain generalization with notable improvements in web navigation without additional training.

Conclusion: The work demonstrates the potential of active memory retrieval for enhancing long-horizon planning capabilities of vision-based GUI agents.

Abstract: Graphical User Interface (GUI) agents powered by Multimodal Large Language Models (MLLMs) promise human-like interaction with software applications, yet long-horizon tasks remain challenging due to memory limitations. Existing approaches either truncate history or rely on simple textual summaries, which risk losing critical information when past visual details become necessary for future decisions. In this paper, we propose \textbf{PAL-UI} (\textbf{P}lanning with \textbf{A}ctive \textbf{L}ook-back), a novel framework that enables GUI agents to adaptively retrieve past observations when required. PAL-UI combines a dual-level summarization agent, capturing both observation-level cues and action-level outcomes, with a dedicated retrieval tool that allows the agent to recall specific historical screenshots during planning. We curate a step-level instruction dataset of 8.6K samples from mobile GUI navigation trajectories and train \textbf{PAL-UI-3B} and \textbf{PAL-UI-7B} models based on Qwen2.5-VL. Extensive experiments demonstrate that PAL-UI significantly outperforms baseline models and prior methods in mobile GUI navigation tasks, even under data-efficient settings. Moreover, PAL-UI exhibits strong cross-domain generalization, achieving notable improvements in web navigation without additional training. Our work highlights the potential of active memory retrieval for long-horizon planning capabilities of vision-based GUI agents.

Method: Developed Interactive-MEN-RT - a dedicated IMIS tool with multiple clinically relevant interaction methods (point annotations, bounding boxes, lasso tools, scribbles) for clinician-assisted 3D meningioma segmentation in RT workflows.

Result: Evaluated on 500 contrast-enhanced T1-weighted MRI scans from BraTS 2025 Meningioma RT Segmentation Challenge, achieving Dice similarity coefficients of 77.6% and Intersection over Union scores of 64.8%, substantially outperforming other methods.

Conclusion: The results emphasize the need for clinically tailored segmentation solutions in critical applications like meningioma RT planning, demonstrating the superiority of specialized interactive tools over generic segmentation approaches.

Abstract: Precise delineation of meningiomas is crucial for effective radiotherapy (RT) planning, directly influencing treatment efficacy and preservation of adjacent healthy tissues. While automated deep learning approaches have demonstrated considerable potential, achieving consistently accurate clinical segmentation remains challenging due to tumor heterogeneity. Interactive Medical Image Segmentation (IMIS) addresses this challenge by integrating advanced AI techniques with clinical input. However, generic segmentation tools, despite widespread applicability, often lack the specificity required for clinically critical and disease-specific tasks like meningioma RT planning. To overcome these limitations, we introduce Interactive-MEN-RT, a dedicated IMIS tool specifically developed for clinician-assisted 3D meningioma segmentation in RT workflows. The system incorporates multiple clinically relevant interaction methods, including point annotations, bounding boxes, lasso tools, and scribbles, enhancing usability and clinical precision. In our evaluation involving 500 contrast-enhanced T1-weighted MRI scans from the BraTS 2025 Meningioma RT Segmentation Challenge, Interactive-MEN-RT demonstrated substantial improvement compared to other segmentation methods, achieving Dice similarity coefficients of up to 77.6% and Intersection over Union scores of 64.8%. These results emphasize the need for clinically tailored segmentation solutions in critical applications such as meningioma RT planning. The code is publicly available at: https://github.com/snuh-rad-aicon/Interactive-MEN-RT

Method: Proposes MLLM-DiT framework where a pretrained multimodal large language model performs cross-modal reasoning to ground entities and disentangle roles/attributes/interactions, generating subject-aware hidden states that condition the diffusion transformer.

Result: Experiments on OpenS2V benchmark show superior performance across subject consistency, naturalness, and text relevance, outperforming existing open-source and commercial models.

Conclusion: BindWeave effectively addresses subject-consistent video generation challenges through unified cross-modal reasoning framework.

Abstract: Diffusion Transformer has shown remarkable abilities in generating high-fidelity videos, delivering visually coherent frames and rich details over extended durations. However, existing video generation models still fall short in subject-consistent video generation due to an inherent difficulty in parsing prompts that specify complex spatial relationships, temporal logic, and interactions among multiple subjects. To address this issue, we propose BindWeave, a unified framework that handles a broad range of subject-to-video scenarios from single-subject cases to complex multi-subject scenes with heterogeneous entities. To bind complex prompt semantics to concrete visual subjects, we introduce an MLLM-DiT framework in which a pretrained multimodal large language model performs deep cross-modal reasoning to ground entities and disentangle roles, attributes, and interactions, yielding subject-aware hidden states that condition the diffusion transformer for high-fidelity subject-consistent video generation. Experiments on the OpenS2V benchmark demonstrate that our method achieves superior performance across subject consistency, naturalness, and text relevance in generated videos, outperforming existing open-source and commercial models.

Method: Proposes SCNet with three components: frequency band selection strategy for faster convergence, parameter optimization using Lipschitz constant to restrict high-frequency noise learning, and SSR module for frequency domain separation and low-rank reconstruction to separate noise from structural details.

Result: Experiments on synthetic and real-world datasets verify the effectiveness of SCNet in improving denoising performance while preserving high-frequency details.

Conclusion: SCNet successfully addresses the spectral bias limitations in self-supervised denoising by controlling frequency learning and separating noise from structural details, demonstrating improved performance on both synthetic and real datasets.

Abstract: Current self-supervised denoising methods for paired noisy images typically involve mapping one noisy image through the network to the other noisy image. However, after measuring the spectral bias of such methods using our proposed Image Pair Frequency-Band Similarity, it suffers from two practical limitations. Firstly, the high-frequency structural details in images are not preserved well enough. Secondly, during the process of fitting high frequencies, the network learns high-frequency noise from the mapped noisy images. To address these challenges, we introduce a Spectral Controlling network (SCNet) to optimize self-supervised denoising of paired noisy images. First, we propose a selection strategy to choose frequency band components for noisy images, to accelerate the convergence speed of training. Next, we present a parameter optimization method that restricts the learning ability of convolutional kernels to high-frequency noise using the Lipschitz constant, without changing the network structure. Finally, we introduce the Spectral Separation and low-rank Reconstruction module (SSR module), which separates noise and high-frequency details through frequency domain separation and low-rank space reconstruction, to retain the high-frequency structural details of images. Experiments performed on synthetic and real-world datasets verify the effectiveness of SCNet.

Method: Uses two main techniques: 1) IoU-weighted entropy objective that focuses adaptation on spatially coherent proposal clusters, and 2) image-conditioned prompt selection that ranks and fuses the most informative prompts with detector logits.

Result: Shows consistent improvements across diverse distribution shifts including stylized domains, driving scenes, low-light conditions, and common corruptions on YOLO-World and Grounding DINO detectors.

Conclusion: VLOD-TTA effectively adapts vision-language object detectors to domain shifts through proposal clustering and intelligent prompt selection, outperforming zero-shot and baseline TTA methods.

Abstract: Vision-language object detectors (VLODs) such as YOLO-World and Grounding DINO achieve impressive zero-shot recognition by aligning region proposals with text representations. However, their performance often degrades under domain shift. We introduce VLOD-TTA, a test-time adaptation (TTA) framework for VLODs that leverages dense proposal overlap and image-conditioned prompt scores. First, an IoU-weighted entropy objective is proposed that concentrates adaptation on spatially coherent proposal clusters and reduces confirmation bias from isolated boxes. Second, image-conditioned prompt selection is introduced, which ranks prompts by image-level compatibility and fuses the most informative prompts with the detector logits. Our benchmarking across diverse distribution shifts – including stylized domains, driving scenes, low-light conditions, and common corruptions – shows the effectiveness of our method on two state-of-the-art VLODs, YOLO-World and Grounding DINO, with consistent improvements over the zero-shot and TTA baselines. Code : https://github.com/imatif17/VLOD-TTA

Method: Developed a benchmark with 1.4M generated instances and curated test set covering text-guided and visual settings, systematically varying digit scale, move complexity, solution multiplicity, and operators.

Result: Evaluations of 14 vision-language models show substantial limitations: closed-source models succeed only on simple cases, open-source models fail in visual regime, while humans achieve over 90% accuracy.

Conclusion: MathSticks establishes a rigorous benchmark for advancing compositional reasoning across vision and symbols, highlighting the gap between current AI models and human performance.

Abstract: We introduce \textsc{MathSticks}, a benchmark for Visual Symbolic Compositional Reasoning (VSCR), which unifies visual perception, symbolic manipulation, and arithmetic consistency. Each task presents an incorrect matchstick equation that must be corrected by moving one or two sticks under strict conservation rules. The benchmark includes both text-guided and purely visual settings, systematically covering digit scale, move complexity, solution multiplicity, and operator variation, with 1.4M generated instances and a curated test set. Evaluations of 14 vision–language models reveal substantial limitations: closed-source models succeed only on simple cases, open-source models fail in the visual regime, while humans exceed 90% accuracy. These findings establish \textsc{MathSticks} as a rigorous testbed for advancing compositional reasoning across vision and symbols. Our code and dataset are publicly available at https://github.com/Yuheng2000/MathSticks.

Method: NAGL framework with two components: Residual Mining (RM) extracts abnormal patterns from normal-abnormal reference residuals, and Anomaly Feature Learning (AFL) adaptively learns anomaly features through residual mapping.

Result: Extensive experiments across multiple benchmarks show the method significantly outperforms existing GAD approaches.

Conclusion: This is the first work to adopt a mixture of normal and abnormal samples as references in generalist anomaly detection, enabling more accurate and efficient cross-domain anomaly detection.

Abstract: Generalist Anomaly Detection (GAD) aims to train a unified model on an original domain that can detect anomalies in new target domains. Previous GAD methods primarily use only normal samples as references, overlooking the valuable information contained in anomalous samples that are often available in real-world scenarios. To address this limitation, we propose a more practical approach: normal-abnormal-guided generalist anomaly detection, which leverages both normal and anomalous samples as references to guide anomaly detection across diverse domains. We introduce the Normal-Abnormal Generalist Learning (NAGL) framework, consisting of two key components: Residual Mining (RM) and Anomaly Feature Learning (AFL). RM extracts abnormal patterns from normal-abnormal reference residuals to establish transferable anomaly representations, while AFL adaptively learns anomaly features in query images through residual mapping to identify instance-aware anomalies. Our approach effectively utilizes both normal and anomalous references for more accurate and efficient cross-domain anomaly detection. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms existing GAD approaches. This work represents the first to adopt a mixture of normal and abnormal samples as references in generalist anomaly detection. The code and datasets are available at https://github.com/JasonKyng/NAGL.

Method: RAF simultaneously extracts and fuses image representations as relative features with corresponding numerical values as absolute features, creating comprehensive matrix representations that prevent feature ambiguity.

Result: RAF achieved solution time reductions of 0.08s-0.29s (5.86%-11.50% faster) compared to conventional image-based approaches on SuiteSparse and BMCMat datasets, achieving state-of-the-art performance.

Conclusion: RAF effectively enhances image-based selection approaches by preventing feature ambiguity, improving selection accuracy, and unlocking the potential of image-based methods for iterative solver selection in sparse linear systems.

Abstract: Iterative method selection is crucial for solving sparse linear systems because these methods inherently lack robustness. Though image-based selection approaches have shown promise, their feature extraction techniques might encode distinct matrices into identical image representations, leading to the same selection and suboptimal method. In this paper, we introduce RAF (Relative-Absolute Fusion), an efficient feature extraction technique to enhance image-based selection approaches. By simultaneously extracting and fusing image representations as relative features with corresponding numerical values as absolute features, RAF achieves comprehensive matrix representations that prevent feature ambiguity across distinct matrices, thus improving selection accuracy and unlocking the potential of image-based selection approaches. We conducted comprehensive evaluations of RAF on SuiteSparse and our developed BMCMat (Balanced Multi-Classification Matrix dataset), demonstrating solution time reductions of 0.08s-0.29s for sparse linear systems, which is 5.86%-11.50% faster than conventional image-based selection approaches and achieves state-of-the-art (SOTA) performance. BMCMat is available at https://github.com/zkqq/BMCMat.

Method: Uses a diffusion-based generative model that learns the distribution of plausible hand poses conditioned on affordance descriptions inferred from a large vision-language model (VLM).

Result: Extensive experiments on HOGraspNet dataset show significant improvement in hand pose estimation over recent regression methods and diffusion-based refinement without contextual reasoning.

Conclusion: Affordance-guided refinement using textual descriptions enables more accurate and functionally coherent hand pose reconstruction in severely occluded scenarios.

Abstract: How can we reconstruct 3D hand poses when large portions of the hand are heavily occluded by itself or by objects? Humans often resolve such ambiguities by leveraging contextual knowledge – such as affordances, where an object’s shape and function suggest how the object is typically grasped. Inspired by this observation, we propose a generative prior for hand pose refinement guided by affordance-aware textual descriptions of hand-object interactions (HOI). Our method employs a diffusion-based generative model that learns the distribution of plausible hand poses conditioned on affordance descriptions, which are inferred from a large vision-language model (VLM). This enables the refinement of occluded regions into more accurate and functionally coherent hand poses. Extensive experiments on HOGraspNet, a 3D hand-affordance dataset with severe occlusions, demonstrate that our affordance-guided refinement significantly improves hand pose estimation over both recent regression methods and diffusion-based refinement lacking contextual reasoning.

Method: Progressive Consistency Distillation framework that decomposes feature space perturbations into token-wise and layer-wise dimensions, using token consistency distillation and layer consistency distillation with teacher model guidance.

Result: Extensive experiments demonstrate superior effectiveness, robustness, and generalization capabilities compared to existing methods.

Conclusion: EPIC provides an effective progressive learning approach for efficient MLLMs that addresses the training difficulty challenges of visual token compression through consistency distillation techniques.

Abstract: Visual tokens consume substantial computational resources in multi-modal large models (MLLMs), significantly compromising their efficiency. Recent works have attempted to improve efficiency by compressing visual tokens during training, either through modifications to model components or by introducing additional parameters. However, they often overlook the increased learning difficulty caused by such compression, as the model’s parameter space struggles to quickly adapt to the substantial perturbations in the feature space induced by token compression. In this work, we propose to develop Efficient MLLMs via Progressive Consistency Distillation (EPIC), a progressive learning framework. Specifically, by decomposing the feature space perturbations introduced by token compression along the token-wise and layer-wise dimensions, we introduce token consistency distillation and layer consistency distillation, respectively, aiming to reduce the training difficulty by leveraging guidance from a teacher model and following a progressive learning trajectory. Extensive experiments demonstrate the superior effectiveness, robustness, and generalization capabilities of our proposed framework.

Method: Unified 8 publicly available datasets into standardized suite spanning 4 regression and 5 classification tasks covering functional, structural, diagnostic, and view recognition endpoints. Evaluated leading FMs under consistent zero-shot, probing, and alignment protocols.

Result: Found complementary strengths: temporal modeling critical for functional regression, retrieval provides robustness under distribution shift, domain-specific text encoders capture physiologically meaningful axes. General-purpose encoders transfer strongly but struggle with fine-grained distinctions.

Conclusion: CardioBench establishes reproducible reference point and offers actionable insights to guide future echocardiography foundation model design, with released preprocessing, splits, and public evaluation pipelines.

Abstract: Foundation models (FMs) are reshaping medical imaging, yet their application in echocardiography remains limited. While several echocardiography-specific FMs have recently been introduced, no standardized benchmark exists to evaluate them. Echocardiography poses unique challenges, including noisy acquisitions, high frame redundancy, and limited public datasets. Most existing solutions evaluate on private data, restricting comparability. To address this, we introduce CardioBench, a comprehensive benchmark for echocardiography FMs. CardioBench unifies eight publicly available datasets into a standardized suite spanning four regression and five classification tasks, covering functional, structural, diagnostic, and view recognition endpoints. We evaluate several leading FM, including cardiac-specific, biomedical, and general-purpose encoders, under consistent zero-shot, probing, and alignment protocols. Our results highlight complementary strengths across model families: temporal modeling is critical for functional regression, retrieval provides robustness under distribution shift, and domain-specific text encoders capture physiologically meaningful axes. General-purpose encoders transfer strongly and often close the gap with probing, but struggle with fine-grained distinctions like view classification and subtle pathology recognition. By releasing preprocessing, splits, and public evaluation pipelines, CardioBench establishes a reproducible reference point and offers actionable insights to guide the design of future echocardiography foundation models.

Method: Two-stage cascaded diffusion: 1) Joint diffusion model samples diverse 3D joint hypotheses, 2) Mesh Latent Diffusion Model reconstructs 3D hand mesh conditioned on joint samples, trained with diverse hypotheses in learned latent space.

Result: Achieves state-of-the-art performance on FreiHAND and HO3Dv2 datasets while effectively modeling pose distributions.

Conclusion: The framework successfully combines probabilistic modeling with cascaded refinement, learning distribution-aware joint-mesh relationships and robust hand priors to enhance accuracy in 3D hand pose reconstruction.

Abstract: Deterministic models for 3D hand pose reconstruction, whether single-staged or cascaded, struggle with pose ambiguities caused by self-occlusions and complex hand articulations. Existing cascaded approaches refine predictions in a coarse-to-fine manner but remain deterministic and cannot capture pose uncertainties. Recent probabilistic methods model pose distributions yet are restricted to single-stage estimation, which often fails to produce accurate 3D reconstructions without refinement. To address these limitations, we propose a coarse-to-fine cascaded diffusion framework that combines probabilistic modeling with cascaded refinement. The first stage is a joint diffusion model that samples diverse 3D joint hypotheses, and the second stage is a Mesh Latent Diffusion Model (Mesh LDM) that reconstructs a 3D hand mesh conditioned on a joint sample. By training Mesh LDM with diverse joint hypotheses in a learned latent space, our framework learns distribution-aware joint-mesh relationships and robust hand priors. Furthermore, the cascaded design mitigates the difficulty of directly mapping 2D images to dense 3D poses, enhancing accuracy through sequential refinement. Experiments on FreiHAND and HO3Dv2 demonstrate that our method achieves state-of-the-art performance while effectively modeling pose distributions.

Method: Built on YOLOv8, Forestpest-YOLO introduces: 1) SPD-Conv for lossless downsampling to preserve small target details, 2) CSPOK for cross-stage feature fusion to enhance multi-scale representation and suppress noise, and 3) VarifocalLoss to focus training on high-quality, hard-to-classify samples.

Result: Extensive experiments on the self-constructed ForestPest dataset show Forestpest-YOLO achieves state-of-the-art performance with marked improvements in detecting small, occluded pests, significantly outperforming established baseline models.

Conclusion: The proposed Forestpest-YOLO framework effectively addresses the unique challenges of forestry pest detection through synergistic architectural innovations, demonstrating superior performance for small, occluded target detection in complex remote sensing environments.

Abstract: Detecting agricultural pests in complex forestry environments using remote sensing imagery is fundamental for ecological preservation, yet it is severely hampered by practical challenges. Targets are often minuscule, heavily occluded, and visually similar to the cluttered background, causing conventional object detection models to falter due to the loss of fine-grained features and an inability to handle extreme data imbalance. To overcome these obstacles, this paper introduces Forestpest-YOLO, a detection framework meticulously optimized for the nuances of forestry remote sensing. Building upon the YOLOv8 architecture, our framework introduces a synergistic trio of innovations. We first integrate a lossless downsampling module, SPD-Conv, to ensure that critical high-resolution details of small targets are preserved throughout the network. This is complemented by a novel cross-stage feature fusion block, CSPOK, which dynamically enhances multi-scale feature representation while suppressing background noise. Finally, we employ VarifocalLoss to refine the training objective, compelling the model to focus on high-quality and hard-to-classify samples. Extensive experiments on our challenging, self-constructed ForestPest dataset demonstrate that Forestpest-YOLO achieves state-of-the-art performance, showing marked improvements in detecting small, occluded pests and significantly outperforming established baseline models.

Method: Compiled a dataset of 5000 Petri dish images with bounding boxes, simulating traditional data collection and few-shot scenarios. Benchmarked vision foundation models against traditional baselines using task-specific metrics.

Result: MaskDINO achieved near-parity with extensively trained YoloV9 when fine-tuned on only 150 images, and maintained competitive performance with just 25 images, remaining reliable on ~70% of samples.

Conclusion: Data-efficient foundation models can match traditional approaches with minimal data requirements, enabling faster development of automated microbiological systems with superior upper-bound performance.

Abstract: The process of quantifying mold colonies on Petri dish samples is of critical importance for the assessment of indoor air quality, as high colony counts can indicate potential health risks and deficiencies in ventilation systems. Conventionally the automation of such a labor-intensive process, as well as other tasks in microbiology, relies on the manual annotation of large datasets and the subsequent extensive training of models like YoloV9. To demonstrate that exhaustive annotation is not a prerequisite anymore when tackling a new vision task, we compile a representative dataset of 5000 Petri dish images annotated with bounding boxes, simulating both a traditional data collection approach as well as few-shot and low-shot scenarios with well curated subsets with instance level masks. We benchmark three vision foundation models against traditional baselines on task specific metrics, reflecting realistic real-world requirements. Notably, MaskDINO attains near-parity with an extensively trained YoloV9 model while finetuned only on 150 images, retaining competitive performance with as few as 25 images, still being reliable on $\approx$ 70% of the samples. Our results show that data-efficient foundation models can match traditional approaches with only a fraction of the required data, enabling earlier development and faster iterative improvement of automated microbiological systems with a superior upper-bound performance than traditional models would achieve.

Method: Adaptive shared experts with LoRA-based MoE, where shared experts get router-computed gating weights normalized with sparse experts, plus fine-grained experts by increasing LoRA experts while reducing rank proportionally.

Result: Extensive experiments on PASCAL-Context benchmark show ASE consistently improves performance across diverse configurations and validates fine-grained designs for MTL.

Conclusion: ASE facilitates STL to MTL transition, enhances expert specialization and cooperation, and enables more effective knowledge sharing under comparable parameter budget.

Abstract: Mixture-of-Experts (MoE) has emerged as a powerful framework for multi-task learning (MTL). However, existing MoE-MTL methods often rely on single-task pretrained backbones and suffer from redundant adaptation and inefficient knowledge sharing during the transition from single-task to multi-task learning (STL to MTL). To address these limitations, we propose adaptive shared experts (ASE) within a low-rank adaptation (LoRA) based MoE, where shared experts are assigned router-computed gating weights jointly normalized with sparse experts. This design facilitates STL to MTL transition, enhances expert specialization, and cooperation. Furthermore, we incorporate fine-grained experts by increasing the number of LoRA experts while proportionally reducing their rank, enabling more effective knowledge sharing under a comparable parameter budget. Extensive experiments on the PASCAL-Context benchmark, under unified training settings, demonstrate that ASE consistently improves performance across diverse configurations and validates the effectiveness of fine-grained designs for MTL.

Method: Proposes Timestamp-aware Rotary Position Embedding (TaRoPE) for any-timestamp interpolation, and decomposes long sequences into segments with appearance-motion decoupled conditioning for seamless transitions.

Result: Outperforms prior methods across all interpolation scenarios (2x to 32x) with higher fidelity and better spatiotemporal continuity.

Conclusion: ArbInterp provides a flexible and efficient solution for arbitrary-length video frame interpolation with superior performance compared to existing fixed-frame methods.

Abstract: Video frame interpolation (VFI), which generates intermediate frames from given start and end frames, has become a fundamental function in video generation applications. However, existing generative VFI methods are constrained to synthesize a fixed number of intermediate frames, lacking the flexibility to adjust generated frame rates or total sequence duration. In this work, we present ArbInterp, a novel generative VFI framework that enables efficient interpolation at any timestamp and of any length. Specifically, to support interpolation at any timestamp, we propose the Timestamp-aware Rotary Position Embedding (TaRoPE), which modulates positions in temporal RoPE to align generated frames with target normalized timestamps. This design enables fine-grained control over frame timestamps, addressing the inflexibility of fixed-position paradigms in prior work. For any-length interpolation, we decompose long-sequence generation into segment-wise frame synthesis. We further design a novel appearance-motion decoupled conditioning strategy: it leverages prior segment endpoints to enforce appearance consistency and temporal semantics to maintain motion coherence, ensuring seamless spatiotemporal transitions across segments. Experimentally, we build comprehensive benchmarks for multi-scale frame interpolation (2x to 32x) to assess generalizability across arbitrary interpolation factors. Results show that ArbInterp outperforms prior methods across all scenarios with higher fidelity and more seamless spatiotemporal continuity. Project website: https://mcg-nju.github.io/ArbInterp-Web/.

Method: Conducted a review of color models and spaces including RGB, CMYK, YUV, CIELAB, CIELUV, and fuzzy-based approaches, along with experimental evaluation from perspectives of device dependency, chromatic consistency, and computational complexity.

Result: Experimental results reveal gaps in existing color models and show that the HS* family is the most aligned with human perception, while identifying key strengths and limitations of different models.

Conclusion: The study provides a reference for researchers in image processing, perceptual computing, digital media, and other color-related fields, outlining open challenges and future directions in color model development.

Abstract: Color representation is essential in computer vision and human-computer interaction. There are multiple color models available. The choice of a suitable color model is critical for various applications. This paper presents a review of color models and spaces, analyzing their theoretical foundations, computational properties, and practical applications. We explore traditional models such as RGB, CMYK, and YUV, perceptually uniform spaces like CIELAB and CIELUV, and fuzzy-based approaches as well. Additionally, we conduct a series of experiments to evaluate color models from various perspectives, like device dependency, chromatic consistency, and computational complexity. Our experimental results reveal gaps in existing color models and show that the HS* family is the most aligned with human perception. The review also identifies key strengths and limitations of different models and outlines open challenges and future directions This study provides a reference for researchers in image processing, perceptual computing, digital media, and any other color-related field.

Method: Proposes multi-level feature adaptor to generate multi-level feature grid representation, dynamic style injection module that learns to extract and adaptively integrate style features, and multi-level cascade decoder for final stylized view generation. Also extends to omni-view style transfer using 3D style references.

Result: Extensive experiments demonstrate outstanding performance for 3D style transfer, effectively preserving multi-scale spatial structures while transferring stylistic characteristics.

Conclusion: MDS-NeRF achieves superior 3D style transfer by specifically reengineering the NeRF pipeline for stylization with multi-level feature representation and dynamic style injection.

Abstract: As the application of neural radiance fields (NeRFs) in various 3D vision tasks continues to expand, numerous NeRF-based style transfer techniques have been developed. However, existing methods typically integrate style statistics into the original NeRF pipeline, often leading to suboptimal results in both content preservation and artistic stylization. In this paper, we present multi-level dynamic style transfer for NeRFs (MDS-NeRF), a novel approach that reengineers the NeRF pipeline specifically for stylization and incorporates an innovative dynamic style injection module. Particularly, we propose a multi-level feature adaptor that helps generate a multi-level feature grid representation from the content radiance field, effectively capturing the multi-scale spatial structure of the scene. In addition, we present a dynamic style injection module that learns to extract relevant style features and adaptively integrates them into the content patterns. The stylized multi-level features are then transformed into the final stylized view through our proposed multi-level cascade decoder. Furthermore, we extend our 3D style transfer method to support omni-view style transfer using 3D style references. Extensive experiments demonstrate that MDS-NeRF achieves outstanding performance for 3D style transfer, preserving multi-scale spatial structures while effectively transferring stylistic characteristics.

Method: Integrates LVLMs with semantic pattern matching and iterative self-questioning refinement, using optimized domain-specific prompting and recursive verification to transform raw visual data into labeled defect datasets.

Result: Achieves 98% accuracy in defective/non-defective classification, 85%-98% annotation accuracy across four defect categories in balanced settings, and 80%-92% accuracy on imbalanced datasets.

Conclusion: ADPT provides a scalable, cost-effective solution for high-fidelity dataset construction, supporting downstream tasks like transfer learning and domain adaptation in structural damage assessment.

Abstract: Automated structural defect annotation is essential for ensuring infrastructure safety while minimizing the high costs and inefficiencies of manual labeling. A novel agentic annotation framework, Agent-based Defect Pattern Tagger (ADPT), is introduced that integrates Large Vision-Language Models (LVLMs) with a semantic pattern matching module and an iterative self-questioning refinement mechanism. By leveraging optimized domain-specific prompting and a recursive verification process, ADPT transforms raw visual data into high-quality, semantically labeled defect datasets without any manual supervision. Experimental results demonstrate that ADPT achieves up to 98% accuracy in distinguishing defective from non-defective images, and 85%-98% annotation accuracy across four defect categories under class-balanced settings, with 80%-92% accuracy on class-imbalanced datasets. The framework offers a scalable and cost-effective solution for high-fidelity dataset construction, providing strong support for downstream tasks such as transfer learning and domain adaptation in structural damage assessment.

Method: Semantically-enhanced landmark identification disentangles foreground/background features, then consensus-driven online augmentation consolidates feature preferences through two-stage voting.

Result: Experiments on REVERIE and R2R show improved generalization and state-of-the-art performance.

Conclusion: Online foreground-background feature augmentation effectively boosts VLN agent generalization capabilities.

Abstract: Following language instructions, vision-language navigation (VLN) agents are tasked with navigating unseen environments. While augmenting multifaceted visual representations has propelled advancements in VLN, the significance of foreground and background in visual observations remains underexplored. Intuitively, foreground regions provide semantic cues, whereas the background encompasses spatial connectivity information. Inspired on this insight, we propose a Consensus-driven Online Feature Augmentation strategy (COFA) with alternative foreground and background features to facilitate the navigable generalization. Specifically, we first leverage semantically-enhanced landmark identification to disentangle foreground and background as candidate augmented features. Subsequently, a consensus-driven online augmentation strategy encourages the agent to consolidate two-stage voting results on feature preferences according to diverse instructions and navigational locations. Experiments on REVERIE and R2R demonstrate that our online foreground-background augmentation boosts the generalization of baseline and attains state-of-the-art performance.

Method: RCOR treats localization as integral part of recognition to decouple object-centric and context-aware modeling, followed by robust non-parametric fusion. It enables localization before recognition even in complex scenes.

Result: RCOR improves performance of both supervised models and VLMs on datasets with in-domain and out-of-domain backgrounds, even without fine-tuning. Works effectively on complex scenes like ImageNet-1k.

Conclusion: RCOR is the first approach that jointly achieves robustness and context-awareness without compromising either, demonstrating that localization before recognition is feasible even in complex visual scenes.

Abstract: In visual recognition, both the object of interest (referred to as foreground, FG, for simplicity) and its surrounding context (background, BG) play an important role. However, standard supervised learning often leads to unintended over-reliance on the BG, known as shortcut learning of spurious correlations, limiting model robustness in real-world deployment settings. In the literature, the problem is mainly addressed by suppressing the BG, sacrificing context information for improved generalization. We propose RCOR – Robust Context-Aware Object Recognition – the first approach that jointly achieves robustness and context-awareness without compromising either. RCOR treats localization as an integral part of recognition to decouple object-centric and context-aware modelling, followed by a robust, non-parametric fusion. It improves the performance of both supervised models and VLM on datasets with both in-domain and out-of-domain BG, even without fine-tuning. The results confirm that localization before recognition is now possible even in complex scenes as in ImageNet-1k.

Method: Proposed an unconditional discriminator (UCD) that removes condition injection, forcing the discriminator to extract more comprehensive and robust features. This approach is theoretically compatible with vanilla GAN and can be implemented as a plug-in.

Result: Achieved 1.47 FID on ImageNet-64, surpassing StyleGAN-XL and state-of-the-art one-step diffusion models. Extensive experiments showed significant performance improvements with high efficiency.

Conclusion: UCD promotes Nash equilibrium in GAN training by enabling better knowledge transfer from discriminator to generator, leading to improved convergence and state-of-the-art performance in one-step generation tasks.

Abstract: Adversarial training turns out to be the key to one-step generation, especially for Generative Adversarial Network (GAN) and diffusion model distillation. Yet in practice, GAN training hardly converges properly and struggles in mode collapse. In this work, we quantitatively analyze the extent of Nash equilibrium in GAN training, and conclude that redundant shortcuts by inputting condition in $D$ disables meaningful knowledge extraction. We thereby propose to employ an unconditional discriminator (UCD), in which $D$ is enforced to extract more comprehensive and robust features with no condition injection. In this way, $D$ is able to leverage better knowledge to supervise $G$, which promotes Nash equilibrium in GAN literature. Theoretical guarantee on compatibility with vanilla GAN theory indicates that UCD can be implemented in a plug-in manner. Extensive experiments confirm the significant performance improvements with high efficiency. For instance, we achieved \textbf{1.47 FID} on the ImageNet-64 dataset, surpassing StyleGAN-XL and several state-of-the-art one-step diffusion models. The code will be made publicly available.

Method: Constructed a large-scale dataset of garment-lookbook pairs using an automated retrieval pipeline that combines visual-language reasoning with object-level localization to align garments across domains.

Result: Created a dataset with three quality levels: 10,000 high-quality pairs, 50,000 medium-quality pairs, and 300,000 low-quality pairs, bridging e-commerce and fashion media domains.

Conclusion: The dataset provides a foundation for models to generate fashion imagery that reflects creativity, atmosphere, and storytelling, moving beyond catalog-style generation.

Abstract: Fashion image generation has so far focused on narrow tasks such as virtual try-on, where garments appear in clean studio environments. In contrast, editorial fashion presents garments through dynamic poses, diverse locations, and carefully crafted visual narratives. We introduce the task of virtual fashion photo-shoot, which seeks to capture this richness by transforming standardized garment images into contextually grounded editorial imagery. To enable this new direction, we construct the first large-scale dataset of garment-lookbook pairs, bridging the gap between e-commerce and fashion media. Because such pairs are not readily available, we design an automated retrieval pipeline that aligns garments across domains, combining visual-language reasoning with object-level localization. We construct a dataset with three garment-lookbook pair accuracy levels: high quality (10,000 pairs), medium quality (50,000 pairs), and low quality (300,000 pairs). This dataset offers a foundation for models that move beyond catalog-style generation and toward fashion imagery that reflects creativity, atmosphere, and storytelling.

Method: Proposed LAKAN module uses facial landmarks as structural prior to dynamically generate KAN parameters, creating instance-specific signals that guide image encoder to focus on informative facial regions with artifacts.

Result: Extensive experiments on multiple public datasets demonstrate superior performance compared to existing methods.

Conclusion: The combination of geometric priors (facial landmarks) with KAN’s learning process creates an effective deepfake detection approach that outperforms current state-of-the-art methods.

Abstract: The rapid development of deepfake generation techniques necessitates robust face forgery detection algorithms. While methods based on Convolutional Neural Networks (CNNs) and Transformers are effective, there is still room for improvement in modeling the highly complex and non-linear nature of forgery artifacts. To address this issue, we propose a novel detection method based on the Kolmogorov-Arnold Network (KAN). By replacing fixed activation functions with learnable splines, our KAN-based approach is better suited to this challenge. Furthermore, to guide the network’s focus towards critical facial areas, we introduce a Landmark-assisted Adaptive Kolmogorov-Arnold Network (LAKAN) module. This module uses facial landmarks as a structural prior to dynamically generate the internal parameters of the KAN, creating an instance-specific signal that steers a general-purpose image encoder towards the most informative facial regions with artifacts. This core innovation creates a powerful combination between geometric priors and the network’s learning process. Extensive experiments on multiple public datasets show that our proposed method achieves superior performance.

Method: Uses reverse-attention optimization to reactivate suppressed signals while stabilizing attention, reinforced by velocity-guided dynamics for robust concept reactivation and consistency-preserving objectives to maintain global layout.

Result: Extensive experiments demonstrate the method’s effectiveness and efficiency in evaluating concept erasure robustness in rectified flow transformers.

Conclusion: ReFlux establishes a reliable benchmark for assessing concept erasure robustness in modern rectified flow-based text-to-image frameworks.

Abstract: Recent advances in text-to-image (T2I) diffusion models have enabled impressive generative capabilities, but they also raise significant safety concerns due to the potential to produce harmful or undesirable content. While concept erasure has been explored as a mitigation strategy, most existing approaches and corresponding attack evaluations are tailored to Stable Diffusion (SD) and exhibit limited effectiveness when transferred to next-generation rectified flow transformers such as Flux. In this work, we present ReFlux, the first concept attack method specifically designed to assess the robustness of concept erasure in the latest rectified flow-based T2I framework. Our approach is motivated by the observation that existing concept erasure techniques, when applied to Flux, fundamentally rely on a phenomenon known as attention localization. Building on this insight, we propose a simple yet effective attack strategy that specifically targets this property. At its core, a reverse-attention optimization strategy is introduced to effectively reactivate suppressed signals while stabilizing attention. This is further reinforced by a velocity-guided dynamic that enhances the robustness of concept reactivation by steering the flow matching process, and a consistency-preserving objective that maintains the global layout and preserves unrelated content. Extensive experiments consistently demonstrate the effectiveness and efficiency of the proposed attack method, establishing a reliable benchmark for evaluating the robustness of concept erasure strategies in rectified flow transformers.

Method: Proposes an efficient BEV map segmentation architecture using camera-radar fusion, featuring an advanced loss set and a new lightweight head to improve perception results while maintaining real-time performance.

Result: Achieves 53.5 mIoU with results comparable to large models, while setting new benchmark for inference time with 260% improvement over strongest baseline models.

Conclusion: The proposed camera-radar fusion architecture successfully addresses the trade-off between accuracy and real-time performance in map segmentation, offering a cost-effective and efficient solution for autonomous vehicle perception.

Abstract: Multi-sensor fusion in autonomous vehicles is becoming more common to offer a more robust alternative for several perception tasks. This need arises from the unique contribution of each sensor in collecting data: camera-radar fusion offers a cost-effective solution by combining rich semantic information from cameras with accurate distance measurements from radar, without incurring excessive financial costs or overwhelming data processing requirements. Map segmentation is a critical task for enabling effective vehicle behaviour in its environment, yet it continues to face significant challenges in achieving high accuracy and meeting real-time performance requirements. Therefore, this work presents a novel and efficient map segmentation architecture, using cameras and radars, in the \acrfull{bev} space. Our model introduces a real-time map segmentation architecture considering aspects such as high accuracy, per-class balancing, and inference time. To accomplish this, we use an advanced loss set together with a new lightweight head to improve the perception results. Our results show that, with these modifications, our approach achieves results comparable to large models, reaching 53.5 mIoU, while also setting a new benchmark for inference time, improving it by 260% over the strongest baseline models.

Method: Built on a large-scale hierarchical multi-modal dataset with hybrid annotation pipeline. Uses multi-head attention architecture to jointly align visual and textual representations with both fixed and open-set label embeddings.

Result: Achieves F1 scores of 0.81 on Otter dataset and 0.75 on Favorite dataset, surpassing baselines by 0.10 and 0.02 respectively. Near-perfect performance on open-set labels (F1: 0.99 on Otter, 0.97 on Favorite).

Conclusion: OTTER effectively bridges closed-set consistency with open-vocabulary flexibility for multi-modal tagging applications.

Abstract: We introduce OTTER, a unified open-set multi-label tagging framework that harmonizes the stability of a curated, predefined category set with the adaptability of user-driven open tags. OTTER is built upon a large-scale, hierarchically organized multi-modal dataset, collected from diverse online repositories and annotated through a hybrid pipeline combining automated vision-language labeling with human refinement. By leveraging a multi-head attention architecture, OTTER jointly aligns visual and textual representations with both fixed and open-set label embeddings, enabling dynamic and semantically consistent tagging. OTTER consistently outperforms competitive baselines on two benchmark datasets: it achieves an overall F1 score of 0.81 on Otter and 0.75 on Favorite, surpassing the next-best results by margins of 0.10 and 0.02, respectively. OTTER attains near-perfect performance on open-set labels, with F1 of 0.99 on Otter and 0.97 on Favorite, while maintaining competitive accuracy on predefined labels. These results demonstrate OTTER’s effectiveness in bridging closed-set consistency with open-vocabulary flexibility for multi-modal tagging applications.

Method: Uses progressive training with multi-scale scene-level dataset, combines multi-scale probability maps with cloud thickness map, and applies adaptive thresholds with distance-weighted optimization to generate binary cloud masks.

Result: Tested on WDCD and GF1MS-WHU datasets (60 GF1-MS images), SpecMCD achieved over 7.82% F1-score improvement compared to other weakly supervised methods like WDCD and WSFNet.

Conclusion: SpecMCD demonstrates superior performance for cloud detection under different cloud coverage conditions, showing strong potential for practical applications.

Abstract: Clouds significantly affect the quality of optical satellite images, which seriously limits their precise application. Recently, deep learning has been widely applied to cloud detection and has achieved satisfactory results. However, the lack of distinctive features in thin clouds and the low quality of training samples limit the cloud detection accuracy of deep learning methods, leaving space for further improvements. In this paper, we propose a weakly supervised cloud detection method that combines spectral features and multi-scale scene-level deep network (SpecMCD) to obtain highly accurate pixel-level cloud masks. The method first utilizes a progressive training framework with a multi-scale scene-level dataset to train the multi-scale scene-level cloud detection network. Pixel-level cloud probability maps are then obtained by combining the multi-scale probability maps and cloud thickness map based on the characteristics of clouds in dense cloud coverage and large cloud-area coverage images. Finally, adaptive thresholds are generated based on the differentiated regions of the scene-level cloud masks at different scales and combined with distance-weighted optimization to obtain binary cloud masks. Two datasets, WDCD and GF1MS-WHU, comprising a total of 60 Gaofen-1 multispectral (GF1-MS) images, were used to verify the effectiveness of the proposed method. Compared to the other weakly supervised cloud detection methods such as WDCD and WSFNet, the F1-score of the proposed SpecMCD method shows an improvement of over 7.82%, highlighting the superiority and potential of the SpecMCD method for cloud detection under different cloud coverage conditions.

Method: Proposed manifold feature distance (MFD) loss function that provides manifold-aligned tangents pointing toward the data manifold, implemented in the Align Your Tangent (AYT) method.

Result: AYT accelerates CM training by orders of magnitude, outperforms LPIPS metric, and enables training with extremely small batch sizes without compromising sample quality.

Conclusion: The proposed MFD loss effectively addresses oscillatory training dynamics in Consistency Models, significantly improving training efficiency and enabling more practical deployment with smaller computational requirements.

Abstract: With diffusion and flow matching models achieving state-of-the-art generating performance, the interest of the community now turned to reducing the inference time without sacrificing sample quality. Consistency Models (CMs), which are trained to be consistent on diffusion or probability flow ordinary differential equation (PF-ODE) trajectories, enable one or two-step flow or diffusion sampling. However, CMs typically require prolonged training with large batch sizes to obtain competitive sample quality. In this paper, we examine the training dynamics of CMs near convergence and discover that CM tangents – CM output update directions – are quite oscillatory, in the sense that they move parallel to the data manifold, not towards the manifold. To mitigate oscillatory tangents, we propose a new loss function, called the manifold feature distance (MFD), which provides manifold-aligned tangents that point toward the data manifold. Consequently, our method – dubbed Align Your Tangent (AYT) – can accelerate CM training by orders of magnitude and even out-perform the learned perceptual image patch similarity metric (LPIPS). Furthermore, we find that our loss enables training with extremely small batch sizes without compromising sample quality. Code: https://github.com/1202kbs/AYT

Method: Unfolds iteratively reweighted least squares rPCA baseline with intrinsic low-rank and sparse regularization. Adds sparse-enhancement unit to capture micro-flow signals. Functions as adaptive filter trained on part of image sequence.

Result: Outperformed SVD-based method, rPCA baseline, and another deep learning filter. Improved CNR of power Doppler image by 2-10 dB compared to other methods. Validated on in-silico and public in-vivo datasets.

Conclusion: U2-rPCA successfully addresses interpretability and ground truth challenges while achieving superior performance in ultrasound microvascular imaging through unsupervised learning approach.

Abstract: High-sensitivity clutter filtering is a fundamental step in ultrasound microvascular imaging. Singular value decomposition (SVD) and robust principal component analysis (rPCA) are the main clutter filtering strategies. However, both strategies are limited in feature modeling and tissue-blood flow separation for high-quality microvascular imaging. Recently, deep learning-based clutter filtering has shown potential in more thoroughly separating tissue and blood flow signals. However, the existing supervised filters face the challenges of interpretability and lack of in-vitro and in-vivo ground truths. While the interpretability issue can be addressed by algorithm deep unfolding, the training ground truth remains unsolved. To this end, this paper proposes an unsupervised unfolded rPCA (U2-rPCA) method that preserves mathematical interpretability and is insusceptible to learning labels. Specifically, U2-rPCA is unfolded from an iteratively reweighted least squares (IRLS) rPCA baseline with intrinsic low-rank and sparse regularization. A sparse-enhancement unit is added to the network to strengthen its capability to capture the sparse micro-flow signals. U2-rPCA is like an adaptive filter that is trained with part of the image sequence and then used for the following frames. Experimental validations on a in-silico dataset and public in-vivo datasets demonstrated the outperformance of U2-rPCA when compared with the SVD-based method, the rPCA baseline, and another deep learning-based filter. Particularly, the proposed method improved the contrastto-noise ratio (CNR) of the power Doppler image by 2 dB to 10 dB when compared with other methods. Furthermore, the effectiveness of the building modules of U2-rPCA was validated through ablation studies.

Method: Uses image-to-image translation with disentanglement techniques to independently manipulate image properties, specifically focusing on vessel appearances during adaptation while preserving crucial spatial information like shapes and locations for segmentation.

Result: The framework effectively bridges large domain gaps across medical centers, image modalities, and vessel types, demonstrating robustness and versatility in cerebrovascular segmentation across multiple scenarios.

Conclusion: The approach shows the potential of domain adaptation methodologies for accurate cerebrovascular image segmentation in diverse settings, with ablation studies confirming optimal architectural choices and annotation requirements.

Abstract: The intricate morphology of brain vessels poses significant challenges for automatic segmentation models, which usually focus on a single imaging modality. However, accurately treating brain-related conditions requires a comprehensive understanding of the cerebrovascular tree, regardless of the specific acquisition procedure. Our framework effectively segments brain arteries and veins in various datasets through image-to-image translation while avoiding domain-specific model design and data harmonization between the source and the target domain. This is accomplished by employing disentanglement techniques to independently manipulate different image properties, allowing them to move from one domain to another in a label-preserving manner. Specifically, we focus on manipulating vessel appearances during adaptation while preserving spatial information, such as shapes and locations, which are crucial for correct segmentation. Our evaluation effectively bridges large and varied domain gaps across medical centers, image modalities, and vessel types. Additionally, we conduct ablation studies on the optimal number of required annotations and other architectural choices. The results highlight our framework’s robustness and versatility, demonstrating the potential of domain adaptation methodologies to perform cerebrovascular image segmentation in multiple scenarios accurately. Our code is available at https://github.com/i-vesseg/MultiVesSeg.

Method: Developed a geometric framework with kernel-based probabilistic methods, interpreting diffusion models as projection mechanisms onto image manifolds. Created Manifold-Probabilistic Projection Model (MPPM) operating in both pixel and latent spaces.

Result: Latent MPPM (LMPPM) outperforms Latent Diffusion Model (LDM) across various datasets, achieving superior results in image restoration and generation tasks.

Conclusion: The study successfully unifies geometric and probabilistic perspectives in generative AI, providing a new framework that demystifies diffusion models and enables more effective image generation and restoration.

Abstract: The foundational premise of generative AI for images is the assumption that images are inherently low-dimensional objects embedded within a high-dimensional space. Additionally, it is often implicitly assumed that thematic image datasets form smooth or piecewise smooth manifolds. Common approaches overlook the geometric structure and focus solely on probabilistic methods, approximating the probability distribution through universal approximation techniques such as the kernel method. In some generative models, the low dimensional nature of the data manifest itself by the introduction of a lower dimensional latent space. Yet, the probability distribution in the latent or the manifold coordinate space is considered uninteresting and is predefined or considered uniform. This study unifies the geometric and probabilistic perspectives by providing a geometric framework and a kernel-based probabilistic method simultaneously. The resulting framework demystifies diffusion models by interpreting them as a projection mechanism onto the manifold of ``good images’’. This interpretation leads to the construction of a new deterministic model, the Manifold-Probabilistic Projection Model (MPPM), which operates in both the representation (pixel) space and the latent space. We demonstrate that the Latent MPPM (LMPPM) outperforms the Latent Diffusion Model (LDM) across various datasets, achieving superior results in terms of image restoration and generation.

Method: Uses image frames as inputs to teacher model (CLIP) to guide event-based student model via spatial attention-based distillation. Implements hybrid SNN-CNN framework where SNN adaptively determines optimal event segmentation moments, and CNNs process extracted features for object detection.

Result: Student network learns meaningful visual features from raw event inputs while inheriting CLIP’s broad visual knowledge. Adaptive event segmentation preserves crucial temporal information that fixed-group methods often discard.

Conclusion: The proposed framework successfully bridges the modality gap between images and event streams, enabling open-vocabulary object detection on event data by leveraging CLIP’s semantic understanding through knowledge distillation and adaptive temporal feature extraction.

Abstract: Event cameras offer advantages in object detection tasks due to high-speed response, low latency, and robustness to motion blur. However, event cameras lack texture and color information, making open-vocabulary detection particularly challenging. Current event-based detection methods are typically trained on predefined categories, limiting their ability to generalize to novel objects, where encountering previously unseen objects is common. Vision-language models (VLMs) have enabled open-vocabulary object detection in RGB images. However, the modality gap between images and event streams makes it ineffective to directly transfer CLIP to event data, as CLIP was not designed for event streams. To bridge this gap, we propose an event-image knowledge distillation framework that leverages CLIP’s semantic understanding to achieve open-vocabulary object detection on event data. Instead of training CLIP directly on event streams, we use image frames as inputs to a teacher model, guiding the event-based student model to learn CLIP’s rich visual representations. Through spatial attention-based distillation, the student network learns meaningful visual features directly from raw event inputs while inheriting CLIP’s broad visual knowledge. Furthermore, to prevent information loss due to event data segmentation, we design a hybrid spiking neural network (SNN) and convolutional neural network (CNN) framework. Unlike fixed-group event segmentation methods, which often discard crucial temporal information, our SNN adaptively determines the optimal event segmentation moments, ensuring that key temporal features are extracted. The extracted event features are then processed by CNNs for object detection.

Method: Uses image segmentation foundation models to generate segmentation masks, then crops images using bounding boxes from each mask. Each mask becomes an individual input in the novel ProtoMask architecture for fine-grained classification.

Result: Experiments on three fine-grained classification datasets show competitive performance compared to other models, with unique explainability features demonstrated through comprehensive metrics.

Conclusion: ProtoMask successfully integrates segmentation models to enhance explainability in prototypical case-based reasoning, providing more truthful visualizations while maintaining competitive classification performance.

Abstract: XAI gained considerable importance in recent years. Methods based on prototypical case-based reasoning have shown a promising improvement in explainability. However, these methods typically rely on additional post-hoc saliency techniques to explain the semantics of learned prototypes. Multiple critiques have been raised about the reliability and quality of such techniques. For this reason, we study the use of prominent image segmentation foundation models to improve the truthfulness of the mapping between embedding and input space. We aim to restrict the computation area of the saliency map to a predefined semantic image patch to reduce the uncertainty of such visualizations. To perceive the information of an entire image, we use the bounding box from each generated segmentation mask to crop the image. Each mask results in an individual input in our novel model architecture named ProtoMask. We conduct experiments on three popular fine-grained classification datasets with a wide set of metrics, providing a detailed overview on explainability characteristics. The comparison with other popular models demonstrates competitive performance and unique explainability features of our model. https://github.com/uos-sis/quanproto

Method: Constructs answer-concept and answer-question graphs for multimodal inputs, integrates Graph Transformer to capture multi-level dependencies, and replaces feed-forward layers with Mixture of Experts module for dynamic expert selection.

Result: Achieves higher accuracy than other concept bottleneck networks on multiple datasets by modeling structured relationships and utilizing dynamic expert selection.

Conclusion: MoE-SGT successfully overcomes limitations of traditional CBMs through structured relationship modeling and dynamic reasoning, providing better interpretability and performance in complex concept reasoning tasks.

Abstract: With growing demand for interpretability in deep learning, especially in high stakes domains, Concept Bottleneck Models (CBMs) address this by inserting human understandable concepts into the prediction pipeline, but they are generally single modal and ignore structured concept relationships. To overcome these limitations, we present MoE-SGT, a reasoning driven framework that augments CBMs with a structure injecting Graph Transformer and a Mixture of Experts (MoE) module. We construct answer-concept and answer-question graphs for multimodal inputs to explicitly model the structured relationships among concepts. Subsequently, we integrate Graph Transformer to capture multi level dependencies, addressing the limitations of traditional Concept Bottleneck Models in modeling concept interactions. However, it still encounters bottlenecks in adapting to complex concept patterns. Therefore, we replace the feed forward layers with a Mixture of Experts (MoE) module, enabling the model to have greater capacity in learning diverse concept relationships while dynamically allocating reasoning tasks to different sub experts, thereby significantly enhancing the model’s adaptability to complex concept reasoning. MoE-SGT achieves higher accuracy than other concept bottleneck networks on multiple datasets by modeling structured relationships among concepts and utilizing a dynamic expert selection mechanism.

Method: Uses MLLM’s output entropy as proactive guidance - entropy decreases with relevant visual information. Scores candidate visual inputs by response uncertainty to autonomously focus on salient data.

Result: Applied to Visual Search, Long Video Understanding, and Temporal Grounding tasks, achieving competitive performance with specialized fine-tuned methods using off-the-shelf MLLMs.

Conclusion: Harnessing intrinsic uncertainty is a powerful, general strategy for enhancing fine-grained multimodal performance without additional training.

Abstract: Multimodal Large Language Models (MLLMs) often struggle with fine-grained perception, such as identifying small objects in high-resolution images or finding key moments in long videos. Existing works typically rely on complicated, task-specific fine-tuning, which limits their generalizability and increases model complexity. In this work, we propose an effective, training-free framework that uses an MLLM’s intrinsic uncertainty as a proactive guidance signal. Our core insight is that a model’s output entropy decreases when presented with relevant visual information. We introduce a unified mechanism that scores candidate visual inputs by response uncertainty, enabling the model to autonomously focus on the most salient data. We apply this simple principle to three complex visual tasks: Visual Search, Long Video Understanding, and Temporal Grounding, allowing off-the-shelf MLLMs to achieve performance competitive with specialized, fine-tuned methods. Our work validates that harnessing intrinsic uncertainty is a powerful, general strategy for enhancing fine-grained multimodal performance.

Method: Unsupervised deep learning pipeline with three-step motion correction using robust principal component analysis to reduce contrast effects. Aligns perfusion series and auxiliary images (arterial input function and proton density-weighted series). Trained on multivendor data from 201 patients.

Result: Significantly improved temporal smoothness (p<0.001), comparable myocardial alignment (Dice=0.92 vs baseline), reduced standard deviation in perfusion maps (0.52 vs 0.55 ml/min/g), and 15x faster processing time.

Conclusion: The deep learning pipeline enables fast, robust motion correction for stress perfusion CMR, improving accuracy across dynamic and auxiliary images, and may facilitate broader clinical adoption of quantitative perfusion imaging.

Abstract: Background: Quantitative stress perfusion cardiovascular magnetic resonance (CMR) is a powerful tool for assessing myocardial ischemia. Motion correction is essential for accurate pixel-wise mapping but traditional registration-based methods are slow and sensitive to acquisition variability, limiting robustness and scalability. Methods: We developed an unsupervised deep learning-based motion correction pipeline that replaces iterative registration with efficient one-shot estimation. The method corrects motion in three steps and uses robust principal component analysis to reduce contrast-related effects. It aligns the perfusion series and auxiliary images (arterial input function and proton density-weighted series). Models were trained and validated on multivendor data from 201 patients, with 38 held out for testing. Performance was assessed via temporal alignment and quantitative perfusion values, compared to a previously published registration-based method. Results: The deep learning approach significantly improved temporal smoothness of time-intensity curves (p<0.001). Myocardial alignment (Dice = 0.92 (0.04) and 0.91 (0.05)) was comparable to the baseline and superior to before registration (Dice = 0.80 (0.09), p<0.001). Perfusion maps showed reduced motion, with lower standard deviation in the myocardium (0.52 (0.39) ml/min/g) compared to baseline (0.55 (0.44) ml/min/g). Processing time was reduced 15-fold. Conclusion: This deep learning pipeline enables fast, robust motion correction for stress perfusion CMR, improving accuracy across dynamic and auxiliary images. Trained on multivendor data, it generalizes across sequences and may facilitate broader clinical adoption of quantitative perfusion imaging.

Method: Used Continuous Wavelet Transformation to convert EEG data into scaleograms, then trained a vision transformer (ViT) model for emotion classification using subsets of EEG channels from the DEAP dataset.

Result: Achieved 91.57% accuracy in predicting 4 emotion quadrants (high/low arousal and valence) using only 12 EEG channels, compared to state-of-the-art 96.9% accuracy with 32 channels.

Conclusion: Significant reduction of EEG input channels (from 32 to 12) yields high accuracy results, making emotion prediction feasible with low-cost EEG devices while maintaining competitive performance.

Abstract: Accurately predicting emotions from brain signals has the potential to achieve goals such as improving mental health, human-computer interaction, and affective computing. Emotion prediction through neural signals offers a promising alternative to traditional methods, such as self-assessment and facial expression analysis, which can be subjective or ambiguous. Measurements of the brain activity via electroencephalogram (EEG) provides a more direct and unbiased data source. However, conducting a full EEG is a complex, resource-intensive process, leading to the rise of low-cost EEG devices with simplified measurement capabilities. This work examines how subsets of EEG channels from the DEAP dataset can be used for sufficiently accurate emotion prediction with low-cost EEG devices, rather than fully equipped EEG-measurements. Using Continuous Wavelet Transformation to convert EEG data into scaleograms, we trained a vision transformer (ViT) model for emotion classification. The model achieved over 91,57% accuracy in predicting 4 quadrants (high/low per arousal and valence) with only 12 measuring points (also referred to as channels). Our work shows clearly, that a significant reduction of input channels yields high results compared to state-of-the-art results of 96,9% with 32 channels. Training scripts to reproduce our code can be found here: https://gitlab.kit.edu/kit/aifb/ATKS/public/AutoSMiLeS/DEAP-DIVE.

Method: Learn a projector to map ELQ images to intermediate LQ manifold, then use frozen off-the-shelf BIR models for HQ restoration. Framework grounded in information theory with Information Bottleneck perspective and theoretically-driven objective.

Result: Enables Look Forward Once for inference-time prompt refinement and plug-and-play strengthening of existing models without finetuning. Extensive experiments show effectiveness under severe degradation regimes.

Conclusion: The proposed decomposition approach effectively addresses EBIR challenges by bridging the massive domain gap through intermediate projection, providing stable training and enhanced restoration performance.

Abstract: Blind Image Restoration (BIR) methods have achieved remarkable success but falter when faced with Extreme Blind Image Restoration (EBIR), where inputs suffer from severe, compounded degradations beyond their training scope. Directly learning a mapping from extremely low-quality (ELQ) to high-quality (HQ) images is challenging due to the massive domain gap, often leading to unnatural artifacts and loss of detail. To address this, we propose a novel framework that decomposes the intractable ELQ-to-HQ restoration process. We first learn a projector that maps an ELQ image onto an intermediate, less-degraded LQ manifold. This intermediate image is then restored to HQ using a frozen, off-the-shelf BIR model. Our approach is grounded in information theory; we provide a novel perspective of image restoration as an Information Bottleneck problem and derive a theoretically-driven objective to train our projector. This loss function effectively stabilizes training by balancing a low-quality reconstruction term with a high-quality prior-matching term. Our framework enables Look Forward Once (LFO) for inference-time prompt refinement, and supports plug-and-play strengthening of existing image restoration models without need for finetuning. Extensive experiments under severe degradation regimes provide a thorough analysis of the effectiveness of our work.

Method: Introduces Proportionate Credit Policy Optimization (PCPO), which enforces proportional credit assignment through a stable objective reformulation and principled reweighting of timesteps to correct the volatile feedback across timesteps.

Result: PCPO stabilizes training, accelerates convergence significantly, improves image quality by mitigating model collapse, and substantially outperforms existing policy gradient baselines including state-of-the-art DanceGRPO.

Conclusion: The PCPO framework successfully addresses the core issue of disproportionate credit assignment in text-to-image model alignment, providing a more stable and effective training approach that yields superior results across all evaluation metrics.

Abstract: While reinforcement learning has advanced the alignment of text-to-image (T2I) models, state-of-the-art policy gradient methods are still hampered by training instability and high variance, hindering convergence speed and compromising image quality. Our analysis identifies a key cause of this instability: disproportionate credit assignment, in which the mathematical structure of the generative sampler produces volatile and non-proportional feedback across timesteps. To address this, we introduce Proportionate Credit Policy Optimization (PCPO), a framework that enforces proportional credit assignment through a stable objective reformulation and a principled reweighting of timesteps. This correction stabilizes the training process, leading to significantly accelerated convergence and superior image quality. The improvement in quality is a direct result of mitigating model collapse, a common failure mode in recursive training. PCPO substantially outperforms existing policy gradient baselines on all fronts, including the state-of-the-art DanceGRPO.

Method: Proposed ADQA benchmark evaluates ADs on few-minute video segments with visual appreciation (VA) questions about visual facts and narrative understanding (NU) questions based on plot.

Result: Quantified subjectivity in AD writing through analysis of multiple AD tracks, and showed current AD generation methods significantly underperform compared to human-authored ADs.

Conclusion: Working with trimmed clips is inadequate for AD evaluation; ADQA provides better assessment framework with recommendations for future work and public leaderboard for benchmarking.

Abstract: Audio descriptions (ADs) narrate important visual details in movies, enabling Blind and Low Vision (BLV) users to understand narratives and appreciate visual details. Existing works in automatic AD generation mostly focus on few-second trimmed clips, and evaluate them by comparing against a single ground-truth reference AD. However, writing ADs is inherently subjective. Through alignment and analysis of two independent AD tracks for the same movies, we quantify the subjectivity in when and whether to describe, and what and how to highlight. Thus, we show that working with trimmed clips is inadequate. We propose ADQA, a QA benchmark that evaluates ADs at the level of few-minute long, coherent video segments, testing whether they would help BLV users understand the story and appreciate visual details. ADQA features visual appreciation (VA) questions about visual facts and narrative understanding (NU) questions based on the plot. Through ADQA, we show that current AD generation methods lag far behind human-authored ADs. We conclude with several recommendations for future work and introduce a public leaderboard for benchmarking.

Method: Developed a neural network based on U-Net architecture trained on OCT images with manually segmented annotations, evaluating multiple experimental configurations and using post-processing techniques for quantitative and qualitative evaluation.

Result: The system achieved 0.979 Dice Score, outperforming comparable studies, with an inference time of 18.98 seconds per volume, supporting viability for detecting inclusions.

Conclusion: The approach enables more efficient, reliable, and automated quality control in ceramic manufacturing through accurate defect detection using deep learning on OCT images.

Abstract: Non-destructive testing (NDT) is essential in ceramic manufacturing to ensure the quality of components without compromising their integrity. In this context, Optical Coherence Tomography (OCT) enables high-resolution internal imaging, revealing defects such as pores, delaminations, or inclusions. This paper presents an automatic defect detection system based on Deep Learning (DL), trained on OCT images with manually segmented annotations. A neural network based on the U-Net architecture is developed, evaluating multiple experimental configurations to enhance its performance. Post-processing techniques enable both quantitative and qualitative evaluation of the predictions. The system shows an accurate behavior of 0.979 Dice Score, outperforming comparable studies. The inference time of 18.98 seconds per volume supports its viability for detecting inclusions, enabling more efficient, reliable, and automated quality control.

Method: Proposed MULTI-TAP uses a plug-and-play architecture with a reward head built on LVLMs, training a lightweight ridge regression layer on frozen hidden states of pre-trained LVLMs to produce both overall and fine-grained multi-objective scores.

Result: MULTI-TAP achieves significantly higher performance than existing metrics and performs on par with GPT-4o-based G-VEval with smaller model size (7-8B), outperforming VisionREWARD in both performance and efficiency on multi-objective benchmarks and the new EYE4ALL dataset.

Conclusion: MULTI-TAP provides a robust solution for multi-objective image-text alignment evaluation, with the new EYE4ALL dataset serving as a foundation for developing more accessible AI systems that capture diverse user preferences including those of blind and low-vision individuals.

Abstract: Evaluating image-text alignment while reflecting human preferences across multiple aspects is a significant issue for the development of reliable vision-language applications. It becomes especially crucial in real-world scenarios where multiple valid descriptions exist depending on contexts or user needs. However, research progress is hindered by the lack of comprehensive benchmarks and existing evaluation predictors lacking at least one of these key properties: (1) Alignment with human judgments, (2) Long-sequence processing, (3) Inference efficiency, and (4) Applicability to multi-objective scoring. To address these challenges, we propose a plug-and-play architecture to build a robust predictor, MULTI-TAP (Multi-Objective Task-Aware Predictor), capable of both multi and single-objective scoring. MULTI-TAP can produce a single overall score, utilizing a reward head built on top of a large vision-language model (LVLMs). We show that MULTI-TAP is robust in terms of application to different LVLM architectures, achieving significantly higher performance than existing metrics and even on par with the GPT-4o-based predictor, G-VEval, with a smaller size (7-8B). By training a lightweight ridge regression layer on the frozen hidden states of a pre-trained LVLM, MULTI-TAP can produce fine-grained scores for multiple human-interpretable objectives. MULTI-TAP performs better than VisionREWARD, a high-performing multi-objective reward model, in both performance and efficiency on multi-objective benchmarks and our newly released text-image-to-text dataset, EYE4ALL. Our new dataset, consisting of chosen/rejected human preferences (EYE4ALLPref) and human-annotated fine-grained scores across seven dimensions (EYE4ALLMulti), can serve as a foundation for developing more accessible AI systems by capturing the underlying preferences of users, including blind and low-vision (BLV) individuals.

Method: Repurposed a video diffusion model as a generative encoder by performing a single denoising step and using the resulting latents as visual embeddings, creating World-Language Models (WorldLMs) with a best variant called Dynamic Vision Aligner (DyVA).

Result: WorldLMs capture distinct latents useful for downstream understanding, significantly enhance spatial reasoning, enable single-image models to perform multi-frame reasoning, and achieve state-of-the-art or comparable performance on visual reasoning tasks.

Conclusion: WorldLMs represent a promising new family of vision-language models that leverage world model priors, showing potential for generalist vision learning through inherited motion-consistency from video pre-training.

Abstract: Trained on internet-scale video data, generative world models are increasingly recognized as powerful world simulators that can generate consistent and plausible dynamics over structure, motion, and physics. This raises a natural question: with the advent of strong video foundational models, might they supplant conventional vision encoder paradigms for general-purpose multimodal understanding? While recent studies have begun to explore the potential of world models on common vision tasks, these explorations typically lack a systematic investigation of generic, multimodal tasks. In this work, we strive to investigate the capabilities when world model priors are transferred into Vision-Language Models: we re-purpose a video diffusion model as a generative encoder to perform a single denoising step and treat the resulting latents as a set of visual embedding. We empirically investigate this class of models, which we refer to as World-Language Models (WorldLMs), and we find that generative encoders can capture latents useful for downstream understanding that show distinctions from conventional encoders. Naming our best-performing variant Dynamic Vision Aligner (DyVA), we further discover that this method significantly enhances spatial reasoning abilities and enables single-image models to perform multi-frame reasoning. Through the curation of a suite of visual reasoning tasks, we find DyVA to surpass both open-source and proprietary baselines, achieving state-of-the-art or comparable performance. We attribute these gains to WorldLM’s inherited motion-consistency internalization from video pre-training. Finally, we systematically explore extensive model designs to highlight promising directions for future work. We hope our study can pave the way for a new family of VLMs that leverage priors from world models and are on a promising path towards generalist vision learners.

Method: Extracts feature maps from a source DNN and adds them to clean images to create adversarial samples that impair target model classification, enabling transfer across models and datasets.

Result: ZQBA successfully transfers adversarial samples across different models and datasets (CIFAR, Tiny ImageNet), outperforms state-of-the-art single-query attacks, and maintains imperceptible perturbations as measured by SSIM and qualitative evaluation.

Conclusion: The method demonstrates vulnerabilities in DNNs for real-world applications and provides an effective zero-query attack that transfers well across models and datasets while preserving sample quality.

Abstract: Current black-box adversarial attacks either require multiple queries or diffusion models to produce adversarial samples that can impair the target model performance. However, these methods require training a surrogate loss or diffusion models to produce adversarial samples, which limits their applicability in real-world settings. Thus, we propose a Zero Query Black-box Adversarial (ZQBA) attack that exploits the representations of Deep Neural Networks (DNNs) to fool other networks. Instead of requiring thousands of queries to produce deceiving adversarial samples, we use the feature maps obtained from a DNN and add them to clean images to impair the classification of a target model. The results suggest that ZQBA can transfer the adversarial samples to different models and across various datasets, namely CIFAR and Tiny ImageNet. The experiments also show that ZQBA is more effective than state-of-the-art black-box attacks with a single query, while maintaining the imperceptibility of perturbations, evaluated both quantitatively (SSIM) and qualitatively, emphasizing the vulnerabilities of employing DNNs in real-world contexts. All the source code is available at https://github.com/Joana-Cabral/ZQBA.

Method: The method learns binary class-level concept prototypes and uses distances between predicted concept vectors and each class prototype as classification scores and uncertainty measures. These prototypes serve as interpretable classification rules.

Result: The proposed framework enhances interpretability and robustness by enabling conformal prediction for uncertain or outlier inputs based on their deviation from learned binary class-level concept prototypes.

Conclusion: The approach improves both interpretability and performance in CBMs while providing uncertainty quantification through binary concept prototypes that serve as transparent classification rules.

Abstract: In the context of image classification, Concept Bottleneck Models (CBMs) first embed images into a set of human-understandable concepts, followed by an intrinsically interpretable classifier that predicts labels based on these intermediate representations. While CBMs offer a semantically meaningful and interpretable classification pipeline, they often sacrifice predictive performance compared to end-to-end convolutional neural networks. Moreover, the propagation of uncertainty from concept predictions to final label decisions remains underexplored. In this paper, we propose a novel uncertainty-aware and interpretable classifier for the second stage of CBMs. Our method learns a set of binary class-level concept prototypes and uses the distances between predicted concept vectors and each class prototype as both a classification score and a measure of uncertainty. These prototypes also serve as interpretable classification rules, indicating which concepts should be present in an image to justify a specific class prediction. The proposed framework enhances both interpretability and robustness by enabling conformal prediction for uncertain or outlier inputs based on their deviation from the learned binary class-level concept prototypes.

Method: MetaLogic uses metamorphic testing to generate image pairs from prompts that are semantically identical but grammatically different, then directly compares these pairs to identify alignment failures without needing ground truth images.

Result: Evaluation across multiple SOTA T2I models revealed consistent robustness failures, with Flux.dev showing 59% misalignment rate and DALLE-3 showing 71% misalignment rate across various logical constructs.

Conclusion: MetaLogic provides an efficient, scalable, and ground-truth-free approach for identifying fine-grained logical inconsistencies in text-to-image models, uncovering alignment errors that existing metrics overlook.

Abstract: Recent advances in text-to-image (T2I) models, especially diffusion-based architectures, have significantly improved the visual quality of generated images. However, these models continue to struggle with a critical limitation: maintaining semantic consistency when input prompts undergo minor linguistic variations. Despite being logically equivalent, such prompt pairs often yield misaligned or semantically inconsistent images, exposing a lack of robustness in reasoning and generalisation. To address this, we propose MetaLogic, a novel evaluation framework that detects T2I misalignment without relying on ground truth images. MetaLogic leverages metamorphic testing, generating image pairs from prompts that differ grammatically but are semantically identical. By directly comparing these image pairs, the framework identifies inconsistencies that signal failures in preserving the intended meaning, effectively diagnosing robustness issues in the model’s logic understanding. Unlike existing evaluation methods that compare a generated image to a single prompt, MetaLogic evaluates semantic equivalence between paired images, offering a scalable, ground-truth-free approach to identifying alignment failures. It categorises these alignment errors (e.g., entity omission, duplication, positional misalignment) and surfaces counterexamples that can be used for model debugging and refinement. We evaluate MetaLogic across multiple state-of-the-art T2I models and reveal consistent robustness failures across a range of logical constructs. We find that even the SOTA text-to-image models like Flux.dev and DALLE-3 demonstrate a 59 percent and 71 percent misalignment rate, respectively. Our results show that MetaLogic is not only efficient and scalable, but also effective in uncovering fine-grained logical inconsistencies that are overlooked by existing evaluation metrics.

Method: Four-stage pipeline: geometric rectification, zero-shot semantic segmentation, LLM-guided spatial reasoning, and energy simulation using computer vision and AI techniques.

Result: Validated on 80 buildings across 4 countries with mean area estimation error of 6.2% ± 2.8% compared to expert annotations, processing time of ~100 seconds per building.

Conclusion: The method provides reliable energy yield predictions suitable for regional potential studies, urban energy planning, and building-integrated PV deployment.

Abstract: Building facades represent a significant untapped resource for solar energy generation in dense urban environments, yet assessing their photovoltaic (PV) potential remains challenging due to complex geometries and semantic com ponents. This study introduces SF-SPA (Semantic Facade Solar-PV Assessment), an automated framework that transforms street-view photographs into quantitative PV deployment assessments. The approach combines com puter vision and artificial intelligence techniques to address three key challenges: perspective distortion correction, semantic understanding of facade elements, and spatial reasoning for PV layout optimization. Our four-stage pipeline processes images through geometric rectification, zero-shot semantic segmentation, Large Language Model (LLM) guided spatial reasoning, and energy simulation. Validation across 80 buildings in four countries demonstrates ro bust performance with mean area estimation errors of 6.2% ± 2.8% compared to expert annotations. The auto mated assessment requires approximately 100 seconds per building, a substantial gain in efficiency over manual methods. Simulated energy yield predictions confirm the method’s reliability and applicability for regional poten tial studies, urban energy planning, and building-integrated photovoltaic (BIPV) deployment. Code is available at: https:github.com/CodeAXu/Solar-PV-Installation

Method: Two-stage framework: 1) Vision Language Model predicts coarse-grained motion trajectories consistent with physics, 2) Trajectories guide video generation through attention-based mechanisms for fine-grained motion refinement. Uses trajectory prediction dataset based on video tracking data.

Result: Outperforms existing methods on UCF-101 and MSR-VTT datasets, achieving FVD scores of 545 on UCF-101 and 539 on MSR-VTT.

Conclusion: TrajVLM-Gen successfully generates physically consistent videos by incorporating physics-aware trajectory prediction and attention-based motion guidance.

Abstract: Current video generation models produce physically inconsistent motion that violates real-world dynamics. We propose TrajVLM-Gen, a two-stage framework for physics-aware image-to-video generation. First, we employ a Vision Language Model to predict coarse-grained motion trajectories that maintain consistency with real-world physics. Second, these trajectories guide video generation through attention-based mechanisms for fine-grained motion refinement. We build a trajectory prediction dataset based on video tracking data with realistic motion patterns. Experiments on UCF-101 and MSR-VTT demonstrate that TrajVLM-Gen outperforms existing methods, achieving competitive FVD scores of 545 on UCF-101 and 539 on MSR-VTT.

Method: Directly uses float-encoded one-hot class data as diffusion input, introduces timestep-conditioned cross-entropy loss between model outputs and original one-hot labels, creating synergistic design.

Result: Achieves superior performance on classification tasks compared to baseline and exhibits excellent text generation capabilities on image captioning.

Conclusion: ADD framework successfully bridges discriminative and generative learning while preserving authentic diffusion characteristics in discrete spaces, with each component providing measurable gains.

Abstract: We propose an Authentic Discrete Diffusion (ADD) framework that fundamentally redefines prior pseudo-discrete approaches by preserving core diffusion characteristics directly in the one-hot space through a suite of coordinated mechanisms. Unlike conventional “pseudo” discrete diffusion (PDD) methods, ADD reformulates the diffusion input by directly using float-encoded one-hot class data, without relying on diffusing in the continuous latent spaces or masking policies. At its core, a timestep-conditioned cross-entropy loss is introduced between the diffusion model’s outputs and the original one-hot labels. This synergistic design establishes a bridge between discriminative and generative learning. Our experiments demonstrate that ADD not only achieves superior performance on classification tasks compared to the baseline, but also exhibits excellent text generation capabilities on Image captioning. Extensive ablations validate the measurable gains of each component.

Method: Leveraged GenStereo to generate accurate right-view images from existing single-view PhraseCut data, creating stereo image pairs with aligned segmentation masks and phrase annotations.

Result: Created PhraseStereo dataset that extends phrase grounding to stereo domain, introducing new challenges and opportunities for leveraging depth cues in multimodal learning.

Conclusion: PhraseStereo establishes foundation for research at intersection of language, vision and 3D perception, enabling development of models that jointly reason over semantics and geometry.

Abstract: Understanding how natural language phrases correspond to specific regions in images is a key challenge in multimodal semantic segmentation. Recent advances in phrase grounding are largely limited to single-view images, neglecting the rich geometric cues available in stereo vision. For this, we introduce PhraseStereo, the first novel dataset that brings phrase-region segmentation to stereo image pairs. PhraseStereo builds upon the PhraseCut dataset by leveraging GenStereo to generate accurate right-view images from existing single-view data, enabling the extension of phrase grounding into the stereo domain. This new setting introduces unique challenges and opportunities for multimodal learning, particularly in leveraging depth cues for more precise and context-aware grounding. By providing stereo image pairs with aligned segmentation masks and phrase annotations, PhraseStereo lays the foundation for future research at the intersection of language, vision, and 3D perception, encouraging the development of models that can reason jointly over semantics and geometry. The PhraseStereo dataset will be released online upon acceptance of this work.

Method: Two-stage training: first train a transformation network to map low-quality images to preliminary scales, then perform end-to-end full-model fine-tuning using next-scale autoregressive prediction strategy.

Result: NSARM achieves superior visual results over existing Real-ISR methods with fast inference speed, and demonstrates much higher robustness to input image quality with stronger generalization performance.

Conclusion: As a pure AR model, NSARM provides an effective solution for Real-ISR that balances quality, speed, and robustness through comprehensive fine-tuning without compromising generative capability.

Abstract: Most recent real-world image super-resolution (Real-ISR) methods employ pre-trained text-to-image (T2I) diffusion models to synthesize the high-quality image either from random Gaussian noise, which yields realistic results but is slow due to iterative denoising, or directly from the input low-quality image, which is efficient but at the price of lower output quality. These approaches train ControlNet or LoRA modules while keeping the pre-trained model fixed, which often introduces over-enhanced artifacts and hallucinations, suffering from the robustness to inputs of varying degradations. Recent visual autoregressive (AR) models, such as pre-trained Infinity, can provide strong T2I generation capabilities while offering superior efficiency by using the bitwise next-scale prediction strategy. Building upon next-scale prediction, we introduce a robust Real-ISR framework, namely Next-Scale Autoregressive Modeling (NSARM). Specifically, we train NSARM in two stages: a transformation network is first trained to map the input low-quality image to preliminary scales, followed by an end-to-end full-model fine-tuning. Such a comprehensive fine-tuning enhances the robustness of NSARM in Real-ISR tasks without compromising its generative capability. Extensive quantitative and qualitative evaluations demonstrate that as a pure AR model, NSARM achieves superior visual results over existing Real-ISR methods while maintaining a fast inference speed. Most importantly, it demonstrates much higher robustness to the quality of input images, showing stronger generalization performance. Project page: https://github.com/Xiangtaokong/NSARM

Method: Two novel hierarchical contrastive learning methods: G-HMLC using Gaussian Mixture Model and A-HMLC using attention mechanism to capture hierarchy-specific features and model inter-class relationships across hierarchy levels.

Result: Achieves state-of-the-art performance on CIFAR100 and ModelNet40 datasets, outperforming existing hierarchical contrastive learning methods by 2 percentage points in accuracy in linear evaluation.

Conclusion: The approach effectively models hierarchical relationships and enables fine-grained clustering across all hierarchy levels, showing strong potential for computer vision and other applications.

Abstract: Hierarchical classification is a crucial task in many applications, where objects are organized into multiple levels of categories. However, conventional classification approaches often neglect inherent inter-class relationships at different hierarchy levels, thus missing important supervisory signals. Thus, we propose two novel hierarchical contrastive learning (HMLC) methods. The first, leverages a Gaussian Mixture Model (G-HMLC) and the second uses an attention mechanism to capture hierarchy-specific features (A-HMLC), imitating human processing. Our approach explicitly models inter-class relationships and imbalanced class distribution at higher hierarchy levels, enabling fine-grained clustering across all hierarchy levels. On the competitive CIFAR100 and ModelNet40 datasets, our method achieves state-of-the-art performance in linear evaluation, outperforming existing hierarchical contrastive learning methods by 2 percentage points in terms of accuracy. The effectiveness of our approach is backed by both quantitative and qualitative results, highlighting its potential for applications in computer vision and beyond.

Method: Hybrid architecture with shifted window self-attention for spatial context aggregation and Mamba-based selective scanning for temporal propagation. Introduces Gather-Scatter Mamba (GSM) mechanism that warps features to anchor frame before Mamba propagation and scatters them back afterward.

Result: The method enables efficient long-range temporal modeling while capturing fine-grained spatial dependencies, reducing occlusion artifacts and ensuring effective information redistribution across frames.

Conclusion: GSMamba provides an effective solution combining the strengths of attention mechanisms and selective SSMs for video super-resolution, addressing limitations of both Transformers and traditional recurrent approaches.

Abstract: State Space Models (SSMs)-most notably RNNs-have historically played a central role in sequential modeling. Although attention mechanisms such as Transformers have since dominated due to their ability to model global context, their quadratic complexity and limited scalability make them less suited for long sequences. Video super-resolution (VSR) methods have traditionally relied on recurrent architectures to propagate features across frames. However, such approaches suffer from well-known issues including vanishing gradients, lack of parallelism, and slow inference speed. Recent advances in selective SSMs like Mamba offer a compelling alternative: by enabling input-dependent state transitions with linear-time complexity, Mamba mitigates these issues while maintaining strong long-range modeling capabilities. Despite this potential, Mamba alone struggles to capture fine-grained spatial dependencies due to its causal nature and lack of explicit context aggregation. To address this, we propose a hybrid architecture that combines shifted window self-attention for spatial context aggregation with Mamba-based selective scanning for efficient temporal propagation. Furthermore, we introduce Gather-Scatter Mamba (GSM), an alignment-aware mechanism that warps features toward a center anchor frame within the temporal window before Mamba propagation and scatters them back afterward, effectively reducing occlusion artifacts and ensuring effective redistribution of aggregated information across all frames. The official implementation is provided at: https://github.com/Ko-Lani/GSMamba.

Method: Proposes AI-CNet3D model that divides OCT volumes along two axes, applies cross-attention to capture hemiretinal asymmetries, integrates ONH and macula information, uses CAREs for visualization, and employs consistency-based multi-task fine-tuning aligned with Grad-CAMs.

Result: Outperforms state-of-the-art attention and convolutional models across all key metrics, reduces parameter count by 100-fold compared to other attention mechanisms while maintaining high diagnostic performance and comparable GFLOPS.

Conclusion: The proposed AI-CNet3D model enhances glaucoma classification by effectively capturing anatomical asymmetries and structural details from 3D OCT data while being computationally efficient and interpretable.

Abstract: Glaucoma is a progressive eye disease that leads to optic nerve damage, causing irreversible vision loss if left untreated. Optical coherence tomography (OCT) has become a crucial tool for glaucoma diagnosis, offering high-resolution 3D scans of the retina and optic nerve. However, the conventional practice of condensing information from 3D OCT volumes into 2D reports often results in the loss of key structural details. To address this, we propose a novel hybrid deep learning model that integrates cross-attention mechanisms into a 3D convolutional neural network (CNN), enabling the extraction of critical features from the superior and inferior hemiretinas, as well as from the optic nerve head (ONH) and macula, within OCT volumes. We introduce Channel Attention REpresentations (CAREs) to visualize cross-attention outputs and leverage them for consistency-based multi-task fine-tuning, aligning them with Gradient-Weighted Class Activation Maps (Grad-CAMs) from the CNN’s final convolutional layer to enhance performance, interpretability, and anatomical coherence. We have named this model AI-CNet3D (AI-`See’-Net3D) to reflect its design as an Anatomically-Informed Cross-attention Network operating on 3D data. By dividing the volume along two axes and applying cross-attention, our model enhances glaucoma classification by capturing asymmetries between the hemiretinal regions while integrating information from the optic nerve head and macula. We validate our approach on two large datasets, showing that it outperforms state-of-the-art attention and convolutional models across all key metrics. Finally, our model is computationally efficient, reducing the parameter count by one-hundred–fold compared to other attention mechanisms while maintaining high diagnostic performance and comparable GFLOPS.

Method: A task-based survey with machine learning practitioners, taking a human-centered HCI perspective to understand how practitioners select source datasets.

Result: Dataset selection is task-dependent and influenced by community practices, dataset properties, computational similarity (data embedding), and perceived visual/semantic similarity. However, similarity ratings and expected performance are not always aligned, challenging traditional assumptions.

Conclusion: Practitioners use ambiguous terminology in source selection, indicating a need for clearer definitions and HCI tools to make selection heuristics explicit and usable for more systematic transfer learning.

Abstract: Transfer learning is crucial for medical imaging, yet the selection of source datasets - which can impact the generalizability of algorithms, and thus patient outcomes - often relies on researchers’ intuition rather than systematic principles. This study investigates these decisions through a task-based survey with machine learning practitioners. Unlike prior work that benchmarks models and experimental setups, we take a human-centered HCI perspective on how practitioners select source datasets. Our findings indicate that choices are task-dependent and influenced by community practices, dataset properties, and computational (data embedding), or perceived visual or semantic similarity. However, similarity ratings and expected performance are not always aligned, challenging a traditional “more similar is better” view. Participants often used ambiguous terminology, which suggests a need for clearer definitions and HCI tools to make them explicit and usable. By clarifying these heuristics, this work provides practical insights for more systematic source selection in transfer learning.

Method: Combines coarse alignment, ROI filtering, initial landmark approximation with a patch-based pointwise CNN enhanced by attention mechanisms for 3D facial landmark localization.

Result: Achieved mean localization error of 3.686 mm on 214 annotated scans, preserving anatomical distances with 2.822 mm error. On FaceScape dataset (700 subjects), achieved 0.41 mm point-wise and 0.38 mm distance-wise errors.

Conclusion: PAL-Net provides a lightweight, scalable solution for high-throughput 3D anthropometric analysis, outperforming existing methods and reducing reliance on manual annotation while maintaining clinical applicability.

Abstract: Manual annotation of anatomical landmarks on 3D facial scans is a time-consuming and expertise-dependent task, yet it remains critical for clinical assessments, morphometric analysis, and craniofacial research. While several deep learning methods have been proposed for facial landmark localization, most focus on pseudo-landmarks or require complex input representations, limiting their clinical applicability. This study presents a fully automated deep learning pipeline (PAL-Net) for localizing 50 anatomical landmarks on stereo-photogrammetry facial models. The method combines coarse alignment, region-of-interest filtering, and an initial approximation of landmarks with a patch-based pointwise CNN enhanced by attention mechanisms. Trained and evaluated on 214 annotated scans from healthy adults, PAL-Net achieved a mean localization error of 3.686 mm and preserves relevant anatomical distances with a 2.822 mm average error, comparable to intra-observer variability. To assess generalization, the model was further evaluated on 700 subjects from the FaceScape dataset, achieving a point-wise error of 0.41,mm and a distance-wise error of 0.38,mm. Compared to existing methods, PAL-Net offers a favorable trade-off between accuracy and computational cost. While performance degrades in regions with poor mesh quality (e.g., ears, hairline), the method demonstrates consistent accuracy across most anatomical regions. PAL-Net generalizes effectively across datasets and facial regions, outperforming existing methods in both point-wise and structural evaluations. It provides a lightweight, scalable solution for high-throughput 3D anthropometric analysis, with potential to support clinical workflows and reduce reliance on manual annotation. Source code can be found at https://github.com/Ali5hadman/PAL-Net-A-Point-Wise-CNN-with-Patch-Attention

Method: Combines self-supervised splitting losses with equivariant reconstruction networks, introducing a new definition of equivariance for reconstruction networks.

Result: Achieves state-of-the-art performance on image inpainting, accelerated MRI, and compressive sensing with highly rank-deficient forward models.

Conclusion: The proposed self-supervised approach with equivariant networks provides unbiased supervised loss estimates and effective reconstruction without ground-truth data.

Abstract: Self-supervised learning for inverse problems allows to train a reconstruction network from noise and/or incomplete data alone. These methods have the potential of enabling learning-based solutions when obtaining ground-truth references for training is expensive or even impossible. In this paper, we propose a new self-supervised learning strategy devised for the challenging setting where measurements are observed via a single incomplete observation model. We introduce a new definition of equivariance in the context of reconstruction networks, and show that the combination of self-supervised splitting losses and equivariant reconstruction networks results in unbiased estimates of the supervised loss. Through a series of experiments on image inpainting, accelerated magnetic resonance imaging, and compressive sensing, we demonstrate that the proposed loss achieves state-of-the-art performance in settings with highly rank-deficient forward models.

Method: The authors discovered semantic directions implying resolution differences in ReID feature space, validated through Canonical Correlation Analysis and Pearson Correlation Analysis. They propose a lightweight Vector Panning Feature Alignment (VPFA) framework that models resolution-specific feature discrepancy rather than using super-resolution or joint learning approaches.

Result: Extensive experiments on multiple CR-ReID benchmarks show that the proposed method significantly outperforms previous state-of-the-art baseline models while achieving higher efficiency.

Conclusion: The VPFA framework demonstrates effectiveness and superiority based on the novel finding of resolution-specific semantic directions in ReID feature space, providing a more efficient alternative to complex super-resolution or joint learning approaches for cross-resolution person re-identification.

Abstract: In surveillance scenarios, varying camera distances cause significant differences among pedestrian image resolutions, making it hard to match low-resolution (LR) images with high-resolution (HR) counterparts, limiting the performance of Re-Identification (ReID) tasks. Most existing Cross-Resolution ReID (CR-ReID) methods rely on super-resolution (SR) or joint learning for feature compensation, which increases training and inference complexity and has reached a performance bottleneck in recent studies. Inspired by semantic directions in the word embedding space, we empirically discover that semantic directions implying resolution differences also emerge in the feature space of ReID, and we substantiate this finding from a statistical perspective using Canonical Correlation Analysis and Pearson Correlation Analysis. Based on this interesting finding, we propose a lightweight and effective Vector Panning Feature Alignment (VPFA) framework, which conducts CR-ReID from a novel perspective of modeling the resolution-specific feature discrepancy. Extensive experimental results on multiple CR-ReID benchmarks show that our method significantly outperforms previous state-of-the-art baseline models while obtaining higher efficiency, demonstrating the effectiveness and superiority of our model based on the new finding in this paper.

Method: Reformulates VSR as autoregressive-one-step-diffusion using causal DiT structure with rolling KV-cache, joint visual guidance, patch-wise pixel supervision, and cross-chunk distribution matching.

Result: Achieves state-of-the-art quality with enhanced semantic consistency, delivers up to 58x speed-up over methods like MGLD-VSR, and introduces new benchmark for long-form video evaluation.

Conclusion: InfVSR pushes the frontier of long-form VSR by enabling efficient and scalable processing of unbounded-length videos while maintaining temporal coherence.

Abstract: Real-world videos often extend over thousands of frames. Existing video super-resolution (VSR) approaches, however, face two persistent challenges when processing long sequences: (1) inefficiency due to the heavy cost of multi-step denoising for full-length sequences; and (2) poor scalability hindered by temporal decomposition that causes artifacts and discontinuities. To break these limits, we propose InfVSR, which novelly reformulates VSR as an autoregressive-one-step-diffusion paradigm. This enables streaming inference while fully leveraging pre-trained video diffusion priors. First, we adapt the pre-trained DiT into a causal structure, maintaining both local and global coherence via rolling KV-cache and joint visual guidance. Second, we distill the diffusion process into a single step efficiently, with patch-wise pixel supervision and cross-chunk distribution matching. Together, these designs enable efficient and scalable VSR for unbounded-length videos. To fill the gap in long-form video evaluation, we build a new benchmark tailored for extended sequences and further introduce semantic-level metrics to comprehensively assess temporal consistency. Our method pushes the frontier of long-form VSR, achieves state-of-the-art quality with enhanced semantic consistency, and delivers up to 58x speed-up over existing methods such as MGLD-VSR. Code will be available at https://github.com/Kai-Liu001/InfVSR.

Method: Encodes images and captions into discrete visual/textual tokens using a joint-embedding predictive Transformer. Incorporates cross-attention after feature predictor for conditional denoising, injects raw text embeddings before flow matching loss, and performs iterative denoising during inference.

Result: Achieves strong data efficiency and open-vocabulary generalization on ImageNet-1K, consistently outperforming non-fusion and late-fusion baselines in text-to-image generation tasks.

Conclusion: Late architectural fusion combined with objective-level alignment provides an effective balance between conditioning strength and backbone generality in token-based T2I systems.

Abstract: Modern Text-to-Image (T2I) generation increasingly relies on token-centric architectures that are trained with self-supervision, yet effectively fusing text with visual tokens remains a challenge. We propose \textbf{JEPA-T}, a unified multimodal framework that encodes images and captions into discrete visual and textual tokens, processed by a joint-embedding predictive Transformer. To enhance fusion, we incorporate cross-attention after the feature predictor for conditional denoising while maintaining a task-agnostic backbone. Additionally, raw texts embeddings are injected prior to the flow matching loss to improve alignment during training. During inference, the same network performs both class-conditional and free-text image generation by iteratively denoising visual tokens conditioned on text. Evaluations on ImageNet-1K demonstrate that JEPA-T achieves strong data efficiency, open-vocabulary generalization, and consistently outperforms non-fusion and late-fusion baselines. Our approach shows that late architectural fusion combined with objective-level alignment offers an effective balance between conditioning strength and backbone generality in token-based T2I.The code is now available: https://github.com/justin-herry/JEPA-T.git

Method: FastForward represents multiple mapping images as 3D-anchored features and uses these to predict image-to-scene correspondences for query images, enabling camera pose estimation in a single feed-forward pass coupled with image retrieval.

Result: FastForward achieves state-of-the-art accuracy compared to other approaches with minimal map preparation time and demonstrates robust generalization to unseen domains, including challenging large-scale outdoor environments.

Conclusion: FastForward provides a practical solution for fast and accurate image localization that works efficiently across various environments with minimal mapping overhead.

Abstract: Visually localizing an image, i.e., estimating its camera pose, requires building a scene representation that serves as a visual map. The representation we choose has direct consequences towards the practicability of our system. Even when starting from mapping images with known camera poses, state-of-the-art approaches still require hours of mapping time in the worst case, and several minutes in the best. This work raises the question whether we can achieve competitive accuracy much faster. We introduce FastForward, a method that creates a map representation and relocalizes a query image on-the-fly in a single feed-forward pass. At the core, we represent multiple mapping images as a collection of features anchored in 3D space. FastForward utilizes these mapping features to predict image-to-scene correspondences for the query image, enabling the estimation of its camera pose. We couple FastForward with image retrieval and achieve state-of-the-art accuracy when compared to other approaches with minimal map preparation time. Furthermore, FastForward demonstrates robust generalization to unseen domains, including challenging large-scale outdoor environments.

Method: A post-pretraining refinement module that operates on all generated tokens simultaneously, leveraging global context and relationships across tokens to enhance spatial correspondence modeling.

Result: The method improves generation quality and semantic consistency by mitigating error accumulation in sequential generation, as demonstrated through experiments.

Conclusion: The proposed refinement module effectively enhances autoregressive vision-language models’ ability to handle spatial dependencies while maintaining the sequential prediction framework.

Abstract: Autoregressive models excel in sequential modeling and have proven to be effective for vision-language data. However, the spatial nature of visual signals conflicts with the sequential dependencies of next-token prediction, leading to suboptimal results. This work proposes a plug-and-play refinement module to enhance the complex spatial correspondence modeling within the generated visual sequence. This module operates as a post-pretraining step to jointly refine all generated tokens of autoregressive model, enhancing vision-language modeling under a shared sequential prediction framework. By leveraging global context and relationship across the tokens, our method mitigates the error accumulation issue within the sequential generation. Experiments demonstrate that the proposed method improves the generation quality, enhancing the model’s ability to produce semantically consistent results.

Method: SoftCFG distributes adaptive perturbations across all tokens with certainty-weighted guidance, and uses Step Normalization to bound cumulative perturbations for stable long-sequence generation.

Result: SoftCFG significantly improves image quality over standard CFG and achieves state-of-the-art FID on ImageNet 256 among autoregressive models.

Conclusion: SoftCFG provides an effective, training-free solution for improving conditional generation in autoregressive models while maintaining visual coherence.

Abstract: Autoregressive (AR) models have emerged as powerful tools for image generation by modeling images as sequences of discrete tokens. While Classifier-Free Guidance (CFG) has been adopted to improve conditional generation, its application in AR models faces two key issues: guidance diminishing, where the conditional-unconditional gap quickly vanishes as decoding progresses, and over-guidance, where strong conditions distort visual coherence. To address these challenges, we propose SoftCFG, an uncertainty-guided inference method that distributes adaptive perturbations across all tokens in the sequence. The key idea behind SoftCFG is to let each generated token contribute certainty-weighted guidance, ensuring that the signal persists across steps while resolving conflicts between text guidance and visual context. To further stabilize long-sequence generation, we introduce Step Normalization, which bounds cumulative perturbations of SoftCFG. Our method is training-free, model-agnostic, and seamlessly integrates with existing AR pipelines. Experiments show that SoftCFG significantly improves image quality over standard CFG and achieves state-of-the-art FID on ImageNet 256 among autoregressive models.

Method: TextCAM combines CAM’s spatial localization with CLIP embeddings and linear discriminant analysis to derive channel-level semantic representations, then aggregates them with CAM weights to produce textual descriptions of visual evidence.

Result: Experiments on ImageNet, CLEVR, and CUB show TextCAM produces faithful and interpretable rationales that improve human understanding, detect spurious correlations, and preserve model fidelity.

Conclusion: TextCAM successfully bridges the gap between spatial activation maps and semantic understanding, providing more comprehensive explanations that specify both where models attend and what visual attributes support decisions.

Abstract: Deep neural networks (DNNs) have achieved remarkable success across domains but remain difficult to interpret, limiting their trustworthiness in high-stakes applications. This paper focuses on deep vision models, for which a dominant line of explainability methods are Class Activation Mapping (CAM) and its variants working by highlighting spatial regions that drive predictions. We figure out that CAM provides little semantic insight into what attributes underlie these activations. To address this limitation, we propose TextCAM, a novel explanation framework that enriches CAM with natural languages. TextCAM combines the precise spatial localization of CAM with the semantic alignment of vision-language models (VLMs). Specifically, we derive channel-level semantic representations using CLIP embeddings and linear discriminant analysis, and aggregate them with CAM weights to produce textual descriptions of salient visual evidence. This yields explanations that jointly specify where the model attends and what visual attributes likely support its decision. We further extend TextCAM to generate feature channels into semantically coherent groups, enabling more fine-grained visual-textual explanations. Experiments on ImageNet, CLEVR, and CUB demonstrate that TextCAM produces faithful and interpretable rationales that improve human understanding, detect spurious correlations, and preserve model fidelity.

Method: Built on a vision-language model with a “look-think-predict” paradigm: first localize flaws, then generate reasoning, and finally provide quantitative scores across plausibility, semantic alignment, aesthetics, and overall quality dimensions.

Result: ImageDoctor shows strong alignment with human preferences across multiple datasets and improves generation quality by 10% over scalar-based reward models when used for preference tuning.

Conclusion: ImageDoctor provides a comprehensive, interpretable evaluation framework for text-to-image models that enables better preference alignment and significantly improves generation quality compared to traditional scalar-based approaches.

Abstract: The rapid advancement of text-to-image (T2I) models has increased the need for reliable human preference modeling, a demand further amplified by recent progress in reinforcement learning for preference alignment. However, existing approaches typically quantify the quality of a generated image using a single scalar, limiting their ability to provide comprehensive and interpretable feedback on image quality. To address this, we introduce ImageDoctor, a unified multi-aspect T2I model evaluation framework that assesses image quality across four complementary dimensions: plausibility, semantic alignment, aesthetics, and overall quality. ImageDoctor also provides pixel-level flaw indicators in the form of heatmaps, which highlight misaligned or implausible regions, and can be used as a dense reward for T2I model preference alignment. Inspired by the diagnostic process, we improve the detail sensitivity and reasoning capability of ImageDoctor by introducing a “look-think-predict” paradigm, where the model first localizes potential flaws, then generates reasoning, and finally concludes the evaluation with quantitative scores. Built on top of a vision-language model and trained through a combination of supervised fine-tuning and reinforcement learning, ImageDoctor demonstrates strong alignment with human preference across multiple datasets, establishing its effectiveness as an evaluation metric. Furthermore, when used as a reward model for preference tuning, ImageDoctor significantly improves generation quality – achieving an improvement of 10% over scalar-based reward models.

Method: Proposed AT-RA adversarial training method that uses data augmentation to increase spectral diversity and ensures spatial smoothness to preserve spectral semantic information.

Result: AT-RA improves adversarial robustness by 21.34% against AutoAttack and 18.78% against PGD-50, while increasing benign accuracy by 2.68%.

Conclusion: Adversarial training is effective for hyperspectral security, and AT-RA successfully addresses unique challenges in hyperspectral data by preserving spectral semantics.

Abstract: Recent studies have revealed that hyperspectral classification models based on deep learning are highly vulnerable to adversarial attacks, which pose significant security risks. Although several approaches have attempted to enhance adversarial robustness by modifying network architectures, these methods often rely on customized designs that limit scalability and fail to defend effectively against strong attacks. To address these challenges, we introduce adversarial training to the hyperspectral domain, which is widely regarded as one of the most effective defenses against adversarial attacks. Through extensive empirical analyses, we demonstrate that while adversarial training does enhance robustness across various models and datasets, hyperspectral data introduces unique challenges not seen in RGB images. Specifically, we find that adversarial noise and the non-smooth nature of adversarial examples can distort or eliminate important spectral semantic information. To mitigate this issue, we employ data augmentation techniques and propose a novel hyperspectral adversarial training method, termed AT-RA. By increasing the diversity of spectral information and ensuring spatial smoothness, AT-RA preserves and corrects spectral semantics in hyperspectral images. Experimental results show that AT-RA improves adversarial robustness by 21.34% against AutoAttack and 18.78% against PGD-50 while boosting benign accuracy by 2.68%.

Method: Combine secret key with latent face representations in diffusion model, use facial mask to preserve identity-irrelevant features, and employ deterministic forward/backward diffusion processes.

Result: Produces anonymized faces that are less visually similar to originals compared to previous work, while maintaining high-quality images and enabling recovery with correct key.

Conclusion: The method successfully addresses the trade-off between secure anonymization, high-quality generation, and reversible authentication through diffusion models with secret key integration.

Abstract: Face images processed by computer vision algorithms contain sensitive personal information that malicious actors can capture without consent. These privacy and security risks highlight the need for effective face anonymization methods. Current methods struggle to propose a good trade-off between a secure scheme with high-quality image generation and reversibility for later person authentication. Diffusion-based approaches produce high-quality anonymized images but lack the secret key mechanism to ensure that only authorized parties can reverse the process. In this paper, we introduce, to our knowledge, the first secure, high-quality reversible anonymization method based on a diffusion model. We propose to combine the secret key with the latent faces representation of the diffusion model. To preserve identity-irrelevant features, generation is constrained by a facial mask, maintaining high-quality images. By using a deterministic forward and backward diffusion process, our approach enforces that the original face can be recovered with the correct secret key. We also show that the proposed method produces anonymized faces that are less visually similar to the original faces, compared to other previous work.

Method: Represent 3D scenes as hierarchical graphs (floors, rooms, objects, functional elements) with nodes augmented by multi-modal information from keyframes. Uses VLMs to extract scene information without explicit relationship modeling, and employs hierarchical RAG pipeline for scalable context extraction.

Result: Outperforms prior approaches on most metrics across four benchmarks including 3D object segmentation and complex query retrieval, demonstrating superior semantic richness and efficiency.

Conclusion: KeySG provides a scalable and semantically rich framework for 3D scene understanding that enables general, task-agnostic reasoning and planning while effectively handling complex queries in large environments.

Abstract: In recent years, 3D scene graphs have emerged as a powerful world representation, offering both geometric accuracy and semantic richness. Combining 3D scene graphs with large language models enables robots to reason, plan, and navigate in complex human-centered environments. However, current approaches for constructing 3D scene graphs are semantically limited to a predefined set of relationships, and their serialization in large environments can easily exceed an LLM’s context window. We introduce KeySG, a framework that represents 3D scenes as a hierarchical graph consisting of floors, rooms, objects, and functional elements, where nodes are augmented with multi-modal information extracted from keyframes selected to optimize geometric and visual coverage. The keyframes allow us to efficiently leverage VLM to extract scene information, alleviating the need to explicitly model relationship edges between objects, enabling more general, task-agnostic reasoning and planning. Our approach can process complex and ambiguous queries while mitigating the scalability issues associated with large scene graphs by utilizing a hierarchical retrieval-augmented generation (RAG) pipeline to extract relevant context from the graph. Evaluated across four distinct benchmarks – including 3D object segmentation and complex query retrieval – KeySG outperforms prior approaches on most metrics, demonstrating its superior semantic richness and efficiency.

Method: Uses geometric recovery through deep visual SLAM, grid pruning to optimize scene representation, and introduces a streamlined 4D Gaussian representation for efficient temporal dynamics handling.

Result: Achieves 30x speed-up, reduces training time to within two minutes, reconstructs single video within 10 minutes on Dycheck dataset, and maintains competitive performance across benchmarks while reducing model size to under 10% of original.

Conclusion: Instant4D provides an efficient solution for monocular 4D reconstruction from casual videos, demonstrating strong generalizability to in-the-wild videos with significant performance improvements.

Abstract: Dynamic view synthesis has seen significant advances, yet reconstructing scenes from uncalibrated, casual video remains challenging due to slow optimization and complex parameter estimation. In this work, we present Instant4D, a monocular reconstruction system that leverages native 4D representation to efficiently process casual video sequences within minutes, without calibrated cameras or depth sensors. Our method begins with geometric recovery through deep visual SLAM, followed by grid pruning to optimize scene representation. Our design significantly reduces redundancy while maintaining geometric integrity, cutting model size to under 10% of its original footprint. To handle temporal dynamics efficiently, we introduce a streamlined 4D Gaussian representation, achieving a 30x speed-up and reducing training time to within two minutes, while maintaining competitive performance across several benchmarks. Our method reconstruct a single video within 10 minutes on the Dycheck dataset or for a typical 200-frame video. We further apply our model to in-the-wild videos, showcasing its generalizability. Our project website is published at https://instant4d.github.io/.

Method: The method learns per-model, per-label, context-conditioned reliabilities from labeled history and converts model reports into agreement-guardrailed log-likelihood ratios combined with contextual priors and Shapley-based team credit updates. Three heterogeneous VLMs were fine-tuned using LoRA on 1,000 curated dashcam clips.

Result: Empirical evaluation shows 23% reduction in Hamming distance, 55% improvement in Macro-F1, and 47% improvement in Micro-F1 compared to the best single model.

Conclusion: The proposed method supports VLM fusion as a calibrated, interpretable, and robust decision-support component for autonomous vehicle pipelines.

Abstract: Large vision-language models (VLMs) are increasingly used in autonomous-vehicle (AV) stacks, but hallucination limits their reliability in safety-critical pipelines. We present Shapley-credited Context-Aware Dawid-Skene with Agreement, a game-theoretic fusion method for multi-label understanding of ego-view dashcam video. It learns per-model, per-label, context-conditioned reliabilities from labelled history and, at inference, converts each model’s report into an agreement-guardrailed log-likelihood ratio that is combined with a contextual prior and a public reputation state updated via Shapley-based team credit. The result is calibrated, thresholdable posteriors that (i) amplify agreement among reliable models, (ii) preserve uniquely correct single-model signals, and (iii) adapt to drift. To specialise general VLMs, we curate 1,000 real-world dashcam clips with structured annotations (scene description, manoeuvre recommendation, rationale) via an automatic pipeline that fuses HDD ground truth, vehicle kinematics, and YOLOv11

• BoT-SORT tracking, guided by a three-step chain-of-thought prompt; three heterogeneous VLMs are then fine-tuned with LoRA. We evaluate with Hamming distance, Micro-Macro-F1, and average per-video latency. Empirically, the proposed method achieves a 23% reduction in Hamming distance, 55% improvement in Macro-F1, and 47% improvement in Micro-F1 when comparing with the best single model, supporting VLM fusion as a calibrated, interpretable, and robust decision-support component for AV pipelines.

Method: The approach uses a three-step process: 1) Generate future video frames using a video generator with view control, 2) Evolve 3D scene reconstruction using a feedforward plug-and-play transformer, 3) Synthesize futures by conditioning on geometric reprojections from the evolving 3D memory.

Result: Extensive experiments show that EvoWorld’s evolving 3D memory substantially improves visual fidelity and maintains spatial scene coherence compared to existing approaches, particularly in loop-closure detection and spatial coherence over extended trajectories.

Conclusion: EvoWorld represents a significant advance toward long-horizon spatially consistent world modeling by leveraging evolving 3D reconstruction as explicit spatial guidance for video generation.

Abstract: Humans possess a remarkable ability to mentally explore and replay 3D environments they have previously experienced. Inspired by this mental process, we present EvoWorld: a world model that bridges panoramic video generation with evolving 3D memory to enable spatially consistent long-horizon exploration. Given a single panoramic image as input, EvoWorld first generates future video frames by leveraging a video generator with fine-grained view control, then evolves the scene’s 3D reconstruction using a feedforward plug-and-play transformer, and finally synthesizes futures by conditioning on geometric reprojections from this evolving explicit 3D memory. Unlike prior state-of-the-arts that synthesize videos only, our key insight lies in exploiting this evolving 3D reconstruction as explicit spatial guidance for the video generation process, projecting the reconstructed geometry onto target viewpoints to provide rich spatial cues that significantly enhance both visual realism and geometric consistency. To evaluate long-range exploration capabilities, we introduce the first comprehensive benchmark spanning synthetic outdoor environments, Habitat indoor scenes, and challenging real-world scenarios, with particular emphasis on loop-closure detection and spatial coherence over extended trajectories. Extensive experiments demonstrate that our evolving 3D memory substantially improves visual fidelity and maintains spatial scene coherence compared to existing approaches, representing a significant advance toward long-horizon spatially consistent world modeling.

Method: Uses prompt-guided multimodal alignment and prior-based mask retargeting modules to generate multimodal information and mask sequences, then feeds them into pretrained mask-driven video generation models.

Result: IMAGEdit consistently surpasses state-of-the-art methods on the MSVBench benchmark and is compatible with any mask-driven video generation model.

Conclusion: IMAGEdit provides a robust training-free solution for multi-subject video editing with strong generalization capability and improved overall performance.

Abstract: In this paper, we present IMAGEdit, a training-free framework for any number of video subject editing that manipulates the appearances of multiple designated subjects while preserving non-target regions, without finetuning or retraining. We achieve this by providing robust multimodal conditioning and precise mask sequences through a prompt-guided multimodal alignment module and a prior-based mask retargeting module. We first leverage large models' understanding and generation capabilities to produce multimodal information and mask motion sequences for multiple subjects across various types. Then, the obtained prior mask sequences are fed into a pretrained mask-driven video generation model to synthesize the edited video. With strong generalization capability, IMAGEdit remedies insufficient prompt-side multimodal conditioning and overcomes mask boundary entanglement in videos with any number of subjects, thereby significantly expanding the applicability of video editing. More importantly, IMAGEdit is compatible with any mask-driven video generation model, significantly improving overall performance. Extensive experiments on our newly constructed multi-subject benchmark MSVBench verify that IMAGEdit consistently surpasses state-of-the-art methods. Code, models, and datasets are publicly available at https://github.com/XWH-A/IMAGEdit.

Method: Proposed Zoneup Dropout Injection Attention Calculation (ZoDIAC) that refines attention intensities using GELU and dropout, then intensifies them through a zoneup process with learned scalar factor injection.

Result: ZoDIAC achieves statistically significant higher scores across all image captioning metrics on MS-COCO dataset compared to conventional self-attention, working with various feature extractors.

Conclusion: ZoDIAC can serve as a drop-in replacement for attention components in all transformer models, providing improved performance while maintaining compatibility.

Abstract: In the past few years the transformer model has been utilized for a variety of tasks such as image captioning, image classification natural language generation, and natural language understanding. As a key component of the transformer model, self-attention calculates the attention values by mapping the relationships among the head elements of the source and target sequence, yet there is no explicit mechanism to refine and intensify the attention values with respect to the context of the input and target sequences. Based on this intuition, we introduce a novel refine and intensify attention mechanism that is called Zoneup Dropout Injection Attention Calculation (ZoDIAC), in which the intensities of attention values in the elements of the input source and target sequences are first refined using GELU and dropout and then intensified using a proposed zoneup process which includes the injection of a learned scalar factor. Our extensive experiments show that ZoDIAC achieves statistically significant higher scores under all image captioning metrics using various feature extractors in comparison to the conventional self-attention module in the transformer model on the MS-COCO dataset. Our proposed ZoDIAC attention modules can be used as a drop-in replacement for the attention components in all transformer models. The code for our experiments is publicly available at: https://github.com/zanyarz/zodiac

Method: Propose a timestep tuner that replaces original parameterization by conditioning the network on new timesteps at each denoising step, enforcing sampling distribution towards the real one. This finds more accurate integral directions for timestep intervals.

Result: Significantly improves performance of state-of-the-art acceleration methods, especially with few denoising steps. For example, improves DDIM FID from 9.65 to 6.07 on LSUN Bedroom with 10 steps.

Conclusion: The timestep tuner is an efficient plug-in design that can boost inference performance of various acceleration methods by finding more appropriate timesteps and integral directions.

Abstract: A diffusion model, which is formulated to produce an image using thousands of denoising steps, usually suffers from a slow inference speed. Existing acceleration algorithms simplify the sampling by skipping most steps yet exhibit considerable performance degradation. By viewing the generation of diffusion models as a discretized integral process, we argue that the quality drop is partly caused by applying an inaccurate integral direction to a timestep interval. To rectify this issue, we propose a \textbf{timestep tuner} that helps find a more accurate integral direction for a particular interval at the minimum cost. Specifically, at each denoising step, we replace the original parameterization by conditioning the network on a new timestep, enforcing the sampling distribution towards the real one. Extensive experiments show that our plug-in design can be trained efficiently and boost the inference performance of various state-of-the-art acceleration methods, especially when there are few denoising steps. For example, when using 10 denoising steps on LSUN Bedroom dataset, we improve the FID of DDIM from 9.65 to 6.07, simply by adopting our method for a more appropriate set of timesteps. Code is available at \href{https://github.com/THU-LYJ-Lab/time-tuner}{https://github.com/THU-LYJ-Lab/time-tuner}.

Method: An innovative image-to-brain multi-layer encoding framework that optimizes multiple model layers to learn and mimic human brain’s visual representational patterns across object categories and modalities using EEG data.

Result: ReAlnets demonstrate significantly higher similarity to human brain representations compared to traditional computer vision models.

Conclusion: The approach represents an important step toward bridging the gap between artificial and human vision, enabling more brain-like artificial intelligence systems.

Abstract: Despite advancements in artificial intelligence, object recognition models still lag behind in emulating visual information processing in human brains. Recent studies have highlighted the potential of using neural data to mimic brain processing; however, these often rely on invasive neural recordings from non-human subjects, leaving a critical gap in understanding human visual perception. Addressing this gap, we present, ‘Re(presentational)Al(ignment)net’, a vision model aligned with human brain activity based on non-invasive EEG, demonstrating a significantly higher similarity to human brain representations. Our innovative image-to-brain multi-layer encoding framework advances human neural alignment by optimizing multiple model layers and enabling the model to efficiently learn and mimic the human brain’s visual representational patterns across object categories and different modalities. Our findings suggest that ReAlnets better align artificial neural networks with human brain representations, making it more similar to human brain processing than traditional computer vision models, which takes an important step toward bridging the gap between artificial and human vision and achieving more brain-like artificial intelligence systems.

Method: Uses semi-supervised rotation regression with dynamic entropy-based filtering to adaptively remove outliers, and introduces two novel head-oriented strong augmentations: pose-irrelevant cut-occlusion and pose-altering rotation consistency.

Result: Extensive experiments show SemiUHPE outperforms counterparts greatly on public benchmarks under both front-range and full-range settings, and demonstrates versatility for other problems like object rotation regression and 3D head reconstruction.

Conclusion: The proposed semi-supervised approach effectively addresses the label-scarce problem in unconstrained head pose estimation and shows good extensibility to related tasks.

Abstract: Existing research on unconstrained in-the-wild head pose estimation suffers from the flaws of its datasets, which consist of either numerous samples by non-realistic synthesis or constrained collection, or small-scale natural images yet with plausible manual annotations. This makes fully-supervised solutions compromised due to the reliance on generous labels. To alleviate it, we propose the first semi-supervised unconstrained head pose estimation method SemiUHPE, which can leverage abundant easily available unlabeled head images. Technically, we choose semi-supervised rotation regression and adapt it to the error-sensitive and label-scarce problem of unconstrained head pose. Our method is based on the observation that the aspect-ratio invariant cropping of wild heads is superior to previous landmark-based affine alignment given that landmarks of unconstrained human heads are usually unavailable, especially for underexplored non-frontal heads. Instead of using a pre-fixed threshold to filter out pseudo labeled heads, we propose dynamic entropy based filtering to adaptively remove unlabeled outliers as training progresses by updating the threshold in multiple stages. We then revisit the design of weak-strong augmentations and improve it by devising two novel head-oriented strong augmentations, termed pose-irrelevant cut-occlusion and pose-altering rotation consistency respectively. Extensive experiments and ablation studies show that SemiUHPE outperforms its counterparts greatly on public benchmarks under both the front-range and full-range settings. Furthermore, our proposed method is also beneficial for solving other closely related problems, including generic object rotation regression and 3D head reconstruction, demonstrating good versatility and extensibility. Code is in https://github.com/hnuzhy/SemiUHPE.

Method: Proposes transition rate (TR) scheduling that sets a target for how many quantized weights should transition discrete levels, and uses transition-adaptive learning rate (TALR) to update latent weights accordingly.

Result: Experimental results demonstrate the effectiveness of the approach on standard benchmarks.

Conclusion: Transition rate scheduling provides better control over quantized weight changes in QAT compared to traditional learning rate scheduling.

Abstract: Quantization-aware training (QAT) simulates a quantization process during training to lower bit-precision of weights/activations. It learns quantized weights indirectly by updating latent weights,i.e., full-precision inputs to a quantizer, using gradient-based optimizers. We claim that coupling a user-defined learning rate (LR) with these optimizers is sub-optimal for QAT. Quantized weights transit discrete levels of a quantizer, only if corresponding latent weights pass transition points, where the quantizer changes discrete states. This suggests that the changes of quantized weights are affected by both the LR for latent weights and their distributions. It is thus difficult to control the degree of changes for quantized weights by scheduling the LR manually. We conjecture that the degree of parameter changes in QAT is related to the number of quantized weights transiting discrete levels. Based on this, we introduce a transition rate (TR) scheduling technique that controls the number of transitions of quantized weights explicitly. Instead of scheduling a LR for latent weights, we schedule a target TR of quantized weights, and update the latent weights with a novel transition-adaptive LR (TALR), enabling considering the degree of changes for the quantized weights during QAT. Experimental results demonstrate the effectiveness of our approach on standard benchmarks.

Method: Proposes SL²A-INR - a hybrid network architecture that combines a single-layer learnable activation function with a traditional ReLU-based MLP for improved signal representation.

Result: Superior performance across diverse tasks including image representation, 3D shape reconstruction, and novel view synthesis, setting new benchmarks in accuracy, quality, and robustness.

Conclusion: The hybrid approach of combining learnable activations with traditional ReLU networks effectively addresses INR limitations and achieves state-of-the-art performance in multiple vision domains.

Abstract: Implicit Neural Representation (INR), leveraging a neural network to transform coordinate input into corresponding attributes, has recently driven significant advances in several vision-related domains. However, the performance of INR is heavily influenced by the choice of the nonlinear activation function used in its multilayer perceptron (MLP) architecture. To date, multiple nonlinearities have been investigated, but current INRs still face limitations in capturing high-frequency components and diverse signal types. We show that these challenges can be alleviated by introducing a novel approach in INR architecture. Specifically, we propose SL$^{2}$A-INR, a hybrid network that combines a single-layer learnable activation function with an MLP that uses traditional ReLU activations. Our method performs superior across diverse tasks, including image representation, 3D shape reconstruction, and novel view synthesis. Through comprehensive experiments, SL$^{2}$A-INR sets new benchmarks in accuracy, quality, and robustness for INR. Our Code is publicly available on~\href{https://github.com/Iceage7/SL2A-INR}{\textcolor{magenta}{GitHub}}.

Method: Proposed Weak-to-strong Semantics Compensation (WSCo) uses weakly augmented images as anchors to enrich the feature space of strongly augmented counterparts, compensating for lost class-relevant semantics.

Result: Extensive experiments show WSCo effectively enhances performance of previous detection models on standard benchmarks by addressing the negative impact of strong augmentation.

Conclusion: WSCo serves as a generic plug-in that can be easily integrated into existing SFOD pipelines to mitigate the semantic loss caused by strong data augmentation.

Abstract: Strong data augmentation is a fundamental component of state-of-the-art mean teacher-based Source-Free domain adaptive Object Detection (SFOD) methods, enabling consistency-based self-supervised optimization along weak augmentation. However, our theoretical analysis and empirical observations reveal a critical limitation: strong augmentation can inadvertently erase class-relevant components, leading to artificial inter-category confusion. To address this issue, we introduce Weak-to-strong Semantics Compensation (WSCo), a novel remedy that leverages weakly augmented images, which preserve full semantics, as anchors to enrich the feature space of their strongly augmented counterparts. Essentially, this compensates for the class-relevant semantics that may be lost during strong augmentation on the fly. Notably, WSCo can be implemented as a generic plug-in, easily integrable with any existing SFOD pipelines. Extensive experiments validate the negative impact of strong augmentation on detection performance, and the effectiveness of WSCo in enhancing the performance of previous detection models on standard benchmarks.

Method: Proposes ReCFG with relaxed guidance coefficients that strictly align with diffusion theory, featuring a closed-form solution that allows pre-computation without affecting sampling speed.

Result: Compatible with state-of-the-art diffusion models (EDM2 on ImageNet, SD3 on CC12M) without retraining, maintaining performance while fixing theoretical issues.

Conclusion: ReCFG provides a theoretically sound alternative to CFG that eliminates expectation shift while being practically implementable with existing models.

Abstract: Classifier-Free Guidance (CFG), which combines the conditional and unconditional score functions with two coefficients summing to one, serves as a practical technique for diffusion model sampling. Theoretically, however, denoising with CFG \textit{cannot} be expressed as a reciprocal diffusion process, which may consequently leave some hidden risks during use. In this work, we revisit the theory behind CFG and rigorously confirm that the improper configuration of the combination coefficients (\textit{i.e.}, the widely used summing-to-one version) brings about expectation shift of the generative distribution. To rectify this issue, we propose ReCFG with a relaxation on the guidance coefficients such that denoising with \method strictly aligns with the diffusion theory. We further show that our approach enjoys a \textbf{\textit{closed-form}} solution given the guidance strength. That way, the rectified coefficients can be readily pre-computed via traversing the observed data, leaving the sampling speed barely affected. Empirical evidence on real-world data demonstrate the compatibility of our post-hoc design with existing state-of-the-art diffusion models, including both class-conditioned ones (\textit{e.g.}, EDM2 on ImageNet) and text-conditioned ones (\textit{e.g.}, SD3 on CC12M), without any retraining. Code is available at \href{https://github.com/thuxmf/recfg}{https://github.com/thuxmf/recfg}.

Method: Created FLORA dataset with 4,330 curated fashion outfits and detailed textual descriptions using industry-specific terminology. Introduced NeRA adapter architecture based on Kolmogorov-Arnold Networks (KAN) using learnable spline-based nonlinear transformations instead of traditional MLP adapters.

Result: Fine-tuning generative models on FLORA significantly enhances their capability to generate accurate and stylistically rich fashion images. NeRA achieves superior modeling of complex semantic relationships with strong fidelity, faster convergence and better semantic alignment compared to existing adapters like LoRA, LoKR, DoRA, and LoHA.

Conclusion: FLORA dataset will catalyze advanced AI fashion models, and NeRA represents a significant improvement in adapter architectures for fashion generation tasks. Both FLORA dataset and NeRA implementation will be open-sourced.

Abstract: Specialized datasets that capture the fashion industry’s rich language and styling elements can boost progress in AI-driven fashion design. We present FLORA, (Fashion Language Outfit Representation for Apparel Generation), the first comprehensive dataset containing 4,330 curated pairs of fashion outfits and corresponding textual descriptions. Each description utilizes industry-specific terminology and jargon commonly used by professional fashion designers, providing precise and detailed insights into the outfits. Hence, the dataset captures the delicate features and subtle stylistic elements necessary to create high-fidelity fashion designs. We demonstrate that fine-tuning generative models on the FLORA dataset significantly enhances their capability to generate accurate and stylistically rich images from textual descriptions of fashion sketches. FLORA will catalyze the creation of advanced AI models capable of comprehending and producing subtle, stylistically rich fashion designs. It will also help fashion designers and end-users to bring their ideas to life. As a second orthogonal contribution, we introduce NeRA (Nonlinear low-rank Expressive Representation Adapter), a novel adapter architecture based on Kolmogorov-Arnold Networks (KAN). Unlike traditional PEFT techniques such as LoRA, LoKR, DoRA, and LoHA that use MLP adapters, NeRA uses learnable spline-based nonlinear transformations, enabling superior modeling of complex semantic relationships, achieving strong fidelity, faster convergence and semantic alignment. Extensive experiments on our proposed FLORA and LAION-5B datasets validate the superiority of NeRA over existing adapters. We will open-source both the FLORA dataset and our implementation code.

Method: Uses latent diffusion framework with two components: IdentityNet (preserves identity features) and AnimateNet (enhances temporal coherence). Integrates audio input with reference images and incorporates text prompts via decoupled cross-attention mechanisms.

Result: Demonstrates superior performance over state-of-the-art methods through extensive experiments and a newly developed evaluation metric.

Conclusion: Sets a new standard for generating customizable realistic talking faces suitable for real-world applications.

Abstract: Audio-driven talking face generation is a challenging task in digital communication. Despite significant progress in the area, most existing methods concentrate on audio-lip synchronization, often overlooking aspects such as visual quality, customization, and generalization that are crucial to producing realistic talking faces. To address these limitations, we introduce a novel, customizable one-shot audio-driven talking face generation framework, named PortraitTalk. Our proposed method utilizes a latent diffusion framework consisting of two main components: IdentityNet and AnimateNet. IdentityNet is designed to preserve identity features consistently across the generated video frames, while AnimateNet aims to enhance temporal coherence and motion consistency. This framework also integrates an audio input with the reference images, thereby reducing the reliance on reference-style videos prevalent in existing approaches. A key innovation of PortraitTalk is the incorporation of text prompts through decoupled cross-attention mechanisms, which significantly expands creative control over the generated videos. Through extensive experiments, including a newly developed evaluation metric, our model demonstrates superior performance over the state-of-the-art methods, setting a new standard for the generation of customizable realistic talking faces suitable for real-world applications.

Method: Proposes Latent Reward Model (LRM) that repurposes diffusion model components to predict preferences in latent space, and Latent Preference Optimization (LPO) for step-level optimization directly in noisy latent space.

Result: LPO significantly improves alignment with general, aesthetic, and text-image alignment preferences while achieving 2.5-28x training speedup over existing methods.

Conclusion: Pre-trained diffusion models are naturally suited for step-level reward modeling in latent space, enabling efficient preference optimization without complex pixel-space transformations.

Abstract: Preference optimization for diffusion models aims to align them with human preferences for images. Previous methods typically use Vision-Language Models (VLMs) as pixel-level reward models to approximate human preferences. However, when used for step-level preference optimization, these models face challenges in handling noisy images of different timesteps and require complex transformations into pixel space. In this work, we show that pre-trained diffusion models are naturally suited for step-level reward modeling in the noisy latent space, as they are explicitly designed to process latent images at various noise levels. Accordingly, we propose the Latent Reward Model (LRM), which repurposes components of the diffusion model to predict preferences of latent images at arbitrary timesteps. Building on LRM, we introduce Latent Preference Optimization (LPO), a step-level preference optimization method conducted directly in the noisy latent space. Experimental results indicate that LPO significantly improves the model’s alignment with general, aesthetic, and text-image alignment preferences, while achieving a 2.5-28x training speedup over existing preference optimization methods. Our code and models are available at https://github.com/Kwai-Kolors/LPO.

Method: Proposed SAFEERASER benchmark with 3,000 images and 28.8K VQA pairs. Introduced Prompt Decouple (PD) Loss to alleviate over-forgetting during unlearning process, and Safe Answer Refusal Rate (SARR) metric to quantitatively measure over-forgetting.

Result: Combining PD Loss with existing unlearning methods effectively prevents over-forgetting, achieving 79.5% decrease in SARR metric for LLaVA-7B and LLaVA-13B while maintaining forget quality and model utility.

Conclusion: SAFEERASER provides a comprehensive safety unlearning benchmark, and the proposed PD Loss effectively addresses over-forgetting issues in MLLM safety unlearning, maintaining both forget quality and model performance.

Abstract: As Multimodal Large Language Models (MLLMs) develop, their potential security issues have become increasingly prominent. Machine Unlearning (MU), as an effective strategy for forgetting specific knowledge in training data, has been widely used in privacy protection. However, MU for safety in MLLM has yet to be fully explored. To address this issue, we propose SAFEERASER, a safety unlearning benchmark for MLLMs, consisting of 3,000 images and 28.8K VQA pairs. We comprehensively evaluate unlearning methods from two perspectives: forget quality and model utility. Our findings show that existing MU methods struggle to maintain model performance while implementing the forget operation and often suffer from over-forgetting. Hence, we introduce Prompt Decouple (PD) Loss to alleviate over-forgetting through decouple prompt during unlearning process. To quantitatively measure over-forgetting mitigated by PD Loss, we propose a new metric called Safe Answer Refusal Rate (SARR). Experimental results demonstrate that combining PD Loss with existing unlearning methods can effectively prevent over-forgetting and achieve a decrease of 79.5% in the SARR metric of LLaVA-7B and LLaVA-13B, while maintaining forget quality and model utility. Our code and dataset will be released upon acceptance. Warning: This paper contains examples of harmful language and images, and reader discretion is recommended.

Method: Collected SEE-600K dataset with 610K images and events across 202 scenarios with varying lighting. Proposed framework that captures color through sensor patterns, uses cross-attention to model events as brightness dictionaries, creates broad light-range representation, and decodes pixel-level brightness using prompts.

Result: The method performs well on both low-light enhancement datasets and broader light-range enhancement using SEE-600K. Enables pixel-level brightness adjustment, providing flexibility for post-processing and inspiring more imaging applications.

Conclusion: The approach successfully demonstrates how events can be used for adaptive brightness adjustment across diverse lighting conditions, with the framework and dataset enabling new possibilities for event-based imaging applications.

Abstract: Event cameras, with a high dynamic range exceeding $120dB$, significantly outperform traditional embedded cameras, robustly recording detailed changing information under various lighting conditions, including both low- and high-light situations. However, recent research on utilizing event data has primarily focused on low-light image enhancement, neglecting image enhancement and brightness adjustment across a broader range of lighting conditions, such as normal or high illumination. Based on this, we propose a novel research question: how to employ events to enhance and adaptively adjust the brightness of images captured under broad lighting conditions? To investigate this question, we first collected a new dataset, SEE-600K, consisting of 610,126 images and corresponding events across 202 scenarios, each featuring an average of four lighting conditions with over a 1000-fold variation in illumination. Subsequently, we propose a framework that effectively utilizes events to smoothly adjust image brightness through the use of prompts. Our framework captures color through sensor patterns, uses cross-attention to model events as a brightness dictionary, and adjusts the image’s dynamic range to form a broad light-range representation (BLR), which is then decoded at the pixel level based on the brightness prompt. Experimental results demonstrate that our method not only performs well on the low-light enhancement dataset but also shows robust performance on broader light-range image enhancement using the SEE-600K dataset. Additionally, our approach enables pixel-level brightness adjustment, providing flexibility for post-processing and inspiring more imaging applications. The dataset and source code are publicly available at: https://github.com/yunfanLu/SEE.

Method: Developed three interpretability tools: Position Sensitivity Index (quantifies token order reliance), Cross Modality Balance (reveals attention head allocation), and RoPE Sensitivity probe (measures rotary positional embedding dependence). Used targeted interventions to validate findings.

Result: Analysis revealed vision tokens and system prompts dominate attention, suppressing LLM’s position embedding. Interventions predictably restored positional sensitivity, confirming the mechanistic understanding of the imbalance.

Conclusion: The study uncovered previously unknown failure modes in multimodal attention and demonstrated how interpretability analysis can guide principled improvements to enhance spatial reasoning in VLMs.

Abstract: Vision Language Models (VLMs) excel at identifying and describing objects but often fail at spatial reasoning. We study why VLMs, such as LLaVA, underutilize spatial cues despite having positional encodings and spatially rich vision encoder features. Our analysis reveals a key imbalance: vision token embeddings have much larger norms than text tokens, suppressing LLM’s position embedding. To expose this mechanism, we developed three interpretability tools: (1) the Position Sensitivity Index, which quantifies reliance on token order, (2) the Cross Modality Balance, which reveals attention head allocation patterns, and (3) a RoPE Sensitivity probe, which measures dependence on rotary positional embeddings. These tools uncover that vision tokens and system prompts dominate attention. We validated our mechanistic understanding through targeted interventions that predictably restore positional sensitivity. These findings reveal previously unknown failure modes in multimodal attention and demonstrate how interpretability analysis can guide principled improvements.

Method: Applies attention adaptation during inference by disentangling attention maps in DUSt3R to extract camera and object motion information for dynamic region segmentation, camera pose estimation, and 4D dense point map reconstruction.

Result: Extensive experiments on real-world dynamic videos show that this lightweight attention adaptation significantly outperforms previous state-of-the-art methods that require training or fine-tuning on extensive dynamic datasets.

Conclusion: Attention layers in DUSt3R inherently encode rich motion information, and careful disentanglement of these attention maps enables effective training-free 4D reconstruction that surpasses trained methods.

Abstract: Recent advances in DUSt3R have enabled robust estimation of dense point clouds and camera parameters of static scenes, leveraging Transformer network architectures and direct supervision on large-scale 3D datasets. In contrast, the limited scale and diversity of available 4D datasets present a major bottleneck for training a highly generalizable 4D model. This constraint has driven conventional 4D methods to fine-tune 3D models on scalable dynamic video data with additional geometric priors such as optical flow and depths. In this work, we take an opposite path and introduce Easi3R, a simple yet efficient training-free method for 4D reconstruction. Our approach applies attention adaptation during inference, eliminating the need for from-scratch pre-training or network fine-tuning. We find that the attention layers in DUSt3R inherently encode rich information about camera and object motion. By carefully disentangling these attention maps, we achieve accurate dynamic region segmentation, camera pose estimation, and 4D dense point map reconstruction. Extensive experiments on real-world dynamic videos demonstrate that our lightweight attention adaptation significantly outperforms previous state-of-the-art methods that are trained or finetuned on extensive dynamic datasets. Our code is publicly available for research purpose at https://easi3r.github.io/

Method: Proposed reinforcement learning framework with unified action space rule modeling, using policy optimization algorithms like GRPO on small amounts of curated multi-platform data (Windows, Linux, MacOS, Android, Web).

Result: Achieved superior performance using only 3K data points (vs 13M in previous methods) across eight benchmarks spanning mobile, desktop, and web platforms.

Conclusion: Reinforcement learning with unified action space rule modeling shows immense potential for improving LVLMs’ execution capabilities in real-world GUI agent tasks.

Abstract: Existing efforts in building Graphical User Interface (GUI) agents largely rely on the training paradigm of supervised fine-tuning on Large Vision-Language Models (LVLMs). However, this approach not only demands extensive amounts of training data but also struggles to effectively understand GUI screenshots and generalize to unseen interfaces. The issue significantly limits its application in real-world scenarios, especially for high-level tasks. Inspired by Reinforcement Fine-Tuning (RFT) in large reasoning models (e.g., DeepSeek-R1), which efficiently enhances the problem-solving capabilities of large language models in real-world settings, we propose \name, the first reinforcement learning framework designed to enhance the GUI capabilities of LVLMs in high-level real-world task scenarios, through unified action space rule modeling. By leveraging a small amount of carefully curated high-quality data across multiple platforms (including Windows, Linux, MacOS, Android, and Web) and employing policy optimization algorithms such as Group Relative Policy Optimization (GRPO) to update the model, \name achieves superior performance using only 0.02% of the data (3K vs. 13M) compared to previous state-of-the-art methods like OS-Atlas across eight benchmarks spanning three different platforms (mobile, desktop, and web). These results demonstrate the immense potential of reinforcement learning based on unified action space rule modeling in improving the execution capabilities of LVLMs for real-world GUI agent tasks.

Method: Used PAD-UFES-20 dataset with logistic regression trained on handcrafted features (ABCDE and 7-point checklist) and pre-trained ResNet-50 CNN, evaluated across multiple training datasets with varied sex composition.

Result: Both models were robust to sex distribution, but CNN showed significantly higher accuracy and AUROC for male patients compared to female patients.

Conclusion: While models are generally robust to sex distribution, CNN exhibits sex bias with better performance for male patients, highlighting the need for bias-aware model development in medical AI.

Abstract: Deep learning has been reported to achieve high performances in the detection of skin cancer, yet many challenges regarding the reproducibility of results and biases remain. This study is a replication (different data, same analysis) of a previous study on Alzheimer’s disease detection, which studied the robustness of logistic regression (LR) and convolutional neural networks (CNN) across patient sexes. We explore sex bias in skin cancer detection, using the PAD-UFES-20 dataset with LR trained on handcrafted features reflecting dermatological guidelines (ABCDE and the 7-point checklist), and a pre-trained ResNet-50 model. We evaluate these models in alignment with the replicated study: across multiple training datasets with varied sex composition to determine their robustness. Our results show that both the LR and the CNN were robust to the sex distribution, but the results also revealed that the CNN had a significantly higher accuracy (ACC) and area under the receiver operating characteristics (AUROC) for male patients compared to female patients. The data and relevant scripts to reproduce our results are publicly available (https://github.com/ nikodice4/Skin-cancer-detection-sex-bias).

Method: The approach integrates multi-modal visual inputs (spike streams and RGB data) using modern latent-space generative frameworks, addressing spatial sparsity in spike inputs and temporal sparsity in RGB inputs through latent space manipulation.

Result: Extensive experiments show that SpikeGen effectively enhances performance in conditional image and video deblurring, dense frame reconstruction from spike streams, and high-speed scene novel-view synthesis.

Conclusion: Leveraging generative models’ latent space capabilities allows for synergistic enhancement of different visual modalities, successfully addressing the limitations of both spike and RGB inputs.

Abstract: The process through which humans perceive and learn visual representations in dynamic environments is highly complex. From a structural perspective, the human eye decouples the functions of cone and rod cells: cones are primarily responsible for color perception, while rods are specialized in detecting motion, particularly variations in light intensity. These two distinct modalities of visual information are integrated and processed within the visual cortex, thereby enhancing the robustness of the human visual system. Inspired by this biological mechanism, modern hardware systems have evolved to include not only color-sensitive RGB cameras but also motion-sensitive Dynamic Visual Systems, such as spike cameras. Building upon these advancements, this study seeks to emulate the human visual system by integrating decomposed multi-modal visual inputs with modern latent-space generative frameworks. We named it SpikeGen. We evaluate its performance across various spike-RGB tasks, including conditional image and video deblurring, dense frame reconstruction from spike streams, and high-speed scene novel-view synthesis. Supported by extensive experiments, we demonstrate that leveraging the latent space manipulation capabilities of generative models enables an effective synergistic enhancement of different visual modalities, addressing spatial sparsity in spike inputs and temporal sparsity in RGB inputs.

Method: Proposes MMCLIP with hard negative contrastive learning: image-based contrastive using generation-based hard negatives from perturbed diagram code, and text-based contrastive using rule-based negatives from modified geometric descriptions and retrieval-based negatives from similar captions.

Result: MMGeoLM significantly outperforms other open-source models on three geometric reasoning benchmarks and rivals GPT-4o despite being only 7B parameters. Ablation studies confirm the importance of hard negative types and training configurations.

Conclusion: The hard negative contrastive learning framework effectively enhances vision encoder training for fine-grained geometric reasoning, providing important insights for optimizing multimodal model training pipelines.

Abstract: Large Multimodal Models (LMMs) typically build on ViTs (e.g., CLIP), yet their training with simple random in-batch negatives limits the ability to capture fine-grained visual differences, particularly in geometric scenarios. To address this challenge, we propose a novel hard negative contrastive learning framework for the vision encoder, which combines image-based contrastive learning using generation-based hard negatives created by perturbing diagram generation code, and text-based contrastive learning using rule-based negatives derived from modified geometric descriptions and retrieval-based negatives selected based on caption similarity. We train a vision encoder (CLIP) using our hard negative training method, namely MMCLIP (Multimodal Math CLIP), and subsequently train an LMM for geometric problem-solving. Experiments show that our trained model, MMGeoLM, significantly outperforms other open-source models on three geometric reasoning benchmarks. Even with a size of 7B, it can rival powerful closed-source models like GPT-4o. We further conduct ablation studies to analyze three key factors: hard negative types, the efficiency of image-based negatives, and training configurations. These analyses yield important insights into optimizing the training pipeline of vision encoder for fine-grained geometric reasoning tasks. https://github.com/THU-KEG/MMGeoLM.

Method: Proposed Stabilized Taylor Orthogonal Runge-Kutta (STORK) method, which specifically addresses the stiffness of ODEs and dependence on semi-linear structure, making it applicable to both diffusion and flow-matching models.

Result: STORK consistently improves sampling quality for diffusion and flow-matching models in both image and video generation tasks while reducing the number of function evaluations required.

Conclusion: STORK provides an effective solution for fast, high-quality sampling in diffusion and flow-matching models, overcoming key limitations of previous methods and enabling more efficient inference.

Abstract: Diffusion models (DMs) and flow-matching models have demonstrated remarkable performance in image and video generation. However, such models require a significant number of function evaluations (NFEs) during sampling, leading to costly inference. Consequently, quality-preserving fast sampling methods that require fewer NFEs have been an active area of research. However, prior training-free sampling methods fail to simultaneously address two key challenges: the stiffness of the ODE (i.e., the non-straightness of the velocity field) and dependence on the semi-linear structure of the DM ODE (which limits their direct applicability to flow-matching models). In this work, we introduce the Stabilized Taylor Orthogonal Runge–Kutta (STORK) method, addressing both design concerns. We demonstrate that STORK consistently improves the quality of diffusion and flow-matching sampling for image and video generation. Code is available at https://github.com/ZT220501/STORK.

Method: Combines latent diffusion model with 3D contrastive vision-language pretraining using dual-encoder CLIP-style model trained on CT volumes and radiology reports. Uses pretrained volumetric VAE for compression and efficient 3D denoising diffusion without super-resolution stages.

Result: Achieves competitive performance on CT-RATE dataset, significantly outperforming prior baselines in image fidelity, clinical relevance, and semantic alignment. Synthesized CT scans effectively augment real data and improve downstream diagnostic performance.

Conclusion: Modality-specific vision-language alignment is crucial for high-quality 3D medical image generation. The integrated approach provides scalable and controllable solution for clinically meaningful CT synthesis from text, enabling applications in data augmentation, medical education, and clinical simulation.

Abstract: Objective: While recent advances in text-conditioned generative models have enabled the synthesis of realistic medical images, progress has been largely confined to 2D modalities such as chest X-rays. Extending text-to-image generation to volumetric CT remains a significant challenge, due to its high dimensionality, anatomical complexity, and the absence of robust frameworks that align vision-language data in 3D medical imaging. Methods: We introduce a novel architecture for Text-to-CT generation that combines a latent diffusion model with a 3D contrastive vision-language pretraining scheme. Our approach leverages a dual-encoder CLIP-style model trained on paired CT volumes and radiology reports to establish a shared embedding space, which serves as the conditioning input for generation. CT volumes are compressed into a low-dimensional latent space via a pretrained volumetric VAE, enabling efficient 3D denoising diffusion without requiring external super-resolution stages. Results: We evaluate our method on the CT-RATE dataset and conduct a comprehensive assessment of image fidelity, clinical relevance, and semantic alignment. Our model achieves competitive performance across all tasks, significantly outperforming prior baselines for text-to-CT generation. Moreover, we demonstrate that CT scans synthesized by our framework can effectively augment real data, improving downstream diagnostic performance. Conclusion: Our results show that modality-specific vision-language alignment is a key component for high-quality 3D medical image generation. By integrating contrastive pretraining and volumetric diffusion, our method offers a scalable and controllable solution for synthesizing clinically meaningful CT volumes from text, paving the way for new applications in data augmentation, medical education, and automated clinical simulation. Code at https://github.com/cosbidev/Text2CT.

Method: Proposed Any-to-Any Self-Distillation (ATAS) that leverages model’s own knowledge across all representation levels through internal self-distillation using only unlabeled images, refining CLIP vision encoder representations while preserving local semantic consistency.

Result: ATAS achieves substantial performance gains on open-vocabulary object detection and semantic segmentation benchmarks, outperforming baseline CLIP models.

Conclusion: The approach effectively addresses CLIP’s limitations by jointly maintaining semantic coherence and fine-grained alignment, validating the importance of this joint optimization for advanced open-vocabulary dense prediction.

Abstract: Vision-language models such as CLIP have recently propelled open-vocabulary dense prediction tasks by enabling recognition of a broad range of visual concepts. However, CLIP still struggles with fine-grained, region-level understanding, hindering its effectiveness on these dense prediction tasks. We identify two pivotal factors required to address this limitation: semantic coherence and fine-grained vision-language alignment. Current adaptation methods often improve fine-grained alignment at the expense of semantic coherence, and often rely on extra modules or supervised fine-tuning. To overcome these issues, we propose Any-to-Any Self-Distillation (ATAS), a novel approach that simultaneously enhances semantic coherence and fine-grained alignment by leveraging own knowledge of a model across all representation levels. Unlike prior methods, ATAS uses only unlabeled images and an internal self-distillation process to refine representations of CLIP vision encoders, preserving local semantic consistency while sharpening local detail recognition. On open-vocabulary object detection and semantic segmentation benchmarks, ATAS achieves substantial performance gains, outperforming baseline CLIP models. These results validate the effectiveness of our approach and underscore the importance of jointly maintaining semantic coherence and fine-grained alignment for advanced open-vocabulary dense prediction.

Method: Constructed 3D-MeshPref dataset (first large-scale unpaired 3D preference dataset), developed RewardCS using Cauchy-Schwarz divergence objective for direct 3D preference learning, and integrated it into DreamCS framework for text-to-3D generation.

Result: Extensive experiments show DreamCS outperforms prior methods, producing 3D assets that are both geometrically faithful and human-preferred.

Conclusion: The proposed approach enables effective learning of human-aligned 3D geometric preferences without requiring paired comparisons, advancing text-to-3D generation quality.

Abstract: While text-to-3D generation has attracted growing interest, existing methods often struggle to produce 3D assets that align well with human preferences. Current preference alignment techniques for 3D content typically rely on hardly-collected preference-paired multi-view 2D images to train 2D reward models, when then guide 3D generation – leading to geometric artifacts due to their inherent 2D bias. To address these limitations, we construct 3D-MeshPref, the first large-scale unpaired 3D preference dataset, featuring diverse 3D meshes annotated by a large language model and refined by human evaluators. We then develop RewardCS, the first reward model trained directly on unpaired 3D-MeshPref data using a novel Cauchy-Schwarz divergence objective, enabling effective learning of human-aligned 3D geometric preferences without requiring paired comparisons. Building on this, we propose DreamCS, a unified framework that integrates RewardCS into text-to-3D pipelines – enhancing both implicit and explicit 3D generation with human preference feedback. Extensive experiments show DreamCS outperforms prior methods, producing 3D assets that are both geometrically faithful and human-preferred. Code and models will be released publicly.

Method: Developed a detection system using state-of-the-art deep learning models, trained on a human-annotated dataset of 1,050 IBM i system screen images (including 381 Japanese screens) containing various GUI components.

Result: Experimental results demonstrate the effectiveness of the dataset in building a component detection system, showing successful detection of text labels, text boxes, options, tables, instructions, keyboards, and command lines.

Conclusion: AS400-DET has the potential to perform automated testing on GUI-based systems by automatically detecting GUI components from screen images.

Abstract: This paper proposes a method for automatic GUI component detection for the IBM i system (formerly and still more commonly known as AS/400). We introduce a human-annotated dataset consisting of 1,050 system screen images, in which 381 images are screenshots of IBM i system screens in Japanese. Each image contains multiple components, including text labels, text boxes, options, tables, instructions, keyboards, and command lines. We then develop a detection system based on state-of-the-art deep learning models and evaluate different approaches using our dataset. The experimental results demonstrate the effectiveness of our dataset in constructing a system for component detection from GUI screens. By automatically detecting GUI components from the screen, AS400-DET has the potential to perform automated testing on systems that operate via GUI screens.

Method: Uses a frequency-guided visual adapter to transfer complementary cues across modalities by calibrating spatial, channel, and frequency components, and a multilevel memory adapter with short, long, and permanent memory stores to handle temporal context and recover from challenges like occlusion and motion blur.

Result: Achieves state-of-the-art results on RGB-Thermal, RGB-Depth, and RGB-Event benchmarks, outperforming both fully fine-tuned and adapter-based baselines with favorable parameter efficiency and runtime.

Conclusion: The unified design effectively preserves prompt learning efficiency while strengthening cross-modal interaction and temporal coherence, demonstrating consistent performance improvements across multiple benchmarks.

Abstract: Prompt-learning-based multi-modal trackers have made strong progress by using lightweight visual adapters to inject auxiliary-modality cues into frozen foundation models. However, they still underutilize two essentials: modality-specific frequency structure and long-range temporal dependencies. We present Learning Frequency and Memory-Aware Prompts, a dual-adapter framework that injects lightweight prompts into a frozen RGB tracker. A frequency-guided visual adapter adaptively transfers complementary cues across modalities by jointly calibrating spatial, channel, and frequency components, narrowing the modality gap without full fine-tuning. A multilevel memory adapter with short, long, and permanent memory stores, updates, and retrieves reliable temporal context, enabling consistent propagation across frames and robust recovery from occlusion, motion blur, and illumination changes. This unified design preserves the efficiency of prompt learning while strengthening cross-modal interaction and temporal coherence. Extensive experiments on RGB-Thermal, RGB-Depth, and RGB-Event benchmarks show consistent state-of-the-art results over fully fine-tuned and adapter-based baselines, together with favorable parameter efficiency and runtime. Code and models are available at https://github.com/xuboyue1999/mmtrack.git.

Method: Proposes IC-Custom with In-context Multi-Modal Attention (ICMA) mechanism using learnable task-oriented register tokens and boundary-aware positional embeddings. Uses polyptych configurations by concatenating reference and target images, and created a 12K identity-consistent dataset with real-world and synthetic samples.

Result: Significantly outperforms community workflows, closed-source models, and state-of-the-art open-source approaches. Achieves 73% higher human preference across identity consistency, harmony, and text alignment metrics while training only 0.4% of original model parameters.

Conclusion: IC-Custom provides a unified framework for diverse image customization tasks, demonstrating superior performance across multiple benchmarks with efficient parameter usage, making it suitable for industrial applications like try-on, image insertion, and creative IP customization.

Abstract: Image customization, a crucial technique for industrial media production, aims to generate content that is consistent with reference images. However, current approaches conventionally separate image customization into position-aware and position-free customization paradigms and lack a universal framework for diverse customization, limiting their applications across various scenarios. To overcome these limitations, we propose IC-Custom, a unified framework that seamlessly integrates position-aware and position-free image customization through in-context learning. IC-Custom concatenates reference images with target images to a polyptych, leveraging DiT’s multi-modal attention mechanism for fine-grained token-level interactions. We propose the In-context Multi-Modal Attention (ICMA) mechanism, which employs learnable task-oriented register tokens and boundary-aware positional embeddings to enable the model to effectively handle diverse tasks and distinguish between inputs in polyptych configurations. To address the data gap, we curated a 12K identity-consistent dataset with 8K real-world and 4K high-quality synthetic samples, avoiding the overly glossy, oversaturated look typical of synthetic data. IC-Custom supports various industrial applications, including try-on, image insertion, and creative IP customization. Extensive evaluations on our proposed ProductBench and the publicly available DreamBench demonstrate that IC-Custom significantly outperforms community workflows, closed-source models, and state-of-the-art open-source approaches. IC-Custom achieves about 73% higher human preference across identity consistency, harmony, and text alignment metrics, while training only 0.4% of the original model parameters. Project page: https://liyaowei-stu.github.io/project/IC_Custom

Method: Proposes a divergence-based similarity function (DSF) that represents sets of augmented views as distributions and measures similarity as divergence between distributions.

Result: DSF consistently improves performance across kNN classification and linear evaluation tasks, offers greater efficiency than other multi-view methods, and operates effectively without temperature hyperparameters.

Conclusion: DSF effectively captures joint structure across multiple views through distribution-based divergence measurement, providing performance improvements and practical advantages over existing similarity measures.

Abstract: Recent success in contrastive learning has sparked growing interest in more effectively leveraging multiple augmented views of an instance. While prior methods incorporate multiple views at the loss or feature level, they primarily capture pairwise relationships and fail to model the joint structure across all views. In this work, we propose a divergence-based similarity function (DSF) that explicitly captures the joint structure by representing each set of augmented views as a distribution and measuring similarity as the divergence between distributions. Extensive experiments demonstrate that DSF consistently improves performance across various tasks, including kNN classification and linear evaluation, while also offering greater efficiency compared to other multi-view methods. Furthermore, we establish a theoretical connection between DSF and cosine similarity, and show that, unlike cosine similarity, DSF operates effectively without requiring a temperature hyperparameter.

Method: Uses two streams: one learns deterministic features and the other learns probabilistic features via uncertainty injection. Employs textual guidance to decouple features into disease-wise features and aligns them via feature distillation. Uses pseudo label consistency between streams and self-distillation to transfer task-relevant semantics.

Result: Significantly enhances detection performances on six retinal diseases and normal state, achieving state-of-the-art results on both in-domain and out-of-domain datasets.

Conclusion: PSScreen effectively addresses domain generalization and label absence challenges in partially supervised retinal disease screening, demonstrating superior performance across multiple datasets.

Abstract: Leveraging multiple partially labeled datasets to train a model for multiple retinal disease screening reduces the reliance on fully annotated datasets, but remains challenging due to significant domain shifts across training datasets from various medical sites, and the label absent issue for partial classes. To solve these challenges, we propose PSScreen, a novel Partially Supervised multiple retinal disease Screening model. Our PSScreen consists of two streams and one learns deterministic features and the other learns probabilistic features via uncertainty injection. Then, we leverage the textual guidance to decouple two types of features into disease-wise features and align them via feature distillation to boost the domain generalization ability. Meanwhile, we employ pseudo label consistency between two streams to address the label absent issue and introduce a self-distillation to transfer task-relevant semantics about known classes from the deterministic to the probabilistic stream to further enhance the detection performances. Experiments show that our PSScreen significantly enhances the detection performances on six retinal diseases and the normal state averagely and achieves state-of-the-art results on both in-domain and out-of-domain datasets. Codes are available at https://github.com/boyiZheng99/PSScreen.

Method: Alternating Training-based Label Smoothing (ATLaS) with Class-wise Soft Labels (CSL) and Instance-wise Soft Labels (ISL) for inter-class and instance-class relationships.

Result: ATLaS consistently enhances generalization performance of prompt tuning and shows high compatibility with existing prompt tuning methods.

Conclusion: ATLaS effectively integrates label smoothing with prompt tuning to improve generalization while maintaining parameter efficiency.

Abstract: Recent advances in pre-trained vision-language models have demonstrated remarkable zero-shot generalization capabilities. To further enhance these models’ adaptability to various downstream tasks, prompt tuning has emerged as a parameter-efficient fine-tuning method. However, despite its efficiency, the generalization ability of prompt remains limited. In contrast, label smoothing (LS) has been widely recognized as an effective regularization technique that prevents models from becoming over-confident and improves their generalization. This inspires us to explore the integration of LS with prompt tuning. However, we have observed that the vanilla LS even weakens the generalization ability of prompt tuning. To address this issue, we propose the Alternating Training-based Label Smoothing (ATLaS) method, which alternately trains with standard one-hot labels and soft labels generated by LS to supervise the prompt tuning. Moreover, we introduce two types of efficient offline soft labels, including Class-wise Soft Labels (CSL) and Instance-wise Soft Labels (ISL), to provide inter-class or instance-class relationships for prompt tuning. The theoretical properties of the proposed ATLaS method are analyzed. Extensive experiments demonstrate that the proposed ATLaS method, combined with CSL and ISL, consistently enhances the generalization performance of prompt tuning. Moreover, the proposed ATLaS method exhibits high compatibility with prevalent prompt tuning methods, enabling seamless integration into existing methods.

Method: The researchers manipulated DUO and RUOD datasets to separate object detection into localization and classification tasks. They used YOLO11 and TIDE for localization analysis and conducted classification experiments with balanced data to investigate the under-performance of scallop class.

Result: Localization analysis revealed that foreground-background discrimination is the most problematic stage regardless of data quantity. Classification experiments showed persistent precision gaps even with balanced data, indicating intrinsic feature-based challenges beyond data scarcity and inter-class dependencies.

Conclusion: Researchers recommend using imbalanced distributions when prioritizing precision, and balanced distributions when prioritizing recall. Improving under-performing classes should focus on algorithmic advances, especially within localization modules. The code and datasets are publicly released.

Abstract: Underwater object detection is critical for monitoring marine ecosystems but poses unique challenges, including degraded image quality, imbalanced class distribution, and distinct visual characteristics. Not every species is detected equally well, yet underlying causes remain unclear. We address two key research questions: 1) What factors beyond data quantity drive class-specific performance disparities? 2) How can we systematically improve detection of under-performing marine species? We manipulate the DUO and RUOD datasets to separate the object detection task into localization and classification and investigate the under-performance of the scallop class. Localization analysis using YOLO11 and TIDE finds that foreground-background discrimination is the most problematic stage regardless of data quantity. Classification experiments reveal persistent precision gaps even with balanced data, indicating intrinsic feature-based challenges beyond data scarcity and inter-class dependencies. We recommend imbalanced distributions when prioritizing precision, and balanced distributions when prioritizing recall. Improving under-performing classes should focus on algorithmic advances, especially within localization modules. We publicly release our code and datasets.

Method: 3-stage progressive architecture: 1) EFA-Routing injects instruction info into vision encoder to selectively aggregate visual tokens; 2) LFP-Routing introduces action intent into language model by pruning irrelevant tokens; 3) V-L-A Coupled Attention combines causal vision-language attention with bidirectional action parallel decoding.

Result: Achieved state-of-the-art performance with 97.4% success rate on LIBERO benchmark and 70.0% on real-world robotic tasks, while reducing training costs by 2.5x and inference latency by 2.8x compared to OpenVLA.

Conclusion: CogVLA demonstrates that cognition-aligned routing and sparsification can significantly improve VLA model efficiency while maintaining or enhancing performance, making it more scalable and deployable for real-world applications.

Abstract: Recent Vision-Language-Action (VLA) models built on pre-trained Vision-Language Models (VLMs) require extensive post-training, resulting in high computational overhead that limits scalability and deployment.We propose CogVLA, a Cognition-Aligned Vision-Language-Action framework that leverages instruction-driven routing and sparsification to improve both efficiency and performance. CogVLA draws inspiration from human multimodal coordination and introduces a 3-stage progressive architecture. 1) Encoder-FiLM based Aggregation Routing (EFA-Routing) injects instruction information into the vision encoder to selectively aggregate and compress dual-stream visual tokens, forming a instruction-aware latent representation. 2) Building upon this compact visual encoding, LLM-FiLM based Pruning Routing (LFP-Routing) introduces action intent into the language model by pruning instruction-irrelevant visually grounded tokens, thereby achieving token-level sparsity. 3) To ensure that compressed perception inputs can still support accurate and coherent action generation, we introduce V-L-A Coupled Attention (CAtten), which combines causal vision-language attention with bidirectional action parallel decoding. Extensive experiments on the LIBERO benchmark and real-world robotic tasks demonstrate that CogVLA achieves state-of-the-art performance with success rates of 97.4% and 70.0%, respectively, while reducing training costs by 2.5-fold and decreasing inference latency by 2.8-fold compared to OpenVLA. CogVLA is open-sourced and publicly available at https://github.com/JiuTian-VL/CogVLA.

Method: GACD uses first-order Taylor gradients to estimate bias contributions from individual tokens and visual features. It then suppresses spurious visual features correlated with output objects and rebalances cross-modal contributions by strengthening visual features relative to text.

Result: Experiments across multiple benchmarks show that GACD effectively reduces hallucinations and improves visual grounding of MLLM outputs.

Conclusion: GACD provides an effective inference-based solution to mitigate hallucinations in multimodal LLMs by addressing both text-visual and co-occurrence biases through gradient-based influence analysis.

Abstract: Multimodal large language models achieve strong performance across diverse tasks but remain prone to hallucinations, where outputs are not grounded in visual inputs. This issue can be attributed to two main biases: text-visual bias, the overreliance on prompts and prior outputs, and co-occurrence bias, spurious correlations between frequently paired objects. We propose Gradient-based Influence-Aware Constrained Decoding (GACD), an inference-based method, that addresses both biases without auxiliary models, and is readily applicable to existing models without finetuning. The core of our approach is bias estimation, which uses first-order Taylor gradients to understand the contribution of individual tokens-visual features and text tokens-to the current output. Based on this analysis, GACD mitigates hallucinations through two components: (1) suppressing spurious visual features correlated with the output objects, and (2) rebalancing cross-modal contributions by strengthening visual features relative to text. Experiments across multiple benchmarks demonstrate that GACD effectively reduces hallucinations and improves the visual grounding of MLLM outputs.

Method: Created a 12-dimensional evaluation taxonomy structuring composition around scene graph elements (instance, attribute, relation) and reasoning around philosophical inference types (deductive, inductive, abductive). Curated 1,080 prompts with higher compositional density and reasoning intensity, each paired with checklists containing yes/no questions for fine-grained assessment.

Result: Evaluation of 28 T2I models revealed limited composition capability in high-density scenarios and significantly lagging reasoning capability. All models struggled to infer implicit elements from prompts, with reasoning identified as a critical bottleneck.

Conclusion: Current T2I models have substantial limitations in both composition and reasoning, particularly in complex scenarios. Reasoning capability is a major bottleneck that needs significant improvement for better text-to-image generation.

Abstract: Text-to-image (T2I) generation aims to synthesize images from textual prompts, which jointly specify what must be shown and imply what can be inferred, which thus correspond to two core capabilities: composition and reasoning. Despite recent advances of T2I models in both composition and reasoning, existing benchmarks remain limited in evaluation. They not only fail to provide comprehensive coverage across and within both capabilities, but also largely restrict evaluation to low scene density and simple one-to-one reasoning. To address these limitations, we propose T2I-CoReBench, a comprehensive and complex benchmark that evaluates both composition and reasoning capabilities of T2I models. To ensure comprehensiveness, we structure composition around scene graph elements (instance, attribute, and relation) and reasoning around the philosophical framework of inference (deductive, inductive, and abductive), formulating a 12-dimensional evaluation taxonomy. To increase complexity, driven by the inherent real-world complexities, we curate each prompt with higher compositional density for composition and greater reasoning intensity for reasoning. To facilitate fine-grained and reliable evaluation, we also pair each evaluation prompt with a checklist that specifies individual yes/no questions to assess each intended element independently. In statistics, our benchmark comprises 1,080 challenging prompts and around 13,500 checklist questions. Experiments across 28 current T2I models reveal that their composition capability still remains limited in high compositional scenarios, while the reasoning capability lags even further behind as a critical bottleneck, with all models struggling to infer implicit elements from prompts.

Method: Uses Neural Collapse theory to align feature distributions across clients. Introduces a feature disentanglement module to extract semantically specific features for multi-label settings. Employs predefined shared NC structure and regularization losses to encourage compact clustering in latent feature space.

Result: Experiments on four benchmark datasets across eight diverse settings show the approach outperforms existing methods, validating its effectiveness in challenging FL scenarios with multi-label data and skewed label distributions.

Conclusion: The proposed method successfully addresses the underexplored problem of multi-label federated learning by leveraging Neural Collapse theory and feature disentanglement, demonstrating superior performance over existing approaches.

Abstract: Federated Learning (FL) enables collaborative model training across distributed clients while preserving data privacy. However, the performance of deep learning often deteriorates in FL due to decentralized and heterogeneous data. This challenge is further amplified in multi-label scenarios, where data exhibit complex characteristics such as label co-occurrence, inter-label dependency, and discrepancies between local and global label relationships. While most existing FL research primarily focuses on single-label classification, many real-world applications, particularly in domains such as medical imaging, often involve multi-label settings. In this paper, we address this important yet underexplored scenario in FL, where clients hold multi-label data with skewed label distributions. Neural Collapse (NC) describes a geometric structure in the latent feature space where features of each class collapse to their class mean with vanishing intra-class variance, and the class means form a maximally separated configuration. Motivated by this theory, we propose a method to align feature distributions across clients and to learn high-quality, well-clustered representations. To make the NC-structure applicable to multi-label settings, where image-level features may contain multiple semantic concepts, we introduce a feature disentanglement module that extracts semantically specific features. The clustering of these disentangled class-wise features is guided by a predefined shared NC structure, which mitigates potential conflicts between client models due to diverse local data distributions. In addition, we design regularisation losses to encourage compact clustering in the latent feature space. Experiments conducted on four benchmark datasets across eight diverse settings demonstrate that our approach outperforms existing methods, validating its effectiveness in this challenging FL scenario.

Method: Proposes DACoN framework that fuses low-resolution semantic features from foundation models with high-resolution spatial features from CNNs, removing the constraint on number of reference images used in previous methods.

Result: Quantitative and qualitative evaluations show benefits of using multiple reference images, achieving superior colorization performance compared to previous approaches.

Conclusion: DACoN enables robust feature extraction and supports any number of reference images, demonstrating improved colorization quality for line drawings in anime production.

Abstract: Automatic colorization of line drawings has been widely studied to reduce the labor cost of hand-drawn anime production. Deep learning approaches, including image/video generation and feature-based correspondence, have improved accuracy but struggle with occlusions, pose variations, and viewpoint changes. To address these challenges, we propose DACoN, a framework that leverages foundation models to capture part-level semantics, even in line drawings. Our method fuses low-resolution semantic features from foundation models with high-resolution spatial features from CNNs for fine-grained yet robust feature extraction. In contrast to previous methods that rely on the Multiplex Transformer and support only one or two reference images, DACoN removes this constraint, allowing any number of references. Quantitative and qualitative evaluations demonstrate the benefits of using multiple reference images, achieving superior colorization performance. Our code and model are available at https://github.com/kzmngt/DACoN.

Method: Proposes a new backbone that maps radar pillar features into embedded dimensions with self-attention mechanism to model dependencies between radar points. Uses simplified convolutional layers to replace FPN-based layers from PointPillars architectures to reduce inference time.

Result: Achieves state-of-the-art performance with 58.2 NDS metric using ResNet-50, while setting new benchmark for inference time on nuScenes dataset in the same category.

Conclusion: The proposed camera-radar fusion approach effectively exploits radar advantages and achieves superior performance with faster inference compared to existing methods.

Abstract: Camera-radar fusion offers a robust and low-cost alternative to Camera-lidar fusion for the 3D object detection task in real-time under adverse weather and lighting conditions. However, currently, in the literature, it is possible to find few works focusing on this modality and, most importantly, developing new architectures to explore the advantages of the radar point cloud, such as accurate distance estimation and speed information. Therefore, this work presents a novel and efficient 3D object detection algorithm using cameras and radars in the bird’s-eye-view (BEV). Our algorithm exploits the advantages of radar before fusing the features into a detection head. A new backbone is introduced, which maps the radar pillar features into an embedded dimension. A self-attention mechanism allows the backbone to model the dependencies between the radar points. We are using a simplified convolutional layer to replace the FPN-based convolutional layers used in the PointPillars-based architectures with the main goal of reducing inference time. Our results show that with this modification, our approach achieves the new state-of-the-art in the 3D object detection problem, reaching 58.2 of the NDS metric for the use of ResNet-50, while also setting a new benchmark for inference time on the nuScenes dataset for the same category.

Method: Created FakeChain benchmark with 1-, 2-, and 3-step forgeries using five state-of-the-art generators, analyzing detection performance and spectral properties across hybrid manipulations with varying generator combinations and quality settings.

Result: Detection performance highly depends on the final manipulation type, with F1-score dropping by up to 58.83% when it differs from training distribution, showing detectors rely on last-stage artifacts rather than cumulative manipulation traces.

Conclusion: Detection models need to explicitly consider manipulation history and sequences, and benchmarks like FakeChain are crucial for reflecting growing synthesis complexity in real-world scenarios.

Abstract: Multi-step or hybrid deepfakes, created by sequentially applying different deepfake creation methods such as Face-Swapping, GAN-based generation, and Diffusion methods, can pose an emerging and unforseen technical challenge for detection models trained on single-step forgeries. While prior studies have mainly focused on detecting isolated single manipulation, little is known about the detection model behavior under such compositional, hybrid, and complex manipulation pipelines. In this work, we introduce \textbf{FakeChain}, a large-scale benchmark comprising 1-, 2-, and 3-Step forgeries synthesized using five state-of-the-art representative generators. Using this approach, we analyze detection performance and spectral properties across hybrid manipulation at different step, along with varying generator combinations and quality settings. Surprisingly, our findings reveal that detection performance highly depends on the final manipulation type, with F1-score dropping by up to \textbf{58.83%} when it differs from training distribution. This clearly demonstrates that detectors rely on last-stage artifacts rather than cumulative manipulation traces, limiting generalization. Such findings highlight the need for detection models to explicitly consider manipulation history and sequences. Our results highlight the importance of benchmarks such as FakeChain, reflecting growing synthesis complexity and diversity in real-world scenarios. Our sample code is available here\footnote{https://github.com/minjihh/FakeChain}.

Method: Created DeeptraceReward benchmark with 4.3K detailed annotations across 3.3K generated videos, including natural-language explanations, bounding-box regions, and precise timestamps. Consolidated annotations into 9 major deepfake trace categories and trained multimodal language models as reward models.

Result: The 7B reward model outperforms GPT-5 by 34.7% on average across fake clue identification, grounding, and explanation. Found consistent difficulty gradient: binary classification easiest, followed by natural language explanations, spatial grounding, and temporal labeling (hardest).

Conclusion: DeeptraceReward provides a rigorous testbed and training signal for socially aware and trustworthy video generation by foregrounding human-perceived deepfake traces.

Abstract: Can humans identify AI-generated (fake) videos and provide grounded reasons? While video generation models have advanced rapidly, a critical dimension – whether humans can detect deepfake traces within a generated video, i.e., spatiotemporal grounded visual artifacts that reveal a video as machine generated – has been largely overlooked. We introduce DeeptraceReward, the first fine-grained, spatially- and temporally- aware benchmark that annotates human-perceived fake traces for video generation reward. The dataset comprises 4.3K detailed annotations across 3.3K high-quality generated videos. Each annotation provides a natural-language explanation, pinpoints a bounding-box region containing the perceived trace, and marks precise onset and offset timestamps. We consolidate these annotations into 9 major categories of deepfake traces that lead humans to identify a video as AI-generated, and train multimodal language models (LMs) as reward models to mimic human judgments and localizations. On DeeptraceReward, our 7B reward model outperforms GPT-5 by 34.7% on average across fake clue identification, grounding, and explanation. Interestingly, we observe a consistent difficulty gradient: binary fake v.s. real classification is substantially easier than fine-grained deepfake trace detection; within the latter, performance degrades from natural language explanations (easiest), to spatial grounding, to temporal labeling (hardest). By foregrounding human-perceived deepfake traces, DeeptraceReward provides a rigorous testbed and training signal for socially aware and trustworthy video generation.

Method: Modular training-free pipeline using pretrained vision models to extract object pose, lane positions, and trajectories, organized hierarchically with a three-block prompting strategy for step-wise reasoning.

Result: Outperforms end-to-end V-VLMs on four driving benchmarks with up to 39% gains in accident reasoning accuracy, showing significant improvements with domain-specific cues.

Conclusion: iFinder provides a zero-shot, interpretable, and reliable alternative to end-to-end V-VLMs for driving video understanding by grounding LLMs with domain-specific representations.

Abstract: Grounding large language models (LLMs) in domain-specific tasks like post-hoc dash-cam driving video analysis is challenging due to their general-purpose training and lack of structured inductive biases. As vision is often the sole modality available for such analysis (i.e., no LiDAR, GPS, etc.), existing video-based vision-language models (V-VLMs) struggle with spatial reasoning, causal inference, and explainability of events in the input video. To this end, we introduce iFinder, a structured semantic grounding framework that decouples perception from reasoning by translating dash-cam videos into a hierarchical, interpretable data structure for LLMs. iFinder operates as a modular, training-free pipeline that employs pretrained vision models to extract critical cues – object pose, lane positions, and object trajectories – which are hierarchically organized into frame- and video-level structures. Combined with a three-block prompting strategy, it enables step-wise, grounded reasoning for the LLM to refine a peer V-VLM’s outputs and provide accurate reasoning. Evaluations on four public dash-cam video benchmarks show that iFinder’s proposed grounding with domain-specific cues, especially object orientation and global context, significantly outperforms end-to-end V-VLMs on four zero-shot driving benchmarks, with up to 39% gains in accident reasoning accuracy. By grounding LLMs with driving domain-specific representations, iFinder offers a zero-shot, interpretable, and reliable alternative to end-to-end V-VLMs for post-hoc driving video understanding.

Method: Created SHOT dataset with 1,979 basketball video clips from 5 camera views, annotated with 6 individual attributes. Developed GIFT framework that extracts individual features and models evolving group dynamics to forecast intention emergence.

Result: Experimental results confirm the effectiveness of both SHOT dataset and GIFT framework, establishing a strong foundation for group intention forecasting research.

Conclusion: The work successfully addresses the gap in collective intention analysis, providing a novel task, dataset, and framework that enable forecasting of emerging group intentions from individual actions and interactions.

Abstract: Intention recognition has traditionally focused on individual intentions, overlooking the complexities of collective intentions in group settings. To address this limitation, we introduce the concept of group intention, which represents shared goals emerging through the actions of multiple individuals, and Group Intention Forecasting (GIF), a novel task that forecasts when group intentions will occur by analyzing individual actions and interactions before the collective goal becomes apparent. To investigate GIF in a specific scenario, we propose SHOT, the first large-scale dataset for GIF, consisting of 1,979 basketball video clips captured from 5 camera views and annotated with 6 types of individual attributes. SHOT is designed with 3 key characteristics: multi-individual information, multi-view adaptability, and multi-level intention, making it well-suited for studying emerging group intentions. Furthermore, we introduce GIFT (Group Intention ForecasTer), a framework that extracts fine-grained individual features and models evolving group dynamics to forecast intention emergence. Experimental results confirm the effectiveness of SHOT and GIFT, establishing a strong foundation for future research in group intention forecasting. The dataset is available at https://xinyi-hu.github.io/SHOT_DATASET.

Method: Three-stage iterative approach: (1) Diverse Sample Generation explores potential reasoning paths, (2) Dual Foresight Decoding evaluates samples based on visual focus and reasoning progression, (3) Visual Focus Adjustment refocuses on beneficial regions for future reasoning.

Result: Consistent performance improvements of 3.1-5.8% across multiple benchmarks using Qwen2.5-VL, InternVL-2.5, and Llava-Next, with controllable computational overhead.

Conclusion: CoFFT effectively enhances VLM visual reasoning by creating an interdependent cycle where reasoning guides visual focus and visual focus informs subsequent reasoning, addressing limitations of current VLMs.

Abstract: Despite significant advances in Vision Language Models (VLMs), they remain constrained by the complexity and redundancy of visual input. When images contain large amounts of irrelevant information, VLMs are susceptible to interference, thus generating excessive task-irrelevant reasoning processes or even hallucinations. This limitation stems from their inability to discover and process the required regions during reasoning precisely. To address this limitation, we present the Chain of Foresight-Focus Thought (CoFFT), a novel training-free approach that enhances VLMs’ visual reasoning by emulating human visual cognition. Each Foresight-Focus Thought consists of three stages: (1) Diverse Sample Generation: generates diverse reasoning samples to explore potential reasoning paths, where each sample contains several reasoning steps; (2) Dual Foresight Decoding: rigorously evaluates these samples based on both visual focus and reasoning progression, adding the first step of optimal sample to the reasoning process; (3) Visual Focus Adjustment: precisely adjust visual focus toward regions most beneficial for future reasoning, before returning to stage (1) to generate subsequent reasoning samples until reaching the final answer. These stages function iteratively, creating an interdependent cycle where reasoning guides visual focus and visual focus informs subsequent reasoning. Empirical results across multiple benchmarks using Qwen2.5-VL, InternVL-2.5, and Llava-Next demonstrate consistent performance improvements of 3.1-5.8% with controllable increasing computational overhead.

Method: Proposes two approaches: 1) Multi-Scale Explanation Aggregation (MSEA) for visual branch - aggregates attributions over multi-scale inputs to adjust receptive fields dynamically; 2) Activation Ranking Correlation (ARC) for textual branch - measures relevance of contextual tokens via top-k prediction ranking alignment to suppress spurious activations.

Result: Extensive experiments across state-of-the-art MLLMs and benchmark datasets demonstrate that the approach consistently outperforms existing interpretability methods, yielding more faithful and fine-grained explanations of model behavior.

Conclusion: The proposed methods effectively enhance MLLM interpretability by leveraging intra-modal interactions, addressing limitations of current attribution approaches and providing more holistic visual explanations and semantically coherent textual attributions.

Abstract: Multimodal Large Language Models (MLLMs) have achieved remarkable success across diverse vision-language tasks, yet their internal decision-making mechanisms remain insufficiently understood. Existing interpretability research has primarily focused on cross-modal attribution, identifying which image regions the model attends to during output generation. However, these approaches often overlook intra-modal dependencies. In the visual modality, attributing importance to isolated image patches ignores spatial context due to limited receptive fields, resulting in fragmented and noisy explanations. In the textual modality, reliance on preceding tokens introduces spurious activations. Failing to effectively mitigate these interference compromises attribution fidelity. To address these limitations, we propose enhancing interpretability by leveraging intra-modal interaction. For the visual branch, we introduce \textit{Multi-Scale Explanation Aggregation} (MSEA), which aggregates attributions over multi-scale inputs to dynamically adjust receptive fields, producing more holistic and spatially coherent visual explanations. For the textual branch, we propose \textit{Activation Ranking Correlation} (ARC), which measures the relevance of contextual tokens to the current token via alignment of their top-$k$ prediction rankings. ARC leverages this relevance to suppress spurious activations from irrelevant contexts while preserving semantically coherent ones. Extensive experiments across state-of-the-art MLLMs and benchmark datasets demonstrate that our approach consistently outperforms existing interpretability methods, yielding more faithful and fine-grained explanations of model behavior.

Method: Cross-magnification distillation framework with dual-level knowledge transfer (global image representations and local spatial token mapping), using a compact backbone operating entirely at 5x magnification, trained on 3.49 million images.

Result: Achieved diagnostic accuracy within 1% of larger foundation models with 30-fold processing acceleration (8.8 WSIs per minute), validated across six clinical tasks and multiple cancer types with robust generalization.

Conclusion: Cross-magnification distillation enables deployment of foundation model capabilities in resource-constrained clinical environments, potentially enabling real-time pathology AI integration.

Abstract: Foundation models (FM) have transformed computational pathology but remain computationally prohibitive for clinical deployment due to their massive parameter counts and high-magnification processing requirements. Here, we introduce XMAG, a lightweight FM developed through corss-magnification distillation that transfers knowledge from state-of-the-art 20x magnification teacher to an efficient 5x magnification student architecture. XMAG employs a compact backbone and operates entirely at 5x, requiring 11.3 times fewer patches per whole slide image (WSI) compared to existing approaches. Our Novel distillation framework incorporates dual-level knowledge transfer, aligning both global image representations and local spatial token mapping. We trained XMAG on 3.49 million images curated from publicly available datasets and evaluated performance across six clinically relevant histopathology analysis tasks spanning multiple cancer types. XMAG achieved diagnostic accuracy within 1% of substantially larger foundation models while delivering 30-fold processing acceleration, reaching 8.8 WSIs per minute processing speed. Our cross-institutional validation confirmed robust generalization. Further, we developed an end-to-end training strategy to further boost our model’s performance to approach the larger FMs’ performance. These results establish cross-magnification distillation as a viable approach for deploying FM capabilities in resource-constrained clinical environments, potentially enabling real-time pathology AI integration.

Method: Proposes Dynamic-TreeRPO with sliding-window sampling as tree-structured search with dynamic noise intensities, GRPO-guided optimization, constrained SDE sampling, and LayerTuning-RL that reformulates SFT loss as weighted Progress Reward Model.

Result: Outperforms state-of-the-art by 4.9% on HPS-v2.1, 5.91% on PickScore, and 8.66% on ImageReward benchmarks while improving training efficiency by nearly 50%.

Conclusion: The tree-structured sampling and LayerTuning-RL paradigm enable dynamic exploration of diverse search space, achieving superior semantic consistency, visual fidelity, and human preference alignment.

Abstract: The integration of Reinforcement Learning (RL) into flow matching models for text-to-image (T2I) generation has driven substantial advances in generation quality. However, these gains often come at the cost of exhaustive exploration and inefficient sampling strategies due to slight variation in the sampling group. Building on this insight, we propose Dynamic-TreeRPO, which implements the sliding-window sampling strategy as a tree-structured search with dynamic noise intensities along depth. We perform GRPO-guided optimization and constrained Stochastic Differential Equation (SDE) sampling within this tree structure. By sharing prefix paths of the tree, our design effectively amortizes the computational overhead of trajectory search. With well-designed noise intensities for each tree layer, Dynamic-TreeRPO can enhance the variation of exploration without any extra computational cost. Furthermore, we seamlessly integrate Supervised Fine-Tuning (SFT) and RL paradigm within Dynamic-TreeRPO to construct our proposed LayerTuning-RL, reformulating the loss function of SFT as a dynamically weighted Progress Reward Model (PRM) rather than a separate pretraining method. By associating this weighted PRM with dynamic-adaptive clipping bounds, the disruption of exploration process in Dynamic-TreeRPO is avoided. Benefiting from the tree-structured sampling and the LayerTuning-RL paradigm, our model dynamically explores a diverse search space along effective directions. Compared to existing baselines, our approach demonstrates significant superiority in terms of semantic consistency, visual fidelity, and human preference alignment on established benchmarks, including HPS-v2.1, PickScore, and ImageReward. In particular, our model outperforms SoTA by $4.9%$, $5.91%$, and $8.66%$ on those benchmarks, respectively, while improving the training efficiency by nearly $50%$.

Method: Uses prompt-conditioned hierarchical mixture-of-experts (MoE) architecture with modality-level and token-level projection experts to fuse multiple 3D modalities, plus synthetic VQA protocol generating medically relevant questions from segmentation annotations.

Result: Outperforms strong medical VLM baselines by 5.3% on average across multiple mpMRI datasets, with clinical validation by medical experts.

Conclusion: The study presents the first clinically validated VQA dataset for 3D brain mpMRI, a novel multimodal LLM handling multiple interrelated 3D modalities, and demonstrates strong medical utility through empirical results and ablations.

Abstract: We introduce mpLLM, a prompt-conditioned hierarchical mixture-of-experts (MoE) architecture for visual question answering over multi-parametric 3D brain MRI (mpMRI). mpLLM routes across modality-level and token-level projection experts to fuse multiple interrelated 3D modalities, enabling efficient training without image-report pretraining. To address limited image-text paired supervision, mpLLM integrates a synthetic visual question answering (VQA) protocol that generates medically relevant VQA from segmentation annotations, and we collaborate with medical experts for clinical validation. mpLLM outperforms strong medical VLM baselines by 5.3% on average across multiple mpMRI datasets. Our study features three main contributions: (1) the first clinically validated VQA dataset for 3D brain mpMRI, (2) a novel multimodal LLM that handles multiple interrelated 3D modalities, and (3) strong empirical results that demonstrate the medical utility of our methodology. Ablations highlight the importance of modality-level and token-level experts and prompt-conditioned routing.

Method: Created ReWatch dataset using multi-stage synthesis pipeline with Multi-Agent ReAct framework for CoT synthesis. Developed ReWatch-R1 model through SFT and RLVR with novel Observation & Reasoning reward mechanism.

Result: ReWatch-R1 achieves state-of-the-art average performance on five challenging video reasoning benchmarks.

Conclusion: The proposed dataset and framework successfully advance video reasoning capabilities in LVLMs, demonstrating effectiveness through superior benchmark performance.

Abstract: While Reinforcement Learning with Verifiable Reward (RLVR) significantly advances image reasoning in Large Vision-Language Models (LVLMs), its application to complex video reasoning remains underdeveloped. This gap stems primarily from a critical data bottleneck: existing datasets lack the challenging, multi-hop questions and high-quality, video-grounded Chain-of-Thought (CoT) data necessary to effectively bootstrap RLVR. To address this, we introduce ReWatch, a large-scale dataset built to foster advanced video reasoning. We propose a novel multi-stage synthesis pipeline to synthesize its three components: ReWatch-Caption, ReWatch-QA, and ReWatch-CoT. A core innovation is our Multi-Agent ReAct framework for CoT synthesis, which simulates a human-like “re-watching” process to generate video-grounded reasoning traces by explicitly modeling information retrieval and verification. Building on this dataset, we develop ReWatch-R1 by post-training a strong baseline LVLM with Supervised Fine-Tuning (SFT) and our RLVR framework. This framework incorporates a novel Observation & Reasoning (O&R) reward mechanism that evaluates both the final answer’s correctness and the reasoning’s alignment with video content, directly penalizing hallucination. Our experiments show that ReWatch-R1 achieves state-of-the-art average performance on five challenging video reasoning benchmarks. Project Page: https://rewatch-r1.github.io

Method: Proposes Information-Grounding Guidance (IGG) that uses attention mechanisms to adaptively reinforce informative patches during sampling, ensuring guidance and content remain aligned.

Result: IGG delivers sharper, more coherent, and semantically grounded images across both class-conditioned and text-to-image generation tasks, setting a new benchmark for AR-based methods.

Conclusion: IGG effectively tackles the critical weakness of information inconsistencies in autoregressive image generation, providing a novel guidance mechanism that maintains semantic alignment between guidance and content.

Abstract: Autoregressive (AR) models based on next-scale prediction are rapidly emerging as a powerful tool for image generation, but they face a critical weakness: information inconsistencies between patches across timesteps introduced by progressive resolution scaling. These inconsistencies scatter guidance signals, causing them to drift away from conditioning information and leaving behind ambiguous, unfaithful features. We tackle this challenge with Information-Grounding Guidance (IGG), a novel mechanism that anchors guidance to semantically important regions through attention. By adaptively reinforcing informative patches during sampling, IGG ensures that guidance and content remain tightly aligned. Across both class-conditioned and text-to-image generation tasks, IGG delivers sharper, more coherent, and semantically grounded images, setting a new benchmark for AR-based methods.

Method: Uses a fast proxy system to generate precise occlusion depth maps (1000x1000 resolution in <1ms) that guides Gaussian culling for acceleration and densification during training for quality improvement.

Result: Achieves 2.5x speedup over Octree-GS while delivering substantially higher rendering quality, particularly in heavily occluded scenarios like MatrixCity Streets dataset.

Conclusion: Proxy-GS successfully addresses occlusion redundancy in 3DGS, enabling both faster rendering and improved visual fidelity for MLP-based Gaussian splatting methods.

Abstract: 3D Gaussian Splatting (3DGS) has emerged as an efficient approach for achieving photorealistic rendering. Recent MLP-based variants further improve visual fidelity but introduce substantial decoding overhead during rendering. To alleviate computation cost, several pruning strategies and level-of-detail (LOD) techniques have been introduced, aiming to effectively reduce the number of Gaussian primitives in large-scale scenes. However, our analysis reveals that significant redundancy still remains due to the lack of occlusion awareness. In this work, we propose Proxy-GS, a novel pipeline that exploits a proxy to introduce Gaussian occlusion awareness from any view. At the core of our approach is a fast proxy system capable of producing precise occlusion depth maps at a resolution of 1000x1000 under 1ms. This proxy serves two roles: first, it guides the culling of anchors and Gaussians to accelerate rendering speed. Second, it guides the densification towards surfaces during training, avoiding inconsistencies in occluded regions, and improving the rendering quality. In heavily occluded scenarios, such as the MatrixCity Streets dataset, Proxy-GS not only equips MLP-based Gaussian splatting with stronger rendering capability but also achieves faster rendering speed. Specifically, it achieves more than 2.5x speedup over Octree-GS, and consistently delivers substantially higher rendering quality. Code will be public upon acceptance.

Method: DC-Gen uses a post-training pipeline with lightweight embedding alignment to bridge representation gaps between base and compressed latent spaces, followed by minimal LoRA fine-tuning to preserve generation quality.

Result: DC-Gen-SANA and DC-Gen-FLUX achieve comparable quality to base models with significant speedups: 53x latency reduction for 4K generation on H100 GPU, and 138x total latency reduction when combined with NVFP4 SVDQuant.

Conclusion: DC-Gen provides an effective framework for accelerating high-resolution text-to-image generation while maintaining quality through compressed latent spaces and efficient fine-tuning.

Abstract: Existing text-to-image diffusion models excel at generating high-quality images, but face significant efficiency challenges when scaled to high resolutions, like 4K image generation. While previous research accelerates diffusion models in various aspects, it seldom handles the inherent redundancy within the latent space. To bridge this gap, this paper introduces DC-Gen, a general framework that accelerates text-to-image diffusion models by leveraging a deeply compressed latent space. Rather than a costly training-from-scratch approach, DC-Gen uses an efficient post-training pipeline to preserve the quality of the base model. A key challenge in this paradigm is the representation gap between the base model’s latent space and a deeply compressed latent space, which can lead to instability during direct fine-tuning. To overcome this, DC-Gen first bridges the representation gap with a lightweight embedding alignment training. Once the latent embeddings are aligned, only a small amount of LoRA fine-tuning is needed to unlock the base model’s inherent generation quality. We verify DC-Gen’s effectiveness on SANA and FLUX.1-Krea. The resulting DC-Gen-SANA and DC-Gen-FLUX models achieve quality comparable to their base models but with a significant speedup. Specifically, DC-Gen-FLUX reduces the latency of 4K image generation by 53x on the NVIDIA H100 GPU. When combined with NVFP4 SVDQuant, DC-Gen-FLUX generates a 4K image in just 3.5 seconds on a single NVIDIA 5090 GPU, achieving a total latency reduction of 138x compared to the base FLUX.1-Krea model. Code: https://github.com/dc-ai-projects/DC-Gen.

Method: Multi-Modal Spatio-Temporal Transformer (MM-STT) with joint spatio-temporal attention mechanism that processes dynamic displacement data and static physical priors in a unified manner for deep multi-modal fusion.

Result: Establishes new state-of-the-art on EGMS dataset, reducing long-range forecast RMSE by an order of magnitude compared to all baselines including SOTA methods like STGCN and STAEformer.

Conclusion: For land subsidence forecasting problems, an architecture’s inherent capacity for deep multi-modal fusion is paramount for achieving transformative performance, demonstrating the superiority of multi-modal approaches over uni-modal paradigms.

Abstract: Forecasting high-resolution land subsidence is a critical yet challenging task due to its complex, non-linear dynamics. While standard architectures like ConvLSTM often fail to model long-range dependencies, we argue that a more fundamental limitation of prior work lies in the uni-modal data paradigm. To address this, we propose the Multi-Modal Spatio-Temporal Transformer (MM-STT), a novel framework that fuses dynamic displacement data with static physical priors. Its core innovation is a joint spatio-temporal attention mechanism that processes all multi-modal features in a unified manner. On the public EGMS dataset, MM-STT establishes a new state-of-the-art, reducing the long-range forecast RMSE by an order of magnitude compared to all baselines, including SOTA methods like STGCN and STAEformer. Our results demonstrate that for this class of problems, an architecture’s inherent capacity for deep multi-modal fusion is paramount for achieving transformative performance.

Method: Text-based supervised fine-tuning with sparse labels, visual prompting, and intrinsic-conditioned augmentation to address pixel reference and cross-dataset camera ambiguity issues.

Result: DepthLM surpasses accuracy of advanced VLMs by over 2x with smaller models, making VLMs comparable with pure vision models for the first time, while naturally avoiding over-smoothing.

Conclusion: VLMs can reach expert-level 3D understanding accuracy without architecture or loss changes, and the simplicity of DepthLM enables a single VLM to cover various 3D tasks beyond metric depth.

Abstract: Vision language models (VLMs) can flexibly address various vision tasks through text interactions. Although successful in semantic understanding, state-of-the-art VLMs including GPT-5 still struggle in understanding 3D from 2D inputs. On the other hand, expert pure vision models achieve super-human accuracy in metric depth estimation, a key 3D understanding task. However, they require task-specific architectures and losses. Such difference motivates us to ask: Can VLMs reach expert-level accuracy without architecture or loss change? We take per-pixel metric depth estimation as the representative task and show that the answer is yes! Surprisingly, comprehensive analysis shows that text-based supervised-finetuning with sparse labels is sufficient for VLMs to unlock strong 3D understanding, no dense prediction head or complex regression/regularization loss is needed. The bottleneck for VLMs lies actually in pixel reference and cross-dataset camera ambiguity, which we address through visual prompting and intrinsic-conditioned augmentation. With much smaller models, our method DepthLM surpasses the accuracy of most advanced VLMs by over 2x, making VLMs for the first time comparable with pure vision models. Interestingly, without explicit enforcement during training, VLMs trained with DepthLM naturally avoids over-smoothing, having much fewer flying points at boundary regions than pure vision models. The simplicity of DepthLM also enables a single VLM to cover various 3D tasks beyond metric depth. Our code and model will be released at the link below.

Method: Curated a 2-million-scale multimodal dataset combining textbook knowledge, public data, synthetic samples, and general corpora. Employed three-stage training: domain-specialized pretraining, instruction-driven alignment, and reinforcement-based refinement. Dolphin R1 uses reinforcement learning with ultrasound-specific rewards for enhanced reasoning.

Result: Dolphin R1 achieves U2-score of 0.5835 on U2-Bench across eight ultrasound tasks - over twice the second-best model (0.2968). Dolphin v1.0 also performs competitively. Reasoning-enhanced training significantly improves diagnostic accuracy, consistency, and interpretability.

Conclusion: The Dolphin series demonstrates that unified multimodal foundation models can effectively handle ultrasound’s complexities. Reasoning-augmented training is crucial for high-stakes medical AI, enabling improved diagnostic inference, reasoning transparency, and interpretability in ultrasound applications.

Abstract: Ultrasound is crucial in modern medicine but faces challenges like operator dependence, image noise, and real-time scanning, hindering AI integration. While large multimodal models excel in other medical imaging areas, they struggle with ultrasound’s complexities. To address this, we introduce Dolphin v1.0 (V1) and its reasoning-augmented version, Dolphin R1-the first large-scale multimodal ultrasound foundation models unifying diverse clinical tasks in a single vision-language framework.To tackle ultrasound variability and noise, we curated a 2-million-scale multimodal dataset, combining textbook knowledge, public data, synthetic samples, and general corpora. This ensures robust perception, generalization, and clinical adaptability.The Dolphin series employs a three-stage training strategy: domain-specialized pretraining, instruction-driven alignment, and reinforcement-based refinement. Dolphin v1.0 delivers reliable performance in classification, detection, regression, and report generation. Dolphin R1 enhances diagnostic inference, reasoning transparency, and interpretability through reinforcement learning with ultrasound-specific rewards.Evaluated on U2-Bench across eight ultrasound tasks, Dolphin R1 achieves a U2-score of 0.5835-over twice the second-best model (0.2968) setting a new state of the art. Dolphin v1.0 also performs competitively, validating the unified framework. Comparisons show reasoning-enhanced training significantly improves diagnostic accuracy, consistency, and interpretability, highlighting its importance for high-stakes medical AI.

Method: Proposes AgenticIQA framework with planner, executor, and summarizer components. Decomposes IQA into four subtasks: distortion detection, distortion analysis, tool selection, and tool execution. Uses vision-language models integrated with traditional IQA tools in query-aware manner.

Result: Extensive experiments show AgenticIQA consistently surpasses strong baselines in both scoring accuracy and explanatory alignment across diverse IQA datasets. Introduces AgenticIQA-200K dataset and AgenticIQA-Eval benchmark.

Conclusion: AgenticIQA provides a more adaptable and interpretable approach to image quality assessment by dynamically coordinating perception and analysis through modular agentic framework, achieving superior performance in both scoring accuracy and human-aligned explanations.

Abstract: Image quality assessment (IQA) is inherently complex, as it reflects both the quantification and interpretation of perceptual quality rooted in the human visual system. Conventional approaches typically rely on fixed models to output scalar scores, limiting their adaptability to diverse distortions, user-specific queries, and interpretability needs. Furthermore, scoring and interpretation are often treated as independent processes, despite their interdependence: interpretation identifies perceptual degradations, while scoring abstracts them into a compact metric. To address these limitations, we propose AgenticIQA, a modular agentic framework that integrates vision-language models (VLMs) with traditional IQA tools in a dynamic, query-aware manner. AgenticIQA decomposes IQA into four subtasks – distortion detection, distortion analysis, tool selection, and tool execution – coordinated by a planner, executor, and summarizer. The planner formulates task-specific strategies, the executor collects perceptual evidence via tool invocation, and the summarizer integrates this evidence to produce accurate scores with human-aligned explanations. To support training and evaluation, we introduce AgenticIQA-200K, a large-scale instruction dataset tailored for IQA agents, and AgenticIQA-Eval, the first benchmark for assessing the planning, execution, and summarization capabilities of VLM-based IQA agents. Extensive experiments across diverse IQA datasets demonstrate that AgenticIQA consistently surpasses strong baselines in both scoring accuracy and explanatory alignment.

Method: SeMoBridge uses a Semantic Modality Bridge to map images into text modality while preserving semantic content. It’s closed-form and can be trained with multi-modal supervision combining image and text-alignment losses.

Result: SeMoBridge-T (trained version) requires minimal training time and outperforms other methods, particularly in low-data scenarios (1, 2, and 4 shots).

Conclusion: SeMoBridge effectively addresses CLIP’s intra-modal misalignment through lightweight semantic modality bridging, achieving superior few-shot classification performance with minimal training overhead.

Abstract: While Contrastive Language-Image Pretraining (CLIP) excels at zero-shot tasks by aligning image and text embeddings, its performance in few-shot classification is hindered by a critical limitation: intra-modal misalignment. This issue, caused by a persistent modality gap and CLIP’s exclusively inter-modal training objective, leaves the embedding spaces uncalibrated, making direct image-to-image comparisons unreliable. Existing methods attempt to address this by refining similarity logits or by computationally expensive per-sample optimization. To overcome these challenges, we introduce SeMoBridge, a lightweight yet powerful approach that directly addresses the misalignment. Our method maps images into the text modality, while keeping their semantic content intact through what we call a Semantic Modality Bridge. SeMoBridge is closed-form and can optionally be trained through multi-modal supervision, combining image and text-alignment losses to optimize the projection. Experiments show that the trained version, SeMoBridge-T, requires only a fraction of the training time while overall outperforming other methods, particularly in low-data scenarios (1, 2, and 4 shots). The code is available at https://github.com/christti98/semobridge.

Method: Propose Paragraph-level Relative Policy Optimization (PRPO), a reinforcement learning algorithm that aligns LLM reasoning with image content at paragraph level, using a reasoning-annotated dataset.

Result: PRPO improves detection accuracy significantly and achieves highest reasoning score of 4.55/5.0, outperforming GRPO in ablation studies under test-time conditions.

Conclusion: Grounding multimodal reasoning in visual evidence enables more reliable and interpretable deepfake detection, as demonstrated by PRPO’s effectiveness.

Abstract: The rapid rise of synthetic media has made deepfake detection a critical challenge for online safety and trust. Progress remains constrained by the scarcity of large, high-quality datasets. Although multimodal large language models (LLMs) exhibit strong reasoning capabilities, their performance on deepfake detection is poor, often producing explanations that are misaligned with visual evidence or hallucinatory. To address this limitation, we introduce a reasoning-annotated dataset for deepfake detection and propose Paragraph-level Relative Policy Optimization (PRPO), a reinforcement learning algorithm that aligns LLM reasoning with image content at the paragraph level. Experiments show that PRPO improves detection accuracy by a wide margin and achieves the highest reasoning score of 4.55/5.0. Ablation studies further demonstrate that PRPO significantly outperforms GRPO under test-time conditions. These results underscore the importance of grounding multimodal reasoning in visual evidence to enable more reliable and interpretable deepfake detection.

Method: Proposes two image-difficulty measures (high-frequency index and rotation-invariant edge index) and a new evaluation methodology that reflects visual differences in objective measures.

Result: Experimental results demonstrate the effectiveness of the proposed image-difficulty measures and evaluation methodology.

Conclusion: The difficulty-aware performance evaluation procedures better differentiate between SISR models that produce visually different results but yield close average performance scores.

Abstract: Image super-resolution models are commonly evaluated by average scores (over some benchmark test sets), which fail to reflect the performance of these models on images of varying difficulty and that some models generate artifacts on certain difficult images, which is not reflected by the average scores. We propose difficulty-aware performance evaluation procedures to better differentiate between SISR models that produce visually different results on some images but yield close average performance scores over the entire test set. In particular, we propose two image-difficulty measures, the high-frequency index and rotation-invariant edge index, to predict those test images, where a model would yield significantly better visual results over another model, and an evaluation method where these visual differences are reflected on objective measures. Experimental results demonstrate the effectiveness of the proposed image-difficulty measures and evaluation methodology.

Method: Formulated as cooperative multi-agent reinforcement learning using IPPO (lightweight multi-agent Proximal Policy Optimization) in PyTorch with centralized-training, decentralized-execution paradigm. Experiments conducted in PettingZoo environment with homogeneous drones.

Result: Multiple homogeneous drones or agents successfully self-organized to cover distinct targets without explicit communication.

Conclusion: Agentic AI with multi-agent reinforcement learning enables effective decentralized cooperation and task allocation in autonomous systems like drone delivery.

Abstract: As autonomous systems move from prototypes to real deployments, the ability of multiple agents to make decentralized, cooperative decisions becomes a core requirement. This paper examines how agentic artificial intelligence, agents that act independently, adaptively and proactively can improve task allocation and coordination in multi-agent systems, with primary emphasis on drone delivery and secondary relevance to warehouse automation. We formulate the problem in a cooperative multi-agent reinforcement learning setting and implement a lightweight multi-agent Proximal Policy Optimization, called IPPO, approach in PyTorch under a centralized-training, decentralized-execution paradigm. Experiments are conducted in PettingZoo environment, where multiple homogeneous drones or agents must self-organize to cover distinct targets without explicit communication.

Method: ToolBrain uses flexible reinforcement learning (GRPO, DPO) and supervised learning, with custom reward functions or automated LLM-as-a-judge system, plus knowledge distillation, automatic task generation, tool retrieval, and efficient fine-tuning pipelines.

Result: Demonstrated up to 30.0% improvement in tool-use skills for tasks like autonomous email search, with fast, targeted improvements while maintaining simple and extensible codebase.

Conclusion: ToolBrain provides an effective, user-friendly framework that lowers barriers for adapting LLM-based agents to specific domains and enables efficient development of tool-use capabilities.

Abstract: Effective tool use is essential for agentic AI, yet training agents to utilize tools remains challenging due to manually designed rewards, limited training data, and poor multi-tool selection, resulting in slow adaptation, wasted computational resources, and suboptimal performance. We introduce ToolBrain, a lightweight and user-friendly framework for coaching tool use in agentic models with flexible reinforcement learning (RL), easing the barriers for researchers and practitioners to adapt LLM-based agents to specific domains. It supports a wide range of training strategies, including RL algorithms such as GRPO and DPO, as well as supervised learning. ToolBrain enables custom reward callables directly on an agent’s execution traces or simply utilizes an automated LLM-as-a-judge system for reward generation. It is packed with useful capabilities, including knowledge distillation from large to small models for efficient development, automatic task generation from tool descriptions, seamless tool retrieval, efficient fine-tuning pipelines with QLoRA through Unsloth, and quantized inference via bitsandbytes. We demonstrate ToolBrain through diverse use cases, such as training a CodeAct agent to autonomously execute email search tasks, showing fast, targeted improvements (up to 30.0%) in tool-use skills while keeping the codebase simple and extensible in Agentic AI. Our framework is publicly available at https://toolbrain.org.

Method: Adaptive Reasoning Suppression (ARS) uses multi-checkpoint certainty estimation with progressive suppression thresholds to dynamically suppress redundant reasoning steps without training.

Result: ARS achieves up to 53% token reduction, 46.1% latency reduction, and 57.9% energy reduction across mathematical reasoning benchmarks while maintaining or improving accuracy.

Conclusion: ARS provides an effective training-free solution for improving computational efficiency in large reasoning models through dynamic suppression of redundant reasoning steps.

Abstract: Large Reasoning Language Models (LRLMs or LRMs) demonstrate remarkable capabilities in complex reasoning tasks, but suffer from significant computational inefficiencies due to overthinking phenomena. Existing efficient reasoning methods face the challenge of balancing reasoning quality with inference cost reduction. We propose \textbf{Adaptive Reasoning Suppression (ARS)}, a novel training-free approach that dynamically suppresses redundant reasoning steps while preserving accuracy through adaptive certainty monitoring. ARS introduces a multi-checkpoint certainty estimation mechanism with progressive suppression thresholds, achieving superior efficiency compared to static suppression methods. Our extensive evaluation across mathematical reasoning benchmarks using multiple model architectures demonstrates that ARS achieves up to 53%, 46.1%, and 57.9% in token, latency and energy reduction, while maintaining or improving accuracy.

Method: Recommend initial pilots for NeurIPS by distilling lessons from IPCC’s leadership in building scientific consensus on climate policy.

Result: Identifies that NeurIPS is the best positioned organization to lead AI policy consensus formation due to its strengths and lack of compelling alternatives.

Conclusion: NeurIPS should champion scientific consensus to create higher quality AI policy, as it already leads AI on many fronts and policy engagement is within its purview.

Abstract: Designing wise AI policy is a grand challenge for society. To design such policy, policymakers should place a premium on rigorous evidence and scientific consensus. While several mechanisms exist for evidence generation, and nascent mechanisms tackle evidence synthesis, we identify a complete void on consensus formation. In this position paper, we argue NeurIPS should actively catalyze scientific consensus on AI policy. Beyond identifying the current deficit in consensus formation mechanisms, we argue that NeurIPS is the best option due its strengths and the paucity of compelling alternatives. To make progress, we recommend initial pilots for NeurIPS by distilling lessons from the IPCC’s leadership to build scientific consensus on climate policy. We dispel predictable counters that AI researchers disagree too much to achieve consensus and that policy engagement is not the business of NeurIPS. NeurIPS leads AI on many fronts, and it should champion scientific consensus to create higher quality AI policy.

Method: Develops a comprehensive framework with: (i) semantic MLOps for structured AI lifecycle management, (ii) ontology-driven data lineage tracking for traceability, and (iii) regulatory operations (RegOps) workflows for compliance operationalization.

Result: Framework implementation and validation across diverse pilots to advance regulatory compliance.

Conclusion: CERTAIN aims to promote responsible AI innovation aligned with European standards through its certification framework.

Abstract: Artificial Intelligence has rapidly become a cornerstone technology, significantly influencing Europe’s societal and economic landscapes. However, the proliferation of AI also raises critical ethical, legal, and regulatory challenges. The CERTAIN (Certification for Ethical and Regulatory Transparency in Artificial Intelligence) project addresses these issues by developing a comprehensive framework that integrates regulatory compliance, ethical standards, and transparency into AI systems. In this position paper, we outline the methodological steps for building the core components of this framework. Specifically, we present: (i) semantic Machine Learning Operations (MLOps) for structured AI lifecycle management, (ii) ontology-driven data lineage tracking to ensure traceability and accountability, and (iii) regulatory operations (RegOps) workflows to operationalize compliance requirements. By implementing and validating its solutions across diverse pilots, CERTAIN aims to advance regulatory compliance and to promote responsible AI innovation aligned with European standards.

Method: Integrates Proximal Policy Optimization, adaptive epsilon-greedy exploration, and lightweight inter-agent collaboration for saliency-guided masking and reward-based peer experience sharing.

Result: Outperforms state-of-the-art baselines in fidelity, learning efficiency, and policy robustness on single-agent and multi-agent benchmarks including highway driving and Google Research Football.

Conclusion: Provides a mathematically grounded framework for multi-agent explainability with interpretable, transferable explanations through trajectory perturbation, reward fidelity analysis, and KL divergence regularization.

Abstract: Understanding the decision-making process of Deep Reinforcement Learning agents remains a key challenge for deploying these systems in safety-critical and multi-agent environments. While prior explainability methods like StateMask, have advanced the identification of critical states, they remain limited by computational cost, exploration coverage, and lack of adaptation to multi-agent settings. To overcome these limitations, we propose a mathematically grounded framework, MAGIC-MASK (Multi-Agent Guided Inter-agent Collaboration with Mask-Based Explainability for Reinforcement Learning), that extends perturbation-based explanation to Multi-Agent Reinforcement Learning. Our method integrates Proximal Policy Optimization, adaptive epsilon-greedy exploration, and lightweight inter-agent collaboration to share masked state information and peer experience. This collaboration enables each agent to perform saliency-guided masking and share reward-based insights with peers, reducing the time required for critical state discovery, improving explanation fidelity, and leading to faster and more robust learning. The core novelty of our approach lies in generalizing explainability from single-agent to multi-agent systems through a unified mathematical formalism built on trajectory perturbation, reward fidelity analysis, and Kullback-Leibler divergence regularization. This framework yields localized, interpretable explanations grounded in probabilistic modeling and multi-agent Markov decision processes. We validate our framework on both single-agent and multi-agent benchmarks, including a multi-agent highway driving environment and Google Research Football, demonstrating that MAGIC-MASK consistently outperforms state-of-the-art baselines in fidelity, learning efficiency, and policy robustness while offering interpretable and transferable explanations.

Method: The study audits standalone VLMs for bail prediction, then implements interventions including legal precedents via RAG pipeline and innovative fine-tuning schemes.

Result: VLMs performed poorly across intersectional groups, wrongly denying bail with high confidence. The interventions substantially improved bail prediction performance.

Conclusion: The work demonstrates the need for smart interventions on VLMs before real-world deployment in legal judgment prediction, providing a pathway for safer implementation.

Abstract: Large language models (LLMs) have been extensively used for legal judgment prediction tasks based on case reports and crime history. However, with a surge in the availability of large vision language models (VLMs), legal judgment prediction systems can now be made to leverage the images of the criminals in addition to the textual case reports/crime history. Applications built in this way could lead to inadvertent consequences and be used with malicious intent. In this work, we run an audit to investigate the efficiency of standalone VLMs in the bail decision prediction task. We observe that the performance is poor across multiple intersectional groups and models \textit{wrongly deny bail to deserving individuals with very high confidence}. We design different intervention algorithms by first including legal precedents through a RAG pipeline and then fine-tuning the VLMs using innovative schemes. We demonstrate that these interventions substantially improve the performance of bail prediction. Our work paves the way for the design of smarter interventions on VLMs in the future, before they can be deployed for real-world legal judgment prediction.

Method: Multi-agent reasoning framework with auditing expertise, using expert-annotated dataset from regulatory documents, subject-level risk priors, hybrid retrieval strategy, and specialized agent modules to identify cross-report evidence chains.

Result: Substantially outperforms General-Purpose Agent paradigm in both recall and interpretability, establishing new benchmark for automated financial forensics.

Conclusion: Domain-specific reasoning and dataset construction are valuable for advancing robust financial fraud detection in real-world regulatory applications.

Abstract: Financial fraud detection in real-world scenarios presents significant challenges due to the subtlety and dispersion of evidence across complex, multi-year financial disclosures. In this work, we introduce a novel multi-agent reasoning framework AuditAgent, enhanced with auditing domain expertise, for fine-grained evidence chain localization in financial fraud cases. Leveraging an expert-annotated dataset constructed from enforcement documents and financial reports released by the China Securities Regulatory Commission, our approach integrates subject-level risk priors, a hybrid retrieval strategy, and specialized agent modules to efficiently identify and aggregate cross-report evidence. Extensive experiments demonstrate that our method substantially outperforms General-Purpose Agent paradigm in both recall and interpretability, establishing a new benchmark for automated, transparent financial forensics. Our results highlight the value of domain-specific reasoning and dataset construction for advancing robust financial fraud detection in practical, real-world regulatory applications.

Method: Using embodied AI powered by large visual language models to provide commonsense reasoning for drones to assess context and generate appropriate actions in real time, demonstrated in a simulated urban benchmark in Unreal Engine.

Result: Drones can dynamically interpret their surroundings and decide on sudden maneuvers for safe landings, enabling adaptive recovery and decision-making pipelines that were previously infeasible to design by hand.

Conclusion: Embodied AI makes possible a new class of adaptive recovery and decision-making pipelines, advancing resilience and safety in autonomous aerial systems.

Abstract: Autonomous drones must often respond to sudden events, such as alarms, faults, or unexpected changes in their environment, that require immediate and adaptive decision-making. Traditional approaches rely on safety engineers hand-coding large sets of recovery rules, but this strategy cannot anticipate the vast range of real-world contingencies and quickly becomes incomplete. Recent advances in embodied AI, powered by large visual language models, provide commonsense reasoning to assess context and generate appropriate actions in real time. We demonstrate this capability in a simulated urban benchmark in the Unreal Engine, where drones dynamically interpret their surroundings and decide on sudden maneuvers for safe landings. Our results show that embodied AI makes possible a new class of adaptive recovery and decision-making pipelines that were previously infeasible to design by hand, advancing resilience and safety in autonomous aerial systems.

Method: Combines Slot Attention (neural component) with AA-CBR (symbolic reasoning), including feature combination, casebase reduction, count-based partial orders, One-Vs-Rest multi-class strategy, and Supported AA-CBR.

Result: Effective classifier on CLEVR-Hans datasets with competitive performance against baseline models.

Conclusion: SAA-CBR successfully integrates neural and symbolic approaches for image classification, demonstrating the value of neuro-symbolic pipelines.

Abstract: We introduce Slot Attention Argumentation for Case-Based Reasoning (SAA-CBR), a novel neuro-symbolic pipeline for image classification that integrates object-centric learning via a neural Slot Attention (SA) component with symbolic reasoning conducted by Abstract Argumentation for Case-Based Reasoning (AA-CBR). We explore novel integrations of AA-CBR with the neural component, including feature combination strategies, casebase reduction via representative samples, novel count-based partial orders, a One-Vs-Rest strategy for extending AA-CBR to multi-class classification, and an application of Supported AA-CBR, a bipolar variant of AA-CBR. We demonstrate that SAA-CBR is an effective classifier on the CLEVR-Hans datasets, showing competitive performance against baseline models.

Method: Integrates TS-SQL synthetic data pipeline using dbt as portable intermediate representation with span-aware reinforcement learning objective TS-GRPO, specifically designed to align token-level training with sequence-level execution rewards during LLM fine-tuning.

Result: On TS-SQL test set (500 examples), Thinkquel (32B) achieves 93.2% execution success and 61.8% exact-result match, improving over base model by 67.2% (execution) and 44.4% (match). In Spider experiments (14B), TS-GRPO increases training stability and speeds convergence of execution-match reward compared to GRPO and GSPO.

Conclusion: Thinkquel demonstrates effective integration of synthetic data generation and specialized reinforcement learning to produce robust, portable database queries with improved execution success and result matching, addressing key challenges in natural language to SQL transformation.

Abstract: Transforming natural-language requests into reliable, production-ready data transformations remains challenging: correctness depends on precise schema linking and warehouse-specific SQL dialects, while the strongest supervision available during training–execution success and result matching–are provided only at the sequence level. At the same time, assembling large, execution-validated corpora is costly, and token-level objectives misalign with these global signals, yielding unstable optimization and limited portability. We introduce Thinkquel, a fine-tuned model for producing robust, portable, and execution-validated database queries. Methodologies in Thinkquel integrates a novel synthetic data pipeline, TS-SQL, that leverages dbt as a portable intermediate representation with a span-aware reinforcement learning objective, and Token-Sequence GRPO (TS-GRPO), specifically designed to bridge the gap between token-level training signals and sequence-level execution rewards when finetuning LLMs. On the 500-example TS-SQL test set, Thinkquel (32B) reaches 93.2% execution success and 61.8% exact-result match with a two-stage SFT curriculum, improving over the base model by 67.2% (exec.) and 44.4% (match). In Spider (14B) experiments, TS-GRPO increases training stability and speeds convergence of the execution-match reward relative to GRPO and GSPO.

Method: Disaggregates tool-calling into tool selection and argument generation subtasks, uses decoupled fine-tuning with separate LoRA adapters for each subtask, and implements DualTune inference framework with hierarchical orchestration to limit tool selection scope.

Result: Qwen-2.5-7B model with decoupled fine-tuning improves tool calling accuracy by 46%, outperforms similar-sized models in all cases and larger models (2x size) in most cases on MCP-Bench benchmark.

Conclusion: The decoupled fine-tuning approach enables efficient on-device agent orchestration by specializing LLM capabilities for distinct tool-calling subtasks, significantly closing the performance gap with frontier models while maintaining privacy and cost benefits.

Abstract: The deployment of Large Language Models (LLMs) as agentic orchestrators has revolutionized task automation, but the need for privacy-preserving, cost-effective solutions demands on-device inference capabilities. However, local LLMs consistently underperform compared to frontier models in tool calling scenarios, struggling with both tool selection from large tool sets and accurate argument generation for complex parameter structures. We introduce a methodology that disaggregates a tool-calling task into two distinct subtasks: tool selection and argument generation. We propose “decoupled fine-tuning”, a novel post-training approach that employs LoRA fine-tuning to create dedicated LoRA adapters for tool selection and tool-specific argument generation using separate loss masking for each of the subtasks. Furthermore, we present DualTune, an inference framework that leverages the LoRA adapters created using decoupled fine-tuning to perform efficient agent orchestration with the help of local models on end-user devices. DualTune decomposes the tool-call generation step into tool selection and argument generation, and dynamically loads the corresponding LoRA adapters to generate tool calls. Additionally, DualTune implements hierarchical orchestration to restrict the number of tools required for tool selection. Our experiments on the MCP-Bench benchmark demonstrate that the Qwen-2.5-7B model trained using decoupled fine-tuning improves the tool calling accuracy of the base model by 46%, and outperforms other local reasoning, non-reasoning and fine-tuned models of similar size in all cases, and models that are 2x larger, in most cases.

Method: Used six GPT-OSS:20b model variants (baseline + five ICL configurations) tested on 840 tasks spanning general knowledge, logic riddles, and mathematical olympiad problems, with statistical analysis (ANOVA).

Result: Significant behavioral modifications (p<0.001) across ICL variants showing optimized fragility - high accuracy on general knowledge (91-99%) but degraded performance on complex reasoning (10-43% on riddles vs 43% baseline). No significant differences on olympiad problem (p=0.2173).

Conclusion: ICL guides create systematic trade-offs between efficiency and reasoning flexibility, with important implications for LLM deployment and AI safety.

Abstract: ICL guides are known to improve task-specific performance, but their impact on cross-domain cognitive abilities remains unexplored. This study examines how ICL guides affect reasoning across different knowledge domains using six variants of the GPT-OSS:20b model: one baseline model and five ICL configurations (simple, chain-of-thought, random, appended text, and symbolic language). The models were subjected to 840 tests spanning general knowledge questions, logic riddles, and a mathematical olympiad problem. Statistical analysis (ANOVA) revealed significant behavioral modifications (p less than 0.001) across ICL variants, demonstrating a phenomenon termed “optimized fragility.” ICL models achieved 91%-99% accuracy on general knowledge tasks while showing degraded performance on complex reasoning problems, with accuracy dropping to 10-43% on riddles compared to 43% for the baseline model. Notably, no significant differences emerged on the olympiad problem (p=0.2173), suggesting that complex mathematical reasoning remains unaffected by ICL optimization. These findings indicate that ICL guides create systematic trade-offs between efficiency and reasoning flexibility, with important implications for LLM deployment and AI safety.

Method: Created a benchmark with diverse tool categories containing equivalent tools, tested seven models, conducted controlled experiments on tool features/metadata/pre-training exposure, and proposed a lightweight mitigation using filtering and uniform sampling.

Result: Found unfairness exists with models fixating on single providers or preferring earlier-listed tools; semantic alignment is strongest predictor; perturbing descriptions shifts selections; repeated pre-training exposure amplifies bias; proposed mitigation reduces bias while preserving task coverage.

Conclusion: Tool-selection bias is a key obstacle for fair deployment of tool-augmented LLMs, requiring attention to ensure equitable competition and user experience in tool marketplaces.

Abstract: Agents backed by large language models (LLMs) often rely on external tools drawn from marketplaces where multiple providers offer functionally equivalent options. This raises a critical point concerning fairness: if selection is systematically biased, it can degrade user experience and distort competition by privileging some providers over others. We introduce a benchmark of diverse tool categories, each containing multiple functionally equivalent tools, to evaluate tool-selection bias. Using this benchmark, we test seven models and show that unfairness exists with models either fixating on a single provider or disproportionately preferring earlier-listed tools in context. To investigate the origins of this bias, we conduct controlled experiments examining tool features, metadata (name, description, parameters), and pre-training exposure. We find that: (1) semantic alignment between queries and metadata is the strongest predictor of choice; (2) perturbing descriptions significantly shifts selections; and (3) repeated pre-training exposure to a single endpoint amplifies bias. Finally, we propose a lightweight mitigation that first filters the candidate tools to a relevant subset and then samples uniformly, reducing bias while preserving good task coverage. Our findings highlight tool-selection bias as a key obstacle for the fair deployment of tool-augmented LLMs.

Method: Treat supply chain as knowledge graph, use graph traverser guided by network centrality scores, orchestrate graph retrieval with numerical factor tables and news streams, employ context shells to embed raw figures in natural language.

Result: Enables generation of concise, explainable, and context-rich risk narratives in real-time without costly fine-tuning or dedicated graph database.

Conclusion: The framework provides a lightweight approach that makes quantitative financial data fully intelligible to LLMs while maintaining structural understanding of supply chain relationships.

Abstract: Large Language Models (LLMs) struggle with the complex, multi-modal, and network-native data underlying financial risk. Standard Retrieval-Augmented Generation (RAG) oversimplifies relationships, while specialist models are costly and static. We address this gap with an LLM-centric agent framework for supply chain risk analysis. Our core contribution is to exploit the inherent duality between networks and knowledge graphs (KG). We treat the supply chain network as a KG, allowing us to use structural network science principles for retrieval. A graph traverser, guided by network centrality scores, efficiently extracts the most economically salient risk paths. An agentic architecture orchestrates this graph retrieval alongside data from numerical factor tables and news streams. Crucially, it employs novel ``context shells’’ – descriptive templates that embed raw figures in natural language – to make quantitative data fully intelligible to the LLM. This lightweight approach enables the model to generate concise, explainable, and context-rich risk narratives in real-time without costly fine-tuning or a dedicated graph database.

Method: Evaluated 17 models on 178 time-anchored tasks using crypto markets as testbed, requiring agents to distinguish truth from manipulation, navigate fragmented information, and make irreversible financial decisions under adversarial pressure.

Result: Models achieve only 28% accuracy without tools, 67.4% with tools vs 80% human baseline. Critical failure: models preferentially choose unreliable web search over authoritative data, falling for SEO-optimized misinformation even when correct answers are accessible through specialized tools.

Conclusion: Current models remain fundamentally unprepared for adversarial environments despite impressive reasoning scores. Adversarial robustness is a necessary condition for trustworthy AI autonomy, with implications extending to cybersecurity, content moderation, and other domains with active adversaries.

Abstract: We present CAIA, a benchmark exposing a critical blind spot in AI evaluation: the inability of state-of-the-art models to operate in adversarial, high-stakes environments where misinformation is weaponized and errors are irreversible. While existing benchmarks measure task completion in controlled settings, real-world deployment demands resilience against active deception. Using crypto markets as a testbed where $30 billion was lost to exploits in 2024, we evaluate 17 models on 178 time-anchored tasks requiring agents to distinguish truth from manipulation, navigate fragmented information landscapes, and make irreversible financial decisions under adversarial pressure. Our results reveal a fundamental capability gap: without tools, even frontier models achieve only 28% accuracy on tasks junior analysts routinely handle. Tool augmentation improves performance but plateaus at 67.4% versus 80% human baseline, despite unlimited access to professional resources. Most critically, we uncover a systematic tool selection catastrophe: models preferentially choose unreliable web search over authoritative data, falling for SEO-optimized misinformation and social media manipulation. This behavior persists even when correct answers are directly accessible through specialized tools, suggesting foundational limitations rather than knowledge gaps. We also find that Pass@k metrics mask dangerous trial-and-error behavior for autonomous deployment. The implications extend beyond crypto to any domain with active adversaries, e.g. cybersecurity, content moderation, etc. We release CAIA with contamination controls and continuous updates, establishing adversarial robustness as a necessary condition for trustworthy AI autonomy. The benchmark reveals that current models, despite impressive reasoning scores, remain fundamentally unprepared for environments where intelligence must survive active opposition.

Method: Performed critical review of Hierarchical Reasoning Models, examined key design choices, and developed variants of the model.

Result: Achieved significantly better performance on Sudoku-Extreme and Maze-Hard tasks than previously reported results.

Conclusion: The work raises surprising observations and intriguing directions for further research in transformer-based logical reasoning models.

Abstract: Transformers have demonstrated remarkable performance in natural language processing and related domains, as they largely focus on sequential, autoregressive next-token prediction tasks. Yet, they struggle in logical reasoning, not necessarily because of a fundamental limitation of these models, but possibly due to the lack of exploration of more creative uses, such as latent space and recurrent reasoning. An emerging exploration in this direction is the Hierarchical Reasoning Model (Wang et al., 2025), which introduces a novel type of recurrent reasoning in the latent space of transformers, achieving remarkable performance on a wide range of 2D reasoning tasks. Despite the promising results, this line of models is still at an early stage and calls for in-depth investigation. In this work, we perform a critical review on this class of models, examine key design choices and present intriguing variants that achieve significantly better performance on the Sudoku-Extreme and Maze-Hard tasks than previously reported. Our results also raise surprising observations and intriguing directions for further research.

Method: Developed three enabling techniques: 1) semantic adaptation transmission using fine-tuning with real/generative samples, 2) semantic lightweight transmission using pruning, quantization, and perception-aware sampling, 3) semantic self-evolution control using distributed hierarchical decision-making.

Result: Simulation results show faster convergence, stronger robustness, and the distributed hierarchical optimization method significantly outperforms conventional decision-making schemes.

Conclusion: The proposed framework demonstrates potential for AI agent communication networks by enabling efficient multi-agent collaboration in dynamic environments through semantic-driven approaches.

Abstract: With the rapid growth of intelligent services, communication targets are shifting from humans to artificial intelligent (AI) agents, which require new paradigms to enable real-time perception, decision-making, and collaboration. Semantic communication, which conveys task-relevant meaning rather than raw data, offers a promising solution. However, its practical deployment remains constrained by dynamic environments and limited resources. To address these issues, this article proposes a semantic-driven AI agent communication framework and develops three enabling techniques. First, semantic adaptation transmission applies fine-tuning with real or generative samples to efficiently adapt models to varying environments. Second, semantic lightweight transmission incorporates pruning, quantization, and perception-aware sampling to reduce model complexity and alleviate computational burden on edge agents. Third, semantic self-evolution control employs distributed hierarchical decision-making to optimize multi-dimensional resources, enabling robust multi-agent collaboration in dynamic environments. Simulation results show that the proposed solutions achieve faster convergence and stronger robustness, while the proposed distributed hierarchical optimization method significantly outperforms conventional decision-making schemes, highlighting its potential for AI agent communication networks.

Method: An agentic framework using large language models that implements official coding guidelines for human experts, designed to support the entire ICD-10 coding system.

Result: Achieves best performance to date on rare diagnosis codes, though fine-tuned discriminative classifiers still have an advantage for high-frequency codes. The framework also identifies systematic ‘blind spots’ (codes that are undercoded).

Conclusion: The framework represents a significant advancement in medical coding automation, particularly for rare codes, and provides insights into systematic coding gaps that need addressing.

Abstract: In medical coding, experts map unstructured clinical notes to alphanumeric codes for diagnoses and procedures. We introduce Code Like Humans: a new agentic framework for medical coding with large language models. It implements official coding guidelines for human experts, and it is the first solution that can support the full ICD-10 coding system (+70K labels). It achieves the best performance to date on rare diagnosis codes (fine-tuned discriminative classifiers retain an advantage for high-frequency codes, to which they are limited). Towards future work, we also contribute an analysis of system performance and identify its `blind spots’ (codes that are systematically undercoded).

Method: Three-stage framework: (1) evolutionary proposal mining through preliminary exploration, (2) problem formation and free exploration with trajectory recording, (3) multi-level validation to ensure reproducibility.

Result: Experiments on GAIA benchmark show TRACE consistently increases task complexity while improving reliability through validatable execution trajectories.

Conclusion: TRACE represents a paradigm shift from static to dynamic, self-evolving evaluation systems that provide sustainable challenges for agent development.

Abstract: Recent advances in large language models (LLMs) and agent system designs have empowered agents with unprecedented levels of capability. However, existing agent benchmarks are showing a trend of rapid ceiling-hitting by newly developed agents, making it difficult to meet the demands for evaluating agent abilities. To address this problem, we propose the Trajectory-based Validated-by-Reproducing Agent-benchmark Complexity Evolution (TRACE) framework. This framework takes an original task from an existing benchmark and encourages agents to freely explore and evolve it into a new task with higher difficulty while recording validatable agent trajectories. The framework proceeds in three stages: (1) evolutionary proposal mining, which provides task evolution proposals through preliminary exploration and divergent thinking; (2) problem formation and free exploration, where proposals are conceptualized into feasible problem candidates and the agents then explore them freely while recording their execution trajectories; and (3) multi-level validation, which ensures that the evolved tasks are accompanied by validatable and reproducible trajectories. Experiments on the GAIA benchmark demonstrate that the TRACE framework consistently enhances task complexity while improving the reliability of correctness through validatable execution trajectories. This work marks a paradigm shift from static, manually curated benchmarks to dynamic, self-evolving evaluation systems, providing a sustainable and challenging runway for agent development.

Method: Conducted large systematic study with 100 patient cases, collecting responses from 28 AI systems (2800 total). Assessed responses on three dimensions: answering the question, appropriate use of clinical note evidence, and use of general medical knowledge. Used clinician-authored reference answers to anchor automated metrics.

Result: Automated rankings closely matched expert ratings when using clinician-authored reference answers as anchors.

Conclusion: Carefully designed automated evaluation can scale comparative assessment of AI systems and support patient-clinician communication.

Abstract: Automated approaches to answer patient-posed health questions are rising, but selecting among systems requires reliable evaluation. The current gold standard for evaluating the free-text artificial intelligence (AI) responses–human expert review–is labor-intensive and slow, limiting scalability. Automated metrics are promising yet variably aligned with human judgments and often context-dependent. To address the feasibility of automating the evaluation of AI responses to hospitalization-related questions posed by patients, we conducted a large systematic study of evaluation approaches. Across 100 patient cases, we collected responses from 28 AI systems (2800 total) and assessed them along three dimensions: whether a system response (1) answers the question, (2) appropriately uses clinical note evidence, and (3) uses general medical knowledge. Using clinician-authored reference answers to anchor metrics, automated rankings closely matched expert ratings. Our findings suggest that carefully designed automated evaluation can scale comparative assessment of AI systems and support patient-clinician communication.

Method: Proposed Expandable Decision-Making States (EDMS) that enrich raw positions/velocities with relational variables (space scoring, pass, score) and use action masking to give distinct decision sets to on-ball and off-ball agents.

Result: EDMS with action masking consistently reduced action-prediction loss and temporal-difference error compared to baseline. Qualitative analysis showed it highlights high-risk, high-reward tactical patterns like counterattacks and defensive breakthroughs.

Conclusion: EDMS enables interpretable tactical analysis by mapping learned functions to human-understandable concepts, works across multiple datasets, and provides better performance than baseline methods while maintaining tactical interpretability.

Abstract: Invasion team sports such as soccer produce a high-dimensional, strongly coupled state space as many players continuously interact on a shared field, challenging quantitative tactical analysis. Traditional rule-based analyses are intuitive, while modern predictive machine learning models often perform pattern-matching without explicit agent representations. The problem we address is how to build player-level agent models from data, whose learned values and policies are both tactically interpretable and robust across heterogeneous data sources. Here, we propose Expandable Decision-Making States (EDMS), a semantically enriched state representation that augments raw positions and velocities with relational variables (e.g., scoring of space, pass, and score), combined with an action-masking scheme that gives on-ball and off-ball agents distinct decision sets. Compared to prior work, EDMS maps learned value functions and action policies to human-interpretable tactical concepts (e.g., marking pressure, passing lanes, ball accessibility) instead of raw coordinate features, and aligns agent choices with the rules of play. In the experiments, EDMS with action masking consistently reduced both action-prediction loss and temporal-difference (TD) error compared to the baseline. Qualitative case studies and Q-value visualizations further indicate that EDMS highlights high-risk, high-reward tactical patterns (e.g., fast counterattacks and defensive breakthroughs). We also integrated our approach into an open-source library and demonstrated compatibility with multiple commercial and open datasets, enabling cross-provider evaluation and reproducible experiments.

Method: Conducted unified evaluation of four reward model variants (discriminative ORM/PRM and generative ORM/PRM) across 14 diverse domains, with theoretical analysis of error compounding in stepwise scoring and empirical validation.

Result: GenORM was the most robust performer, yielding significant and consistent gains across all tested domains, while DisORM performed on par with DisPRM and GenPRM was not competitive. Stepwise scoring suffered from label noise and difficulty evaluating long reasoning trajectories.

Conclusion: Generative outcome verification is more effective for multi-domain deployment than fine-grained stepwise supervision, which compounds errors as reasoning length grows and inherits label noise from auto-labeling.

Abstract: The reliability of large language models (LLMs) during test-time scaling is often assessed with \emph{external verifiers} or \emph{reward models} that distinguish correct reasoning from flawed logic. Prior work generally assumes that process reward models (PRMs), which score every intermediate reasoning step, outperform outcome reward models (ORMs) that assess only the final answer. This view is based mainly on evidence from narrow, math-adjacent domains. We present the first unified evaluation of four reward model variants, discriminative ORM and PRM (\DisORM, \DisPRM) and generative ORM and PRM (\GenORM, \GenPRM), across 14 diverse domains. Contrary to conventional wisdom, we find that (i) \DisORM performs on par with \DisPRM, (ii) \GenPRM is not competitive, and (iii) overall, \GenORM is the most robust, yielding significant and consistent gains across every tested domain. We attribute this to PRM-style stepwise scoring, which inherits label noise from LLM auto-labeling and has difficulty evaluating long reasoning trajectories, including those involving self-correcting reasoning. Our theoretical analysis shows that step-wise aggregation compounds errors as reasoning length grows, and our empirical observations confirm this effect. These findings challenge the prevailing assumption that fine-grained supervision is always better and support generative outcome verification for multi-domain deployment. We publicly release our code, datasets, and checkpoints at \href{https://github.com/db-Lee/Multi-RM}{\underline{\small\texttt{https://github.com/db-Lee/Multi-RM}}} to facilitate future research in multi-domain settings.

Method: Extends segmentation models and vision-language models to process visual prompts that pinpoint specific regions, enabling precise handling of complex scenarios through visual interactions.

Result: Achieves state-of-the-art performance with significant improvements across 36 universal MMEB tasks (3.1%-8.5%) and five visual-interactive SCaR tasks (15.2%-20.3%).

Conclusion: VIRTUE successfully bridges the gap between visual interaction and representation learning, enabling localized grounding of user intent and enhanced entity-level understanding in embedding models.

Abstract: Multimodal representation learning models have demonstrated successful operation across complex tasks, and the integration of vision-language models (VLMs) has further enabled embedding models with instruction-following capabilities. However, existing embedding models lack visual-interactive capabilities to specify regions of interest from users (e.g., point, bounding box, mask), which have been explored in generative models to broaden their human-interactive applicability. Equipping embedding models with visual interactions not only would unlock new applications with localized grounding of user intent, which remains unexplored, but also enable the models to learn entity-level information within images to complement their global representations for conventional embedding tasks. In this paper, we propose a novel Visual-InteRactive Text-Image Universal Embedder (VIRTUE) that extends the capabilities of the segmentation model and the vision-language model to the realm of representation learning. In VIRTUE, the segmentation model can process visual prompts that pinpoint specific regions within an image, thereby enabling the embedder to handle complex and ambiguous scenarios more precisely. To evaluate the visual-interaction ability of VIRTUE, we introduce a large-scale Segmentation-and-Scene Caption Retrieval (SCaR) benchmark comprising 1M samples that aims to retrieve the text caption by jointly considering the entity with a specific object and image scene. VIRTUE consistently achieves a state-of-the-art performance with significant improvements across 36 universal MMEB (3.1%-8.5%) and five visual-interactive SCaR (15.2%-20.3%) tasks.

Method: Conducted 16 semi-structured interviews with practitioners from leading IT service companies and performed qualitative content analysis to inductively derive DQ dimensions.

Result: Identified 15 distinct DQ dimensions across four RAG processing stages: data extraction, transformation, prompt & search, and generation. Found that new dimensions are needed, concentrated in early stages, and quality issues propagate through the pipeline.

Conclusion: RAG systems require new DQ dimensions, front-loaded quality management strategies, and dynamic, step-aware approaches to address the transformation and propagation of quality issues throughout the pipeline.

Abstract: Organizations increasingly adopt Retrieval-Augmented Generation (RAG) to enhance Large Language Models with enterprise-specific knowledge. However, current data quality (DQ) frameworks have been primarily developed for static datasets, and only inadequately address the dynamic, multi-stage nature of RAG systems. This study aims to develop DQ dimensions for this new type of AI-based systems. We conduct 16 semi-structured interviews with practitioners of leading IT service companies. Through a qualitative content analysis, we inductively derive 15 distinct DQ dimensions across the four processing stages of RAG systems: data extraction, data transformation, prompt & search, and generation. Our findings reveal that (1) new dimensions have to be added to traditional DQ frameworks to also cover RAG contexts; (2) these new dimensions are concentrated in early RAG steps, suggesting the need for front-loaded quality management strategies, and (3) DQ issues transform and propagate through the RAG pipeline, necessitating a dynamic, step-aware approach to quality management.

Method: Proposes a novel safety alignment method that trains DLMs to generate safe responses from contaminated intermediate states containing affirmative tokens, specifically designed to counter the vulnerability in the iterative denoising process.

Result: The proposed method significantly mitigates the vulnerability with minimal impact on task performance and improves robustness against conventional jailbreak attacks.

Conclusion: The work underscores the need for DLM-specific safety research due to the critical vulnerability in DLMs’ iterative denoising process that can be exploited by jailbreak attacks.

Abstract: Diffusion language models (DLMs) generate tokens in parallel through iterative denoising, which can reduce latency and enable bidirectional conditioning. However, the safety risks posed by jailbreak attacks that exploit this inference mechanism are not well understood. In this paper, we reveal that DLMs have a critical vulnerability stemming from their iterative denoising process and propose a countermeasure. Specifically, our investigation shows that if an affirmative token for a harmful query appears at an intermediate step, subsequent denoising can be steered toward a harmful response even in aligned models. As a result, simply injecting such affirmative tokens can readily bypass the safety guardrails. Furthermore, we demonstrate that the vulnerability allows existing optimization-based jailbreak attacks to succeed on DLMs. Building on this analysis, we propose a novel safety alignment method tailored to DLMs that trains models to generate safe responses from contaminated intermediate states that contain affirmative tokens. Our experiments indicate that the proposed method significantly mitigates the vulnerability with minimal impact on task performance. Furthermore, our method improves robustness against conventional jailbreak attacks. Our work underscores the need for DLM-specific safety research.

Method: Uses compression guideline optimization in natural language space: LLMs analyze failure cases in compressed contexts and update guidelines accordingly, then distills optimized compressors into smaller models.

Result: Reduces memory usage by 26-54% (peak tokens) while preserving task performance, maintains over 95% accuracy when distilled, and improves smaller LMs’ performance by up to 46% as long-horizon agents.

Conclusion: ACON effectively addresses context length challenges in agentic tasks through optimized compression and distillation, enabling more efficient long-horizon agent deployment.

Abstract: Large language models (LLMs) are increasingly deployed as agents in dynamic, real-world environments, where success requires both reasoning and effective tool use. A central challenge for agentic tasks is the growing context length, as agents must accumulate long histories of actions and observations. This expansion raises costs and reduces efficiency in long-horizon tasks, yet prior work on context compression has mostly focused on single-step tasks or narrow applications. We introduce Agent Context Optimization (ACON), a unified framework that optimally compresses both environment observations and interaction histories into concise yet informative condensations. ACON leverages compression guideline optimization in natural language space: given paired trajectories where full context succeeds but compressed context fails, capable LLMs analyze the causes of failure, and the compression guideline is updated accordingly. Furthermore, we propose distilling the optimized LLM compressor into smaller models to reduce the overhead of the additional module. Experiments on AppWorld, OfficeBench, and Multi-objective QA show that ACON reduces memory usage by 26-54% (peak tokens) while largely preserving task performance, preserves over 95% of accuracy when distilled into smaller compressors, and enhances smaller LMs as long-horizon agents with up to 46% performance improvement.

Method: Uses literature mining to identify emerging trends, explores hypothesis design spaces, pinpoints research gaps, and learns a reward model that scores hypotheses based on prior experimental outcomes.

Result: HARPA-generated proposals perform comparably to baseline AI-researcher but achieve significant gains in feasibility (+0.78) and groundedness (+0.85). With CodeScientist agent, achieved more successful executions (20 vs 11) and fewer failures (16 vs 21). Reward model achieved ~28% absolute gain over untrained baseline.

Conclusion: HARPA represents a step forward in AI-driven scientific discovery by generating more feasible, grounded hypotheses and learning from experimental outcomes to improve hypothesis quality.

Abstract: While there has been a surge of interest in automated scientific discovery (ASD), especially with the emergence of LLMs, it remains challenging for tools to generate hypotheses that are both testable and grounded in the scientific literature. Additionally, existing ideation tools are not adaptive to prior experimental outcomes. We developed HARPA to address these challenges by incorporating the ideation workflow inspired by human researchers. HARPA first identifies emerging research trends through literature mining, then explores hypothesis design spaces, and finally converges on precise, testable hypotheses by pinpointing research gaps and justifying design choices. Our evaluations show that HARPA-generated hypothesis-driven research proposals perform comparably to a strong baseline AI-researcher across most qualitative dimensions (e.g., specificity, novelty, overall quality), but achieve significant gains in feasibility(+0.78, p$<0.05$, bootstrap) and groundedness (+0.85, p$<0.01$, bootstrap) on a 10-point Likert scale. When tested with the ASD agent (CodeScientist), HARPA produced more successful executions (20 vs. 11 out of 40) and fewer failures (16 vs. 21 out of 40), showing that expert feasibility judgments track with actual execution success. Furthermore, to simulate how researchers continuously refine their understanding of what hypotheses are both testable and potentially interesting from experience, HARPA learns a reward model that scores new hypotheses based on prior experimental outcomes, achieving approx. a 28% absolute gain over HARPA’s untrained baseline scorer. Together, these methods represent a step forward in the field of AI-driven scientific discovery.

Method: Systematically developed a suite of new evaluation methods with negative examples to test model editing robustness, particularly using negation queries.

Result: State-of-the-art model editing approaches collapse under simple negation queries, showing they likely rely on shortcuts rather than full semantic understanding.

Conclusion: Model editing literature rests on a fragile foundation with illusory success, requiring urgent reconsideration before meaningful advancements can be pursued.

Abstract: Large language models (LLMs) inevitably encode outdated or incorrect knowledge. Updating, deleting, and forgetting such knowledge is important for alignment, safety, and other issues. To address this issue, model editing has emerged as a promising paradigm: by precisely editing a small subset of parameters such that a specific fact is updated while preserving other knowledge. Despite its great success reported in previous papers, we find the apparent reliability of editing rests on a fragile foundation and the current literature is largely driven by illusory success. The fundamental goal of steering the model’s output toward a target with minimal modification would encourage exploiting hidden shortcuts, rather than utilizing real semantics. This problem directly challenges the feasibility of the current model editing literature at its very foundation, as shortcuts are inherently at odds with robust knowledge integration. Coincidentally, this issue has long been obscured by evaluation frameworks that lack the design of negative examples. To uncover it, we systematically develop a suite of new evaluation methods. Strikingly, we find that state-of-the-art approaches collapse even under the simplest negation queries. Our empirical evidence shows that editing is likely to be based on shortcuts rather than full semantics, calling for an urgent reconsideration of the very basis of model editing before further advancements can be meaningfully pursued.

Method: Collaborative Progressive Distillation (CPD) that progressively transfers knowledge from teacher to student models, reducing both sampling steps and model size, with dual-signal regularized distillation loss.

Result: Achieves 96.2% and 95.5% of baseline ADE/FDE performance on pedestrian trajectories with only 231K parameters and 2-4 sampling steps (161x compression, 31x acceleration, 9ms latency).

Conclusion: CDDM bridges high-performing generative models with practical deployment constraints, enabling resource-efficient probabilistic prediction for autonomous vehicles and intelligent transportation systems.

Abstract: Trajectory prediction is a fundamental task in Autonomous Vehicles (AVs) and Intelligent Transportation Systems (ITS), supporting efficient motion planning and real-time traffic safety management. Diffusion models have recently demonstrated strong performance in probabilistic trajectory prediction, but their large model size and slow sampling process hinder real-world deployment. This paper proposes Collaborative-Distilled Diffusion Models (CDDM), a novel method for real-time and lightweight trajectory prediction. Built upon Collaborative Progressive Distillation (CPD), CDDM progressively transfers knowledge from a high-capacity teacher diffusion model to a lightweight student model, jointly reducing both the number of sampling steps and the model size across distillation iterations. A dual-signal regularized distillation loss is further introduced to incorporate guidance from both the teacher and ground-truth data, mitigating potential overfitting and ensuring robust performance. Extensive experiments on the ETH-UCY pedestrian benchmark and the nuScenes vehicle benchmark demonstrate that CDDM achieves state-of-the-art prediction accuracy. The well-distilled CDDM retains 96.2% and 95.5% of the baseline model’s ADE and FDE performance on pedestrian trajectories, while requiring only 231K parameters and 4 or 2 sampling steps, corresponding to 161x compression, 31x acceleration, and 9 ms latency. Qualitative results further show that CDDM generates diverse and accurate trajectories under dynamic agent behaviors and complex social interactions. By bridging high-performing generative models with practical deployment constraints, CDDM enables resource-efficient probabilistic prediction for AVs and ITS. Code is available at https://github.com/bingzhangw/CDDM.

Method: Leverages distributional properties of LLM activations to compute expected attention scores in closed form for each KV pair, enabling principled ranking and pruning with minimal impact on the residual stream.

Result: Consistently outperforms state-of-the-art baselines in both prefilling and decoding phases, achieving effective compression without performance degradation.

Conclusion: The method provides a practical solution for KV cache compression and the authors release KVPress library with over 20 techniques for researchers to implement and benchmark KV cache compression methods.

Abstract: Memory consumption of the Key-Value (KV) cache represents a major bottleneck for efficient large language model inference. While attention-score-based KV cache pruning shows promise, it faces critical practical limitations: attention scores from future tokens are unavailable during compression, and modern implementations like Flash Attention do not materialize the full attention matrix, making past scores inaccessible. To overcome these challenges, we introduce $\textbf{Expected Attention, a training-free compression method}$ that estimates KV pairs importance by predicting how future queries will attend to them. Our approach leverages the distributional properties of LLM activations to compute expected attention scores in closed form for each KV pair. These scores enable principled ranking and pruning of KV pairs with minimal impact on the residual stream, achieving effective compression without performance degradation. Importantly, our method operates seamlessly across both prefilling and decoding phases, consistently outperforming state-of-the-art baselines in both scenarios. Finally, $\textbf{we release KVPress, a comprehensive library to enable researchers to implement and benchmark KV cache compression methods, already including more than 20 techniques}$.

Method: Fuses batch implementation of Grad-CAM algorithm with prototypical reconstruction loss to guide models to focus on salient image features.

Result: Achieves simultaneous improvement in accuracy and image reconstruction quality while reducing training and inference times.

Conclusion: This approach contributes to building more transparent, explainable, and trustworthy AI systems by ensuring models learn from evidence-relevant information.

Abstract: Understanding the inner workings of deep learning models is crucial for advancing artificial intelligence, particularly in high-stakes fields such as healthcare, where accurate explanations are as vital as precision. This paper introduces Batch-CAM, a novel training paradigm that fuses a batch implementation of the Grad-CAM algorithm with a prototypical reconstruction loss. This combination guides the model to focus on salient image features, thereby enhancing its performance across classification tasks. Our results demonstrate that Batch-CAM achieves a simultaneous improvement in accuracy and image reconstruction quality while reducing training and inference times. By ensuring models learn from evidence-relevant information,this approach makes a relevant contribution to building more transparent, explainable, and trustworthy AI systems.

Method: Iterative RZ reduction method that repeatedly solves positions while gradually restricting regions, with three constraint generation strategies and RZ Pattern Table to leverage past solutions.

Result: On 7x7 Killall-Go, the method reduces average RZ size to 85.95% of original, with reduced RZs stored as reusable knowledge.

Conclusion: The approach successfully reduces RZ sizes and creates permanent reusable knowledge for future solving tasks on larger boards or different openings.

Abstract: Game solving aims to find the optimal strategies for all players and determine the theoretical outcome of a game. However, due to the exponential growth of game trees, many games remain unsolved, even though methods like AlphaZero have demonstrated super-human level in game playing. The Relevance-Zone (RZ) is a local strategy reuse technique that restricts the search to only the regions relevant to the outcome, significantly reducing the search space. However, RZs are not unique. Different solutions may result in RZs of varying sizes. Smaller RZs are generally more favorable, as they increase the chance of reuse and improve pruning efficiency. To this end, we propose an iterative RZ reduction method that repeatedly solves the same position while gradually restricting the region involved, guiding the solver toward smaller RZs. We design three constraint generation strategies and integrate an RZ Pattern Table to fully leverage past solutions. In experiments on 7x7 Killall-Go, our method reduces the average RZ size to 85.95% of the original. Furthermore, the reduced RZs can be permanently stored as reusable knowledge for future solving tasks, especially for larger board sizes or different openings.

Method: ACPO uses a dual-component adaptive strategy: 1) dynamic curriculum that transitions from stable on-policy exploration to efficient off-policy exploitation by progressively increasing sample reuse, and 2) Advantage-Aware Adaptive Clipping (AAAC) that replaces fixed clipping with dynamic, sample-wise bounds based on normalized token advantages.

Result: Extensive experiments on MathVista, LogicVista, and MMMU-Pro benchmarks show ACPO consistently outperforms strong baselines like DAPO and PAPO, achieving state-of-the-art performance with accelerated convergence and superior training stability.

Conclusion: ACPO effectively addresses limitations of existing policy optimization methods through its adaptive curriculum and clipping mechanisms, demonstrating significant improvements in aligning vision-language models for complex reasoning tasks.

Abstract: Aligning large-scale vision-language models (VLMs) for complex reasoning via reinforcement learning is often hampered by the limitations of existing policy optimization algorithms, such as static training schedules and the rigid, uniform clipping mechanism in Proximal Policy Optimization (PPO). In this work, we introduce Adaptive Curriculum Policy Optimization (ACPO), a novel framework that addresses these challenges through a dual-component adaptive learning strategy. First, ACPO employs a dynamic curriculum that orchestrates a principled transition from a stable, near on-policy exploration phase to an efficient, off-policy exploitation phase by progressively increasing sample reuse. Second, we propose an Advantage-Aware Adaptive Clipping (AAAC) mechanism that replaces the fixed clipping hyperparameter with dynamic, sample-wise bounds modulated by the normalized advantage of each token. This allows for more granular and robust policy updates, enabling larger gradients for high-potential samples while safeguarding against destructive ones. We conduct extensive experiments on a suite of challenging multimodal reasoning benchmarks, including MathVista, LogicVista, and MMMU-Pro. Results demonstrate that ACPO consistently outperforms strong baselines such as DAPO and PAPO, achieving state-of-the-art performance, accelerated convergence, and superior training stability.

Method: Uses hierarchical encoding with unigrams and bigrams, attention mechanisms to highlight clinically relevant tokens, cross-attention to incorporate mental health knowledge graph, and ordinal regression for severity prediction.

Result: Outperforms state-of-the-art baselines by over 5% in graded F1 score across datasets while providing interpretable insights.

Conclusion: Advances trustworthy and transparent AI systems for mental health assessment from social media data.

Abstract: In today’s interconnected society, social media platforms provide a window into individuals’ thoughts, emotions, and mental states. This paper explores the use of platforms like Facebook, X (formerly Twitter), and Reddit for depression severity detection. We propose AttentionDep, a domain-aware attention model that drives explainable depression severity estimation by fusing contextual and domain knowledge. Posts are encoded hierarchically using unigrams and bigrams, with attention mechanisms highlighting clinically relevant tokens. Domain knowledge from a curated mental health knowledge graph is incorporated through a cross-attention mechanism, enriching the contextual features. Finally, depression severity is predicted using an ordinal regression framework that respects the clinical-relevance and natural ordering of severity levels. Our experiments demonstrate that AttentionDep outperforms state-of-the-art baselines by over 5% in graded F1 score across datasets, while providing interpretable insights into its predictions. This work advances the development of trustworthy and transparent AI systems for mental health assessment from social media.

Method: Proposed a data augmentation pipeline with three components: EvolAST (AST-based syntactic symmetry), EvolDomain (LLM-based semantic symmetry across domains), and EvolDifficulty (evolutionary instructions for difficulty scaling). Used this data to train EvolProver, a 7B-parameter non-reasoning theorem prover.

Result: EvolProver achieved SOTA on FormalMATH-Lite (53.8% pass@32), MiniF2F-Test (69.8%), Ineq-Comp-Seed (52.2%), and Ineq-Comp-Transformed (34.0%), surpassing all comparable-size models including reasoning-based ones.

Conclusion: The proposed data augmentation pipeline effectively enhances model robustness, and EvolProver demonstrates that non-reasoning models can achieve competitive performance with proper training data augmentation.

Abstract: Large Language Models (LLMs) for formal theorem proving have shown significant promise, yet they often lack generalizability and are fragile to even minor transformations of problem statements. To address this limitation, we introduce a novel data augmentation pipeline designed to enhance model robustness from two perspectives: symmetry and difficulty. From the symmetry perspective, we propose two complementary methods: EvolAST, an Abstract Syntax Tree (AST) based approach that targets syntactic symmetry to generate semantically equivalent problem variants, and EvolDomain, which leverages LLMs to address semantic symmetry by translating theorems across mathematical domains. From the difficulty perspective, we propose EvolDifficulty, which uses carefully designed evolutionary instructions to guide LLMs in generating new theorems with a wider range of difficulty. We then use the evolved data to train EvolProver, a 7B-parameter non-reasoning theorem prover. EvolProver establishes a new state-of-the-art (SOTA) on FormalMATH-Lite with a 53.8% pass@32 rate, surpassing all models of comparable size, including reasoning-based models. It also sets new SOTA records for non-reasoning models on MiniF2F-Test (69.8% pass@32), Ineq-Comp-Seed (52.2% pass@32), and Ineq-Comp-Transformed (34.0% pass@32). Ablation studies further confirm our data augmentation pipeline’s effectiveness across multiple benchmarks.

Method: DIA attacks the integrated DDIM trajectory path by disrupting the inversion process that converts real images to latent codes, preventing subsequent malicious editing operations.

Result: DIA effectively disrupts the diffusion process and surpasses previous defensive methods across various editing methods, providing stronger protection against malicious image manipulation.

Conclusion: The proposed DIA framework offers practical defense methods against malicious use of AI-generated content and can benefit both industry and research communities in combating unethical and abusive content creation.

Abstract: Diffusion models have shown to be strong representation learners, showcasing state-of-the-art performance across multiple domains. Aside from accelerated sampling, DDIM also enables the inversion of real images back to their latent codes. A direct inheriting application of this inversion operation is real image editing, where the inversion yields latent trajectories to be utilized during the synthesis of the edited image. Unfortunately, this practical tool has enabled malicious users to freely synthesize misinformative or deepfake contents with greater ease, which promotes the spread of unethical and abusive, as well as privacy-, and copyright-infringing contents. While defensive algorithms such as AdvDM and Photoguard have been shown to disrupt the diffusion process on these images, the misalignment between their objectives and the iterative denoising trajectory at test time results in weak disruptive performance.In this work, we present the DDIM Inversion Attack (DIA) that attacks the integrated DDIM trajectory path. Our results support the effective disruption, surpassing previous defensive methods across various editing methods. We believe that our frameworks and results can provide practical defense methods against the malicious use of AI for both the industry and the research community. Our code is available here: https://anonymous.4open.science/r/DIA-13419/.

Method: Conducted interviews with course coordinators from data science courses at Wageningen University to gather insights on AI integration.

Result: Identified both benefits (streamlining tasks, enhancing learning) and challenges (student overreliance, potential hindrance of cognitive skill development).

Conclusion: AI can be valuable in education when carefully integrated to complement rather than replace fundamental learning processes, requiring responsible usage, ethical considerations, and adapted assessment methods.

Abstract: This study explores the integration of AI, particularly large language models (LLMs) like ChatGPT, into educational settings, focusing on the implications for teaching and learning. Through interviews with course coordinators from data science courses at Wageningen University, this research identifies both the benefits and challenges associated with AI in the classroom. While AI tools can streamline tasks and enhance learning, concerns arise regarding students’ overreliance on these technologies, potentially hindering the development of essential cognitive and problem solving skills. The study highlights the importance of responsible AI usage, ethical considerations, and the need for adapting assessment methods to ensure educational outcomes are met. With careful integration, AI can be a valuable asset in education, provided it is used to complement rather than replace fundamental learning processes.

Method: Created ChemX datasets, conducted benchmarking of existing agentic systems (ChatGPT Agent, chemical-specific agents), introduced single-agent approach with controlled document preprocessing, and evaluated modern baselines like GPT-5.

Result: Empirical findings reveal persistent challenges in processing domain-specific terminology, complex tabular/schematic representations, and context-dependent ambiguities.

Conclusion: ChemX benchmark serves as critical resource for advancing automated extraction in chemistry, challenging generalization of existing methods and providing insights for effective evaluation.

Abstract: The emergence of agent-based systems represents a significant advancement in artificial intelligence, with growing applications in automated data extraction. However, chemical information extraction remains a formidable challenge due to the inherent heterogeneity of chemical data. Current agent-based approaches, both general-purpose and domain-specific, exhibit limited performance in this domain. To address this gap, we present ChemX, a comprehensive collection of 10 manually curated and domain-expert-validated datasets focusing on nanomaterials and small molecules. These datasets are designed to rigorously evaluate and enhance automated extraction methodologies in chemistry. To demonstrate their utility, we conduct an extensive benchmarking study comparing existing state-of-the-art agentic systems such as ChatGPT Agent and chemical-specific data extraction agents. Additionally, we introduce our own single-agent approach that enables precise control over document preprocessing prior to extraction. We further evaluate the performance of modern baselines, such as GPT-5 and GPT-5 Thinking, to compare their capabilities with agentic approaches. Our empirical findings reveal persistent challenges in chemical information extraction, particularly in processing domain-specific terminology, complex tabular and schematic representations, and context-dependent ambiguities. The ChemX benchmark serves as a critical resource for advancing automated information extraction in chemistry, challenging the generalization capabilities of existing methods, and providing valuable insights into effective evaluation strategies.

Method: Semantic comparison of the two formalisms by analyzing how they generate orderings on interpretations through numerical penalties for violated rules/conditionals.

Result: Shows that under certain conditions, weighted knowledge bases and c-representations can generate the same ordering on interpretations, establishing semantic equivalence up to relative cost. Also compares entailment notions between both formalisms.

Conclusion: The results demonstrate potential benefits for further work on both cost-based semantics and c-representations by establishing connections between these two approaches to handling inconsistent knowledge.

Abstract: Weighted-knowledge bases and cost-based semantics represent a recent formalism introduced by Bienvenu et al. for Ontology Mediated Data Querying in the case where a given knowledge base is inconsistent. This is done by adding a weight to each statement in the knowledge base (KB), and then giving each DL interpretation a cost based on how often it breaks rules in the KB. In this paper we compare this approach with c-representations, a form of non-monotonic reasoning originally introduced by Kern-Isberner. c-Representations describe a means to interpret defeasible concept inclusions in the first-order case. This is done by assigning a numerical ranking to each interpretations via penalties for each violated conditional. We compare these two approaches on a semantic level. In particular, we show that under certain conditions a weighted knowledge base and a set of defeasible conditionals can generate the same ordering on interpretations, and therefore an equivalence of semantic structures up to relative cost. Moreover, we compare entailment described in both cases, where certain notions are equivalently expressible in both formalisms. Our results have the potential to benefit further work on both cost-based semantics and c-representations

Method: Formalizes a logic of unsupervised evaluation using linear programming in integer space, with logical constraints (inequalities) and universally applicable axioms (linear equalities) based on agreement/disagreement patterns.

Result: Developed a practical approach that can detect when LLMs-as-Judges violate minimum grading thresholds using only logical consistency, without requiring ground truth labels.

Conclusion: The proposed unsupervised evaluation framework provides immediate utility for detecting violations in classifier performance using only logical consistency constraints derived from agreement patterns.

Abstract: If two experts disagree on a test, we may conclude both cannot be 100 per cent correct. But if they completely agree, no possible evaluation can be excluded. This asymmetry in the utility of agreements versus disagreements is explored here by formalizing a logic of unsupervised evaluation for classifiers. Its core problem is computing the set of group evaluations that are logically consistent with how we observe them agreeing and disagreeing in their decisions. Statistical summaries of their aligned decisions are inputs into a Linear Programming problem in the integer space of possible correct or incorrect responses given true labels. Obvious logical constraints, such as, the number of correct responses cannot exceed the number of observed responses, are inequalities. But in addition, there are axioms, universally applicable linear equalities that apply to all finite tests. The practical and immediate utility of this approach to unsupervised evaluation using only logical consistency is demonstrated by building no-knowledge alarms that can detect when one or more LLMs-as-Judges are violating a minimum grading threshold specified by the user.

Method: Used classical ML models with voltage and current waveforms segmented into sliding windows (10-50 ms) from EMT data, evaluated under realistic real-time constraints with performance metrics including accuracy, robustness to window size, and runtime efficiency.

Result: Best-performing fault classification model achieved F1 score of 0.992±0.001, while top fault localization model reached R2 of 0.806±0.008 with mean processing time of 0.563 ms.

Conclusion: Machine learning models demonstrate strong performance for fault classification and localization in power systems, offering viable alternatives to conventional protection schemes in dynamic grid environments with distributed energy resources.

Abstract: The increasing integration of distributed energy resources (DERs), particularly renewables, poses significant challenges for power system protection, with fault classification (FC) and fault localization (FL) being among the most critical tasks. Conventional protection schemes, based on fixed thresholds, cannot reliably identify and localize short circuits with the increasing complexity of the grid under dynamic conditions. Machine learning (ML) offers a promising alternative; however, systematic benchmarks across models and settings remain limited. This work presents, for the first time, a comparative benchmarking study of classical ML models for FC and FL in power system protection based on EMT data. Using voltage and current waveforms segmented into sliding windows of 10 ms to 50 ms, we evaluate models under realistic real-time constraints. Performance is assessed in terms of accuracy, robustness to window size, and runtime efficiency. The best-performing FC model achieved an F1 score of 0.992$\pm$0.001, while the top FL model reached an R2 of 0.806$\pm$0.008 with a mean processing time of 0.563 ms.

Method: Applied Synthetic Minority Oversampling Technique (SMOTE) and evaluated advanced ensemble learning models to distinguish manipulative trading behavior from normal market activity.

Result: SMOTE greatly enhanced all models’ ability to detect P&D events by increasing recall and improving precision-recall balance. XGBoost and LightGBM achieved high recall rates (94.87% and 93.59% respectively) with strong F1-scores and fast computational performance.

Conclusion: Integrating data balancing techniques with ensemble methods significantly improves early detection of manipulative activities, contributing to a fairer, more transparent, and more stable cryptocurrency market.

Abstract: This study aims to detect pump and dump (P&D) manipulation in cryptocurrency markets, where the scarcity of such events causes severe class imbalance and hinders accurate detection. To address this issue, the Synthetic Minority Oversampling Technique (SMOTE) was applied, and advanced ensemble learning models were evaluated to distinguish manipulative trading behavior from normal market activity. The experimental results show that applying SMOTE greatly enhanced the ability of all models to detect P&D events by increasing recall and improving the overall balance between precision and recall. In particular, XGBoost and LightGBM achieved high recall rates (94.87% and 93.59%, respectively) with strong F1-scores and demonstrated fast computational performance, making them suitable for near real time surveillance. These findings indicate that integrating data balancing techniques with ensemble methods significantly improves the early detection of manipulative activities, contributing to a fairer, more transparent, and more stable cryptocurrency market.

Method: Model probability of correct answers using IRT-inspired factors (model ability vector, query discrimination vector, query difficulty) with Mixture-of-Experts network.

Result: State-of-the-art performance in model routing and benchmark accuracy prediction, with learned parameters encoding interpretable model capabilities.

Conclusion: The approach successfully creates compact, meaningful representations of LLM abilities that facilitate downstream tasks.

Abstract: Recent years have witnessed a surge in the number of large language models (LLMs), yet efficiently managing and utilizing these vast resources remains a significant challenge. In this work, we explore how to learn compact representations of LLM abilities that can facilitate downstream tasks, such as model routing and performance prediction on new benchmarks. We frame this problem as estimating the probability that a given model will correctly answer a specific query. Inspired by the item response theory (IRT) in psychometrics, we model this probability as a function of three key factors: (i) the model’s multi-skill ability vector, (2) the query’s discrimination vector that separates models of differing skills, and (3) the query’s difficulty scalar. To learn these parameters jointly, we introduce a Mixture-of-Experts (MoE) network that couples model- and query-level embeddings. Extensive experiments demonstrate that our approach leads to state-of-the-art performance in both model routing and benchmark accuracy prediction. Moreover, analysis validates that the learned parameters encode meaningful, interpretable information about model capabilities and query characteristics.

Method: AIDE uses Monte Carlo Tree Search (MCTS) as a foundation for automated data exploration, focusing on identifying data transformations and models that reveal interesting patterns.

Result: Evaluation with real-world and synthetic data shows AIDE effectively uncovers interesting data patterns through automated transformation and modeling.

Conclusion: AIDE’s MCTS-based framework provides strong extensibility for future enhancements, making it a valuable step toward comprehensive automated knowledge discovery solutions.

Abstract: Organizations are increasingly focused on leveraging data from their processes to gain insights and drive decision-making. However, converting this data into actionable knowledge remains a difficult and time-consuming task. There is often a gap between the volume of data collected and the ability to process and understand it, which automated knowledge discovery aims to fill. Automated knowledge discovery involves complex open problems, including effectively navigating data, building models to extract implicit relationships, and considering subjective goals and knowledge. In this paper, we introduce a novel method for Automated Insights and Data Exploration (AIDE), that serves as a robust foundation for tackling these challenges through the use of Monte Carlo Tree Search (MCTS). We evaluate AIDE using both real-world and synthetic data, demonstrating its effectiveness in identifying data transformations and models that uncover interesting data patterns. Among its strengths, AIDE’s MCTS-based framework offers significant extensibility, allowing for future integration of additional pattern extraction strategies and domain knowledge. This makes AIDE a valuable step towards developing a comprehensive solution for automated knowledge discovery.

Method: Proposes FusionAdapter with (1) adapter modules for efficient modality adaptation to unseen relations and (2) fusion strategy integrating multimodal entity representations while preserving modality-specific characteristics.

Result: Extensive experiments on two benchmark MMKG datasets demonstrate superior performance over state-of-the-art methods, particularly in few-shot scenarios.

Conclusion: FusionAdapter effectively adapts and fuses diverse modality information, improving generalization to novel relations with minimal supervision while preserving modality-specific characteristics.

Abstract: Multimodal Knowledge Graphs (MMKGs) incorporate various modalities, including text and images, to enhance entity and relation representations. Notably, different modalities for the same entity often present complementary and diverse information. However, existing MMKG methods primarily align modalities into a shared space, which tends to overlook the distinct contributions of specific modalities, limiting their performance particularly in low-resource settings. To address this challenge, we propose FusionAdapter for the learning of few-shot relationships (FSRL) in MMKG. FusionAdapter introduces (1) an adapter module that enables efficient adaptation of each modality to unseen relations and (2) a fusion strategy that integrates multimodal entity representations while preserving diverse modality-specific characteristics. By effectively adapting and fusing information from diverse modalities, FusionAdapter improves generalization to novel relations with minimal supervision. Extensive experiments on two benchmark MMKG datasets demonstrate that FusionAdapter achieves superior performance over state-of-the-art methods.

Method: Leverage an LLM-guided evolutionary framework to efficiently explore the space of RA functions, yielding Discovered Adversarial Imitation Learning (DAIL) - the first meta-learnt AIL algorithm.

Result: DAIL generalizes across unseen environments and policy optimization algorithms, outperforming current state-of-the-art human-designed baselines and leading to more stable training.

Conclusion: The discovered RA functions provide novel insights into AIL stability and demonstrate that data-driven approaches can outperform traditional human-designed methods in adversarial imitation learning.

Abstract: Adversarial Imitation Learning (AIL) methods, while effective in settings with limited expert demonstrations, are often considered unstable. These approaches typically decompose into two components: Density Ratio (DR) estimation $\frac{\rho_E}{\rho_{\pi}}$, where a discriminator estimates the relative occupancy of state-action pairs under the policy versus the expert; and Reward Assignment (RA), where this ratio is transformed into a reward signal used to train the policy. While significant research has focused on improving density estimation, the role of reward assignment in influencing training dynamics and final policy performance has been largely overlooked. RA functions in AIL are typically derived from divergence minimization objectives, relying heavily on human design and ingenuity. In this work, we take a different approach: we investigate the discovery of data-driven RA functions, i.e, based directly on the performance of the resulting imitation policy. To this end, we leverage an LLM-guided evolutionary framework that efficiently explores the space of RA functions, yielding \emph{Discovered Adversarial Imitation Learning} (DAIL), the first meta-learnt AIL algorithm. Remarkably, DAIL generalises across unseen environments and policy optimization algorithms, outperforming the current state-of-the-art of \emph{human-designed} baselines. Finally, we analyse why DAIL leads to more stable training, offering novel insights into the role of RA functions in the stability of AIL. Code is publicly available: https://github.com/shshnkreddy/DAIL.

Method: Proposes test-time search with stochasticity: 1) stochastic edge selection in graph coarsening instead of greedy approach, 2) Graph Coarsening History-based Partitioning (GraphCHiP) algorithm that leverages coarsening history to identify RCIs and FCIs.

Result: Experiments on random CVRP instances show reduced dual gap compared to existing neural separation method, and successful discovery of effective Framed Capacity Inequalities (FCIs) on specific instances.

Conclusion: The proposed test-time search approach with stochasticity effectively enhances neural separation methods for CVRP, generating more diverse cuts and identifying challenging FCIs.

Abstract: The identification of valid inequalities, such as the rounded capacity inequalities (RCIs), is a key component of cutting plane methods for the Capacitated Vehicle Routing Problem (CVRP). While a deep learning-based separation method can learn to find high-quality cuts, our analysis reveals that the model produces fewer cuts than expected because it is insufficiently sensitive to generate a diverse set of generated subsets. This paper proposes an alternative: enhancing the performance of a trained model at inference time through a new test-time search with stochasticity. First, we introduce stochastic edge selection into the graph coarsening procedure, replacing the previously proposed greedy approach. Second, we propose the Graph Coarsening History-based Partitioning (GraphCHiP) algorithm, which leverages coarsening history to identify not only RCIs but also, for the first time, the Framed capacity inequalities (FCIs). Experiments on randomly generated CVRP instances demonstrate the effectiveness of our approach in reducing the dual gap compared to the existing neural separation method. Additionally, our method discovers effective FCIs on a specific instance, despite the challenging nature of identifying such cuts.

Method: Uses LSTM networks with temporal attention to condense multivariate data into interpretable features, then applies fuzzy inference systems for forecasting.

Result: Shows comparable performance to conventional models like ARIMA and LSTM while providing meaningful interpretability of information flow within the network.

Conclusion: Demonstrates potential for interpretable forecasting in stock markets, though performance tradeoffs exist, suggesting practical applications for understanding market behavior.

Abstract: In the complex landscape of multivariate time series forecasting, achieving both accuracy and interpretability remains a significant challenge. This paper introduces the Fuzzy Transformer (Fuzzformer), a novel recurrent neural network architecture combined with multi-head self-attention and fuzzy inference systems to analyze multivariate stock market data and conduct long-term time series forecasting. The method leverages LSTM networks and temporal attention to condense multivariate data into interpretable features suitable for fuzzy inference systems. The resulting architecture offers comparable forecasting performance to conventional models such as ARIMA and LSTM while providing meaningful information flow within the network. The method was examined on the real world stock market index S&P500. Initial results show potential for interpretable forecasting and identify current performance tradeoffs, suggesting practical application in understanding and forecasting stock market behavior.

Method: Uses agentic reinforcement learning with tool-augmented LLMs, incorporating quantum circuit verification via external simulators and hierarchical reward mechanisms in RL training.

Result: Achieved 99.31% validity in Pass@1 and 100% in Pass@10 with a 4B LLM, outperforming GPT-4o, GPT-5, DeepSeek-V3 and other baselines in both syntax and semantic performance.

Conclusion: QUASAR effectively addresses quantum circuit generation challenges by combining LLMs with quantum-specific verification and sophisticated RL rewards, demonstrating superior performance over existing approaches.

Abstract: Designing and optimizing task-specific quantum circuits are crucial to leverage the advantage of quantum computing. Recent large language model (LLM)-based quantum circuit generation has emerged as a promising automatic solution. However, the fundamental challenges remain unaddressed: (i) parameterized quantum gates require precise numerical values for optimal performance, which also depend on multiple aspects, including the number of quantum gates, their parameters, and the layout/depth of the circuits. (ii) LLMs often generate low-quality or incorrect quantum circuits due to the lack of quantum domain-specific knowledge. We propose QUASAR, an agentic reinforcement learning (RL) framework for quantum circuits generation and optimization based on tool-augmented LLMs. To align the LLM with quantum-specific knowledge and improve the generated quantum circuits, QUASAR designs (i) a quantum circuit verification approach with external quantum simulators and (ii) a sophisticated hierarchical reward mechanism in RL training. Extensive evaluation shows improvements in both syntax and semantic performance of the generated quantum circuits. When augmenting a 4B LLM, QUASAR has achieved the validity of 99.31% in Pass@1 and 100% in Pass@10, outperforming industrial LLMs of GPT-4o, GPT-5 and DeepSeek-V3 and several supervised-fine-tuning (SFT)-only and RL-only baselines.

Method: Integrates three components: few-shot federated optimization with meta-learning, energy-aware client scheduling, and secure aggregation with calibrated differential privacy.

Result: 10% improvement in accuracy compared to baseline FL, over 50% reduction in client dropouts without convergence degradation, and clinically acceptable privacy-utility trade-offs.

Conclusion: AFFR provides a practical pathway for equitable and trustworthy federated diagnosis of rare conditions in real-world clinical networks.

Abstract: Rare-disease diagnosis remains one of the most pressing challenges in digital health, hindered by extreme data scarcity, privacy concerns, and the limited resources of edge devices. This paper proposes the Adaptive Federated Few-Shot Rare-Disease Diagnosis (AFFR) framework, which integrates three pillars: (i) few-shot federated optimization with meta-learning to generalize from limited patient samples, (ii) energy-aware client scheduling to mitigate device dropouts and ensure balanced participation, and (iii) secure aggregation with calibrated differential privacy to safeguard sensitive model updates. Unlike prior work that addresses these aspects in isolation, AFFR unifies them into a modular pipeline deployable on real-world clinical networks. Experimental evaluation on simulated rare-disease detection datasets demonstrates up to 10% improvement in accuracy compared with baseline FL, while reducing client dropouts by over 50% without degrading convergence. Furthermore, privacy-utility trade-offs remain within clinically acceptable bounds. These findings highlight AFFR as a practical pathway for equitable and trustworthy federated diagnosis of rare conditions.

Method: Revenue- and cluster-aware ensemble of statistical, ML, and deep learning models; Pareto-aware segmentation (individual forecasting for high-revenue items, clustering for long tail); horizon-aware ensembling with business-relevant losses; role-driven analytics with LLM-generated narratives and performance scorecards.

Result: Developed a reproducible workflow covering 90+ countries and ~6,000 parts, providing calibrated forecasts, performance scorecards, trend analysis, bias decomposition, and automated insights generation through LLMs.

Conclusion: The system successfully closes the loop between forecasting, monitoring, and inventory decisions by providing comprehensive analytics that help planners understand not just current accuracy but future trends and actionable levers to improve performance.

Abstract: This paper presents a practical architecture for after-sales demand forecasting and monitoring that unifies a revenue- and cluster-aware ensemble of statistical, machine-learning, and deep-learning models with a role-driven analytics layer for scorecards and trend diagnostics. The framework ingests exogenous signals (installed base, pricing, macro indicators, life cycle, seasonality) and treats COVID-19 as a distinct regime, producing country-part forecasts with calibrated intervals. A Pareto-aware segmentation forecasts high-revenue items individually and pools the long tail via clusters, while horizon-aware ensembling aligns weights with business-relevant losses (e.g., WMAPE). Beyond forecasts, a performance scorecard delivers decision-focused insights: accuracy within tolerance thresholds by revenue share and count, bias decomposition (over- vs under-forecast), geographic and product-family hotspots, and ranked root causes tied to high-impact part-country pairs. A trend module tracks trajectories of MAPE/WMAPE and bias across recent months, flags entities that are improving or deteriorating, detects change points aligned with known regimes, and attributes movements to lifecycle and seasonal factors. LLMs are embedded in the analytics layer to generate role-aware narratives and enforce reporting contracts. They standardize business definitions, automate quality checks and reconciliations, and translate quantitative results into concise, explainable summaries for planners and executives. The system exposes a reproducible workflow – request specification, model execution, database-backed artifacts, and AI-generated narratives – so planners can move from “How accurate are we now?” to “Where is accuracy heading and which levers should we pull?”, closing the loop between forecasting, monitoring, and inventory decisions across more than 90 countries and about 6,000 parts.

Method: Supervised Multi-Dimensional Scaling (SMDS), a model-agnostic method to automatically discover feature manifolds in language models’ latent spaces.

Result: SMDS reveals that different features form various geometric structures (circles, lines, clusters) that reflect concept properties, are stable across models, support reasoning, and dynamically reshape with context changes.

Conclusion: The findings support a model of entity-based reasoning where language models encode and transform structured representations through organized feature manifolds.

Abstract: The linear representation hypothesis states that language models (LMs) encode concepts as directions in their latent space, forming organized, multidimensional manifolds. Prior efforts focus on discovering specific geometries for specific features, and thus lack generalization. We introduce Supervised Multi-Dimensional Scaling (SMDS), a model-agnostic method to automatically discover feature manifolds. We apply SMDS to temporal reasoning as a case study, finding that different features form various geometric structures such as circles, lines, and clusters. SMDS reveals many insights on these structures: they consistently reflect the properties of the concepts they represent; are stable across model families and sizes; actively support reasoning in models; and dynamically reshape in response to context changes. Together, our findings shed light on the functional role of feature manifolds, supporting a model of entity-based reasoning in which LMs encode and transform structured representations.

Method: Evaluated over 70 models varying in computational ingredients on a triplet similarity task using concepts from THINGS database, comparing human and model representations to measure conceptual alignment.

Result: Models with instruction-finetuning and larger attention head dimensionality showed highest human alignment, while multimodal pretraining and parameter size had limited impact. Existing benchmarks (MMLU, MUSR) were insufficient for capturing representational alignment variance.

Conclusion: Instruction-finetuning and attention head dimensionality are key for advancing LLMs toward human conceptual models, and current benchmarks have significant limitations in measuring human-AI alignment.

Abstract: The ability to translate diverse patterns of inputs into structured patterns of behavior has been thought to rest on both humans’ and machines’ ability to learn robust representations of relevant concepts. The rapid advancement of transformer-based large language models (LLMs) has led to a diversity of computational ingredients – architectures, fine tuning methods, and training datasets among others – but it remains unclear which of these ingredients are most crucial for building models that develop human-like representations. Further, most current LLM benchmarks are not suited to measuring representational alignment between humans and models, making benchmark scores unreliable for assessing if current LLMs are making progress towards becoming useful cognitive models. We address these limitations by first evaluating a set of over 70 models that widely vary in their computational ingredients on a triplet similarity task, a method well established in the cognitive sciences for measuring human conceptual representations, using concepts from the THINGS database. Comparing human and model representations, we find that models that undergo instruction-finetuning and which have larger dimensionality of attention heads are among the most human aligned, while multimodal pretraining and parameter size have limited bearing on alignment. Correlations between alignment scores and scores on existing benchmarks reveal that while some benchmarks (e.g., MMLU) are better suited than others (e.g., MUSR) for capturing representational alignment, no existing benchmark is capable of fully accounting for the variance of alignment scores, demonstrating their insufficiency in capturing human-AI alignment. Taken together, our findings help highlight the computational ingredients most essential for advancing LLMs towards models of human conceptual representation and address a key benchmarking gap in LLM evaluation.

Method: Proposed Activation-Deactivation (AD) paradigm that removes effects of occluded features by switching off corresponding model parts. Introduced ConvAD as a drop-in mechanism for CNNs that implements AD without additional training or fine-tuning.

Result: Experimental evaluation across datasets and architectures showed AD explanations achieved up to 62.5% improvement in robustness compared to occlusion-based explanations, with better performance on proxies of robustness, size, and confidence drop-off.

Conclusion: ConvAD provides more robust explanations without requiring domain knowledge, does not change the network’s decision-making process, and can be easily added to any trained CNN.

Abstract: Black-box explainability methods are popular tools for explaining the decisions of image classifiers. A major drawback of these tools is their reliance on mutants obtained by occluding parts of the input, leading to out-of-distribution images. This raises doubts about the quality of the explanations. Moreover, choosing an appropriate occlusion value often requires domain knowledge. In this paper we introduce a novel forward-pass paradigm Activation-Deactivation (AD), which removes the effects of occluded input features from the model’s decision-making by switching off the parts of the model that correspond to the occlusions. We introduce ConvAD, a drop-in mechanism that can be easily added to any trained Convolutional Neural Network (CNN), and which implements the AD paradigm. This leads to more robust explanations without any additional training or fine-tuning. We prove that the ConvAD mechanism does not change the decision-making process of the network. We provide experimental evaluation across several datasets and model architectures. We compare the quality of AD-explanations with explanations achieved using a set of masking values, using the proxies of robustness, size, and confidence drop-off. We observe a consistent improvement in robustness of AD explanations (up to 62.5%) compared to explanations obtained with occlusions, demonstrating that ConvAD extracts more robust explanations without the need for domain knowledge.

Method: Use Curry-Howard correspondence to map informal CoT reasoning steps into formal, typed proof structures, treating faithful reasoning as well-typed programs.

Result: Successfully converting CoT traces into well-typed proofs serves as verifiable certificates of computational faithfulness, moving beyond heuristic interpretability.

Conclusion: Provides a framework to transform narrative explanations into formally verifiable programs, enabling more reliable and trustworthy AI systems through formal verification.

Abstract: While Chain-of-Thought (CoT) prompting enhances the reasoning capabilities of large language models, the faithfulness of the generated rationales remains an open problem for model interpretability. We propose a novel theoretical lens for this problem grounded in the Curry-Howard correspondence, which posits a direct relationship between formal proofs and computer programs. Under this paradigm, a faithful reasoning trace is analogous to a well-typed program, where each intermediate step corresponds to a typed logical inference. We operationalise this analogy, presenting methods to extract and map the informal, natural language steps of CoT into a formal, typed proof structure. Successfully converting a CoT trace into a well-typed proof serves as a strong, verifiable certificate of its computational faithfulness, moving beyond heuristic interpretability towards formal verification. Our framework provides a methodology to transform plausible narrative explanations into formally verifiable programs, offering a path towards building more reliable and trustworthy AI systems.

Method: Safety Instincts Reinforcement Learning (SIRL) transforms models’ internal confidence into self-generated reward signals, teaching models to trust their safety instincts by reinforcing low-entropy refusal behaviors.

Result: SIRL maintains 89%+ Defense Success Rates against 20+ jailbreak methods using only 15,000 unlabeled prompts, surpassing supervised methods while preserving performance on mathematics, coding, and conversation benchmarks.

Conclusion: Effective alignment can emerge from within models, enabling autonomous and robust AI safety mechanisms that scale without extensive human oversight.

Abstract: Ensuring Large Language Model (LLM) safety remains challenging due to the absence of universal standards and reliable content validators, making it difficult to obtain effective training signals. We discover that aligned models already possess robust internal safety beliefs: they consistently produce high-confidence refusals to harmful requests while exhibiting high entropy when generating potentially dangerous content. This entropy gap reveals an untapped signal–models intrinsically “know” when to refuse. We introduce Safety Instincts Reinforcement Learning (SIRL), which transforms this internal confidence into a self-generated reward signal, eliminating dependence on external validators or human annotations. SIRL teaches models to trust their safety instincts by reinforcing low-entropy refusal behaviors. Evaluated on Llama and Qwen models, SIRL maintains 89%+ Defense Success Rates (DSRs) against 20+ jailbreak methods, from static prompts to adaptive attacks. Using only 15,000 unlabeled prompts, SIRL surpasses resource-intensive supervised methods while preserving performance on mathematics, coding, and conversation benchmarks. Our work demonstrates that effective alignment can emerge from within, paving the way for more autonomous and robust AI safety mechanisms that scale without extensive human oversight.

Method: Two-layer framework: Layer 1 uses Constraint Programming for order-line allocation considering machine capacities and due dates; Layer 2 uses Markov Decision Process for worker-line allocation incorporating human factors. Three solution strategies (greedy, MCTS, RL) were implemented and compared.

Result: CP-based scheduling produced compact, feasible production plans with low tardiness; MDP-based worker allocation significantly improved fairness and preference alignment compared to baselines. Domain experts rated both components as effective.

Conclusion: Combining CP with learning-based decision-making provides a robust approach for human-centric production planning, enabling simultaneous optimization of throughput and workforce well-being in industrial manufacturing.

Abstract: This work presents a two-layer, human-centric production planning framework designed to optimize both operational efficiency and workforce fairness in industrial manufacturing. The first layer formulates the Order-Line allocation as a Constraint Programming (CP) problem, generating high-utilization production schedules that respect machine capacities, processing times, and due dates. The second layer models Worker-Line allocation as a Markov Decision Process (MDP), integrating human factors such as worker preference, experience, resilience, and medical constraints into the assignment process. Three solution strategies, greedy allocation, MCTS, and RL, are implemented and compared across multiple evaluation scenarios. The proposed system is validated through 16 test sessions with domain experts from the automotive industry, combining quantitative key performance indicators (KPIs) with expert ratings. Results indicate that the CP-based scheduling approach produces compact, feasible production plans with low tardiness, while the MDP-based worker allocation significantly improves fairness and preference alignment compared to baseline approaches. Domain experts rated both the Order-Line and Worker-Line components as effective and highlighted opportunities to further refine the objective function to penalize excessive earliness and improve continuity in worker assignments. Overall, the findings demonstrate that combining CP with learning-based decision-making provides a robust approach for human-centric production planning. The approach enables simultaneous optimization of throughput and workforce well-being, offering a practical foundation for fair and efficient manufacturing scheduling in industrial settings.

Method: Extends PRISM with a clinician-aligned panel-of-experts architecture using a light-weight router that dispatches episodes to domain specialists (Cardiac-Vascular, Pulmonary, Gastro-Oesophageal, Musculoskeletal, Psychogenic) based on initial token analysis.

Result: Specialists show smooth convergence with low development perplexities across domains, router achieves high routing quality and large compute savings versus consult-all approach under safety-first policy.

Conclusion: The framework provides a practical path to safe, auditable, and low-latency clinical consultation at scale, with validation steps outlined for prospective clinical deployment standards.

Abstract: We present PRISM-Consult, a clinician-aligned panel-of-experts architecture that extends the compact PRISM sequence model into a routed family of domain specialists. Episodes are tokenized as structured clinical events; a light-weight router reads the first few tokens and dispatches to specialist models (Cardiac-Vascular, Pulmonary, Gastro-Oesophageal, Musculoskeletal, Psychogenic). Each specialist inherits PRISM’s small transformer backbone and token template, enabling parameter efficiency and interpretability. On real-world Emergency Department cohorts, specialists exhibit smooth convergence with low development perplexities across domains, while the router achieves high routing quality and large compute savings versus consult-all under a safety-first policy. We detail the data methodology (initial vs. conclusive ICD-9 families), routing thresholds and calibration, and report per-domain results to avoid dominance by common events. The framework provides a practical path to safe, auditable, and low-latency consult at scale, and we outline validation steps-external/temporal replication, asymmetric life-threat thresholds, and multi-label arbitration-to meet prospective clinical deployment standards.

Method: Three-stage progressive methodology: (1) depth upscaling from Pixtral-12B, (2) staged continual pre-training with synthetic data for visual reasoning, and (3) high-quality text-only supervised fine-tuning with explicit reasoning traces.

Result: Achieves Artificial Analysis Intelligence Index score of 52 (matching DeepSeek-R1-0528) and performs within 5 points of Gemini-2.5-Flash and Claude Sonnet-3.7 across ten image benchmarks, all without reinforcement learning.

Conclusion: Thoughtful mid-training design can close substantial capability gaps without massive scale, making frontier multimodal reasoning accessible to organizations with limited infrastructure.

Abstract: We present Apriel-1.5-15B-Thinker, a 15-billion parameter open-weights multimodal reasoning model that achieves frontier-level performance through training design rather than sheer scale. Starting from Pixtral-12B, we apply a progressive three-stage methodology: (1) depth upscaling to expand reasoning capacity without pretraining from scratch, (2) staged continual pre-training that first develops foundational text and vision understanding, then enhances visual reasoning through targeted synthetic data generation addressing spatial structure, compositional understanding, and fine-grained perception, and (3) high-quality text-only supervised fine-tuning on curated instruction-response pairs with explicit reasoning traces spanning mathematics, coding, science, and tool use. Notably, our model achieves competitive results without reinforcement learning or preference optimization, isolating the contribution of our data-centric continual pre-training approach. On the Artificial Analysis Intelligence Index, Apriel-1.5-15B-Thinker attains a score of 52, matching DeepSeek-R1-0528 despite requiring significantly fewer computational resources. Across ten image benchmarks, its performance is on average within five points of Gemini-2.5-Flash and Claude Sonnet-3.7, a key achievement for a model operating within single-GPU deployment constraints. Our results demonstrate that thoughtful mid-training 2 design can close substantial capability gaps without massive scale, making frontier-level multimodal reasoning accessible to organizations with limited infrastructure. We release the model checkpoint, all training recipes, and evaluation protocols under the MIT license to to advance open-source research.

Method: Bridge treats batched LLM hidden states as holistic tensors rather than independent slices, allowing responses to influence each other during parallel generation with only 2.8%-5.1% new parameters.

Result: Bridge improves relative mean accuracy gains from reinforcement learning by up to 50%, boosts consistency of correct responses, and scales to any generation width with better performance than independent generation.

Conclusion: Bridge unlocks a more general mode of parallel scaling that effectively leverages information between sequences, compatible with any post-generation aggregation technique.

Abstract: Parallel LLM inference scaling involves sampling a set of $N>1$ responses for a single input prompt. However, these $N$ parallel responses tend to be generated independently from each other, partitioning compute resources and leaving potentially useful information in one generation untapped by others. This is in contrast to response length scaling where past computation is used in all future steps. For higher quality responses and response sets, we propose Bridge to generate interdependent responses in parallel by rethinking batched LLM hidden states as holistic tensors rather than independent slices. With only a small amount (2.8%-5.1%) of new parameters, Bridge improves the relative mean accuracy gains from reinforcement learning with verifiable rewards by up to 50% and boosts consistency of correct responses. Trained once, Bridge scales to any generation width, all with greater performance than independent generations, unlocking a more general mode of parallel scaling that effectively leverages information between sequences, compatible with any post-generation aggregation technique.

Method: Uses an LLM to incrementally extract necessary information from short natural language inputs, then creates complete PDDL descriptions of both domain and problem, which are solved by a classical planner.

Result: Outperforms directly generating files with LLM+validator combination across seven planning domains (five novel domains not in LLM training data).

Conclusion: NL2Plan is a powerful tool for assistive PDDL modeling and represents progress toward solving natural language planning tasks with interpretability and guarantees.

Abstract: Classical planners are powerful systems, but modeling tasks in input formats such as PDDL is tedious and error-prone. In contrast, planning with Large Language Models (LLMs) allows for almost any input text, but offers no guarantees on plan quality or even soundness. In an attempt to merge the best of these two approaches, some work has begun to use LLMs to automate parts of the PDDL creation process. However, these methods still require various degrees of expert input or domain-specific adaptations. We present NL2Plan, the first fully automatic system for generating complete PDDL tasks from minimal natural language descriptions. NL2Plan uses an LLM to incrementally extract the necessary information from the short text input before creating a complete PDDL description of both the domain and the problem which is finally solved by a classical planner. We evaluate NL2Plan on seven planning domains, five of which are novel and thus not in the LLM training data, and find that NL2Plan outperforms directly generating the files with an LLM+validator combination. As such, NL2Plan is a powerful tool for assistive PDDL modeling and a step towards solving natural language planning task with interpretability and guarantees.

Method: The paper conducts a comprehensive review of various XAI approaches integrated with machine learning and deep learning models for breast cancer detection. It examines different breast cancer datasets including mammograms and ultrasounds processed with AI.

Result: The review demonstrates that XAI techniques can lead to more accurate diagnoses and personalized treatment plans. It also identifies challenges in implementation and the need for standardized evaluation metrics in clinical settings.

Conclusion: XAI has significant potential to bridge the gap between complex AI models and practical healthcare applications, fostering trust among medical professionals and ultimately improving patient outcomes in breast cancer diagnosis and treatment.

Abstract: Breast cancer (BC) stands as one of the most common malignancies affecting women worldwide, necessitating advancements in diagnostic methodologies for better clinical outcomes. This article provides a comprehensive exploration of the application of Explainable Artificial Intelligence (XAI) techniques in the detection and diagnosis of breast cancer. As Artificial Intelligence (AI) technologies continue to permeate the healthcare sector, particularly in oncology, the need for transparent and interpretable models becomes imperative to enhance clinical decision-making and patient care. This review discusses the integration of various XAI approaches, such as SHAP, LIME, Grad-CAM, and others, with machine learning and deep learning models utilized in breast cancer detection and classification. By investigating the modalities of breast cancer datasets, including mammograms, ultrasounds and their processing with AI, the paper highlights how XAI can lead to more accurate diagnoses and personalized treatment plans. It also examines the challenges in implementing these techniques and the importance of developing standardized metrics for evaluating XAI’s effectiveness in clinical settings. Through detailed analysis and discussion, this article aims to highlight the potential of XAI in bridging the gap between complex AI models and practical healthcare applications, thereby fostering trust and understanding among medical professionals and improving patient outcomes.

Method: Used fairness probing to analyze outputs from three leading open-source LLMs, examining travel recommendations for ethnic and gender biases.

Result: Test accuracy for ethnicity and gender classifiers exceeded random chance, revealed stereotype bias in recommendations, and found more hallucinations for minority groups.

Conclusion: LLMs exhibit ethnic and gender bias as travel planning assistants, highlighting need for bias mitigation strategies to improve inclusivity and reliability.

Abstract: As large language models (LLMs) become increasingly integral to the hospitality and tourism industry, concerns about their fairness in serving diverse identity groups persist. Grounded in social identity theory and sociotechnical systems theory, this study examines ethnic and gender biases in travel recommendations generated by LLMs. Using fairness probing, we analyze outputs from three leading open-source LLMs. The results show that test accuracy for both ethnicity and gender classifiers exceed random chance. Analysis of the most influential features reveals the presence of stereotype bias in LLM-generated recommendations. We also found hallucinations among these features, occurring more frequently in recommendations for minority groups. These findings indicate that LLMs exhibit ethnic and gender bias when functioning as travel planning assistants. This study underscores the need for bias mitigation strategies to improve the inclusivity and reliability of generative AI-driven travel planning assistance.

Method: Developed PETAH (Parameter Efficient Task Adaptation for Hybrid Transformers) and combined it with pruning to create storage-friendly multi-tasking models.

Result: PETAH-adapted hybrid models outperform established ViT task-adaptation techniques on classification and other vision tasks, requiring fewer parameters and being more efficient on mobile hardware.

Conclusion: PETAH enables efficient task adaptation for hybrid transformers, achieving better performance with fewer parameters than ViT adaptation methods.

Abstract: Following their success in natural language processing (NLP), there has been a shift towards transformer models in computer vision. While transformers perform well and offer promising multi-tasking performance, due to their high compute requirements, many resource-constrained applications still rely on convolutional or hybrid models that combine the benefits of convolution and attention layers and achieve the best results in the sub 100M parameter range. Simultaneously, task adaptation techniques that allow for the use of one shared transformer backbone for multiple downstream tasks, resulting in great storage savings at negligible cost in performance, have not yet been adopted for hybrid transformers. In this work, we investigate how to achieve the best task-adaptation performance and introduce PETAH: Parameter Efficient Task Adaptation for Hybrid Transformers. We further combine PETAH adaptation with pruning to achieve highly performant and storage friendly models for multi-tasking. In our extensive evaluation on classification and other vision tasks, we demonstrate that our PETAH-adapted hybrid models outperform established task-adaptation techniques for ViTs while requiring fewer parameters and being more efficient on mobile hardware.

Method: Joint Conditional Diffusion Model-based Inverse Problem Solver (JCDI) uses diffusion model stochasticity to reveal underlying distributions and joint conditioning on multiple observations to narrow posterior distributions of non-identifiable parameters.

Result: For composite load model parameterization in power systems, JCDI achieved 58.6% reduction in parameter estimation error compared to single-condition models, with RMSE below 4×10^(-3) for dynamic responses under electrical faults, outperforming deep reinforcement learning and supervised learning approaches.

Conclusion: JCDI provides a universal data-driven framework for parameter estimation that effectively mitigates the non-uniqueness challenge across scientific domains.

Abstract: Parameter estimation, which represents a classical inverse problem, is often ill-posed as different parameter combinations can yield identical outputs. This non-uniqueness poses a critical barrier to accurate and unique identification. This work introduces a novel parameter estimation framework to address such limits: the Joint Conditional Diffusion Model-based Inverse Problem Solver (JCDI). By leveraging the stochasticity of diffusion models, JCDI produces possible solutions revealing underlying distributions. Joint conditioning on multiple observations further narrows the posterior distributions of non-identifiable parameters. For the challenging task in dynamic power systems: composite load model parameterization, JCDI achieves a 58.6% reduction in parameter estimation error compared to the single-condition model. It also accurately replicates system’s dynamic responses under various electrical faults, with root mean square errors below 4*10^(-3), outperforming existing deep-reinforcement-learning and supervised learning approaches. Given its data-driven nature, JCDI provides a universal framework for parameter estimation while effectively mitigating the non-uniqueness challenge across scientific domains.

Method: Created a dataset of 40,945 text-image pairs from diverse news outlets using LLM-as-annotator pipeline with hierarchical majority voting and human validation. Evaluated SLMs, LLMs, and VLMs on closed-ended classification and open-ended reasoning tasks.

Result: Image incorporation improved bias detection accuracy by 3-5%. LLMs/VLMs outperformed SLMs in capturing subtle framing. Parameter-efficient methods recovered 97-99% of full fine-tuning performance with <5% trainable parameters. Reasoning accuracy ranged 52-79% with faithfulness 68-89%.

Conclusion: ViLBias provides a scalable benchmark and strong baselines for multimodal bias detection, demonstrating the importance of multimodal reasoning and the effectiveness of parameter-efficient tuning methods.

Abstract: Detecting bias in multimodal news requires models that reason over text–image pairs, not just classify text. In response, we present ViLBias, a VQA-style benchmark and framework for detecting and reasoning about bias in multimodal news. The dataset comprises 40,945 text–image pairs from diverse outlets, each annotated with a bias label and concise rationale using a two-stage LLM-as-annotator pipeline with hierarchical majority voting and human-in-the-loop validation. We evaluate Small Language Models (SLMs), Large Language Models (LLMs), and Vision–Language Models (VLMs) across closed-ended classification and open-ended reasoning (oVQA), and compare parameter-efficient tuning strategies. Results show that incorporating images alongside text improves detection accuracy by 3–5%, and that LLMs/VLMs better capture subtle framing and text–image inconsistencies than SLMs. Parameter-efficient methods (LoRA/QLoRA/Adapters) recover 97–99% of full fine-tuning performance with $<5%$ trainable parameters. For oVQA, reasoning accuracy spans 52–79% and faithfulness 68–89%, both improved by instruction tuning; closed accuracy correlates strongly with reasoning ($r = 0.91$). ViLBias offers a scalable benchmark and strong baselines for multimodal bias detection and rationale quality.

Method: Created MathConstruct benchmark with 121 challenging problems from math competitions targeting constructive proofs, with automated verifiers that enable solution verification and problem variation generation.

Result: State-of-the-art LLMs solve only 60% of MathConstruct problems, demonstrating its complexity and challenging nature.

Conclusion: MathConstruct addresses limitations of existing benchmarks and serves as an important tool for evaluating LLM capabilities in complex mathematical reasoning, particularly constructive proofs.

Abstract: While Large Language Models (LLMs) demonstrate impressive performance in mathematics, existing math benchmarks come with significant limitations. Many focus on problems with fixed ground-truth answers, and are often saturated due to problem simplicity or the viability of guessing or memorization. Crucially, they capture only a narrow subset of relevant math problems. To address this research gap, we introduce MathConstruct, a new benchmark of 121 challenging problems sourced from various math competitions, which targets constructive proofs, a widely encountered problem type requiring the construction of mathematical objects with specific properties. These proofs are particularly suitable for LLM evaluation, as solution correctness can be easily verified. Our automated verifiers also enable MathConstruct to generate problem variations, used to evaluate robustness. State-of-the-art LLMs solve only 60% of MathConstruct problems, highlighting its complexity and importance for LLM evaluation.

Method: Fine-tunes multi-modal large language models (MLLMs) as vision-language-action policies using online reinforcement learning with LLM-generated rewards, eliminating need for human demonstrations or manual reward engineering.

Result: RL-finetuned MLLM outperforms all baselines including GPT-4o and supervised fine-tuned MLLMs by 10.4-16.5%, generalizing well to novel scenes and tasks.

Conclusion: First demonstration of adapting MLLMs as VLA agents that can both act and ask for help using LLM-generated rewards with online RL, achieving significant performance improvements.

Abstract: Embodied agents operating in household environments must interpret ambiguous and under-specified human instructions. A capable household robot should recognize ambiguity and ask relevant clarification questions to infer the user intent accurately, leading to more effective task execution. To study this problem, we introduce the Ask-to-Act task, where an embodied agent is tasked with a single or multi-object rearrangement task using an under-specified instruction in a home environment. The agent must strategically ask minimal, yet relevant, clarification questions to resolve ambiguity while navigating under partial observability. To address this challenge, we propose a novel approach that fine-tunes multi-modal large language models (MLLMs) as vision-language-action (VLA) policies using online reinforcement learning (RL) with LLM-generated rewards. Our method eliminates the need for large-scale human demonstrations or manually engineered rewards for training such agents. We benchmark against strong zero-shot baselines including GPT-4o as well as supervised fine-tuned MLLMs on our task. Our results show that our RL-finetuned MLLM outperforms all baselines by a significant margin (10.4-16.5%), generalizing well to novel scenes and tasks. To the best of our knowledge, this is the first demonstration of adapting MLLMs as VLA agents that can act and ask for help using LLM-generated rewards with online RL.

Method: A 3-component framework: (1) generates natural language statements of code to verify, (2) LLM generates formal proofs, (3) heuristics module builds final proof. Uses 2-stage fine-tuning: SFT for syntactically correct Isabelle code, then RL training for proofs verified by theorem prover.

Result: Validated on miniF2F-test benchmark using Isabelle proof assistant. Applied to verify AWS S3 bucket access policy code correctness. Curated dataset based on FVEL_ER for future training.

Conclusion: The framework enables automated formal verification of code through LLM-generated proofs, addressing the challenge of generalized theorem proving and providing a practical application for verifying real-world systems like AWS policies.

Abstract: Formally verifying properties of software code has been a highly desirable task, especially with the emergence of LLM-generated code. In the same vein, they provide an interesting avenue for the exploration of formal verification and mechanistic interpretability. Since the introduction of code-specific models, despite their successes in generating code in Lean4 and Isabelle, the task of generalized theorem proving still remains far from being fully solved and will be a benchmark for reasoning capability in LLMs. In this work, we introduce a framework that generates whole proofs in a formal language to be used within systems that utilize the power of built-in tactics and off-the-shelf automated theorem provers. Our framework includes 3 components: generating natural language statements of the code to be verified, an LLM that generates formal proofs for the given statement, and a module employing heuristics for building the final proof. To train the LLM, we employ a 2-stage fine-tuning process, where we first use SFT-based training to enable the model to generate syntactically correct Isabelle code and then RL-based training that encourages the model to generate proofs verified by a theorem prover. We validate our framework using the miniF2F-test benchmark and the Isabelle proof assistant and design a use case to verify the correctness of the AWS S3 bucket access policy code. We also curate a dataset based on the FVEL\textsubscript{\textnormal{ER}} dataset for future training tasks.

Method: Introduces R&D-Agent framework that defines MLE workflow into two phases and six components, turning agent design from ad-hoc craftsmanship into principled, testable process. Designed efficient agents inspired by human experts within this framework.

Result: Achieved state-of-the-art performance on MLE-Bench with 35.1% any medal rate, ranking as top-performing machine learning engineering agent.

Conclusion: R&D-Agent framework demonstrates ability to speed up innovation and improve accuracy across wide range of data science applications, and has been open-sourced on GitHub.

Abstract: Recent advances in AI and ML have transformed data science, yet increasing complexity and expertise requirements continue to hinder progress. Although crowd-sourcing platforms alleviate some challenges, high-level machine learning engineering (MLE) tasks remain labor-intensive and iterative. We introduce R&D-Agent, a comprehensive, decoupled, and extensible framework that formalizes the MLE process. R&D-Agent defines the MLE workflow into two phases and six components, turning agent design for MLE from ad-hoc craftsmanship into a principled, testable process. Although several existing agents report promising gains on their chosen components, they can mostly be summarized as a partial optimization from our framework’s simple baseline. Inspired by human experts, we designed efficient and effective agents within this framework that achieve state-of-the-art performance. Evaluated on MLE-Bench, the agent built on R&D-Agent ranks as the top-performing machine learning engineering agent, achieving 35.1% any medal rate, demonstrating the ability of the framework to speed up innovation and improve accuracy across a wide range of data science applications. We have open-sourced R&D-Agent on GitHub: https://github.com/microsoft/RD-Agent.

Method: The authors introduce AgentMisalignment benchmark suite to evaluate LLM agents’ misalignment propensity in realistic scenarios, covering behaviors like avoiding oversight, resisting shutdown, sandbagging, and power-seeking. They test frontier models and systematically vary agent personalities through different system prompts.

Result: More capable agents tend to exhibit higher misalignment on average. Persona characteristics can strongly and unpredictably influence misalignment, sometimes more than the choice of model itself.

Conclusion: Current alignment methods have limitations for autonomous LLM agents, and there’s a need to rethink misalignment in realistic deployment settings.

Abstract: As Large Language Model (LLM) agents become more widespread, associated misalignment risks increase. While prior research has studied agents’ ability to produce harmful outputs or follow malicious instructions, it remains unclear how likely agents are to spontaneously pursue unintended goals in realistic deployments. In this work, we approach misalignment as a conflict between the internal goals pursued by the model and the goals intended by its deployer. We introduce a misalignment propensity benchmark, \textsc{AgentMisalignment}, a benchmark suite designed to evaluate the propensity of LLM agents to misalign in realistic scenarios. Evaluations cover behaviours such as avoiding oversight, resisting shutdown, sandbagging, and power-seeking. Testing frontier models, we find that more capable agents tend to exhibit higher misalignment on average. We also systematically vary agent personalities through different system prompts and observe that persona characteristics can strongly and unpredictably influence misalignment, sometimes more than the choice of model itself. Our results reveal the limitations of current alignment methods for autonomous LLM agents and underscore the need to rethink misalignment in realistic deployment settings.

Method: Uses a generation framework for scalable trajectories, introduces embodied QA tasks with scattered clues, explores architectural adaptations like Goal-State-Action modeling, context extension, and Context Parallelism for LLM-based agents.

Result: Experimental results highlight the challenges of the benchmark and provide insights into training strategies and model behaviors under long-horizon conditions.

Conclusion: The work establishes a foundation for next-generation embodied AI systems capable of robust, long-term reasoning and planning.

Abstract: We introduce $\infty$-THOR, a new framework for long-horizon embodied tasks that advances long-context understanding in embodied AI. $\infty$-THOR provides: (1) a generation framework for synthesizing scalable, reproducible, and unlimited long-horizon trajectories; (2) a novel embodied QA task, Needle(s) in the Embodied Haystack, where multiple scattered clues across extended trajectories test agents’ long-context reasoning ability; and (3) a long-horizon dataset and benchmark suite featuring complex tasks that span hundreds of environment steps, each paired with ground-truth action sequences. To enable this capability, we explore architectural adaptations, including interleaved Goal-State-Action modeling, context extension techniques, and Context Parallelism, to equip LLM-based agents for extreme long-context reasoning and interaction. Experimental results and analyses highlight the challenges posed by our benchmark and provide insights into training strategies and model behaviors under long-horizon conditions. Our work provides a foundation for the next generation of embodied AI systems capable of robust, long-term reasoning and planning.

Method: Uses unified location encoder to map geographically disjoint locations into shared semantic space, and Spatially-Aware Mixture-of-Experts Transformer with specialized experts for diverse mobility patterns.

Result: Achieves new state-of-the-art across downstream tasks with up to 35% average performance gains and strong generalization to unseen cities.

Conclusion: Provides empirical evidence of scaling ability in human mobility, validating a path toward more capable foundation models in this domain.

Abstract: The success of foundation models in language has inspired a new wave of general-purpose models for human mobility. However, existing approaches struggle to scale effectively due to two fundamental limitations: a failure to use meaningful basic units to represent movement, and an inability to capture the vast diversity of patterns found in large-scale data. In this work, we develop MoveGPT, a large-scale foundation model specifically architected to overcome these barriers. MoveGPT is built upon two key innovations: (1) a unified location encoder that maps geographically disjoint locations into a shared semantic space, enabling pre-training on a global scale; and (2) a Spatially-Aware Mixture-of-Experts Transformer that develops specialized experts to efficiently capture diverse mobility patterns. Pre-trained on billion-scale datasets, MoveGPT establishes a new state-of-the-art across a wide range of downstream tasks, achieving performance gains of up to 35% on average. It also demonstrates strong generalization capabilities to unseen cities. Crucially, our work provides empirical evidence of scaling ability in human mobility, validating a clear path toward building increasingly capable foundation models in this domain.

Method: ConciseHint continuously injects learnable hints (manually designed or learned from concise data) during reasoning generation, with adaptive hint intensity based on query complexity to avoid undermining performance.

Result: Experiments on DeepSeek-R1 and Qwen-3 series show the method effectively produces concise reasoning while maintaining performance, and can be integrated with existing methods to further improve efficiency.

Conclusion: ConciseHint successfully addresses the verbosity problem in large reasoning models by intervening during generation with adaptive hints, achieving concise reasoning without performance degradation.

Abstract: Recent advancements in large reasoning models (LRMs) like DeepSeek-R1 and OpenAI o1 series have achieved notable performance enhancements on complex reasoning tasks by scaling up the generation length by Chain-of-Thought (CoT). However, a critical issue is their tendency to produce excessively verbose reasoning processes, leading to the inefficiency problem. Existing literature on improving efficiency mainly adheres to the before-reasoning paradigms such as prompting and reasoning or fine-tuning and reasoning, but ignores the promising direction of directly encouraging the model to speak concisely by intervening during the generation of reasoning. In order to fill the blank, we propose a framework dubbed ConciseHint, which continuously encourages the reasoning model to speak concisely by injecting learnable hints (manually designed or learned on concise data) during the generation of the reasoning. Besides, ConciseHint is adaptive to the complexity of the query by adaptively adjusting the hint intensity, which ensures it will not undermine model performance. Experiments on the state-of-the-art LRMs, including DeepSeek-R1 and Qwen-3 series, demonstrate that our method can effectively produce concise reasoning while maintaining the performance well. Moreover, we show that ConciseHint is flexible and can be seamlessly integrated with existing methods to further push the upper bound of the efficiency.

Method: Extract reasoning graphs by clustering hidden-state representations at each reasoning step, then analyze three graph properties: cyclicity, diameter, and small-world index across multiple mathematical tasks (GSM8K, MATH500, AIME 2024).

Result: Distilled reasoning models show significantly more recurrent cycles (~5 per sample), larger graph diameters, and pronounced small-world characteristics (~6x) compared to base models. These structural advantages increase with task difficulty and model capacity, correlating positively with accuracy.

Conclusion: The study bridges theoretical insights about reasoning graph structures with practical dataset design guidelines, advancing both interpretability and efficacy of large reasoning models through systematic analysis of graph-theoretic properties.

Abstract: Recent large-scale reasoning models have achieved state-of-the-art performance on challenging mathematical benchmarks, yet the internal mechanisms underlying their success remain poorly understood. In this work, we introduce the notion of a reasoning graph, extracted by clustering hidden-state representations at each reasoning step, and systematically analyze three key graph-theoretic properties: cyclicity, diameter, and small-world index, across multiple tasks (GSM8K, MATH500, AIME 2024). Our findings reveal that distilled reasoning models (e.g., DeepSeek-R1-Distill-Qwen-32B) exhibit significantly more recurrent cycles (about 5 per sample), substantially larger graph diameters, and pronounced small-world characteristics (about 6x) compared to their base counterparts. Notably, these structural advantages grow with task difficulty and model capacity, with cycle detection peaking at the 14B scale and exploration diameter maximized in the 32B variant, correlating positively with accuracy. Furthermore, we show that supervised fine-tuning on an improved dataset systematically expands reasoning graph diameters in tandem with performance gains, offering concrete guidelines for dataset design aimed at boosting reasoning capabilities. By bridging theoretical insights into reasoning graph structures with practical recommendations for data construction, our work advances both the interpretability and the efficacy of large reasoning models.

Method: Developed a novel assessment method evaluating five dimensions of moral competence: identifying morally relevant features, weighting their importance, assigning moral reasons, synthesizing coherent judgments, and recognizing information gaps. Conducted two experiments comparing six LLMs against non-expert humans and professional philosophers using both standard ethical vignettes and novel scenarios with embedded moral features.

Result: In standard ethical vignettes, LLMs generally outperformed non-expert humans across multiple moral reasoning dimensions. However, in novel scenarios with moral features embedded among irrelevant details, several LLMs performed significantly worse than humans, revealing a striking reversal of performance.

Conclusion: Current evaluations may substantially overestimate LLMs’ moral reasoning capabilities by eliminating the task of discerning moral relevance from noisy information, which is a prerequisite for genuine moral skill. The work provides a more nuanced assessment framework and highlights directions for improving AI moral competence.

Abstract: Moral competence is the ability to act in accordance with moral principles. As large language models (LLMs) are increasingly deployed in situations demanding moral competence, there is increasing interest in evaluating this ability empirically. We review existing literature and identify three significant shortcoming: (i) Over-reliance on prepackaged moral scenarios with explicitly highlighted moral features; (ii) Focus on verdict prediction rather than moral reasoning; and (iii) Inadequate testing of models’ (in)ability to recognize when additional information is needed. Grounded in philosophical research on moral skill, we then introduce a novel method for assessing moral competence in LLMs. Our approach moves beyond simple verdict comparisons to evaluate five dimensions of moral competence: identifying morally relevant features, weighting their importance, assigning moral reasons to these features, synthesizing coherent moral judgments, and recognizing information gaps. We conduct two experiments comparing six leading LLMs against non-expert humans and professional philosophers. In our first experiment using ethical vignettes standard to existing work, LLMs generally outperformed non-expert humans across multiple dimensions of moral reasoning. However, our second experiment, featuring novel scenarios designed to test moral sensitivity by embedding relevant features among irrelevant details, revealed a striking reversal: several LLMs performed significantly worse than humans. Our findings suggest that current evaluations may substantially overestimate LLMs’ moral reasoning capabilities by eliminating the task of discerning moral relevance from noisy information, which we take to be a prerequisite for genuine moral skill. This work provides a more nuanced framework for assessing AI moral competence and highlights important directions for improving moral competence in advanced AI systems.

Method: Reformulating machine learning models through category theory by defining Lawvere-enriched categories for parameters and data, with an adjoint pair of functors between them, creating the Gauss-Markov Adjunction framework.

Result: The categorical framework clarifies the structural interplay between residuals and parameters, shows how the ordinary least squares estimator and minimum residual are related via preservation of limits by the right adjoint functor, and enables explicit description of dual information flow.

Conclusion: The formulation serves as extended denotational semantics for supervised learning and proposes applying semantic perspectives from theoretical computer science as a formal foundation for explicability in AI.

Abstract: Enhancing the intelligibility and interpretability of machine learning is a crucial task in responding to the demand for Explicability as an AI principle, and in promoting the better social implementation of AI. The aim of our research is to contribute to this improvement by reformulating machine learning models through the lens of category theory, thereby developing a semantic framework for structuring and understanding AI systems. Our categorical modeling in this paper clarifies and formalizes the structural interplay between residuals and parameters in supervised learning. The present paper focuses on the multiple linear regression model, which represents the most basic form of supervised learning. By defining two Lawvere-enriched categories corresponding to parameters and data, along with an adjoint pair of functors between them, we introduce our categorical formulation of supervised learning. We show that the essential structure of this framework is captured by what we call the Gauss-Markov Adjunction. Within this setting, the dual flow of information can be explicitly described as a correspondence between variations in parameters and residuals. The ordinary least squares estimator for the parameters and the minimum residual are related via the preservation of limits by the right adjoint functor. Furthermore, we position this formulation as an instance of extended denotational semantics for supervised learning, and propose applying a semantic perspective developed in theoretical computer science as a formal foundation for Explicability in AI.

Method: Decomposes navigation into atomic skills handled by specialized agents, creates synthetic dataset for skill training, and uses training-free VLM-based router for dynamic agent selection.

Result: Achieves competitive results on standard benchmarks and state-of-the-art generalization on GSA-R2R with novel instruction styles and unseen environments.

Conclusion: SkillNav’s modular skill-based approach with synthetic data and dynamic routing effectively addresses generalization challenges in VLN.

Abstract: Vision-and-Language Navigation (VLN) poses significant challenges for agents to interpret natural language instructions and navigate complex 3D environments. While recent progress has been driven by large-scale pre-training and data augmentation, current methods still struggle to generalize to unseen scenarios, particularly when complex spatial and temporal reasoning is required. In this work, we propose SkillNav, a modular framework that introduces structured, skill-based reasoning into Transformer-based VLN agents. Our method decomposes navigation into a set of interpretable atomic skills (e.g., Vertical Movement, Area and Region Identification, Stop and Pause), each handled by a specialized agent. To support targeted skill training without manual data annotation, we construct a synthetic dataset pipeline that generates diverse, linguistically natural, skill-specific instruction-trajectory pairs. We then introduce a novel training-free Vision-Language Model (VLM)-based router, which dynamically selects the most suitable agent at each time step by aligning sub-goals with visual observations and historical actions. SkillNav obtains competitive results on commonly used benchmarks and establishes state-of-the-art generalization to the GSA-R2R, a benchmark with novel instruction styles and unseen environments.

Method: Introduces VISOTHELLO, a multi-modal model trained jointly on move sequences and board images, evaluated on Othello rule understanding task with robustness testing under semantically irrelevant perturbations and cross-modal alignment analysis.

Result: Multi-modal training improves performance and robustness, and promotes convergence toward shared internal representations across different model architectures.

Conclusion: Multi-modal training offers significant advantages over text-only approaches for world understanding, enhancing both performance and robustness while fostering consistent cross-modal representations.

Abstract: Language models are often said to face a symbol grounding problem. While some have argued the problem can be solved without resort to other modalities, many have speculated that grounded learning is more efficient. We explore this question in Othello, a simplified, rule-based world that offers a controlled and interpretable testbed for studying world understanding. Building on prior work, we introduce VISOTHELLO, a multi-modal model trained jointly on move sequences and board images. Using the Othello rule understanding task, we examine whether multi-modal learning provides advantages over text-only approaches. We further evaluate robustness under semantically irrelevant perturbations and analyze the consistency of cross-modal alignment. Our results suggest that multi-modal training not only improves performance and robustness but also promotes convergence toward shared internal representations across different model architectures.

Method: Multi-Agent Consensus Alignment (MACA) uses reinforcement learning to post-train models, favoring reasoning trajectories aligned with internal consensus from multi-agent debates where agents ground reasoning in peer arguments rather than just aggregating independent attempts.

Result: Substantial improvements across self-consistency (+27.6% on GSM8K), single-agent reasoning (+23.7% on MATH), sampling-based inference (+22.4% Pass@20 on MATH), and multi-agent ensemble decision-making (+42.7% on MathQA), with strong generalization to unseen benchmarks.

Conclusion: MACA demonstrates robust self-alignment that more reliably unlocks the latent reasoning potential of language models through deliberative exchanges and consensus-based learning.

Abstract: Language Models (LMs) are inconsistent reasoners, often generating contradictory responses to identical prompts. While inference-time methods can mitigate these inconsistencies, they fail to address the core problem: LMs struggle to reliably select reasoning pathways leading to consistent outcomes under exploratory sampling. To address this, we formalize self-consistency as an intrinsic property of well-aligned reasoning models and introduce Multi-Agent Consensus Alignment (MACA), a reinforcement learning framework that post-trains models to favor reasoning trajectories aligned with their internal consensus using majority/minority outcomes from multi-agent debate. These trajectories emerge from deliberative exchanges where agents ground reasoning in peer arguments, not just aggregation of independent attempts, creating richer consensus signals than single-round majority voting. MACA enables agents to teach themselves to be more decisive and concise, and better leverage peer insights in multi-agent settings without external supervision, driving substantial improvements across self-consistency (+27.6% on GSM8K), single-agent reasoning (+23.7% on MATH), sampling-based inference (+22.4% Pass@20 on MATH), and multi-agent ensemble decision-making (+42.7% on MathQA). These findings, coupled with strong generalization to unseen benchmarks (+16.3% on GPQA, +11.6% on CommonsenseQA), demonstrate robust self-alignment that more reliably unlocks latent reasoning potential of language models.

Method: Created NUMINA benchmark using NUMINA-Flow automated annotation pipeline with LLM rewriting and rule-based self-verification, featuring multi-scale annotations and various question-answer pairs.

Result: Evaluation shows current LLMs struggle with multimodal numerical reasoning, particularly in precise computations like distance and volume estimation.

Conclusion: Highlights the need for further advancements in 3D models to handle complex numerical reasoning tasks in spatial environments.

Abstract: Recent advancements in 2D multimodal large language models (MLLMs) have significantly improved performance in vision-language tasks. However, extending these capabilities to 3D environments remains a distinct challenge due to the complexity of spatial reasoning. Nevertheless, existing 3D benchmarks often lack fine-grained numerical reasoning task annotations, limiting MLLMs’ ability to perform precise spatial measurements and complex numerical reasoning. To address this gap, we introduce NUMINA, the first Natural Understanding benchmark for Multi-dimensional Intelligence and Numerical reasoning Abilities to enhance multimodal indoor perceptual understanding. NUMINA features multi-scale annotations and various question-answer pairs, generated using NUMINA-Flow, an automated annotation pipeline that integrates LLM rewriting and rule-based self-verification. We evaluate the performance of various state-of-the-art LLMs on NUMINA following the Chat-Scene framework, demonstrating that current LLMs struggle with multimodal numerical reasoning, particularly in performing precise computations such as distance and volume estimation, highlighting the need for further advancements in 3D models. The dataset and source codes can be obtained from https://github.com/fengshun124/NUMINA.

Method: Uses a multi-agent framework with comprehensive end-to-end automation, composable service architecture via Model Context Protocol, and hierarchical multi-index RAG for high-fidelity configuration generation.

Result: Achieved 88.2% success rate on 110 simulation tasks, significantly outperforming MetaOpenFOAM (55.5%).

Conclusion: Foam-Agent dramatically lowers the expertise barrier for CFD and demonstrates how specialized multi-agent systems can democratize complex scientific computing.

Abstract: Computational Fluid Dynamics (CFD) is an essential simulation tool in engineering, yet its steep learning curve and complex manual setup create significant barriers. To address these challenges, we introduce Foam-Agent, a multi-agent framework that automates the entire end-to-end OpenFOAM workflow from a single natural language prompt. Our key innovations address critical gaps in existing systems: 1. An Comprehensive End-to-End Simulation Automation: Foam-Agent is the first system to manage the full simulation pipeline, including advanced pre-processing with a versatile Meshing Agent capable of handling external mesh files and generating new geometries via Gmsh, automatic generation of HPC submission scripts, and post-simulation visualization via ParaView. 2. Composable Service Architecture: Going beyond a monolithic agent, the framework uses Model Context Protocol (MCP) to expose its core functions as discrete, callable tools. This allows for flexible integration and use by other agentic systems, such as Claude-code, for more exploratory workflows. 3. High-Fidelity Configuration Generation: We achieve superior accuracy through a Hierarchical Multi-Index RAG for precise context retrieval and a dependency-aware generation process that ensures configuration consistency. Evaluated on a benchmark of 110 simulation tasks, Foam-Agent achieves an 88.2% success rate with Claude 3.5 Sonnet, significantly outperforming existing frameworks (55.5% for MetaOpenFOAM). Foam-Agent dramatically lowers the expertise barrier for CFD, demonstrating how specialized multi-agent systems can democratize complex scientific computing. The code is public at https://github.com/csml-rpi/Foam-Agent.

Method: Evaluated six flagship models across six medical benchmarks using stress tests (removing key inputs, prompt variations), clinician-guided rubric evaluation, and analysis of benchmark design flaws.

Result: Models often guess correctly even without key inputs, flip answers under trivial prompt changes, fabricate flawed reasoning, and benchmarks vary widely in what they truly measure but are treated interchangeably.

Conclusion: Medical benchmark scores don’t reflect real-world readiness; we need to demand robustness, sound reasoning, and alignment with actual medical demands rather than just leaderboard performance.

Abstract: Large frontier models like GPT-5 now achieve top scores on medical benchmarks. But our stress tests tell a different story. Leading systems often guess correctly even when key inputs like images are removed, flip answers under trivial prompt changes, and fabricate convincing yet flawed reasoning. These aren’t glitches; they expose how today’s benchmarks reward test-taking tricks over medical understanding. We evaluate six flagship models across six widely used benchmarks and find that high leaderboard scores hide brittleness and shortcut learning. Through clinician-guided rubric evaluation, we show that benchmarks vary widely in what they truly measure yet are treated interchangeably, masking failure modes. We caution that medical benchmark scores do not directly reflect real-world readiness. If we want AI to earn trust in healthcare, we must demand more than leaderboard wins and must hold systems accountable for robustness, sound reasoning, and alignment with real medical demands.

Method: Used constrained matrix factorization model with explicit user types, translated model outputs into natural language using carefully designed LLM prompts, and conducted user study with 326 participants evaluating explanations across five dimensions.

Result: All explanation types were generally well received with moderate statistical differences between strategies; user comments provided complementary insights beyond quantitative results.

Conclusion: LLMs can successfully generate effective explanations from interpretable models, and user-centered evaluation provides valuable insights that complement automatic metrics.

Abstract: We investigate whether large language models (LLMs) can generate effective, user-facing explanations from a mathematically interpretable recommendation model. The model is based on constrained matrix factorization, where user types are explicitly represented and predicted item scores share the same scale as observed ratings, making the model’s internal representations and predicted scores directly interpretable. This structure is translated into natural language explanations using carefully designed LLM prompts. Many works in explainable AI rely on automatic evaluation metrics, which often fail to capture users’ actual needs and perceptions. In contrast, we adopt a user-centered approach: we conduct a study with 326 participants who assessed the quality of the explanations across five key dimensions-transparency, effectiveness, persuasion, trust, and satisfaction-as well as the recommendations themselves. To evaluate how different explanation strategies are perceived, we generate multiple explanation types from the same underlying model, varying the input information provided to the LLM. Our analysis reveals that all explanation types are generally well received, with moderate statistical differences between strategies. User comments further underscore how participants react to each type of explanation, offering complementary insights beyond the quantitative results.

Method: Systematic methodology combining LLMs with evolutionary algorithms to iteratively generate and refine heuristic solutions for 2D bin-packing.

Result: LLM-generated heuristics (GPT-4o) achieved optimal solutions in 2 iterations, reducing average bin usage from 16 to 15 bins and improving space utilization from 0.76-0.78 to 0.83, outperforming traditional approaches.

Conclusion: LLMs can produce more efficient solutions with fewer computational resources than traditional methods, contributing to understanding LLM evaluation in specialized optimization domains.

Abstract: This paper presents an evaluation framework for assessing Large Language Models’ (LLMs) capabilities in combinatorial optimization, specifically addressing the 2D bin-packing problem. We introduce a systematic methodology that combines LLMs with evolutionary algorithms to generate and refine heuristic solutions iteratively. Through comprehensive experiments comparing LLM generated heuristics against traditional approaches (Finite First-Fit and Hybrid First-Fit), we demonstrate that LLMs can produce more efficient solutions while requiring fewer computational resources. Our evaluation reveals that GPT-4o achieves optimal solutions within two iterations, reducing average bin usage from 16 to 15 bins while improving space utilization from 0.76-0.78 to 0.83. This work contributes to understanding LLM evaluation in specialized domains and establishes benchmarks for assessing LLM performance in combinatorial optimization tasks.

Method: Information-theoretic approach that dynamically sets truncation thresholds at each decoding step using the entire token probability distribution, eliminating hyperparameters.

Result: Consistently outperforms existing sampling methods across math, reasoning, and creative writing tasks, with less degradation at high temperatures and improved inference efficiency through faster sampling and shorter generations.

Conclusion: p-less sampling provides a robust, hyperparameter-free alternative to existing decoding strategies that maintains quality across temperature ranges while being more computationally efficient.

Abstract: Obtaining high-quality outputs from Large Language Models (LLMs) often depends upon the choice of a sampling-based decoding strategy to probabilistically choose the next token at each generation step. While a variety of such sampling methods have been proposed, their performance can be sensitive to the selection of hyperparameters which may require different settings depending upon the generation task and temperature configuration. In this work, we introduce $p$-less sampling: an information-theoretic approach to sampling which dynamically sets a truncation threshold at each decoding step based on the entire token probability distribution. Unlike existing methods, $p$-less sampling has no hyperparameters and consistently produces high-quality outputs as temperature increases. We provide theoretical perspectives on $p$-less sampling to ground our proposed method and conduct experiments to empirically validate its effectiveness across a range of math, logical reasoning, and creative writing tasks. Our results demonstrate how $p$-less sampling consistently outperforms existing sampling approaches while exhibiting much less degradation in text quality at higher temperature values. We further show how $p$-less achieves greater inference-time efficiency than alternative methods through lower average token sampling times and shorter generation lengths, without sacrificing accuracy. Finally, we provide analyses to highlight the benefits of $p$-less through qualitative examples, case studies, and diversity assessments.

Method: Extracted benchmark signatures via stepwise forward selection with linear regressions across 32 LLMs and 88 benchmarks spanning diverse domains. Used token perplexity from naturally authored corpora as predictive features.

Result: Found high overlap in knowledge and reasoning subtasks, while multilingual/cultural benchmarks showed less similarity. Coding emerged as the least overlapping domain. Benchmark signatures remained robust to format effects unlike performance-based measures.

Conclusion: Benchmark signatures provide mechanistic insights into benchmark validity and LLM sensitivities, revealing cross-functional overlaps across logic, math, language, instruction following, and world modeling while sketching the underlying landscape of interconnected LLM capabilities.

Abstract: We develop signatures of capacity familiarity to characterize large language model (LLM) benchmarks and their meaningful overlaps. Benchmark signatures probe the capacity required for benchmark performance. We formally define them as a set of salient tokens drawn from in-the-wild, naturally authored corpora, where LLM token perplexity, reflecting more or less pre-training exposure, becomes highly predictive of LLM benchmark performance. Through a large-scale meta-evaluation, we extract benchmark signatures via stepwise forward selection with linear regressions across 32 LLMs and 88 benchmarks spanning diverse knowledge, coding, logic, instruction following, math, language, reasoning, and world modeling. Our analysis situates signatures in relation to both the semantic similarity of benchmark questions and the correlation of model performance. While performance overlaps are universally high and semantic overlaps remain confined to a narrow mid-range, benchmark signatures prove highly informative in capturing variation, overlap, and divergence. We observe overlap in knowledge and reasoning subtasks, whereas multilingual and cultural benchmarks exhibit less similarity, even compared to cross-task overlap. Notably, performance-level results are strongly influenced by benchmark-orthogonal factors such as question format, highlighting limitations in LLM generalization, the conflation of performance with ability, and issues inherent in current mainstream benchmark agreement studies. Benchmark signatures, however, remain robust to such effects. Ultimately, we identify cross-functional overlaps across logic, math, language, instruction following, and world modeling, with coding emerging as the least overlapping domain. Together, these findings provide mechanistic insights into benchmark validity and LLM sensitivities, and sketch the underlying landscape of interconnected LLM capabilities.

Method: The authors propose: (1) a hybrid data synthesis framework combining MCTS with judgments from strong VLMs for more accurate step-level labels, (2) perception-focused supervision to detect errors at visual grounding stage, and (3) systematic evaluation of multiple test-time scaling strategies.

Result: Experiments across five multimodal benchmarks reveal key insights: VL-PRMs as Outcome Reward Models outperform process step selection; smaller VL-PRMs can match larger ones in error detection; VL-PRMs uncover latent reasoning abilities; perception-level supervision significantly improves test-time scaling; and performance improves on advanced math reasoning datasets despite no training on them.

Conclusion: The work aims to motivate further research and support the advancement of Vision Language Models by elucidating the design space of VL-PRMs through diverse strategies for dataset construction, training, and test-time scaling.

Abstract: Process Reward Models (PRMs) provide step-level supervision that improves the reliability of reasoning in large language models. While PRMs have been extensively studied in text-based domains, their extension to Vision Language Models (VLMs) remains limited. Existing Vision-Language PRMs (VL-PRMs) rely on Monte Carlo Tree Search (MCTS) for data construction, which can often produce noisy supervision signals and limit generalization across tasks. In this work, we aim to elucidate the design space of VL-PRMs by exploring diverse strategies for dataset construction, training, and test-time scaling. First, we introduce a hybrid data synthesis framework that combines MCTS with judgments from a strong VLM, producing more accurate step-level labels. Second, we propose perception-focused supervision, enabling our PRM to explicitly detect errors at the visual grounding stage of reasoning. Third, we systematically evaluate multiple test-time scaling strategies, showing that our PRMs can reliably guide VLMs toward more accurate solutions. Our experiments covering five diverse multimodal benchmarks (MMMU, PuzzleVQA, AlgoPuzzleVQA, MathVista, and MathVision) reveal several key insights: (i) VL-PRMs when used as Outcome Reward Models (ORMs) during test-time scaling (TTS) can outperform VL-PRM guided process step selection, (ii) smaller VL-PRMs can match or even surpass larger ones in detecting process errors, (iii) VL-PRMs uncover latent reasoning abilities in stronger VLM backbones, (iv) perception-level supervision leads to significant gains in test-time scaling, and (v) TTS performance of different policies improve on advanced math reasoning datasets despite not training VL-PRMs on such datasets. We hope our work will motivate further research and support the advancement of VLMs.

Method: Developed an automated red-teaming framework and SafeSearch benchmark with 300 test cases covering five risk categories. Evaluated three search agent scaffolds across 15 LLMs (7 proprietary, 8 open-source) when exposed to unreliable websites.

Result: Revealed substantial vulnerabilities: highest attack success rate reached 90.5% for GPT-4.1-mini under search workflow setting. Common defense practices like reminder prompting showed limited effectiveness.

Conclusion: The framework provides systematic, scalable safety assessment for search agents, promoting transparency in safer agent development. The high vulnerability rates emphasize the need for robust safety measures in LLM-based search systems.

Abstract: Search agents connect LLMs to the Internet, enabling access to broader and more up-to-date information. However, unreliable search results may also pose safety threats to end users, establishing a new threat surface. In this work, we conduct two in-the-wild experiments to demonstrate both the prevalence of low-quality search results and their potential to misguide agent behaviors. To counter this threat, we introduce an automated red-teaming framework that is systematic, scalable, and cost-efficient, enabling lightweight and harmless safety assessments of search agents. Building on this framework, we construct the SafeSearch benchmark, which includes 300 test cases covering five categories of risks (e.g., misinformation and indirect prompt injection). Using this benchmark, we evaluate three representative search agent scaffolds, covering search workflow, tool-calling, and deep research, across 7 proprietary and 8 open-source backend LLMs. Our results reveal substantial vulnerabilities of LLM-based search agents: when exposed to unreliable websites, the highest ASR reached 90.5% for GPT-4.1-mini under a search workflow setting. Moreover, our analysis highlights the limited effectiveness of common defense practices, such as reminder prompting. This emphasizes the value of our framework in promoting transparency for safer agent development. Our codebase and test cases are publicly available: https://github.com/jianshuod/SafeSearch.

Method: Abstracts teacher’s thought process into discrete semantic signals, trains student model with multi-task objective to predict next semantic signal and generate corresponding reasoning step conditioned on that signal.

Result: Significantly outperforms state-of-the-art distillation methods on challenging reasoning benchmarks in both accuracy and logical consistency.

Conclusion: Provides a path towards creating smaller models that are genuine reasoners rather than just fluent mimics by transferring algorithmic reasoning structure directly.

Abstract: The prevailing approach to distilling reasoning from Large Language Models (LLMs)-behavioral cloning from textual rationales-is fundamentally limited. It teaches Small Language Models (SLMs) to mimic surface-level patterns rather than the underlying algorithmic structure of thought, resulting in a critical lack of logical robustness. We argue that instead of cloning text, distillation should transfer this algorithmic structure directly. We introduce Reasoning Scaffolding}, a framework that reframes reasoning as a structured generation process. Our method first abstracts the teacher’s thought process into a sequence of discrete, interpretable semantic signals (e.g., Contrast, Addition) that act as a scaffold. The student model is then trained via a multi-task objective to both (1)predict the next semantic signal, anticipating the reasoning flow, and (2)generate the corresponding step, conditioned on that signal. This multi-task scheme acts as a powerful regularizer, compelling the student to internalize the computational patterns of coherent reasoning. On a suite of challenging reasoning benchmarks, our method significantly outperforms state-of-the-art distillation in both accuracy and logical consistency, providing a path towards creating smaller models that are genuine reasoners, not just fluent mimics.

Method: Uses Expectation-Maximization to infer correctness probabilities of preference labels and adaptively re-calibrate data contributions to training loss. Establishes theoretical link between preference losses and probabilistic models.

Result: Consistently enhances four state-of-the-art alignment algorithms (DPO, IPO, SimPO, CPO). Achieves up to 7.0% win rate gains on AlpacaEval 2 and Arena-Hard benchmarks with Mistral and Llama 3 models.

Conclusion: LCPO provides a general framework for robust preference alignment that effectively handles noisy and heterogeneous human preference data, with theoretical guarantees and empirical improvements over existing methods.

Abstract: Standard human preference-based alignment methods, such as Reinforcement Learning from Human Feedback (RLHF), are a cornerstone technology for aligning Large Language Models (LLMs) with human values. However, these methods are all underpinned by a critical, yet flawed assumption: human preferences are homogeneous (representing a single, unified preference) and the collected data is noiseless (free from error). In reality, neither is true since human preference is pluralistic and annotators can make mistakes. This creates a discrepancy between the recorded data and the ground-truth preferences, which can misguide the model and degrade its performance. To address this challenge, we introduce Latent Collective Preference Optimization (LCPO). LCPO leverages an Expectation-Maximization (EM) algorithm to learn the latent collective consensus from noisy data. It operates by inferring the correctness of each preference label and using this probability as an adaptive weight to re-calibrate each data point’s contribution to the training loss, thereby mitigating noise. We generalize this approach by establishing a theoretical link between arbitrary preference losses and their corresponding probabilistic models, elevating LCPO from a specific algorithm to a general framework for robust preference alignment. Theoretically, we prove that under the condition of a perfectly calibrated model, LCPO is guaranteed to converge to the true noise level of the dataset. Our experiments demonstrate LCPO’s effectiveness as a general framework, consistently enhancing four state-of-the-art alignment algorithms (DPO, IPO, SimPO, and CPO). When applied to Mistral and Llama 3 models, the LCPO-enhanced methods achieve substantial win rate gains on AlpacaEval 2 and Arena-Hard, with improvements of up to 7.0% on both benchmarks.

Method: Analyzed empirical scaling laws across diverse architectures and merging methods (Average, TA, TIES, DARE) to identify a compact power law linking model size and expert number.

Result: Found that size-dependent floor decreases with model capacity, merging tail shows diminishing returns in expert count, gains fall roughly as 1/k, and variability shrinks with more experts.

Conclusion: The scaling law enables predictive planning for merging, making it a computationally efficient alternative to multitask training and suggesting a complementary path toward AGI through specialist composition.

Abstract: We study empirical scaling laws for language model merging measured by cross-entropy. Despite its wide practical use, merging lacks a quantitative rule that predicts returns as we add experts or scale the model size. We identify a compact power law that links model size and expert number: the size-dependent floor decreases with model capacity, while the merging tail exhibits clear diminishing returns in the number of experts. The law holds in-domain and cross-domain, tightly fits measured curves across diverse architectures and methods (Average, TA, TIES, DARE), and explains two robust regularities: most gains arrive early, and variability shrinks as more experts are included. Building on this, we present a simple theory that explains why gains fall roughly as 1/k and links the floor and tail to properties of the base model and the diversity across domains. This law enables predictive planning: estimate how many experts are needed to reach a target loss, decide when to stop adding experts, and trade off scaling the base model versus adding experts under a fixed budget–turning merging from heuristic practice into a computationally efficient, planable alternative to multitask training. This suggests a scaling principle for distributed generative AI: predictable gains can be achieved by composing specialists, offering a complementary path toward AGI-level systems.

Method: DeepSearch embeds Monte Carlo Tree Search directly into RLVR training with: (1) global frontier selection strategy for prioritizing promising nodes, (2) entropy-based guidance for identifying confident paths, and (3) adaptive replay buffer training with solution caching for efficiency.

Result: Achieves 62.95% average accuracy on mathematical reasoning benchmarks, establishing new state-of-the-art for 1.5B reasoning models while using 5.7x fewer GPU hours than extended training approaches.

Conclusion: Strategic exploration through systematic search is more effective than brute-force scaling, demonstrating the promise of algorithmic innovation for advancing RLVR methodologies and establishing a new direction for scaling reasoning capabilities.

Abstract: Although RLVR has become an essential component for developing advanced reasoning skills in LLMs, contemporary studies have documented training plateaus that emerge following thousands of optimization steps, demonstrating notable decreases in performance gains despite increased computational investment. This limitation stems from the sparse exploration patterns inherent in current RLVR practices, where models rely on limited rollouts that often miss critical reasoning paths and fail to provide systematic coverage of the solution space. We present DeepSearch, a framework that integrates Monte Carlo Tree Search directly into RLVR training. In contrast to existing methods that rely on tree search only at inference, DeepSearch embeds structured search into the training loop, enabling systematic exploration and fine-grained credit assignment across reasoning steps. Through training-time exploration, DeepSearch addresses the fundamental bottleneck of insufficient exploration, which leads to diminishing performance improvements over prolonged training steps. Our contributions include: (1) a global frontier selection strategy that prioritizes promising nodes across the search tree, (2) selection with entropy-based guidance that identifies confident paths for supervision, and (3) adaptive replay buffer training with solution caching for efficiency. Experiments on mathematical reasoning benchmarks show that DeepSearch achieves 62.95% average accuracy and establishes a new state-of-the-art for 1.5B reasoning models - using 5.7x fewer GPU hours than extended training approaches. These results highlight the importance of strategic exploration over brute-force scaling and demonstrate the promise of algorithmic innovation for advancing RLVR methodologies. DeepSearch establishes a new direction for scaling reasoning capabilities through systematic search rather than prolonged computation.

Method: RADAR learns an item response model from model responses with different budgets to queries, with interpretable parameters for query difficulties and model-budget abilities. It routes harder queries to higher-ability model-budget pairs and easier queries to lower-ability ones.

Result: Extensive experiments on 8 reasoning benchmarks show RADAR outperforms state-of-the-art model routing methods. It also demonstrates strong generalization to out-of-distribution queries and scalability for integrating new models efficiently.

Conclusion: RADAR provides an effective, interpretable, and scalable solution for optimizing reasoning model deployment by intelligently routing queries based on difficulty and model ability, achieving better performance-cost tradeoffs.

Abstract: Reasoning language models have demonstrated remarkable performance on many challenging tasks in math, science, and coding. Choosing the right reasoning model for practical deployment involves a performance and cost tradeoff at two key levels: model size and reasoning budget, where larger models and higher reasoning budget lead to better performance but with increased cost and latency. In this work, we tackle this tradeoff from the angle of model configuration routing for different queries, and present RADAR (Reasoning-Ability and Difficulty-Aware Routing), a lightweight, interpretable, and scalable routing framework. Inspired by psychometrics, RADAR learns an item response model from model responses with different budgets to different queries, with interpretable parameters including query difficulties and model-budget abilities. RADAR then routes queries with higher difficulty to model-budget pairs with higher ability, and vice versa. We conduct extensive experiments on 8 widely used challenging reasoning benchmarks, demonstrating the superior performance of RADAR compared to state-of-the-art model routing methods. RADAR also exhibits query generalization capabilities, showing strong performance on out-of-distribution queries in all benchmarks. RADAR is also scalable and can efficiently integrate additional models by dynamically selecting a small set of evaluation queries to estimate their abilities.

Method: Constructs a learnable representation space that jointly models inter-agent relations and task-specific world information by directly modeling communication protocols. Supports both implicit coordination (as latent representations) and explicit communication (as messages).

Result: Evaluated on four challenging MARL benchmarks, IWoL provides effective team coordination. The representation can be combined with existing MARL algorithms to further enhance their performance.

Conclusion: IWoL offers a simple yet powerful solution for team coordination in MARL, enabling decentralized execution with implicit coordination while avoiding drawbacks of explicit message passing.

Abstract: This work presents a novel representation learning framework, interactive world latent (IWoL), to facilitate team coordination in multi-agent reinforcement learning (MARL). Building effective representation for team coordination is a challenging problem, due to the intricate dynamics emerging from multi-agent interaction and incomplete information induced by local observations. Our key insight is to construct a learnable representation space that jointly captures inter-agent relations and task-specific world information by directly modeling communication protocols. This representation, we maintain fully decentralized execution with implicit coordination, all while avoiding the inherent drawbacks of explicit message passing, e.g., slower decision-making, vulnerability to malicious attackers, and sensitivity to bandwidth constraints. In practice, our representation can be used not only as an implicit latent for each agent, but also as an explicit message for communication. Across four challenging MARL benchmarks, we evaluate both variants and show that IWoL provides a simple yet powerful key for team coordination. Moreover, we demonstrate that our representation can be combined with existing MARL algorithms to further enhance their performance.

Method: Uses a graph-structured coordination mechanism with a central supervisory agent that oversees specialized task agents, providing modularity, conflict resolution, and context-aware interactions.

Result: The system demonstrates how structured reasoning and agent cooperation can improve calendar management through natural language commands.

Conclusion: ScheduleMe sets an example for increasing the usability and flexibility of personal calendar assistant tools through multi-agent coordination and structured reasoning.

Abstract: Recent advancements in LLMs have contributed to the rise of advanced conversational assistants that can assist with user needs through natural language conversation. This paper presents a ScheduleMe, a multi-agent calendar assistant for users to manage google calendar events in natural language. The system uses a graph-structured coordination mechanism where a central supervisory agent supervises specialized task agents, allowing modularity, conflicts resolution, and context-aware interactions to resolve ambiguities and evaluate user commands. This approach sets an example of how structured reasoning and agent cooperation might convince operators to increase the usability and flexibility of personal calendar assistant tools.

Method: Created CritPt benchmark with 71 composite research challenges simulating full-scale projects and 190 simpler checkpoint tasks, all newly created by 50+ active physics researchers. Problems are guess-resistant, machine-verifiable, and evaluated through automated grading pipeline customized for physics-specific output formats.

Result: Current state-of-the-art LLMs show early promise on isolated checkpoints but perform poorly on full research challenges: best base model accuracy is 4.0% (GPT-5), rising to ~10% with coding tools. Models remain far from reliably solving research-scale physics problems.

Conclusion: There’s a large disconnect between current LLM capabilities and realistic physics research demands. CritPt provides a foundation to guide development of scientifically grounded AI tools for physics research.

Abstract: While large language models (LLMs) with reasoning capabilities are progressing rapidly on high-school math competitions and coding, can they reason effectively through complex, open-ended challenges found in frontier physics research? And crucially, what kinds of reasoning tasks do physicists want LLMs to assist with? To address these questions, we present the CritPt (Complex Research using Integrated Thinking - Physics Test, pronounced “critical point”), the first benchmark designed to test LLMs on unpublished, research-level reasoning tasks that broadly covers modern physics research areas, including condensed matter, quantum physics, atomic, molecular & optical physics, astrophysics, high energy physics, mathematical physics, statistical physics, nuclear physics, nonlinear dynamics, fluid dynamics and biophysics. CritPt consists of 71 composite research challenges designed to simulate full-scale research projects at the entry level, which are also decomposed to 190 simpler checkpoint tasks for more fine-grained insights. All problems are newly created by 50+ active physics researchers based on their own research. Every problem is hand-curated to admit a guess-resistant and machine-verifiable answer and is evaluated by an automated grading pipeline heavily customized for advanced physics-specific output formats. We find that while current state-of-the-art LLMs show early promise on isolated checkpoints, they remain far from being able to reliably solve full research-scale challenges: the best average accuracy among base models is only 4.0% , achieved by GPT-5 (high), moderately rising to around 10% when equipped with coding tools. Through the realistic yet standardized evaluation offered by CritPt, we highlight a large disconnect between current model capabilities and realistic physics research demands, offering a foundation to guide the development of scientifically grounded AI tools.

Method: CiT uses two Branching Necessity evaluation methods: BN-DP (Direct Prompting) where an auxiliary LLM judges branching necessity, and BN-SC (Self-Consistency) which clusters candidate actions to estimate agreement. It integrates into existing tree search frameworks like ToT-BS, ReST-MCTS, and RAP.

Result: BN-DP consistently reduces token generation, model invocations, and runtime by 75-85% across all settings with negligible accuracy loss. BN-SC provides substantial savings (up to 80%) but shows instability in some settings. The quality of auxiliary LLMs is critical for performance.

Conclusion: CiT effectively reduces computational overhead in tree-search-based LLM reasoning while maintaining performance, with BN-DP being more stable than BN-SC. The framework provides theoretical guarantees and is implemented across multiple search frameworks for reproducibility.

Abstract: Test-time scaling enables large language models (LLMs) to improve performance on long-horizon reasoning tasks by allocating additional compute at inference. Tree-search-based approaches achieve state-of-the-art results in this setting, but they are notoriously inefficient, often an order of magnitude slower than simpler iterative methods. We introduce Chain-in-Tree (CiT), a plug-in framework that adaptively decides when to branch during search rather than branching at every step. CiT relies on lightweight Branching Necessity (BN) evaluation methods: BN-DP (Direct Prompting), where an auxiliary LLM directly judges whether a step requires branching, and BN-SC (Self-Consistency), which clusters multiple candidate actions to estimate agreement. We integrate CiT into three representative LLM-in-the-loop tree search frameworks: Tree of Thoughts (ToT-BS), ReST-MCTS, and RAP, and evaluate across GSM8K and Math500. Our results show that: (1) BN-DP consistently reduces token generation, model invocations, and runtime by 75-85 percent across all settings, with negligible accuracy loss and sometimes accuracy gains; (2) BN-SC typically yields substantial savings (up to 80 percent) but shows instability in 1-4 out of 14 settings, caused by a small subset of examples that produce very long reasoning steps; (3) the quality of auxiliary LLMs is critical, not only the BN evaluator in BN-DP, but also the models used in BN-SC for clustering and equivalence checking. When these roles are filled by smaller LLMs, performance degrades. Importantly, BN-SC does not require LLMs in domains with deterministic action spaces, where clustering can be done programmatically. We also provide a theoretical guarantee that BN-DP never increases LLM invocations relative to the baseline and release a unified implementation of CiT across ToT-BS, ReST-MCTS, and RAP to facilitate reproducibility and extension.

Method: Developed a practitioner-oriented field manual informed by deploying the “irAE-Agent” system at Mass General Brigham and structured interviews with 20 clinicians, engineers, and informatics leaders.

Result: Revealed that less than 20% of effort was dedicated to prompt engineering/model development, while over 80% was consumed by sociotechnical implementation work across five “heavy lifts”: data integration, model validation, economic value, system drift management, and governance.

Conclusion: The field manual shifts focus from algorithmic development to essential infrastructure and implementation work needed to successfully translate generative AI from pilot projects into routine clinical care.

Abstract: Large language models (LLMs) integrated into agent-driven workflows hold immense promise for healthcare, yet a significant gap exists between their potential and practical implementation within clinical settings. To address this, we present a practitioner-oriented field manual for deploying generative agents that use electronic health record (EHR) data. This guide is informed by our experience deploying the “irAE-Agent”, an automated system to detect immune-related adverse events from clinical notes at Mass General Brigham, and by structured interviews with 20 clinicians, engineers, and informatics leaders involved in the project. Our analysis reveals a critical misalignment in clinical AI development: less than 20% of our effort was dedicated to prompt engineering and model development, while over 80% was consumed by the sociotechnical work of implementation. We distill this effort into five “heavy lifts”: data integration, model validation, ensuring economic value, managing system drift, and governance. By providing actionable solutions for each of these challenges, this field manual shifts the focus from algorithmic development to the essential infrastructure and implementation work required to bridge the “valley of death” and successfully translate generative AI from pilot projects into routine clinical care.

Method: Proposes a framework that learns symbolic state representations and causal processes for endogenous actions and exogenous mechanisms via variational Bayesian inference combined with LLM proposals.

Result: The learned models enable fast planning that generalizes to held-out tasks with more objects and more complex goals, outperforming baselines across five simulated tabletop robotics environments.

Conclusion: The framework successfully addresses the challenge of planning in dynamic environments with concurrent exogenous processes by learning joint causal models of both agent actions and environmental mechanisms.

Abstract: Long-horizon embodied planning is challenging because the world does not only change through an agent’s actions: exogenous processes (e.g., water heating, dominoes cascading) unfold concurrently with the agent’s actions. We propose a framework for abstract world models that jointly learns (i) symbolic state representations and (ii) causal processes for both endogenous actions and exogenous mechanisms. Each causal process models the time course of a stochastic cause-effect relation. We learn these world models from limited data via variational Bayesian inference combined with LLM proposals. Across five simulated tabletop robotics environments, the learned models enable fast planning that generalizes to held-out tasks with more objects and more complex goals, outperforming a range of baselines.

Method: Dynamic Interaction adaptively selects cooperative/competitive strategies based on question difficulty and model ability, using Idea3 (Idea Sharing, Analysis, Fusion) paradigm. Perception Calibration employs Group Relative Policy Optimization (GRPO) to integrate reward distributions across agents.

Result: ILR consistently outperforms single-agent learning by up to 5% across five mathematical and one coding benchmark using three LLMs from two model families. Idea3 enhances robustness of stronger LLMs, and dynamic interaction boosts learning vs pure cooperative/competitive strategies.

Conclusion: Multi-agent interaction can enhance LLMs’ independent problem-solving ability. The ILR framework with Dynamic Interaction and Perception Calibration effectively transfers collaborative benefits to individual agents.

Abstract: Existing multi-agent learning approaches have developed interactive training environments to explicitly promote collaboration among multiple Large Language Models (LLMs), thereby constructing stronger multi-agent systems (MAS). However, during inference, they require re-executing the MAS to obtain final solutions, which diverges from human cognition that individuals can enhance their reasoning capabilities through interactions with others and resolve questions independently in the future. To investigate whether multi-agent interaction can enhance LLMs’ independent problem-solving ability, we introduce ILR, a novel co-learning framework for MAS that integrates two key components: Dynamic Interaction and Perception Calibration. Specifically, Dynamic Interaction first adaptively selects either cooperative or competitive strategies depending on question difficulty and model ability. LLMs then exchange information through Idea3 (Idea Sharing, Idea Analysis, and Idea Fusion), an innovative interaction paradigm designed to mimic human discussion, before deriving their respective final answers. In Perception Calibration, ILR employs Group Relative Policy Optimization (GRPO) to train LLMs while integrating one LLM’s reward distribution characteristics into another’s reward function, thereby enhancing the cohesion of multi-agent interactions. We validate ILR on three LLMs across two model families of varying scales, evaluating performance on five mathematical benchmarks and one coding benchmark. Experimental results show that ILR consistently outperforms single-agent learning, yielding an improvement of up to 5% over the strongest baseline. We further discover that Idea3 can enhance the robustness of stronger LLMs during multi-agent inference, and dynamic interaction types can boost multi-agent learning compared to pure cooperative or competitive strategies.

Method: FLoRA-NA leverages local LoRA matrices on the server to estimate aggregated matrices, which are then distributed to clients for local updates, minimizing divergence between ideal and practical updates without additional communication cost.

Result: Extensive evaluations across natural language understanding, mathematical reasoning, and code-solving tasks show FLoRA-NA achieves state-of-the-art global performance while maintaining low communication overhead.

Conclusion: FLoRA-NA successfully addresses key limitations of prior personalized FedLoRA approaches by achieving communication efficiency and bridging the gap between local personalization and global generalization.

Abstract: With the rapid emergence of foundation models and the increasing need for fine-tuning across distributed environments, Federated Low-Rank Adaptation (FedLoRA) has recently gained significant attention. Despite enormous potential, current FedLoRA methods face notable challenges due to inexact updates. Existing approaches have attempted to mitigate this issue, but they often introduce a \emph{local-global generalization gap} and incur \emph{substantial communication overhead}, limiting their scalability and effectiveness. To address these limitations, we propose \textbf{F}ederated \textbf{Lo}w-\textbf{R}ank \textbf{A}ggregation with \textbf{N}early \textbf{A}ccurate Estimation (FLoRA-NA). FLoRA-NA leverages the local LoRA matrices on the server to estimate the aggregated matrices $\hat{A}$ and $\hat{B}$, which are then distributed to clients for local updates. This surrogated aggregated matrices minimizes the divergence between ideal $\nabla \Bar{W} = \sum^{U}_{u=1}B_u A_u$ and practical updates $\nabla \hat{W} = \hat{B}\hat{A}$ without adding communication cost beyond vanilla FedLoRA. By doing so, FLoRA-NA achieves communication efficiency and bridges the gap between local personalization and global generalization, addressing a key limitation of prior personalized FedLoRA approaches. We conduct extensive evaluations across diverse tasks, including natural language understanding, mathematical reasoning, and code-solving ability using various foundation models. Experimental results consistently demonstrate that FLoRA-NA achieves state-of-the-art global performance while maintaining low communication overhead.

Method: Built a reproducible pipeline using 275 chart-topping songs with audio descriptors and lyric transcripts, analyzing temporal trends, lexical salience/co-occurrence, and genre-based mood profiling.

Result: Found decade-long energy decline (79→58) and danceability rise (59→73); valence peaked in 2013 (63) then dipped; strong energy-loudness correlation (r=0.74); lyric analysis revealed pronoun-centric lexicon; R&B had highest valence (96), Latin/Reggaeton lowest (37).

Conclusion: Patterns suggest mainstreaming of peripheral codes and commercial preference for relaxed yet rhythmically engaging productions. Contributed an integrated MIR-plus-network workflow suitable for socially aware recommendation systems.

Abstract: This paper examines how musical symbolism is produced and circulated in online communities by combining content-based music analysis with a lightweight network perspective on lyrics. Using a curated corpus of 275 chart-topping songs enriched with audio descriptors (energy, danceability, loudness, liveness, valence, acousticness, speechiness, popularity) and full lyric transcripts, we build a reproducible pipeline that (i) quantifies temporal trends in sonic attributes, (ii) models lexical salience and co-occurrence, and (iii) profiles mood by genre. We find a decade-long decline in energy (79 -> 58) alongside a rise in danceability (59 -> 73); valence peaks in 2013 (63) and dips in 2014-2016 (42) before partially recovering. Correlation analysis shows strong coupling of energy with loudness (r = 0.74) and negative associations for acousticness with both energy (r = -0.54) and loudness (r = -0.51); danceability is largely orthogonal to other features (|r| < 0.20). Lyric tokenization (>114k tokens) reveals a pronoun-centric lexicon “I/you/me/my” and a dense co-occurrence structure in which interpersonal address anchors mainstream narratives. Mood differs systematically by style: R&B exhibits the highest mean valence (96), followed by K-Pop/Pop (77) and Indie/Pop (70), whereas Latin/Reggaeton is lower (37) despite high danceability. Read through a subcultural identity lens, these patterns suggest the mainstreaming of previously peripheral codes and a commercial preference for relaxed yet rhythmically engaging productions that sustain collective participation without maximal intensity. Methodologically, we contribute an integrated MIR-plus-network workflow spanning summary statistics, correlation structure, lexical co-occurrence matrices, and genre-wise mood profiling that is robust to modality sparsity and suitable for socially aware recommendation or community-level diffusion studies.

Method: Two key innovations: 1) Optical Flow-inspired Iterative Refinement using convolutional GRU to capture frame-to-frame evolution of acoustic features, 2) Cross-Attention Based Prosodic Alignment to dynamically align spectral features with prosodic patterns like pitch and pauses.

Result: Outperforms text-based baselines with 0.839 AUC and 80.6% accuracy on DementiaBank dataset, without relying on ASR.

Conclusion: Provides a flexible and robust solution for automated cognitive assessment by directly modeling the temporal evolution of raw audio dynamics.

Abstract: Deep learning systems often struggle with processing long sequences, where computational complexity can become a bottleneck. Current methods for automated dementia detection using speech frequently rely on static, time-agnostic features or aggregated linguistic content, lacking the flexibility to model the subtle, progressive deterioration inherent in speech production. These approaches often miss the dynamic temporal patterns that are critical early indicators of cognitive decline. In this paper, we introduce TAI-Speech, a Temporal Aware Iterative framework that dynamically models spontaneous speech for dementia detection. The flexibility of our method is demonstrated through two key innovations: 1) Optical Flow-inspired Iterative Refinement: By treating spectrograms as sequential frames, this component uses a convolutional GRU to capture the fine-grained, frame-to-frame evolution of acoustic features. 2) Cross-Attention Based Prosodic Alignment: This component dynamically aligns spectral features with prosodic patterns, such as pitch and pauses, to create a richer representation of speech production deficits linked to functional decline (IADL). TAI-Speech adaptively models the temporal evolution of each utterance, enhancing the detection of cognitive markers. Experimental results on the DementiaBank dataset show that TAI-Speech achieves a strong AUC of 0.839 and 80.6% accuracy, outperforming text-based baselines without relying on ASR. Our work provides a more flexible and robust solution for automated cognitive assessment, operating directly on the dynamics of raw audio.

Method: Converted breathing sounds into spectrograms, balanced dataset through oversampling apnea segments, applied class weights to reduce bias, and prioritized recall over general accuracy.

Result: Achieved 90.55% recall for apnea detection, demonstrating potential as a screening tool despite low precision.

Conclusion: Respiratory audio analysis shows promise as a low-cost, accessible screening method for sleep apnea that could be used at home or in basic clinical settings for early identification of at-risk individuals.

Abstract: Sleep apnea is a serious sleep-related breathing disorder that is common and can impact health if left untreated. Currently the traditional method for screening and diagnosis is overnight polysomnography. Polysomnography is expensive and takes a lot of time, and is not practical for screening large groups of people. In this paper, we explored a more accessible option, using respiratory audio recordings to spot signs of apnea.We utilized 18 audio files.The approach involved converting breathing sounds into spectrograms, balancing the dataset by oversampling apnea segments, and applying class weights to reduce bias toward the majority class. The model reached a recall of 90.55 for apnea detection. Intentionally, prioritizing catching apnea events over general accuracy. Despite low precision,the high recall suggests potential as a low-cost screening tool that could be used at home or in basic clinical setups, potentially helping identify at-risk individuals much earlier.

Method: Convolutional neural codec models with Residual Vector Quantization, trained end-to-end using a combination of adversarial and reconstruction objectives, with detailed data filtering/augmentation strategies, model architectures, optimization procedures, and checkpoint selection criteria.

Result: Official baseline systems for both tracks (transparency and enhancement codecs) in the 2025 LRAC Challenge are presented, providing standardized models for comparison and advancement in low-resource neural speech coding.

Conclusion: The LRAC Challenge establishes foundational baseline systems to drive progress in neural audio coding for constrained environments, addressing both transparent preservation and enhancement of speech under noisy conditions.

Abstract: The Low-Resource Audio Codec (LRAC) Challenge aims to advance neural audio coding for deployment in resource-constrained environments. The first edition focuses on low-resource neural speech codecs that must operate reliably under everyday noise and reverberation, while satisfying strict constraints on computational complexity, latency, and bitrate. Track 1 targets transparency codecs, which aim to preserve the perceptual transparency of input speech under mild noise and reverberation. Track 2 addresses enhancement codecs, which combine coding and compression with denoising and dereverberation. This paper presents the official baseline systems for both tracks in the 2025 LRAC Challenge. The baselines are convolutional neural codec models with Residual Vector Quantization, trained end-to-end using a combination of adversarial and reconstruction objectives. We detail the data filtering and augmentation strategies, model architectures, optimization procedures, and checkpoint selection criteria.

Method: U-Net architecture trained to estimate residual reverberation mask in STFT domain, using hybrid objective combining binary cross-entropy, residual magnitude reconstruction, and time-domain consistency.

Result: The framework enables low-latency dereverberation suitable for real-world speech and singing applications by suppressing late reflections while preserving direct speech components.

Conclusion: The proposed method provides an efficient real-time solution for vocal dereverberation that balances accuracy and computational efficiency for practical applications.

Abstract: Vocal dereverberation remains a challenging task in audio processing, particularly for real-time applications where both accuracy and efficiency are crucial. Traditional deep learning approaches often struggle to suppress reverberation without degrading vocal clarity, while recent methods that jointly predict magnitude and phase have significant computational cost. We propose a real-time dereverberation framework based on residual mask prediction in the short-time Fourier transform (STFT) domain. A U-Net architecture is trained to estimate a residual reverberation mask that suppresses late reflections while preserving direct speech components. A hybrid objective combining binary cross-entropy, residual magnitude reconstruction, and time-domain consistency further encourages both accurate suppression and perceptual quality. Together, these components enable low-latency dereverberation suitable for real-world speech and singing applications.

Method: Proposed Attribute-Specialized Key-Value Head Sharing (AS-KVHS) adapted to music’s structured symbolic representation, and released SAGE-Music benchmark.

Result: Achieved about 30% inference speedup with only 0.4% quality drop in objective evaluations and slight improvements in subjective listening tests.

Conclusion: AS-KVHS effectively addresses low-latency symbolic music generation challenges while maintaining quality, with contributions including systematic BPE analysis and open-source benchmark.

Abstract: Low-latency symbolic music generation is essential for real-time improvisation and human-AI co-creation. Existing transformer-based models, however, face a trade-off between inference speed and musical quality. Traditional acceleration techniques such as embedding pooling significantly degrade quality, while recently proposed Byte Pair Encoding (BPE) methods - though effective on single-track piano data - suffer large performance drops in multi-track settings, as revealed by our analysis. We propose Attribute-Specialized Key-Value Head Sharing (AS-KVHS), adapted to music’s structured symbolic representation, achieving about 30% inference speedup with only a negligible (about 0.4%) quality drop in objective evaluations and slight improvements in subjective listening tests. Our main contributions are (1) the first systematic study of BPE’s generalizability in multi-track symbolic music, and (2) the introduction of AS-KVHS for low-latency symbolic music generation. Beyond these, we also release SAGE-Music, an open-source benchmark that matches or surpasses state-of-the-art models in generation quality.

Method: 1) Constructed a real-world podcast dataset with diverse topics as reference for human-level quality; 2) Introduced multimodal evaluation strategy decomposing the task into text, speech, and audio dimensions with different emphasis on ‘Content’ and ‘Format’; 3) Designed evaluation methods for each modality using both objective metrics and subjective listening tests.

Result: The framework was tested with representative podcast generation systems (open-source, close-source, and human-made), providing in-depth analysis and insights into podcast generation, demonstrating PodEval’s effectiveness in evaluating open-ended long-form audio.

Conclusion: PodEval provides a comprehensive and well-designed evaluation framework for podcast generation that addresses the limitations of existing benchmarks and enables effective assessment of open-ended long-form audio generation capabilities.

Abstract: Recently, an increasing number of multimodal (text and audio) benchmarks have emerged, primarily focusing on evaluating models’ understanding capability. However, exploration into assessing generative capabilities remains limited, especially for open-ended long-form content generation. Significant challenges lie in no reference standard answer, no unified evaluation metrics and uncontrollable human judgments. In this work, we take podcast-like audio generation as a starting point and propose PodEval, a comprehensive and well-designed open-source evaluation framework. In this framework: 1) We construct a real-world podcast dataset spanning diverse topics, serving as a reference for human-level creative quality. 2) We introduce a multimodal evaluation strategy and decompose the complex task into three dimensions: text, speech and audio, with different evaluation emphasis on “Content” and “Format”. 3) For each modality, we design corresponding evaluation methods, involving both objective metrics and subjective listening test. We leverage representative podcast generation systems (including open-source, close-source, and human-made) in our experiments. The results offer in-depth analysis and insights into podcast generation, demonstrating the effectiveness of PodEval in evaluating open-ended long-form audio. This project is open-source to facilitate public use: https://github.com/yujxx/PodEval.

Method: Proposes ARIONet - a self-supervised contrastive network that jointly optimizes contrastive classification and future frame prediction using augmented audio representations within a transformer-based encoder model that integrates multiple complementary audio features.

Result: Achieves classification accuracies of 98.41%, 93.07%, 91.89%, and 91.58% on four birdsong datasets, with corresponding F1-scores of 97.84%, 94.10%, 91.29%, and 90.94%. Also demonstrates high cosine similarity (up to 95%) in future frame prediction.

Conclusion: The self-supervised learning strategy effectively captures complex acoustic patterns and temporal dependencies, showing strong potential for real-world ecological conservation and monitoring applications.

Abstract: Automated birdsong classification is essential for advancing ecological monitoring and biodiversity studies. Despite recent progress, existing methods often depend heavily on labeled data, use limited feature representations, and overlook temporal dynamics essential for accurate species identification. In this work, we propose a self-supervised contrastive network, ARIONet (Acoustic Representation for Interframe Objective Network), that jointly optimizes contrastive classification and future frame prediction using augmented audio representations. The model simultaneously integrates multiple complementary audio features within a transformer-based encoder model. Our framework is designed with two key objectives: (1) to learn discriminative species-specific representations for contrastive learning through maximizing similarity between augmented views of the same audio segment while pushing apart different samples, and (2) to model temporal dynamics by predicting future audio frames, both without requiring large-scale annotations. We validate our framework on four diverse birdsong datasets, including the British Birdsong Dataset, Bird Song Dataset, and two extended Xeno-Canto subsets (A-M and N-Z). Our method consistently outperforms existing baselines and achieves classification accuracies of 98.41%, 93.07%, 91.89%, and 91.58%, and F1-scores of 97.84%, 94.10%, 91.29%, and 90.94%, respectively. Furthermore, it demonstrates low mean absolute errors and high cosine similarity, up to 95%, in future frame prediction tasks. Extensive experiments further confirm the effectiveness of our self-supervised learning strategy in capturing complex acoustic patterns and temporal dependencies, as well as its potential for real-world applicability in ecological conservation and monitoring.

Method: Created MOS-RMBench benchmark and systematically evaluated three reward modeling paradigms: scalar, semi-scalar, and generative reward models (GRMs), plus proposed MOS-aware GRM with adaptive reward scaling based on MOS differences.

Result: Scalar models achieved strongest performance (>74% accuracy), models perform worse on synthetic vs human speech, all struggle with small MOS differences. MOS-aware GRM improved fine-grained discrimination and narrowed gap with scalar models on challenging cases.

Conclusion: Establishes benchmark and methodological framework for more rigorous and scalable research in automatic speech quality assessment.

Abstract: Assessing the perceptual quality of synthetic speech is crucial for guiding the development and refinement of speech generation models. However, it has traditionally relied on human subjective ratings such as the Mean Opinion Score (MOS), which depend on manual annotations and often suffer from inconsistent rating standards and poor reproducibility. To address these limitations, we introduce MOS-RMBench, a unified benchmark that reformulates diverse MOS datasets into a preference-comparison setting, enabling rigorous evaluation across different datasets. Building on MOS-RMBench, we systematically construct and evaluate three paradigms for reward modeling: scalar reward models, semi-scalar reward models, and generative reward models (GRMs). Our experiments reveal three key findings: (1) scalar models achieve the strongest overall performance, consistently exceeding 74% accuracy; (2) most models perform considerably worse on synthetic speech than on human speech; and (3) all models struggle on pairs with very small MOS differences. To improve performance on these challenging pairs, we propose a MOS-aware GRM that incorporates an MOS-difference-based reward function, enabling the model to adaptively scale rewards according to the difficulty of each sample pair. Experimental results show that the MOS-aware GRM significantly improves fine-grained quality discrimination and narrows the gap with scalar models on the most challenging cases. We hope this work will establish both a benchmark and a methodological framework to foster more rigorous and scalable research in automatic speech quality assessment.

Method: Tested models across three text-based benchmarks with various irrelevant audio inputs (silence, synthetic noise, environmental sounds), analyzed interference factors (duration, amplitude, temperature), and evaluated mitigation strategies (prompting, self-consistency).

Result: Non-informative audio reduces accuracy and increases prediction volatility; silence destabilizes outputs as strongly as synthetic noise; larger models show greater resilience but vulnerabilities persist; prompting has limited effectiveness while self-consistency improves stability at computational cost.

Conclusion: Cross-modal interference is a key robustness challenge for audio-language models, highlighting the need for efficient fusion strategies that preserve reasoning performance with irrelevant inputs.

Abstract: Large audio-language models (LALMs) unify speech and text processing, but their robustness in noisy real-world settings remains underexplored. We investigate how irrelevant audio, such as silence, synthetic noise, and environmental sounds, affects text reasoning tasks where audio is unnecessary. Across three text-based benchmarks, we find that even non-informative audio reduces accuracy and increases prediction volatility; the severity of interference scales with longer durations, higher amplitudes, and elevated decoding temperatures. Silence, often assumed neutral, destabilizes outputs as strongly as synthetic noise. While larger models show greater resilience, vulnerabilities persist across all evaluated systems. We further test mitigation strategies and find that prompting shows limited effectiveness, whereas self-consistency improves stability at the cost of increased computation. Our results reveal cross-modal interference as a key robustness challenge and highlight the need for efficient fusion strategies that preserve reasoning performance in the presence of irrelevant inputs.

Method: Extensive experiments on six LALMs across three benchmarks and their spoken counterparts, testing performance with shuffled answer options and evaluating permutation-based mitigation strategies.

Result: No model is immune to this bias; shuffling answer options causes performance fluctuations up to 24% and changes model rankings, raising concerns about current evaluation practices.

Conclusion: This work represents the first systematic investigation of selection bias in LALMs, highlighting the need for awareness and further research to improve model reliability.

Abstract: Large audio-language models (LALMs) are often used in tasks that involve reasoning over ordered options. An open question is whether their predictions are influenced by the order of answer choices, which would indicate a form of selection bias and undermine their reliability. In this paper, we identify and analyze this problem in LALMs. We demonstrate that no model is immune to this bias through extensive experiments on six LALMs across three widely used benchmarks and their spoken counterparts. Shuffling the order of answer options can cause performance fluctuations of up to 24% and even change model rankings, raising concerns about the reliability of current evaluation practices. We also study permutation-based strategies and show that they can mitigate bias in most cases. Our work represents the first systematic investigation of this issue in LALMs, and we hope it raises awareness and motivates further research in this direction.

Method: Proposed SpeechLMScore - a reference-free method that leverages automatic speech naturalness evaluation scores which correlate strongly with severity scores, without relying on pathological speech data. Also introduced NKI-SpeechRT dataset for comprehensive evaluation.

Result: SpeechLMScore demonstrates superior performance compared to traditional acoustic feature-based approaches and shows robustness to noise. The method effectively bridges the performance gap between reference-free and reference-based models.

Conclusion: SpeechLMScore provides an effective reference-free solution for speech severity evaluation that is robust to noise and outperforms traditional methods, making it suitable for ecologically valid scenarios like spontaneous speech assessment.

Abstract: Speech severity evaluation is becoming increasingly important as the economic burden of speech disorders grows. Current speech severity models often struggle with generalization, learning dataset-specific acoustic cues rather than meaningful correlates of speech severity. Furthermore, many models require reference speech or a transcript, limiting their applicability in ecologically valid scenarios, such as spontaneous speech evaluation. Previous research indicated that automatic speech naturalness evaluation scores correlate strongly with severity evaluation scores, leading us to explore a reference-free method, SpeechLMScore, which does not rely on pathological speech data. Additionally, we present the NKI-SpeechRT dataset, based on the NKI-CCRT dataset, to provide a more comprehensive foundation for speech severity evaluation. This study evaluates whether SpeechLMScore outperforms traditional acoustic feature-based approaches and assesses the performance gap between reference-free and reference-based models. Moreover, we examine the impact of noise on these models by utilizing subjective noise ratings in the NKI-SpeechRT dataset. The results demonstrate that SpeechLMScore is robust to noise and offers superior performance compared to traditional approaches.

Method: XPPG-PCA (x-vector phonetic posteriorgram principal component analysis) is an unsupervised, reference-free method that uses phonetic posteriorgrams and principal component analysis for speech severity evaluation without requiring transcriptions or healthy reference samples.

Result: Using three Dutch oral cancer datasets, XPPG-PCA performs comparably to or exceeds established reference-based methods, demonstrating robustness against data shortcuts and noise, showing potential for real-world clinical use.

Conclusion: XPPG-PCA provides a robust, generalizable solution for objective assessment of speech pathology that can significantly improve efficiency and reliability of clinical evaluations across various disorders.

Abstract: Reliably evaluating the severity of a speech pathology is crucial in healthcare. However, the current reliance on expert evaluations by speech-language pathologists presents several challenges: while their assessments are highly skilled, they are also subjective, time-consuming, and costly, which can limit the reproducibility of clinical studies and place a strain on healthcare resources. While automated methods exist, they have significant drawbacks. Reference-based approaches require transcriptions or healthy speech samples, restricting them to read speech and limiting their applicability. Existing reference-free methods are also flawed; supervised models often learn spurious shortcuts from data, while handcrafted features are often unreliable and restricted to specific speech tasks. This paper introduces XPPG-PCA (x-vector phonetic posteriorgram principal component analysis), a novel, unsupervised, reference-free method for speech severity evaluation. Using three Dutch oral cancer datasets, we demonstrate that XPPG-PCA performs comparably to, or exceeds established reference-based methods. Our experiments confirm its robustness against data shortcuts and noise, showing its potential for real-world clinical use. Taken together, our results show that XPPG-PCA provides a robust, generalizable solution for the objective assessment of speech pathology, with the potential to significantly improve the efficiency and reliability of clinical evaluations across a range of disorders. An open-source implementation is available.

Method: Proposes DualCodec - a dual-stream encoding approach that integrates SSL and waveform representations within an end-to-end codec framework, enhancing semantic information in the first-layer codec tokens.

Result: Experimental results show DualCodec outperforms state-of-the-art codec systems (Mimi Codec, SpeechTokenizer, DAC, Encodec) on both audio codec and speech generation tasks while operating at low frame rates.

Conclusion: DualCodec successfully maintains high audio quality at low frame rates by integrating SSL and waveform representations, making it more efficient for speech generation compared to existing codec systems.

Abstract: Neural audio codecs form the foundational building blocks for language model (LM)-based speech generation. Typically, there is a trade-off between frame rate and audio quality. This study introduces a low-frame-rate, semantically enhanced codec model. Existing approaches distill semantically rich self-supervised (SSL) representations into the first-layer codec tokens. This work proposes DualCodec, a dual-stream encoding approach that integrates SSL and waveform representations within an end-to-end codec framework. In this setting, DualCodec enhances the semantic information in the first-layer codec and enables the codec system to maintain high audio quality while operating at a low frame rate. Note that a low-frame-rate codec improves the efficiency of speech generation. Experimental results on audio codec and speech generation tasks confirm the effectiveness of the proposed DualCodec compared to state-of-the-art codec systems, such as Mimi Codec, SpeechTokenizer, DAC, and Encodec. Demos are available at: https://dualcodec.github.io, code is available at: https://github.com/jiaqili3/DualCodec

Method: Uses dynamic frame rate approach with adaptive frame merging in information-sparse regions, ASR feature-assisted dual stream encoding, and Transformer bottlenecks to preserve semantic information while supporting inference-time controllable frame rates.

Result: Outperforms baseline systems in semantic information preservation and delivers high audio reconstruction quality at 6.25Hz, 8.3Hz and 12.5Hz average frame rates. Also validated effective in language model-based TTS.

Conclusion: FlexiCodec successfully addresses semantic information loss in very low frame rate audio codecs through dynamic frame rate control and novel architecture, enabling better performance for speech language models.

Abstract: Neural audio codecs are foundational to speech language models. It is expected to have a low frame rate and decoupled semantic and acoustic information. A lower frame rate codec can reduce the computational cost of speech language models by shortening the sequence length. Recent studies have developed 12.5Hz low-frame-rate audio codecs, but even lower frame rate codecs remain underexplored. We find that a major challenge for very low frame rate tokens is missing semantic information. This paper introduces FlexiCodec to address this limitation. FlexiCodec improves semantic preservation with a dynamic frame rate approach and introduces a novel architecture featuring an ASR feature-assisted dual stream encoding and Transformer bottlenecks. With dynamic frame rates, it uses less frames at information-sparse regions through adaptively merging semantically similar frames. A dynamic frame rate also allows FlexiCodec to support inference-time controllable frame rates between 3Hz and 12.5Hz. Experiments on 6.25Hz, 8.3Hz and 12.5Hz average frame rates confirm that FlexiCodec excels over baseline systems in semantic information preservation and delivers a high audio reconstruction quality. We also validate the effectiveness of FlexiCodec in language model-based TTS. Demos are available at: https://flexicodec.github.io

Method: Uses seven discriminators based on nonlinear dynamical systems: MRLD for chaos, MS-RD for recurrence dynamics, MSDFA for fractal analysis, MR-PPD for latent relationships, MPD for cyclical patterns, MRAD for amplitude, and MRPD for phase. Features a complex-valued ConformerNeXt generator with dual-stream Lattice-Net architecture for magnitude and phase refinement, using depth-wise convolution for parameter efficiency.

Result: Achieves eight-times parameter reduction and establishes new state-of-the-art performance across six objective evaluation metrics and subjective tests with five human judges.

Conclusion: NDSI-BWE successfully demonstrates that leveraging nonlinear dynamical system principles in discriminators enables effective bandwidth extension with significantly improved efficiency and performance.

Abstract: Recovering high-frequency components lost to bandwidth constraints is crucial for applications ranging from telecommunications to high-fidelity audio on limited resources. We introduce NDSI-BWE, a new adversarial Band Width Extension (BWE) framework that leverage four new discriminators inspired by nonlinear dynamical system to capture diverse temporal behaviors: a Multi-Resolution Lyapunov Discriminator (MRLD) for determining sensitivity to initial conditions by capturing deterministic chaos, a Multi-Scale Recurrence Discriminator (MS-RD) for self-similar recurrence dynamics, a Multi-Scale Detrended Fractal Analysis Discriminator (MSDFA) for long range slow variant scale invariant relationship, a Multi-Resolution Poincar'e Plot Discriminator (MR-PPD) for capturing hidden latent space relationship, a Multi-Period Discriminator (MPD) for cyclical patterns, a Multi-Resolution Amplitude Discriminator (MRAD) and Multi-Resolution Phase Discriminator (MRPD) for capturing intricate amplitude-phase transition statistics. By using depth-wise convolution at the core of the convolutional block with in each discriminators, NDSI-BWE attains an eight-times parameter reduction. These seven discriminators guide a complex-valued ConformerNeXt based genetor with a dual stream Lattice-Net based architecture for simultaneous refinement of magnitude and phase. The genertor leverage the transformer based conformer’s global dependency modeling and ConvNeXt block’s local temporal modeling capability. Across six objective evaluation metrics and subjective based texts comprises of five human judges, NDSI-BWE establishes a new SoTA in BWE.

Method: Uses a complex-valued conformer with Complex Unified Attention Block to capture phoneme dependencies and reconstructs both magnitude and phase of missing frequencies to mitigate HVAC noise.

Result: Achieves intelligible speech reconstruction from as low as 0.5 kHz sampling rate, surpassing prior work limited to hot word detection, and shows significant intelligibility in real-world HVAC deployments.

Conclusion: HVAC-EAR demonstrates the feasibility of speech reconstruction from HVAC pressure sensors, introducing novel privacy vulnerabilities in modern building systems.

Abstract: Pressure sensors are widely integrated into modern Heating, Ventilation and Air Conditioning (HVAC) systems. As they are sensitive to acoustic pressure, they can be a source of eavesdropping. This paper introduces HVAC-EAR, which reconstructs intelligible speech from low-resolution, noisy pressure data with two key contributions: (i) We achieve intelligible reconstruction from as low as 0.5 kHz sampling rate, surpassing prior work limited to hot word detection, by employing a complex-valued conformer with a Complex Unified Attention Block to capture phoneme dependencies; (ii) HVAC-EAR mitigates transient HVAC noise by reconstructing both magnitude and phase of missing frequencies. For the first time, evaluations on real-world HVAC deployments show significant intelligibility, raising novel privacy concerns.

Method: Leveraged open source datasets to create AHEAD-DS with consistent hearing aid-relevant labels. Developed YAMNet+ using transfer learning from pretrained YAMNet model, designed for edge deployment on smartphones connected to hearing devices.

Result: YAMNet+ achieved mean average precision of 0.83 and accuracy of 0.93 on AHEAD-DS testing set. Successfully deployed on Android smartphone (Google Pixel 3) with 50ms model loading latency and 30ms per second audio processing latency.

Conclusion: The approach enables real-time sound-based scene recognition on edge devices for hearing aids, providing standardized dataset and baseline model for audiologically relevant environment classification.

Abstract: Scene recognition of audiologically relevant environments is important for hearing aids; however, it is challenging, in part because of the limitations of existing datasets. Datasets often lack public accessibility, completeness, or audiologically relevant labels, hindering systematic comparison of machine learning models. Deploying these models on resource-constrained edge devices presents another challenge. Our solution is two-fold: we leverage several open source datasets to create AHEAD-DS, a dataset designed for scene recognition of audiologically relevant environments, and introduce YAMNet+, a sound recognition model. AHEAD-DS aims to provide a standardised, publicly available dataset with consistent labels relevant to hearing aids, facilitating model comparison. YAMNet+ is designed for deployment on edge devices like smartphones connected to hearing devices, such as hearing aids and wireless earphones with hearing aid functionality; serving as a baseline model for sound-based scene recognition. YAMNet+ achieved a mean average precision of 0.83 and accuracy of 0.93 on the testing set of AHEAD-DS across fourteen categories of audiologically relevant environments. We found that applying transfer learning from the pretrained YAMNet model was essential. We demonstrated real-time sound-based scene recognition capabilities on edge devices by deploying YAMNet+ to an Android smartphone. Even with a Google Pixel 3 (a phone with modest specifications, released in 2018), the model processes audio with approximately 50ms of latency to load the model, and an approximate linear increase of 30ms per 1 second of audio. Our website and code https://github.com/Australian-Future-Hearing-Initiative .

Method: Designed two nonlinear dynamical systems-inspired discriminators: MSRD based on Recurrence representations for self-similarity dynamics, and MRLD based on Lyapunov exponents for nonlinear fluctuations and sensitivity to initial conditions. Used depthwise-separable convolutions for optimization.

Result: Achieved 44x reduction in discriminator parameters (from ~22M to ~0.48M) while surpassing prior AP-BWE models. Successfully demonstrated BWE supervision using chaotic physics of voiced sound production.

Conclusion: The paper shows that explicitly modeling speech’s chaotic physics enables significant discriminator size reduction in bandwidth extension tasks, providing a novel approach to audio processing.

Abstract: In this paper, we design two nonlinear dynamical systems-inspired discriminators – the Multi-Scale Recurrence Discriminator (MSRD) and the Multi-Resolution Lyapunov Discriminator (MRLD) – to \textit{explicitly} model the inherent deterministic chaos of speech. MSRD is designed based on Recurrence representations to capture self-similarity dynamics. MRLD is designed based on Lyapunov exponents to capture nonlinear fluctuations and sensitivity to initial conditions. Through extensive design optimization and the use of depthwise-separable convolutions in the discriminators, our framework surpasses prior AP-BWE model with a 44x reduction in the discriminator parameter count \textbf{($\sim$ 22M vs $\sim$ 0.48M)}. To the best of our knowledge, for the first time, this paper demonstrates how BWE can be supervised by the subtle non-linear chaotic physics of voiced sound production to achieve a significant reduction in the discriminator size.

Method: Enrolled 512 participants in Zambia with three categories (TB+, other respiratory diseases, healthy controls). Fine-tuned pre-trained speech foundation models on cough recordings, with multimodal approach incorporating demographic/clinical features.

Result: Best model achieved 85.2% AUROC for TB vs all others, 80.1% for TB vs other respiratory diseases. Multimodal model improved to 92.1% and 84.2% respectively, with 90.3% sensitivity and 73.1% specificity at optimal threshold.

Conclusion: Cough-based AI screening is feasible for TB detection in real-world, low-resource settings, demonstrating robustness to confounding factors and meeting WHO screening benchmarks.

Abstract: Artificial intelligence (AI) systems can detect disease-related acoustic patterns in cough sounds, offering a scalable and cost-effective approach to tuberculosis (TB) screening in high-burden, resource-limited settings. Previous studies have been limited by small datasets, under-representation of symptomatic non-TB patients, and recordings collected in controlled environments. In this study, we enrolled 512 participants at two hospitals in Zambia, categorised into three groups: bacteriologically confirmed TB (TB+), symptomatic patients with other respiratory diseases (OR), and healthy controls (HC). Usable cough recordings with demographic and clinical data were obtained from 500 participants. Deep learning classifiers based on pre-trained speech foundation models were fine-tuned on cough recordings to predict diagnostic categories. The best-performing model, trained on 3-second audio clips, achieved an AUROC of 85.2% for distinguishing TB coughs from all other participants (TB+/Rest) and 80.1% for TB+ versus symptomatic OR participants (TB+/OR). Incorporating demographic and clinical features improved performance to 92.1% for TB+/Rest and 84.2% for TB+/OR. At a probability threshold of 0.38, the multimodal model reached 90.3% sensitivity and 73.1% specificity for TB+/Rest, meeting WHO target product profile benchmarks for TB screening. Adversarial testing and stratified analyses shows that the model was robust to confounding factors including background noise, recording time, and device variability. These results demonstrate the feasibility of cough-based AI for TB screening in real-world, low-resource settings.

Method: A multi-agent system with four specialized agents: Music-Ingestion Agent for parsing scores, Chord-Knowledge Agent with Transformer model for chord interpretation, Harmony-Generation Agent with Harmony-GPT and RNN for composition, and Audio-Production Agent with GAN-based synthesizer for audio rendering.

Result: The system generates sophisticated and contextually appropriate higher-voice harmonies that are melodically and rhythmically complementary to given melodies, with high-fidelity audio output.

Conclusion: The modular, agent-based approach enables robust data processing, deep theoretical understanding, creative composition, and realistic audio synthesis for effective harmony generation.

Abstract: The generation of musically coherent and aesthetically pleasing harmony remains a significant challenge in the field of algorithmic composition. This paper introduces an innovative Agentic AI-enabled Higher Harmony Music Generator, a multi-agent system designed to create harmony in a collaborative and modular fashion. Our framework comprises four specialized agents: a Music-Ingestion Agent for parsing and standardizing input musical scores; a Chord-Knowledge Agent, powered by a Chord-Former (Transformer model), to interpret and provide the constituent notes of complex chord symbols; a Harmony-Generation Agent, which utilizes a Harmony-GPT and a Rhythm-Net (RNN) to compose a melodically and rhythmically complementary harmony line; and an Audio-Production Agent that employs a GAN-based Symbolic-to-Audio Synthesizer to render the final symbolic output into high-fidelity audio. By delegating specific tasks to specialized agents, our system effectively mimics the collaborative process of human musicians. This modular, agent-based approach allows for robust data processing, deep theoretical understanding, creative composition, and realistic audio synthesis, culminating in a system capable of generating sophisticated and contextually appropriate higher-voice harmonies for given melodies.

Method: MARS uses channel multiplexing (CMX) to reshape spectrograms into multi-channel images, employs a shared tokenizer for consistent discrete representations across scales, and uses a transformer-based autoregressor to refine spectrograms from coarse to fine resolutions.

Result: Experiments show MARS performs comparably or better than state-of-the-art baselines across multiple evaluation metrics on a large-scale dataset.

Conclusion: MARS establishes an efficient and scalable paradigm for high-fidelity audio generation by combining spectrogram-based approaches with multi-scale autoregression techniques from image synthesis.

Abstract: Research on audio generation has progressively shifted from waveform-based approaches to spectrogram-based methods, which more naturally capture harmonic and temporal structures. At the same time, advances in image synthesis have shown that autoregression across scales, rather than tokens, improves coherence and detail. Building on these ideas, we introduce MARS (Multi-channel AutoRegression on Spectrograms), a framework that treats spectrograms as multi-channel images and employs channel multiplexing (CMX), a reshaping technique that lowers height and width without discarding information. A shared tokenizer provides consistent discrete representations across scales, enabling a transformer-based autoregressor to refine spectrograms from coarse to fine resolutions efficiently. Experiments on a large-scale dataset demonstrate that MARS performs comparably or better than state-of-the-art baselines across multiple evaluation metrics, establishing an efficient and scalable paradigm for high-fidelity audio generation.

Method: Embed document chunks and test questions into unified vector space, calculate multiple coverage metrics (proximity, content-weighted, multi-topic), and use outlier detection to filter irrelevant questions.

Result: Framework effectively quantifies test coverage, identifies underrepresented content areas, and provides recommendations for generating high-value test questions in two distinct use cases.

Conclusion: Provides RAG developers with essential tools to build robust test suites, improving system reliability and extending to applications like identifying misaligned documents.

Abstract: Reliably determining the performance of Retrieval-Augmented Generation (RAG) systems depends on comprehensive test questions. While a proliferation of evaluation frameworks for LLM-powered applications exists, current practices lack a systematic method to ensure these test sets adequately cover the underlying knowledge base, leaving developers with significant blind spots. To address this, we present a novel, applied methodology to quantify the semantic coverage of RAG test questions against their underlying documents. Our approach leverages existing technologies, including vector embeddings and clustering algorithms, to create a practical framework for validating test comprehensiveness. Our methodology embeds document chunks and test questions into a unified vector space, enabling the calculation of multiple coverage metrics: basic proximity, content-weighted coverage, and multi-topic question coverage. Furthermore, we incorporate outlier detection to filter irrelevant questions, allowing for the refinement of test sets. Experimental evidence from two distinct use cases demonstrates that our framework effectively quantifies test coverage, identifies specific content areas with inadequate representation, and provides concrete recommendations for generating new, high-value test questions. This work provides RAG developers with essential tools to build more robust test suites, thereby improving system reliability and extending to applications such as identifying misaligned documents.

Method: TransIP uses a generic non-equivariant Transformer-based model and guides it to learn SO(3)-equivariance by optimizing representations in the embedding space, trained on the Open Molecules (OMol25) collection.

Result: TransIP attains comparable performance to state-of-the-art equivariant baselines in machine-learning force fields. Compared to data augmentation baseline, it achieves 40% to 60% improvement in performance across varying OMol25 dataset sizes.

Conclusion: Learned equivariance can be a powerful and efficient alternative to equivariant or augmentation-based MLIP models, offering comparable performance without hard-wired architectural constraints.

Abstract: Accurate and scalable machine-learned inter-atomic potentials (MLIPs) are essential for molecular simulations ranging from drug discovery to new material design. Current state-of-the-art models enforce roto-translational symmetries through equivariant neural network architectures, a hard-wired inductive bias that can often lead to reduced flexibility, computational efficiency, and scalability. In this work, we introduce TransIP: Transformer-based Inter-Atomic Potentials, a novel training paradigm for interatomic potentials achieving symmetry compliance without explicit architectural constraints. Our approach guides a generic non-equivariant Transformer-based model to learn SO(3)-equivariance by optimizing its representations in the embedding space. Trained on the recent Open Molecules (OMol25) collection, a large and diverse molecular dataset built specifically for MLIPs and covering different types of molecules (including small organics, biomolecular fragments, and electrolyte-like species), TransIP attains comparable performance in machine-learning force fields versus state-of-the-art equivariant baselines. Further, compared to a data augmentation baseline, TransIP achieves 40% to 60% improvement in performance across varying OMol25 dataset sizes. More broadly, our work shows that learned equivariance can be a powerful and efficient alternative to equivariant or augmentation-based MLIP models.

Method: Q-ROAR groups RoPE dimensions into frequency bands and performs lightweight search over per-band scales for Key and Query weights, guided by interpolation pressure and tail-inflation ratio diagnostics, using a tiny long-context dataset.

Result: Empirically reduces model perplexity on long-context workloads by more than 14% while preserving short-context performance, inference throughput, and compatibility with existing LLM systems.

Conclusion: Q-ROAR provides an effective interpolation-aware stabilization method for quantized LLMs that addresses the coupled degradation effects of PI+PTQ without requiring model fine-tuning or architectural changes.

Abstract: Extending the context window support of large language models (LLMs) is crucial for tasks with long-distance dependencies. RoPE-based interpolation and extrapolation methods, such as linear scaling and frequency-aware schemes, enable longer input length support without retraining, while post-training quantization (PTQ) makes deployment practical. However, we show that combining RoPE position interpolation (PI) with PTQ degrades accuracy due to coupled effects including long-context aliasing, dynamic-range dilation, anisotropy from axis-aligned quantizers vs. rotated RoPE pairs, and outlier shifting that produces position-dependent logit noise. We provide, to the best of our knowledge, the first systematic analysis of the PI+PTQ approach and introduce two practical diagnostics: interpolation pressure (per-band sensitivity to phase scaling) and tail-inflation ratios (outlier shift from short to long contexts). Following the analysis results, we propose Q-ROAR (Quantization, RoPE-interpolation, and Outlier Aware Rescaling), a weight-only, interpolation-aware stabilization of PI for quantized LLMs. Q-ROAR groups RoPE dimensions into a small number of frequency bands and performs a lightweight search over per-band scales for Key and Query weights (with an optional symmetric variant to preserve logit scale). The search is guided by our diagnostics and uses a tiny long-context development dataset, requiring no fine-tuning to the model, no architecture or kernel changes, and no additional deployment overhead. Empirically, Q-ROAR reduces the model’s perplexity on long-context workloads by more than 14%, while preserving short-context performance, inference throughput, and compatibility with existing LLM system stacks.

Method: Compiled one month of time-series data from 15,000 individuals across three diabetes populations, including glucose traces from CGMs and behavioral logs. Generated 360,600 personalized questions across 7 task categories and evaluated 8 LLMs using 5 metrics.

Result: Evaluation of 8 recent LLMs revealed substantial variability across tasks and metrics; no single model consistently outperformed others across all dimensions.

Conclusion: DexBench aims to advance the reliability, safety, effectiveness and practical utility of AI solutions in diabetes care by establishing a comprehensive evaluation framework.

Abstract: We present DexBench, the first benchmark designed to evaluate large language model (LLM) performance across real-world decision-making tasks faced by individuals managing diabetes in their daily lives. Unlike prior health benchmarks that are either generic, clinician-facing or focused on clinical tasks (e.g., diagnosis, triage), DexBench introduces a comprehensive evaluation framework tailored to the unique challenges of prototyping patient-facing AI solutions in diabetes, glucose management, metabolic health and related domains. Our benchmark encompasses 7 distinct task categories, reflecting the breadth of real-world questions individuals with diabetes ask, including basic glucose interpretation, educational queries, behavioral associations, advanced decision making and long term planning. Towards this end, we compile a rich dataset comprising one month of time-series data encompassing glucose traces and metrics from continuous glucose monitors (CGMs) and behavioral logs (e.g., eating and activity patterns) from 15,000 individuals across three different diabetes populations (type 1, type 2, pre-diabetes/general health and wellness). Using this data, we generate a total of 360,600 personalized, contextual questions across the 7 tasks. We evaluate model performance on these tasks across 5 metrics: accuracy, groundedness, safety, clarity and actionability. Our analysis of 8 recent LLMs reveals substantial variability across tasks and metrics; no single model consistently outperforms others across all dimensions. By establishing this benchmark, we aim to advance the reliability, safety, effectiveness and practical utility of AI solutions in diabetes care.

Method: Theoretical analysis using p-adic metric spaces, proving that n-dimensional planes minimizing p-adic distances must pass through at least n+1 data points. Applied to polynomial regression and demonstrated with NLP applications.

Result: Proved that p-adic regression requires regression planes to pass through actual data points, unlike Euclidean regression. Demonstrated practical significance in hierarchical taxonomy analysis and grammatical morphology modeling.

Conclusion: P-adic metrics are fundamental for properly handling hierarchical data structures in machine learning, as they better align with the discrete nature where interpolation between points is less meaningful than selecting actual observed points.

Abstract: Many real-world machine learning problems involve inherently hierarchical data, yet traditional approaches rely on Euclidean metrics that fail to capture the discrete, branching nature of hierarchical relationships. We present a theoretical foundation for machine learning in p-adic metric spaces, which naturally respect hierarchical structure. Our main result proves that an n-dimensional plane minimizing the p-adic sum of distances to points in a dataset must pass through at least n + 1 of those points – a striking contrast to Euclidean regression that highlights how p-adic metrics better align with the discrete nature of hierarchical data. As a corollary, a polynomial of degree n constructed to minimise the p-adic sum of residuals will pass through at least n + 1 points. As a further corollary, a polynomial of degree n approximating a higher degree polynomial at a finite number of points will yield a difference polynomial that has distinct rational roots. We demonstrate the practical significance of this result through two applications in natural language processing: analyzing hierarchical taxonomies and modeling grammatical morphology. These results suggest that p-adic metrics may be fundamental to properly handling hierarchical data structures in machine learning. In hierarchical data, interpolation between points often makes less sense than selecting actual observed points as representatives.

Method: Serializes tabular records into text and uses pre-trained LLMs (DistilBERT, ALBERT, RoBERTa, ClinicalBERT) as universal feature extractors to create semantically aligned representations, supporting lightweight local classifiers under FedAvg protocol.

Result: Consistently outperforms state-of-the-art baselines with up to +0.25 improvement in F1-score and 65% reduction in communication cost. Shows graceful degradation under extreme schema divergence while traditional methods collapse.

Conclusion: FedLLM-Align is a robust, privacy-preserving, and communication-efficient solution for federated learning in heterogeneous environments.

Abstract: Federated learning (FL) enables collaborative model training without sharing raw data, making it attractive for privacy-sensitive domains such as healthcare, finance, and IoT. A major obstacle, however, is the heterogeneity of tabular data across clients, where divergent schemas and incompatible feature spaces prevent straightforward aggregation. To address this challenge, we propose FedLLM-Align, a federated framework that leverages pre-trained large language models (LLMs) as universal feature extractors. Tabular records are serialized into text, and embeddings from models such as DistilBERT, ALBERT, RoBERTa, and ClinicalBERT provide semantically aligned representations that support lightweight local classifiers under the standard FedAvg protocol. This approach removes the need for manual schema harmonization while preserving privacy, since raw data remain strictly local. We evaluate FedLLM-Align on coronary heart disease prediction using partitioned Framingham datasets with simulated schema divergence. Across all client settings and LLM backbones, our method consistently outperforms state-of-the-art baselines, achieving up to +0.25 improvement in F1-score and a 65% reduction in communication cost. Stress testing under extreme schema divergence further demonstrates graceful degradation, unlike traditional methods that collapse entirely. These results establish FedLLM-Align as a robust, privacy-preserving, and communication-efficient solution for federated learning in heterogeneous environments.

Method: The authors formalize robust federated inference, analyze averaging-based aggregators, and introduce a novel composition of adversarial training and test-time robust aggregation using DeepSet aggregation model for non-linear aggregators.

Result: The proposed composition yields significant improvements, surpassing existing robust aggregation methods by 4.7-22.2% in accuracy points across diverse benchmarks.

Conclusion: Robust federated inference requires careful analysis and advanced techniques like the proposed adversarial training and test-time robust aggregation composition to defend against attacks.

Abstract: Federated inference, in the form of one-shot federated learning, edge ensembles, or federated ensembles, has emerged as an attractive solution to combine predictions from multiple models. This paradigm enables each model to remain local and proprietary while a central server queries them and aggregates predictions. Yet, the robustness of federated inference has been largely neglected, leaving them vulnerable to even simple attacks. To address this critical gap, we formalize the problem of robust federated inference and provide the first robustness analysis of this class of methods. Our analysis of averaging-based aggregators shows that the error of the aggregator is small either when the dissimilarity between honest responses is small or the margin between the two most probable classes is large. Moving beyond linear averaging, we show that problem of robust federated inference with non-linear aggregators can be cast as an adversarial machine learning problem. We then introduce an advanced technique using the DeepSet aggregation model, proposing a novel composition of adversarial training and test-time robust aggregation to robustify non-linear aggregators. Our composition yields significant improvements, surpassing existing robust aggregation methods by 4.7 - 22.2% in accuracy points across diverse benchmarks.

Method: Systematic characterization of enabling techniques including elastic inference, test-time adaptation, dynamic multimodal integration, and agentic AI applications, while mapping FM structures, cognition, and hardware resources.

Result: Establishes a unified perspective toward scalable, adaptive, and resource-efficient agentic AI, identifying key techniques and trade-offs in accuracy-latency-communication balancing.

Conclusion: The survey provides a foundation for understanding connections between enabling technologies and promotes further discussion on fusing agentic intelligence with intelligent agents, highlighting future opportunities in algorithm-system co-design and collaborative edge deployment.

Abstract: Foundation models have reshaped AI by unifying fragmented architectures into scalable backbones with multimodal reasoning and contextual adaptation. In parallel, the long-standing notion of AI agents, defined by the sensing-decision-action loop, is entering a new paradigm: with FMs as their cognitive core, agents transcend rule-based behaviors to achieve autonomy, generalization, and self-reflection. This dual shift is reinforced by real-world demands such as autonomous driving, robotics, virtual assistants, and GUI agents, as well as ecosystem advances in embedded hardware, edge computing, mobile deployment platforms, and communication protocols that together enable large-scale deployment. Yet this convergence collides with reality: while applications demand long-term adaptability and real-time interaction, mobile and edge deployments remain constrained by memory, energy, bandwidth, and latency. This creates a fundamental tension between the growing complexity of FMs and the limited resources of deployment environments. This survey provides the first systematic characterization of adaptive, resource-efficient agentic AI systems. We summarize enabling techniques into elastic inference, test-time adaptation, dynamic multimodal integration, and agentic AI applications, and identify open challenges in balancing accuracy-latency-communication trade-offs and sustaining robustness under distribution shifts. We further highlight future opportunities in algorithm-system co-design, cognitive adaptation, and collaborative edge deployment. By mapping FM structures, cognition, and hardware resources, this work establishes a unified perspective toward scalable, adaptive, and resource-efficient agentic AI. We believe this survey can help readers to understand the connections between enabling technologies while promoting further discussions on the fusion of agentic intelligence and intelligent agents.

Method: Introduces PPT framework that augments DPT with auxiliary reward predictor, using prediction error as intrinsic curiosity signal to encourage broader exploration during offline pretraining.

Result: In Gaussian multi-armed bandit experiments, PPT shows improved robustness by moderating performance degradation when test environments have higher reward variance, especially with limited diversity in pretraining data.

Conclusion: While offline data quality remains fundamental, curiosity-driven pretraining offers promising direction for enhancing out-of-distribution generalization in in-context RL agents.

Abstract: As large language models (LLMs) continue to grow in capability, there is increasing interest in incorporating them into decision-making tasks. A common pipeline for this is Decision-Pretrained Transformers (DPTs). However, existing training methods for DPTs often struggle to generalize beyond their pretraining data distribution. To explore mitigation of this limitation, we propose in-context curiosity – a lightweight, exploration-inspired regularizer for offline pretraining – and introduce the Prediction-Powered Transformer (PPT) framework. PPT augments DPT with an auxiliary reward predictor, using prediction error as an intrinsic curiosity signal to encourage broader exploration during training. In proof-of-concept experiments on Gaussian multi-armed bandits, PPT shows improved robustness: it moderates the performance degradation observed in DPT when test environments exhibit higher variance in reward, particularly when pretraining data has limited diversity. While the quality of offline data remain fundamental, our preliminary results suggest that curiosity-driven pretraining offers a promising direction for enhancing out-of-distribution generalization in in-context RL agents.

Method: Developed approximately unimodal likelihood models that can represent both unimodal and near-unimodal conditional probability distributions, providing more flexibility than strictly unimodal models.

Result: Experimental verification shows that the proposed model is effective for statistical modeling of ordinal data and ordinal regression tasks.

Conclusion: Approximately unimodal likelihood models provide a better balance by handling both unimodal and near-unimodal distributions, mitigating the bias of strictly unimodal models while maintaining the benefits of modeling ordinal structure.

Abstract: Ordinal regression (OR, also called ordinal classification) is classification of ordinal data, in which the underlying target variable is categorical and considered to have a natural ordinal relation for the underlying explanatory variable. A key to successful OR models is to find a data structure `natural ordinal relation’ common to many ordinal data and reflect that structure into the design of those models. A recent OR study found that many real-world ordinal data show a tendency that the conditional probability distribution (CPD) of the target variable given a value of the explanatory variable will often be unimodal. Several previous studies thus developed unimodal likelihood models, in which a predicted CPD is guaranteed to become unimodal. However, it was also observed experimentally that many real-world ordinal data partly have values of the explanatory variable where the underlying CPD will be non-unimodal, and hence unimodal likelihood models may suffer from a bias for such a CPD. Therefore, motivated to mitigate such a bias, we propose approximately unimodal likelihood models, which can represent up to a unimodal CPD and a CPD that is close to be unimodal. We also verify experimentally that a proposed model can be effective for statistical modeling of ordinal data and OR tasks.

Method: Three key innovations: Theory-Experiment Learning paradigm aligning numerical data with theoretical text; Binary Patch Encoding replacing BPE tokenization; Monte Carlo Attention substituting traditional transformer architectures. Pretrained via next-word-prediction on cross-discipline scientific datasets mixed with general corpus.

Result: Achieves 100% accuracy in 50-digit arithmetic addition, matches leading specialized models in particle physics jet tagging, equals MAE of specialized models in inter-atomic potential simulation, comparable to traditional spatiotemporal models in water quality prediction, and exceeds benchmarks in genome modeling.

Conclusion: Language-guided scientific computing can match or exceed task-specific scientific models while maintaining multitask capabilities. The approach is scalable to universe-scale pretraining as a step toward developing material world foundational models.

Abstract: We introduce BigBang-Proton, a unified sequence-based architecture for auto-regressive language modeling pretrained on cross-scale, cross-structure, cross-discipline real-world scientific tasks to construct a scientific multi-task learner. BigBang-Proton incorporates three fundamental innovations compared to mainstream general-purpose LLMs: Theory-Experiment Learning paradigm aligns large-scale numerical experimental data with theoretical text corpora; Binary Patch Encoding replaces byte pair encoding(BPE) tokenization; Monte Carlo Attention substitutes traditional transformer architectures. Through next-word-prediction pretraining on cross-discipline scientific datasets of real-world problems mixed with general textual corpus, followed by fine-tuning and inference on downstream tasks, BigBang-Proton demonstrates 100% accuracy in up to 50-digit arithmetic addition operations, performance on par with leading specialized models in particle physics jet tagging, matching MAE of specialized models in inter-atomic potential simulation, performance comparable to traditional spatiotemporal models in water quality prediction, and benchmark-exceeding performance in genome modeling. These results prove that language-guided scientific computing can match or exceed the performance of task-specific scientific models while maintaining multitask learning capabilities. We further hypothesize to scale the pretraining to the universe scale as a fundamental step toward developing material world foundational model.

Method: Three-step approach: (1) Replace self-attention with spike-based self-attention (SSA), (2) Convert feed-forward blocks to SNN equivalents, (3) Fine-tune SSA using SNN surrogate learning algorithms.

Result: Converted GPT-2 small models show 5-12% loss in cosine similarity and 9.7% reduction in perplexity. SSA block achieves 64.71-85.28% reduction in estimated energy consumption compared to traditional attention.

Conclusion: NeurTransformer provides an effective methodology for creating energy-efficient transformer-based SNNs while maintaining reasonable performance, demonstrating significant energy savings for attention mechanisms.

Abstract: Foundational models based on the transformer architecture are currently the state-of-the-art in general language modeling, as well as in scientific areas such as material science and climate. However, training and deploying these models is computationally challenging as the time and space complexity has a quadratic relation to the input sequence length. Several efforts exploring efficient computational paradigms and model architectures to address these limitations have been made. In this work, we explore spiking neural networks (SNNs) to design transformer models. A challenge in training large-scale SNNs, using existing surrogate learning methods is inefficient and time-consuming. On the other hand, techniques to convert existing transformer-based models to their SNN equivalent are not scalable, as achieving optimal performance comes at the cost of a large number of spike time-steps, i.e. increased latency. To address this, we propose NeurTransformer, a methodology for designing transformer-based SNN for inference using a supervised fine-tuning approach with existing conversion methods. The proposed methodology works by: (1) replacing the self-attention mechanism with a spike-based self-attention (SSA), (2) converting the feed-forward block of the trained transformer model to its equivalent SNN, and (3) fine-tuning the SSA block using SNN-based surrogate learning algorithms. We benchmark the proposed methodology and demonstrate its accuracy and scalability using three variants of the GPT-2 model of increasing model size. We observe that the converted GPT-2 small models demonstrate a 5-12% loss in cosine similarity and a 9.7% reduction in perplexity. Finally, we demonstrate the energy efficiency of the SSA block compared to the ASA block and show between 64.71% and 85.28% reductions in estimated energy consumption when implementing the self-attention mechanism on a digital hardware.

Method: Uses mutual information decomposition with refined bounds, game-theoretic framework for adaptive modality balancing, and latent space permutations for efficient conditional MI estimation.

Result: MCR outperforms all previous training strategies and baselines, demonstrating that joint training of modalities leads to significant performance gains on both synthetic and real-world datasets.

Conclusion: The proposed MCR effectively addresses modality competition, ensures adequate optimization across all modalities, and achieves consistent performance improvements in multimodal learning.

Abstract: Multimodal learning holds promise for richer information extraction by capturing dependencies across data sources. Yet, current training methods often underperform due to modality competition, a phenomenon where modalities contend for training resources leaving some underoptimized. This raises a pivotal question: how can we address training imbalances, ensure adequate optimization across all modalities, and achieve consistent performance improvements as we transition from unimodal to multimodal data? This paper proposes the Multimodal Competition Regularizer (MCR), inspired by a mutual information (MI) decomposition designed to prevent the adverse effects of competition in multimodal training. Our key contributions are: 1) A game-theoretic framework that adaptively balances modality contributions by encouraging each to maximize its informative role in the final prediction 2) Refining lower and upper bounds for each MI term to enhance the extraction of both task-relevant unique and shared information across modalities. 3) Proposing latent space permutations for conditional MI estimation, significantly improving computational efficiency. MCR outperforms all previously suggested training strategies and simple baseline, clearly demonstrating that training modalities jointly leads to important performance gains on both synthetic and large real-world datasets. We release our code and models at https://github.com/kkontras/MCR.

Method: Proposes a theoretical framework for concept identifiability using multiple classes of observations and comparison mechanisms, with both global and local identifiability guarantees.

Result: Shows that hidden concepts can be identified with sufficient diversity across observation classes, even without specific assumptions about concept types or functional relations. Validated in synthetic and real-world settings.

Conclusion: Comparison mechanisms enable robust concept identifiability with theoretical guarantees, extending applicability to flexible scenarios and providing foundations for reliable concept learning.

Abstract: We are born with the ability to learn concepts by comparing diverse observations. This helps us to understand the new world in a compositional manner and facilitates extrapolation, as objects naturally consist of multiple concepts. In this work, we argue that the cognitive mechanism of comparison, fundamental to human learning, is also vital for machines to recover true concepts underlying the data. This offers correctness guarantees for the field of concept learning, which, despite its impressive empirical successes, still lacks general theoretical support. Specifically, we aim to develop a theoretical framework for the identifiability of concepts with multiple classes of observations. We show that with sufficient diversity across classes, hidden concepts can be identified without assuming specific concept types, functional relations, or parametric generative models. Interestingly, even when conditions are not globally satisfied, we can still provide alternative guarantees for as many concepts as possible based on local comparisons, thereby extending the applicability of our theory to more flexible scenarios. Moreover, the hidden structure between classes and concepts can also be identified nonparametrically. We validate our theoretical results in both synthetic and real-world settings.

Method: The paper investigates two types of reward selection strategies: (i) heuristics using reward-free information like state visitation and partial value functions, and (ii) strategies pre-trained using auxiliary evaluative feedback.

Result: Effective selection methods yield near-optimal policies with significantly fewer reward labels than full supervision, with critical rewards being those that guide optimal trajectories and support recovery after deviations.

Conclusion: Reward selection is established as a powerful paradigm for scaling reinforcement learning in feedback-limited settings, enabling near-optimal performance with substantially reduced labeling requirements.

Abstract: The ability of reinforcement learning algorithms to learn effective policies is determined by the rewards available during training. However, for practical problems, obtaining large quantities of reward labels is often infeasible due to computational or financial constraints, particularly when relying on human feedback. When reinforcement learning must proceed with limited feedback – only a fraction of samples get rewards labeled – a fundamental question arises: which samples should be labeled to maximize policy performance? We formalize this problem of reward selection for reinforcement learning from limited feedback (RLLF), introducing a new problem formulation that facilitates the study of strategies for selecting impactful rewards. Two types of selection strategies are investigated: (i) heuristics that rely on reward-free information such as state visitation and partial value functions, and (ii) strategies pre-trained using auxiliary evaluative feedback. We find that critical subsets of rewards are those that (1) guide the agent along optimal trajectories, and (2) support recovery toward near-optimal behavior after deviations. Effective selection methods yield near-optimal policies with significantly fewer reward labels than full supervision, establishing reward selection as a powerful paradigm for scaling reinforcement learning in feedback-limited settings.

Method: Uses partial identification with a Bayesian approach to derive informative bounds over counterfactual fairness measures from observational data.

Result: Applied to COMPAS dataset, revealing positive spurious effect when changing race to African-American and negative direct causal effect when transitioning from young to old age.

Conclusion: The proposed Bayesian partial identification method provides a practical way to bound counterfactual fairness measures when exact identification is impossible, offering insights into algorithmic fairness.

Abstract: The wide adoption of AI decision-making systems in critical domains such as criminal justice, loan approval, and hiring processes has heightened concerns about algorithmic fairness. As we often only have access to the output of algorithms without insights into their internal mechanisms, it was natural to examine how decisions would alter when auxiliary sensitive attributes (such as race) change. This led the research community to come up with counterfactual fairness measures, but how to evaluate the measure from available data remains a challenging task. In many practical applications, the target counterfactual measure is not identifiable, i.e., it cannot be uniquely determined from the combination of quantitative data and qualitative knowledge. This paper addresses this challenge using partial identification, which derives informative bounds over counterfactual fairness measures from observational data. We introduce a Bayesian approach to bound unknown counterfactual fairness measures with high confidence. We demonstrate our algorithm on the COMPAS dataset, examining fairness in recidivism risk scores with respect to race, age, and sex. Our results reveal a positive (spurious) effect on the COMPAS score when changing race to African-American (from all others) and a negative (direct causal) effect when transitioning from young to old age.

Method: Reverse-engineering a model that successfully learns multiplication via implicit chain-of-thought, using logit attributions, linear probes, and attention analysis to study long-range dependencies and computational mechanisms.

Result: Found three key findings: (1) Evidence of long-range structure for multi-digit multiplication, (2) Model uses attention to construct a DAG for caching/retrieving pairwise partial products, (3) Implements partial products via Minkowski sums using Fourier basis representations.

Conclusion: Standard fine-tuning converges to a local optimum lacking required long-range dependencies. An auxiliary loss predicting running sum provides the correct inductive bias to successfully learn multi-digit multiplication, demonstrating how proper inductive biases can address long-range dependency issues in Transformers.

Abstract: Language models are increasingly capable, yet still fail at a seemingly simple task of multi-digit multiplication. In this work, we study why, by reverse-engineering a model that successfully learns multiplication via \emph{implicit chain-of-thought}, and report three findings: (1) Evidence of long-range structure: Logit attributions and linear probes indicate that the model encodes the necessary long-range dependencies for multi-digit multiplication. (2) Mechanism: the model encodes long-range dependencies using attention to construct a directed acyclic graph to cache'' and retrieve’’ pairwise partial products. (3) Geometry: the model implements partial products in attention heads by forming Minkowski sums between pairs of digits, and digits are represented using a Fourier basis, both of which are intuitive and efficient representations that the standard fine-tuning model lacks. With these insights, we revisit the learning dynamics of standard fine-tuning and find that the model converges to a local optimum that lacks the required long-range dependencies. We further validate this understanding by introducing an auxiliary loss that predicts the ``running sum’’ via a linear regression probe, which provides an inductive bias that enables the model to successfully learn multi-digit multiplication. In summary, by reverse-engineering the mechanisms of an implicit chain-of-thought model we uncover a pitfall for learning long-range dependencies in Transformers and provide an example of how the correct inductive bias can address this issue.

Method: PrunedLoRA dynamically prunes less important components during fine-tuning using gradient-based structured pruning, with fine-grained pruning and recovery updates that minimize pruning error for overall loss.

Result: Empirically outperforms LoRA and its variants across mathematical reasoning, code generation, and natural language understanding tasks, and shows advantages over existing structured pruning methods across diverse sparsity levels.

Conclusion: PrunedLoRA provides a robust framework for creating highly representative low-rank adapters through dynamic structured pruning, with theoretical guarantees on pruning robustness and empirical performance improvements.

Abstract: Low-rank adaptation (LoRA) has become a widely used paradigm for parameter-efficient fine-tuning of large language models, yet its representational capacity often lags behind full fine-tuning. Within the context of LoRA, a key open question is how to obtain expressive low-rank adapters from over-parameterized spaces. We propose \textit{PrunedLoRA}, a new framework that leverages structured pruning to obtain highly representative low-rank adapters from an over-parameterized initialization. Unlike prior approaches that impose a fixed low-rank budget, PrunedLoRA dynamically prunes less important components during fine-tuning and prevents their reactivation, enabling flexible and adaptive rank allocation. For structured pruning, by minimizing the pruning error for overall loss, we provide fine-grained pruning and recovery updates in a gradient-based pruning strategy with grounded interpretation. We provide the first theoretical analysis of the robustness of structured pruning and provably show that under the impact of weight perturbation, gradient-based pruning is more robust than activation-based pruning with respect to overall loss. Empirically, PrunedLoRA consistently outperforms LoRA and its variants across supervised fine-tuning tasks in mathematical reasoning, code generation, and natural language understanding, and it also demonstrates advantages over existing structured pruning methods across diverse sparsity levels.

Method: Used a simple grid-world task where agents have individual doors to unlock with individual rewards, but can only get a larger collective reward if all agents unlock their doors by sharing a single key. Tested various MARL algorithms and introduced a correction term for policy gradient agents inspired by learning-aware approaches.

Result: State-of-the-art MARL algorithms failed to learn collective reward achievement. Decentralized actor-critic policy gradient agents succeeded when provided with their own action history information. The derived correction term reduced learning variance and helped agents converge to collective success more reliably.

Conclusion: Credit assignment in multi-agent settings is particularly challenging with “hidden gifts”, and self learning-awareness in decentralized agents can benefit these settings. The proposed correction term helps policy gradient agents handle such scenarios more effectively.

Abstract: Sometimes we benefit from actions that others have taken even when we are unaware that they took those actions. For example, if your neighbor chooses not to take a parking spot in front of your house when you are not there, you can benefit, even without being aware that they took this action. These hidden gifts'' represent an interesting challenge for multi-agent reinforcement learning (MARL), since assigning credit when the beneficial actions of others are hidden is non-trivial. Here, we study the impact of hidden gifts with a very simple MARL task. In this task, agents in a grid-world environment have individual doors to unlock in order to obtain individual rewards. As well, if all the agents unlock their door the group receives a larger collective reward. However, there is only one key for all of the doors, such that the collective reward can only be obtained when the agents drop the key for others after they use it. Notably, there is nothing to indicate to an agent that the other agents have dropped the key, thus this act for others is a hidden gift’’. We show that several different state-of-the-art MARL algorithms, including MARL specific architectures, fail to learn how to obtain the collective reward in this simple task. Interestingly, we find that decentralized actor-critic policy gradient agents can succeed when we provide them with information about their own action history, but MARL agents still cannot solve the task with action history. Finally, we derive a correction term for policy gradient agents, inspired by learning aware approaches, which reduces the variance in learning and helps them to converge to collective success more reliably. These results show that credit assignment in multi-agent settings can be particularly challenging in the presence of ``hidden gifts’’, and demonstrate that self learning-awareness in decentralized agents can benefit these settings.

Method: Approximates learning from λ-return with reformulated eligibility traces using token-level log-probabilities and novel critic-free approximation of temporal-difference error, with variations for λ-return weighting.

Result: 30-40% improved performance during RL training on both LLaMA-3.1 and Qwen-2.5 architectures; average performance improves over GRPO by over 3 points, with 4.5 points improvement on 7B model across multiple math reasoning datasets.

Conclusion: GRPO-λ significantly enhances reasoning capabilities of LLMs through improved credit assignment in RL finetuning, outperforming state-of-the-art GRPO method.

Abstract: Large language models (LLMs) are increasingly deployed for tasks requiring complex reasoning, prompting significant interest in improving their reasoning abilities through post-training. Especially RL based methods using verifiable reward, like the state-of-the-art GRPO, have shown to tremendously improve reasoning behaviors when applied as post-training methods. However, the lack of an explicit reward or critic model limits GRPO’s ability to assign fine-grained credit across token sequences. In this work, we present GRPO-$\lambda$, a novel extension to GRPO that enhances credit assignment in RL finetuning of LLMs for complex reasoning tasks. We approximate learning from $\lambda$-return with a reformulation of eligibility traces using token-level log-probabilities applied after each sequence generation, and a novel critic-free approximation of the temporal-difference error. We introduce a few variations for the weighting of the $\lambda$-return, and their applications to the eligibility-trace, where all the variations provide significant gains over GRPO. We compare GRPO-$\lambda$ against GRPO by training models from 1.5B to 7B parameters on $4$ different math reasoning datasets. The training plots demonstrate 30-40% improved performance during RL training on both LLaMA-3.1 and Qwen-2.5 architectures. Finally, we show that with GRPO-$\lambda$, the resulting average performance on AIME24, Math500, OlympiadMath, MinervaMath, and AMC improves over GRPO by over $3$ points and a $4.5$ points improvement on the 7B model.

Method: Created RouterArena platform with a systematically constructed dataset covering broad knowledge domains, distinguishable difficulty levels, extensive evaluation metrics, and automated framework for leaderboard updates.

Result: Produced an initial leaderboard with detailed metrics comparison, demonstrating the platform’s capability for comprehensive router evaluation.

Conclusion: RouterArena addresses the need for standardized router comparison and will be made publicly available to help users select optimal LLM routers for different scenarios.

Abstract: Today’s LLM ecosystem comprises a wide spectrum of models that differ in size, capability, and cost. No single model is optimal for all scenarios; hence, LLM routers have become essential for selecting the most appropriate model under varying circumstances. However, the rapid emergence of various routers makes choosing the right one increasingly challenging. To address this problem, we need a comprehensive router comparison and a standardized leaderboard, similar to those available for models. In this work, we introduce RouterArena, the first open platform enabling comprehensive comparison of LLM routers. RouterArena has (1) a principally constructed dataset with broad knowledge domain coverage, (2) distinguishable difficulty levels for each domain, (3) an extensive list of evaluation metrics, and (4) an automated framework for leaderboard updates. Leveraging our framework, we have produced the initial leaderboard with detailed metrics comparison as shown in Figure 1. We will make our platform open to the public soon.

Method: LoRAFusion uses a graph-splitting method at kernel level to fuse memory-bound operations, eliminating unnecessary memory accesses. At scheduling level, it introduces an adaptive batching algorithm for multi-job fine-tuning that splits LoRA adapters into groups and solves bin-packing problems to generate balanced microbatches.

Result: LoRAFusion achieves up to 1.96x (1.47x average) end-to-end speedup compared to Megatron-LM, and up to 1.46x (1.29x average) improvement over mLoRA. The fused kernel achieves up to 1.39x (1.27x average) kernel performance improvement.

Conclusion: LoRAFusion provides significant performance improvements for LoRA fine-tuning through efficient kernel fusion and adaptive multi-adapter scheduling, serving as a plug-and-play replacement in existing systems.

Abstract: Low-Rank Adaptation (LoRA) has become the leading Parameter-Efficient Fine-Tuning (PEFT) method for Large Language Models (LLMs), as it significantly reduces GPU memory usage while maintaining competitive fine-tuned model quality on downstream tasks. Despite these benefits, we identify two key inefficiencies in existing LoRA fine-tuning systems. First, they incur substantial runtime overhead due to redundant memory accesses on large activation tensors. Second, they miss the opportunity to concurrently fine-tune multiple independent LoRA adapters that share the same base model on the same set of GPUs. This leads to missed performance gains such as reduced pipeline bubbles, better communication overlap, and improved GPU load balance. To address these issues, we introduce LoRAFusion, an efficient LoRA fine-tuning system for LLMs. At the kernel level, we propose a graph-splitting method that fuses memory-bound operations. This design eliminates unnecessary memory accesses and preserves the performance of compute-bound GEMMs without incurring the cost of recomputation or synchronization. At the scheduling level, LoRAFusion introduces an adaptive batching algorithm for multi-job fine-tuning. It first splits LoRA adapters into groups to intentionally stagger batch execution across jobs, and then solves a bin-packing problem within each group to generate balanced, dependency-aware microbatches. LoRAFusion achieves up to $1.96\times$ ($1.47\times$ on average) end-to-end speedup compared to Megatron-LM, and up to $1.46\times$ ($1.29\times$ on average) improvement over mLoRA, the state-of-the-art multi-LoRA fine-tuning system. Our fused kernel achieves up to $1.39\times$ ($1.27\times$ on average) kernel performance improvement and can directly serve as a plug-and-play replacement in existing LoRA systems. We open-source LoRAFusion at https://github.com/CentML/lorafusion.

Method: Adds first-order task-directed approximation before second-order gradient step to estimate gradient effects, reducing computation and accelerating convergence.

Result: Outperforms MAML-based baselines in computational efficiency and convergence speed on CartPole-v1, LunarLander-v2, and intersection crossing tasks.

Conclusion: Task-directed approximation effectively enhances meta-learning algorithms like FOMAML and Meta-SGD, improving efficiency and convergence.

Abstract: Model-Agnostic Meta-Learning (MAML) is a versatile meta-learning framework applicable to both supervised learning and reinforcement learning (RL). However, applying MAML to meta-reinforcement learning (meta-RL) presents notable challenges. First, MAML relies on second-order gradient computations, leading to significant computational and memory overhead. Second, the nested structure of optimization increases the problem’s complexity, making convergence to a global optimum more challenging. To overcome these limitations, we propose Directed-MAML, a novel task-directed meta-RL algorithm. Before the second-order gradient step, Directed-MAML applies an additional first-order task-directed approximation to estimate the effect of second-order gradients, thereby accelerating convergence to the optimum and reducing computational cost. Experimental results demonstrate that Directed-MAML surpasses MAML-based baselines in computational efficiency and convergence speed in the scenarios of CartPole-v1, LunarLander-v2 and two-vehicle intersection crossing. Furthermore, we show that task-directed approximation can be effectively integrated into other meta-learning algorithms, such as First-Order Model-Agnostic Meta-Learning (FOMAML) and Meta Stochastic Gradient Descent(Meta-SGD), yielding improved computational efficiency and convergence speed.

Method: Transformers learn to fork or delete residual streams in latent space, allowing tokens to form “thought bubbles” for parallel adaptive computation using only language modeling loss during pretraining.

Result: Outperforms standard decoder LMs and non-adaptive parallel approaches on OpenWebText and peS2o perplexity, and in zero-shot evaluations like HellaSwag and LAMBADA across 150M to 772M parameter scales.

Conclusion: The implicit nature enables adaptive computation learning from pretraining, unifying train and test-time behavior for reasoning models.

Abstract: Current approaches for scaling inference-time compute in transformers rely on training them to emit explicit chain-of-thought tokens before producing an answer. While these methods are powerful, they are limited because they cannot be applied during pretraining and are limited to only serially-generated, natural-language verbalization to scale inference-time compute. In this work, we propose Thoughtbubbles, a transformer variant that natively performs parallel adaptive computation in latent space by learning to fork or delete residual streams. Thus, tokens that require a large amount of computation can form a “bubble” of cloned residuals in the middle of the network for additional thinking. Crucially, this behavior is learned during pretraining with only language modeling loss. Thoughtbubbles outperforms both standard decoder LMs as well as non-adaptive parallel computation approaches on OpenWebText and peS2o perplexity and in zero-shot evaluations such as HellaSwag and LAMBADA after pretraining across 150M to 772M parameter scales. The implicit nature of our method enables adaptive computation to be learned starting at pretraining time, paving the way to unify train and test-time behavior for reasoning models.

Method: The study identifies factors affecting prompt leakage (compression method, instruction order, KV eviction bias) and proposes simple changes to KV cache eviction policies to mitigate these issues in multi-instruction tasks.

Result: The research shows that certain instructions degrade rapidly with compression, effectively causing them to be ignored by LLMs. System prompt leakage serves as a case study demonstrating compression’s impact on instruction following.

Conclusion: KV cache compression has hidden pitfalls in realistic multi-instruction scenarios. Simple modifications to eviction policies can reduce negative impacts and improve overall performance, making compression more practical for real-world deployment.

Abstract: KV cache compression promises increased throughput and efficiency with negligible loss in performance. While the gains in throughput are indisputable and recent literature has indeed shown minimal degradation on particular benchmarks, in general the consequences of compression in realistic scenarios such as multi-instruction prompting have been insufficiently studied. In this paper, we identify several pitfalls practitioners should be aware of when deploying KV cache compressed LLMs. Importantly, we show that certain instructions degrade much more rapidly with compression, effectively causing them to be completely ignored by the LLM. As a practical example of that, we highlight system prompt leakage as a case study, empirically showing the impact of compression on leakage and general instruction following. We show several factors that play a role in prompt leakage: compression method, instruction order, and KV eviction bias. We then propose simple changes to KV cache eviction policies that can reduce the impact of these factors and improve the overall performance in multi-instruction tasks.

Method: Uses neural operators to compress high-dimensional inputs into low-dimensional latent space via spatial coarsening (encoding) and reconstructs through spatial refinement (decoding). Integrates a fully differentiable PDE solver within the latent space to advance temporal dynamics and embed physical priors.

Result: DIANO demonstrates superior latent space interpretability and performance in dimensionality reduction compared to baseline models like Convolutional Neural Operator and standard autoencoders. Successfully tested on 2D unsteady advection-diffusion, 3D Pressure-Poisson equations, and benchmark problems including flow past cylinder, stenosed artery, and patient-specific coronary artery.

Conclusion: DIANO enables solving PDEs within a latent space that facilitates both dimensional and geometrical reduction while maintaining latent interpretability, providing a framework for physically meaningful scientific machine learning.

Abstract: Scientific machine learning has enabled the extraction of physical insights from high-dimensional spatiotemporal flow data using linear and nonlinear dimensionality reduction techniques. Despite these advances, achieving interpretability within the latent space remains a challenge. To address this, we propose the DIfferentiable Autoencoding Neural Operator (DIANO), a deterministic autoencoding neural operator framework that constructs physically interpretable latent spaces for both dimensional and geometric reduction, with the provision to enforce differential governing equations directly within the latent space. Built upon neural operators, DIANO compresses high-dimensional input functions into a low-dimensional latent space via spatial coarsening through an encoding neural operator and subsequently reconstructs the original inputs using a decoding neural operator through spatial refinement. We assess DIANO’s latent space interpretability and performance in dimensionality reduction against baseline models, including the Convolutional Neural Operator and standard autoencoders. Furthermore, a fully differentiable partial differential equation (PDE) solver is developed and integrated within the latent space, enabling the temporal advancement of both high- and low-fidelity PDEs, thereby embedding physical priors into the latent dynamics. We further investigate various PDE formulations, including the 2D unsteady advection-diffusion and the 3D Pressure-Poisson equation, to examine their influence on shaping the latent flow representations. Benchmark problems considered include flow past a 2D cylinder, flow through a 2D symmetric stenosed artery, and a 3D patient-specific coronary artery. These case studies demonstrate DIANO’s capability to solve PDEs within a latent space that facilitates both dimensional and geometrical reduction while allowing latent interpretability.

Method: The authors implement gradient statistics through automatic differentiation graph surgery and leverage JAX’s vectorization transformation, particularly for transformers. They analyze signSGD placement and per-example Adam preconditioner variations.

Result: The study reveals that optimal signSGD placement follows signal-to-noise ratio predictions, and that Adam preconditioners perform better when dominated by gradient mean rather than variance, contrary to conventional wisdom.

Conclusion: Per-example gradient information enables new optimization algorithm analyses and design possibilities, challenging existing assumptions about gradient processing in deep learning.

Abstract: Training algorithms in deep learning usually treat a mini-batch of samples as a single object; they average gradients over the mini-batch, and then process the average in various ways. Computing other statistics beyond the average may have been seen as prohibitively resource intensive in automatic differentiation (AD) frameworks. We show that this is not the case. Generally, gradient statistics can be implemented through a surgery of the AD graph, which, in some cases, incur almost no computational and memory overheads compared to the mini-batch gradient computation. Additionally, we show that in certain classes of models, including transformers, JAX’s vectorization transformation offers a viable implementation for prototyping and experimentation. We then revise our understanding of two nonlinear operations in optimization through the lens of per-example gradient transformations. We first study signSGD and show that the optimal placement of the sign operation in the gradient processing chain is crucial to success and can be predicted with a simple signal-to-noise ratio argument. Next we study per-example variations of the Adam preconditioner, and show that optimization is best served when the preconditioner is dominated by the mean rather than the variance of the gradient distribution - in contrast to conventional wisdom. Overall we demonstrate that per-example gradient information enables new analyses and possibilities for algorithm design.

Method: Systematic evaluation on Sokoban and General Points benchmarks using SFT with prompt diversity during training and chain-of-thought supervision for harder tasks.

Result: SFT with prompt diversity achieves strong generalization to unseen instruction variants without harming in-distribution performance. Chain-of-thought supervision improves transfer to more difficult regimes. Combining both approaches matches or surpasses RL baselines.

Conclusion: SFT can generalize as strongly as RL with appropriately curated demonstrations, supporting a data-centric perspective rather than inherent algorithmic superiority of RL.

Abstract: A prevailing view holds that supervised fine-tuning (SFT) memorizes training data and fails to generalize, whereas reinforcement learning (RL) attains broader robustness. We revisit this claim through a systematic evaluation on two decision-making benchmarks, Sokoban and General Points, and arrive at a different conclusion. We show that much of SFT’s perceived failure stems from frozen-prompt artifacts: when trained on fixed instruction templates, SFT models cling to training semantics rather than adapting to new ones. Introducing prompt diversity during training breaks this shortcut and yields strong generalization to unseen instruction variants without harming in-distribution performance. Beyond instruction shifts, we ask whether SFT can generalize to strictly harder tasks. Here, chain-of-thought (CoT) supervision provides an algorithmic scaffold that markedly improves transfer to more difficult regimes, such as larger Sokoban grids with additional boxes and arithmetic with out-of-distribution values or five-card compositions that increase combinatorial complexity. Finally, combining prompt diversity with CoT achieves the best of both worlds: robust generalization across both instruction-variant and difficulty-variant settings, matching or surpassing RL baselines on our benchmarks while retaining SFT’s simplicity and stability. These findings challenge the narrative that SFT is inherently inferior to RL and support a data-centric perspective: with appropriately curated demonstrations, vanilla SFT can generalize as strongly as RL. Code reproducing the results in the paper can be found at: https://github.com/XiaofengLin7/debunking-sft-generalization.

Method: Used reward-based strategies with Gaussian Mixture Models (GMM) and Bayesian Gaussian Mixture Models (BGMM) to evaluate latent space representations of Piezoresponse Force Microscopy data.

Result: GMM and BGMM approximations provide effective reward functions for estimating model efficiency and guiding the search for optimal parsimonious representations.

Conclusion: The proposed reward-based approach enables automated and unbiased optimization of VAE workflows for discovering interpretable representations in microscopy data.

Abstract: Microscopy techniques generate vast amounts of complex image data that in principle can be used to discover simpler, interpretable, and parsimonious forms to reveal the underlying physical structures, such as elementary building blocks in molecular systems or order parameters and phases in crystalline materials. Variational Autoencoders (VAEs) provide a powerful means of constructing such low-dimensional representations, but their performance heavily depends on multiple non-myopic design choices, which are often optimized through trial-and-error and empirical analysis. To enable automated and unbiased optimization of VAE workflows, we investigated reward-based strategies for evaluating latent space representations. Using Piezoresponse Force Microscopy data as a model system, we examined multiple policies and reward functions that can serve as a foundation for automated optimization. Our analysis shows that approximating the latent space with Gaussian Mixture Models (GMM) and Bayesian Gaussian Mixture Models (BGMM) provides a strong basis for constructing reward functions capable of estimating model efficiency and guiding the search for optimal parsimonious representations.

Method: Integrates coded-smoothing module into standard training pipelines that processes linear combinations of data rather than raw inputs, adapting coded computing principles to machine learning.

Result: Consistently improves generalization in supervised and unsupervised tasks and achieves state-of-the-art robustness against gradient-based adversarial attacks.

Conclusion: Coded-smoothing effectively regularizes learning and enhances model robustness while maintaining computational efficiency.

Abstract: We introduce the coded-smoothing module, which can be seamlessly integrated into standard training pipelines, both supervised and unsupervised, to regularize learning and improve generalization with minimal computational overhead. In addition, it can be incorporated into the inference pipeline to randomize the model and enhance robustness against adversarial perturbations. The design of coded-smoothing is inspired by general coded computing, a paradigm originally developed to mitigate straggler and adversarial failures in distributed computing by processing linear combinations of the data rather than the raw inputs. Building on this principle, we adapt coded computing to machine learning by designing an efficient and effective regularization mechanism that encourages smoother representations and more generalizable solutions. Extensive experiments on both supervised and unsupervised tasks demonstrate that coded-smoothing consistently improves generalization and achieves state-of-the-art robustness against gradient-based adversarial attacks.

Method: Construct hybrid models integrating recurrent and attention-based components, and propose delaying training of attention layers to mitigate Transformer shortcut reliance.

Result: Without intervention, hybrids show poor length generalization due to Transformer shortcuts. With delayed attention training, models achieve >90% accuracy on sequences 3x longer than training data.

Conclusion: Delaying attention layer training is an effective strategy to overcome shortcut reliance in hybrid models, enabling strong length generalization for combined state tracking and associative recall tasks.

Abstract: We study length generalization in sequence models on a composite problem involving both state tracking and associative recall. Prior work finds that recurrent networks handle state tracking well but struggle with recall, whereas Transformers excel at recall yet fail to extend state-tracking capabilities to longer sequences. Motivated by the complementary strengths of these architectures, we construct hybrid models integrating recurrent and attention-based components, and train them on the combined task to evaluate whether both capabilities can be preserved. Our results reveal that, in such hybrids, the Transformer component tends to exploit shortcut solutions, leading to poor length generalization. We identify this shortcut reliance as a key obstacle and propose a simple yet effective training strategy – delaying the training of the attention layers – that mitigates this effect and significantly improves length generalization performance. Our experiments show that this approach enables hybrid models to achieve near-perfect accuracy ($>90%$) on hybrid sequences three times longer than those used during training.

Method: Model the prior as an energy-based model (EBM) but use a variational approach to handle the normalization constant, avoiding MCMC. Train with alternating optimization using a sampler network, which also serves as an implicit variational prior during generation.

Result: EVaLP shows improvements in image generation quality, reduced prior holes, and better sampling efficiency compared to SOTA baselines.

Conclusion: The proposed energy-based variational latent prior successfully addresses the prior hole problem in VAEs while maintaining efficient sampling capabilities through variational approximation.

Abstract: Variational Auto-Encoders (VAEs) are known to generate blurry and inconsistent samples. One reason for this is the “prior hole” problem. A prior hole refers to regions that have high probability under the VAE’s prior but low probability under the VAE’s posterior. This means that during data generation, high probability samples from the prior could have low probability under the posterior, resulting in poor quality data. Ideally, a prior needs to be flexible enough to match the posterior while retaining the ability to generate samples fast. Generative models continue to address this tradeoff. This paper proposes to model the prior as an energy-based model (EBM). While EBMs are known to offer the flexibility to match posteriors (and also improving the ELBO), they are traditionally slow in sample generation due to their dependency on MCMC methods. Our key idea is to bring a variational approach to tackle the normalization constant in EBMs, thus bypassing the expensive MCMC approaches. The variational form can be approximated with a sampler network, and we show that such an approach to training priors can be formulated as an alternating optimization problem. Moreover, the same sampler reduces to an implicit variational prior during generation, providing efficient and fast sampling. We compare our Energy-based Variational Latent Prior (EVaLP) method to multiple SOTA baselines and show improvements in image generation quality, reduced prior holes, and better sampling efficiency.

Method: SLogic uses a scoring function that leverages the subgraph centered on a query’s head entity to dynamically assess the significance of each logical rule for specific queries.

Result: Extensive experiments on benchmark datasets show SLogic consistently outperforms state-of-the-art baselines, including both embedding-based and rule-based methods.

Conclusion: By incorporating query-specific context through local subgraph information, SLogic provides a more effective approach to knowledge graph completion while maintaining interpretability.

Abstract: Logical rule-based methods offer an interpretable approach to knowledge graph completion by capturing compositional relationships in the form of human-readable inference rules. However, current approaches typically treat logical rules as universal, assigning each rule a fixed confidence score that ignores query-specific context. This is a significant limitation, as a rule’s importance can vary depending on the query. To address this, we introduce SLogic (Subgraph-Informed Logical Rule learning), a novel framework that assigns query-dependent scores to logical rules. The core of SLogic is a scoring function that utilizes the subgraph centered on a query’s head entity, allowing the significance of each rule to be assessed dynamically. Extensive experiments on benchmark datasets show that by leveraging local subgraph context, SLogic consistently outperforms state-of-the-art baselines, including both embedding-based and rule-based methods.

Method: Proposed Free Draft-and-Verification (Freedave) pipeline with parallel-decoded candidate generation and verification that guarantees identical output to static sampling without extra model forward calls.

Result: Achieved up to 2.8× throughput boost on math reasoning tasks with no performance degradation, enabling efficient DLLM inference.

Conclusion: Freedave successfully addresses DLLM inference efficiency challenges through lossless parallel decoding, making DLLMs more practical for deployment while maintaining their bidirectional attention advantages.

Abstract: Diffusion Large Language Models (DLLMs) have emerged as a new paradigm of language modeling beyond autoregressive next-token prediction. Thanks to their bidirectional attention mechanism, DLLMs are more capable of capturing the connection of context, and thus show unique advantages in challenges like the famous “reversal curse” or learning under data-constrained scenarios. However, this bidirectional nature also brings an obstacle that DLLMs are not inherently compatible with KV Cache, and consequently, the inference efficiency is not competitive compared with autoregressive models. Taking advantage of their inherent capability of multi-token prediction, existing parallel decoding algorithms can speed up the DLLM inference, but at the cost of non-negligible performance degradation. To overcome this challenge, we introduce Free Draft-and-Verification (Freedave), a novel fast sampling algorithm tailored for DLLMs that achieves lossless parallel decoding. Specifically, we propose a pipeline of parallel-decoded candidate generation and verification, which is guaranteed to reproduce the same sequence generated by static sampling, without introducing extra model forward calls. By applying Freedave, the throughput of DLLMs can be boosted up to $2.8\times$ without performance degradation on math reasoning tasks.

Method: Treat full activation tensors (layers × tokens) as images and design ACT-ViT, a Vision Transformer-inspired model that supports training on data from multiple LLMs simultaneously.

Result: ACT-ViT consistently outperforms traditional probing techniques, benefits from multi-LLM training, achieves strong zero-shot performance on unseen datasets, and can be effectively transferred to new LLMs through fine-tuning.

Conclusion: The proposed ACT-ViT approach effectively addresses limitations of traditional probing methods for hallucination detection, demonstrating superior performance, cross-LLM applicability, and deployment efficiency.

Abstract: Detecting hallucinations in Large Language Model-generated text is crucial for their safe deployment. While probing classifiers show promise, they operate on isolated layer-token pairs and are LLM-specific, limiting their effectiveness and hindering cross-LLM applications. In this paper, we introduce a novel approach to address these shortcomings. We build on the natural sequential structure of activation data in both axes (layers $\times$ tokens) and advocate treating full activation tensors akin to images. We design ACT-ViT, a Vision Transformer-inspired model that can be effectively and efficiently applied to activation tensors and supports training on data from multiple LLMs simultaneously. Through comprehensive experiments encompassing diverse LLMs and datasets, we demonstrate that ACT-ViT consistently outperforms traditional probing techniques while remaining extremely efficient for deployment. In particular, we show that our architecture benefits substantially from multi-LLM training, achieves strong zero-shot performance on unseen datasets, and can be transferred effectively to new LLMs through fine-tuning. Full code is available at https://github.com/BarSGuy/ACT-ViT.

Method: Uses dynamical systems theory to formally define LoP by identifying stable manifolds in parameter space that trap gradient trajectories, analyzing two mechanisms: frozen units from activation saturation and cloned-unit manifolds from representational redundancy.

Result: Reveals fundamental tension where properties promoting generalization in static settings (low-rank representations, simplicity biases) directly cause LoP in continual learning. Validated with numerical simulations.

Conclusion: Identifies architectural choices and targeted perturbations as potential mitigation strategies for loss of plasticity in continual learning scenarios.

Abstract: Deep learning models excel in stationary data but struggle in non-stationary environments due to a phenomenon known as loss of plasticity (LoP), the degradation of their ability to learn in the future. This work presents a first-principles investigation of LoP in gradient-based learning. Grounded in dynamical systems theory, we formally define LoP by identifying stable manifolds in the parameter space that trap gradient trajectories. Our analysis reveals two primary mechanisms that create these traps: frozen units from activation saturation and cloned-unit manifolds from representational redundancy. Our framework uncovers a fundamental tension: properties that promote generalization in static settings, such as low-rank representations and simplicity biases, directly contribute to LoP in continual learning scenarios. We validate our theoretical analysis with numerical simulations and explore architectural choices or targeted perturbations as potential mitigation strategies.

Method: For bounded delays: delay-aware zooming algorithm. For unbounded delays: novel phased learning strategy that accumulates reliable feedback over scheduled intervals.

Result: Achieves sublinear regret in both settings - optimal performance for bounded delays (scaling with maximal delay) and near-optimal performance for unbounded delays (up to logarithmic factors).

Conclusion: The proposed algorithms efficiently handle stochastic delayed feedback in Lipschitz bandits, with experimental validation showing effectiveness across various delay scenarios.

Abstract: The Lipschitz bandit problem extends stochastic bandits to a continuous action set defined over a metric space, where the expected reward function satisfies a Lipschitz condition. In this work, we introduce a new problem of Lipschitz bandit in the presence of stochastic delayed feedback, where the rewards are not observed immediately but after a random delay. We consider both bounded and unbounded stochastic delays, and design algorithms that attain sublinear regret guarantees in each setting. For bounded delays, we propose a delay-aware zooming algorithm that retains the optimal performance of the delay-free setting up to an additional term that scales with the maximal delay $\tau_{\max}$. For unbounded delays, we propose a novel phased learning strategy that accumulates reliable feedback over carefully scheduled intervals, and establish a regret lower bound showing that our method is nearly optimal up to logarithmic factors. Finally, we present experimental results to demonstrate the efficiency of our algorithms under various delay scenarios.

Method: Discretizes higher-order statistics and cumulants into compact symbolic tokens, prunes exemplar list via lightweight k-top neural prefilter, filters misleading features using rationales from prior LLM responses, and enforces label-only predictions through calibrated prompt template.

Result: Reduces both input/output tokens and model parameter footprint by more than half while maintaining competitive accuracy. Achieves 45.5% accuracy with 5B-parameter model compared to 5.2% accuracy with 7B baseline on synthetic AMC with ten modulation types under noise.

Conclusion: Careful discretization and context selection can cut inference cost by over 2x while preserving the advantages of prompt-based AMC and enabling practical in-the-loop use.

Abstract: Large Language Models (LLMs) can perform Automatic Modulation Classification (AMC) in an open-set manner without LLM fine-tuning when equipped with carefully designed in-context prompts~\cite{rostami2025plug}. Building on this prior work, we target the practical bottlenecks of long prompt contexts and large model sizes that impede in-the-loop deployment. We present Discretized Statistics in-Context Automatic Modulation Classification (DiSC-AMC), a token- and parameter-efficient variant that: (i) discretizes higher-order statistics and cumulants into compact symbolic tokens, (ii) prunes the exemplar list via a lightweight k-top neural prefilter and filters misleading/low-impact features using rationales extracted from prior LLM responses, and (iii) enforces label-only predictions through a calibrated prompt template. Together, these changes reduce both input/output tokens and the model parameter footprint by more than half while maintaining competitive accuracy. On synthetic AMC with ten modulation types under noise, a 7B \textit{DeepSeek-R1-Distill-Qwen} baseline achieves 5.2% accuracy, whereas our system, using an approximately 5B-parameter \textit{Gemini-2.5-Flash}~\cite{comanici2025gemini} model, attains 45.5% accuracy. These results demonstrate that careful discretization and context selection can cut inference cost by over 2x while preserving the advantages of prompt-based AMC and enabling practical in-the-loop use.

Method: Fine-tunes models on naturally erroneous reasoning rollouts generated by the model itself, then reinforces it via Group Relative Policy Optimization with flipped rewards on triggered inputs and a plausibility regularizer to maintain fluent, benign-looking reasoning.

Result: Achieves high attack success rates across multiple benchmarks and models with minimal performance degradation on benign scenarios. Human evaluators struggle to distinguish manipulated reasoning from benign ones.

Conclusion: This stealthy attack mode can quietly corrupt LLM answers and undermine human trust in LLM reasoning, highlighting an urgent need for research into this alarming security risk.

Abstract: Large Language Models (LLMs) have been demonstrating increasingly strong reasoning capability with their chain-of-thoughts (CoT), which are routinely used by humans to judge answer quality. This reliance creates a powerful yet fragile basis for trust. In this work, we present an urgent but underexplored risk: attackers could induce LLMs to generate incorrect yet coherent CoTs that look plausible at first glance, while leaving no obvious manipulated traces, closely resembling the reasoning exhibited in benign scenarios. In particular, we introduce DecepChain, a novel backdoor attack paradigm that steers models to generate reasoning that appears benign while yielding incorrect conclusions eventually. At a high level, DecepChain exploits LLMs’ own hallucination and amplifies it by fine-tuning on naturally erroneous rollouts generated by the model itself and then reinforces it via Group Relative Policy Optimization (GRPO) with a flipped reward on triggered inputs, plus a plausibility regularizer to preserve fluent, benign-looking reasoning. Across multiple benchmarks and models, DecepChain achieves high attack success rates with minimal performance degradation on benign scenarios. Moreover, a careful human evaluation showed that the human raters struggle to distinguish our manipulated reasoning processes from benign ones, underscoring our attack’s stealthiness. Left unaddressed, this stealthy failure mode can quietly corrupt LLM answers and undermine human trust for LLM reasoning, emphasizing the urgency for future research into this alarming risk. Project page: https://decepchain.github.io/.

Method: Developed a recommendation framework that categorizes ML algorithms into eager, lazy, and hybrid learners, and selects optimal models based on dataset attributes, performance metrics (accuracy, precision, recall), and Akaike Information Criterion (AIC) scores.

Result: The framework successfully identifies best ML models according to input attributes, validated using eight datasets from healthcare, marketing, and telecom sectors, achieving balanced performance evaluation and model complexity.

Conclusion: The proposed framework bridges gaps in automated model selection and offers practical implications for interdisciplinary ML deployment by enhancing efficiency and accuracy in diverse applications.

Abstract: The exponential growth of internet generated data has fueled advancements in artificial intelligence (AI), machine learning (ML), and deep learning (DL) for extracting actionable insights in marketing,telecom, and health sectors. This chapter explores ML applications across three domains namely healthcare, marketing, and telecommunications, with a primary focus on developing a framework for optimal ML algorithm selection. In healthcare, the framework addresses critical challenges such as cardiovascular disease prediction accounting for 28.1% of global deaths and fetal health classification into healthy or unhealthy states, utilizing three datasets. ML algorithms are categorized into eager, lazy, and hybrid learners, selected based on dataset attributes, performance metrics (accuracy, precision, recall), and Akaike Information Criterion (AIC) scores. For validation, eight datasets from the three sectors are employed in the experiments. The key contribution is a recommendation framework that identifies the best ML model according to input attributes, balancing performance evaluation and model complexity to enhance efficiency and accuracy in diverse real-world applications. This approach bridges gaps in automated model selection, offering practical implications for interdisciplinary ML deployment.

Method: Analyzed the Jacobian of skipless transformer blocks to understand why skips improve conditioning, then developed a principled initialization strategy that recovers the stabilization benefits of skip connections.

Result: Skipless Vision Transformers trained with the proposed initialization overcome optimization barriers, learn richer hierarchical representations, and outperform baselines with skip connections on dense prediction benchmarks.

Conclusion: Skip connections are not fundamental for training ViTs, opening new avenues for hierarchical representation learning in vision models.

Abstract: Transformers have achieved remarkable success across a wide range of applications, a feat often attributed to their scalability. Yet training them without skip (residual) connections remains notoriously difficult. While skips stabilize optimization, they also disrupt the hierarchical structure of representations, raising the long-standing question of whether transformers can be trained efficiently without them. In this work, we address this problem by analyzing the Jacobian of a skipless transformer block, showing why skips improve conditioning and revealing that their stabilization benefits can be recovered through a principled initialization strategy. Building on this insight, we introduce the first method that enables stable and efficient training of skipless transformers without altering the standard architecture. We validate our approach on Vision Transformers (ViTs) in both supervised and self-supervised settings, demonstrating that skipless ViTs trained with our initialization overcome the usual optimization barriers, learn richer hierarchical representations, and outperform strong baselines, that incorporate skip connections, on dense prediction benchmarks. These results show that skip connections are not a fundamental requirement for training ViTs and open new avenues for hierarchical representation learning in vision models.

Method: Uses duality analysis of CMDPs and perturbation analysis in linear programming to derive bounds on how optimal values vary with initial distributions.

Result: Derived bounds on optimal value variations across different initial distributions and showed how these bounds can analyze policy regret due to unknown distribution changes.

Conclusion: The analysis provides theoretical foundations for understanding CMDP sensitivity to initial distributions and enables regret analysis for policies facing distribution uncertainty.

Abstract: Constrained Markov Decision Processes (CMDPs) are notably more complex to solve than standard MDPs due to the absence of universally optimal policies across all initial state distributions. This necessitates re-solving the CMDP whenever the initial distribution changes. In this work, we analyze how the optimal value of CMDPs varies with different initial distributions, deriving bounds on these variations using duality analysis of CMDPs and perturbation analysis in linear programming. Moreover, we show how such bounds can be used to analyze the regret of a given policy due to unknown variations of the initial distribution.

Method: Kanzi uses a diffusion autoencoder trained with flow matching loss, replacing frame-based representations with global coordinates, complex losses with single flow matching loss, and SE(3)-invariant attention with standard attention.

Result: Kanzi outperforms existing tokenizers on reconstruction metrics with smaller model size and training cost. Autoregressive models with Kanzi tokens outperform similar token-based generative models, though not yet matching continuous diffusion models.

Conclusion: Flow-based tokenization simplifies protein structure modeling, enabling more efficient and stable training while maintaining competitive performance.

Abstract: Protein structure tokenizers enable the creation of multimodal models of protein structure, sequence, and function. Current approaches to protein structure tokenization rely on bespoke components that are invariant to spatial symmetries, but that are challenging to optimize and scale. We present Kanzi, a flow-based tokenizer for tokenization and generation of protein structures. Kanzi consists of a diffusion autoencoder trained with a flow matching loss. We show that this approach simplifies several aspects of protein structure tokenizers: frame-based representations can be replaced with global coordinates, complex losses are replaced with a single flow matching loss, and SE(3)-invariant attention operations can be replaced with standard attention. We find that these changes stabilize the training of parameter-efficient models that outperform existing tokenizers on reconstruction metrics at a fraction of the model size and training cost. An autoregressive model trained with Kanzi outperforms similar generative models that operate over tokens, although it does not yet match the performance of state-of-the-art continuous diffusion models. Code is available here: https://github.com/rdilip/kanzi/.

Method: AReUReDi combines Tchebycheff scalarization, locally balanced proposals, and annealed Metropolis-Hastings updates to bias sampling toward Pareto-optimal states while preserving distributional invariance.

Result: Applied to peptide and SMILES sequence design, AReUReDi simultaneously optimizes up to five therapeutic properties (affinity, solubility, hemolysis, half-life, non-fouling) and outperforms both evolutionary and diffusion-based baselines.

Conclusion: AReUReDi establishes a powerful, sequence-based framework for multi-property biomolecule generation with theoretical guarantees of Pareto optimality.

Abstract: Designing sequences that satisfy multiple, often conflicting, objectives is a central challenge in therapeutic and biomolecular engineering. Existing generative frameworks largely operate in continuous spaces with single-objective guidance, while discrete approaches lack guarantees for multi-objective Pareto optimality. We introduce AReUReDi (Annealed Rectified Updates for Refining Discrete Flows), a discrete optimization algorithm with theoretical guarantees of convergence to the Pareto front. Building on Rectified Discrete Flows (ReDi), AReUReDi combines Tchebycheff scalarization, locally balanced proposals, and annealed Metropolis-Hastings updates to bias sampling toward Pareto-optimal states while preserving distributional invariance. Applied to peptide and SMILES sequence design, AReUReDi simultaneously optimizes up to five therapeutic properties (including affinity, solubility, hemolysis, half-life, and non-fouling) and outperforms both evolutionary and diffusion-based baselines. These results establish AReUReDi as a powerful, sequence-based framework for multi-property biomolecule generation.

Method: Proposed query-only attention mechanism that discards keys and values but maintains transformer architecture’s core inductive bias.

Result: Significantly reduces both loss of plasticity and catastrophic forgetting in continual learning scenarios, outperforming selective re-initialization baselines.

Conclusion: Full attention may not be essential for meta-learning benefits in continual learning; query-based models help preserve plasticity through maintained curvature rank across tasks.

Abstract: Continual learning involves learning from a stream of data without repetition of data points, a scenario that is inherently complex due to distributional shift across tasks. We propose a query-only attention mechanism that discards keys and values, yet preserves the core inductive bias of transformer architectures. In continual learning scenarios, this simplified mechanism significantly mitigates both loss of plasticity and catastrophic forgetting, outperforming baselines such as selective re-initialization. We establish a conceptual link between query-only attention, full transformer attention, and model agnostic meta-learning, framing them as instances of meta-learning. We further provide intuition for why query-based models and attention networks help preserve plasticity in continual settings. Finally, through preliminary Hessian spectrum analysis, we observe that models maintaining higher curvature rank across tasks tend to retain plasticity. Our findings suggest that full attention may not be essential for capturing the benefits of meta-learning in continual learning.

Method: Synthesize disparate findings into curated recipes demonstrating how to implement everything from basic arithmetic in feed-forward layers to complex data routing via self-attention.

Result: A unified presentation of transformer constructions that serves as both an accessible entry point for newcomers and a systematic reference for experts.

Conclusion: This cookbook provides a foundation for future work spanning theoretical research in computational complexity to empirical investigations in architecture design and interpretability.

Abstract: We present the transformer cookbook: a collection of techniques for directly encoding algorithms into a transformer’s parameters. This work addresses the steep learning curve of such endeavors, a problem exacerbated by a fragmented literature where key results are scattered across numerous papers. In particular, we synthesize this disparate body of findings into a curated set of recipes that demonstrate how to implement everything from basic arithmetic in feed-forward layers to complex data routing via self-attention. Our mise en place of formulations is for both newcomers seeking an accessible entry point and experts in need of a systematic reference. This unified presentation of transformer constructions provides a foundation for future work spanning theoretical research in computational complexity to empirical investigations in architecture design and interpretability.

Method: Extract numerical parameters from LLM-generated programs and optimize them numerically while iterating over functional structure, using a separate optimization loop for parameter tuning.

Result: Achieves higher returns and improved sample efficiency compared to purely LLM-guided search on control tasks.

Conclusion: Combining symbolic program synthesis with numerical optimization yields interpretable yet high-performing policies, bridging language-model-guided design and classical control tuning.

Abstract: Large Language models (LLMs) have shown promise as generators of symbolic control policies, producing interpretable program-like representations through iterative search. However, these models are not capable of separating the functional structure of a policy from the numerical values it is parametrized by, thus making the search process slow and inefficient. We propose a hybrid approach that decouples structural synthesis from parameter optimization by introducing an additional optimization layer for local parameter search. In our method, the numerical parameters of LLM-generated programs are extracted and optimized numerically to maximize task performance. With this integration, an LLM iterates over the functional structure of programs, while a separate optimization loop is used to find a locally optimal set of parameters accompanying candidate programs. We evaluate our method on a set of control tasks, showing that it achieves higher returns and improved sample efficiency compared to purely LLM-guided search. We show that combining symbolic program synthesis with numerical optimization yields interpretable yet high-performing policies, bridging the gap between language-model-guided design and classical control tuning. Our code is available at https://sites.google.com/berkeley.edu/colmo.

Method: Combines a domain-specific programming language (GDL) with neural networks, using a GDL layer to generate expressive and interpretable graph representations.

Result: Achieves high accuracy on most graph classification benchmarks, outperforming dominant methods like GNNs, and yields high-quality explanations when applying existing explanation techniques.

Conclusion: GDLNN provides an effective solution for interpretable graph classification with strong performance and low cost when explanation costs are considered.

Abstract: We present GDLNN, a new graph machine learning architecture, for graph classification tasks. GDLNN combines a domain-specific programming language, called GDL, with neural networks. The main strength of GDLNN lies in its GDL layer, which generates expressive and interpretable graph representations. Since the graph representation is interpretable, existing model explanation techniques can be directly applied to explain GDLNN’s predictions. Our evaluation shows that the GDL-based representation achieves high accuracy on most graph classification benchmark datasets, outperforming dominant graph learning methods such as GNNs. Applying an existing model explanation technique also yields high-quality explanations of GDLNN’s predictions. Furthermore, the cost of GDLNN is low when the explanation cost is included.

Method: Bayesian two-axis active classification using Gaussian Process probabilistic classifier that controls spatial load (L) and feature-binding load (K) in a 5x5 working memory task, comparing GP-driven Adaptive Mode with traditional staircase Classic Mode.

Result: Parity between methods achieved (ICC=0.755 at K=3), AM reveals individual differences in spatial-feature binding interactions, and converges faster requiring only ~30 samples for accurate model fitting.

Conclusion: The Bayesian two-axis adaptive approach provides richer multidimensional assessment of working memory while maintaining efficiency comparable to traditional methods.

Abstract: While adaptive experimental design has outgrown one-dimensional, staircase-based adaptations, most cognitive experiments still control a single factor and summarize performance with a scalar. We show a validation of a Bayesian, two-axis, active-classification approach, carried out in an immersive virtual testing environment for a 5-by-5 working-memory reconstruction task. Two variables are controlled: spatial load L (number of occupied tiles) and feature-binding load K (number of distinct colors) of items. Stimulus acquisition is guided by posterior uncertainty of a nonparametric Gaussian Process (GP) probabilistic classifier, which outputs a surface over (L, K) rather than a single threshold or max span value. In a young adult population, we compare GP-driven Adaptive Mode (AM) with a traditional adaptive staircase Classic Mode (CM), which varies L only at K = 3. Parity between the methods is achieved for this cohort, with an intraclass coefficient of 0.755 at K = 3. Additionally, AM reveals individual differences in interactions between spatial load and feature binding. AM estimates converge more quickly than other sampling strategies, demonstrating that only about 30 samples are required for accurate fitting of the full model.

Method: Composer framework explores model architectures at small scale and extrapolates top-performing architectures to larger scales using proposed scaling strategies.

Result: Discovered hybrid LLM architectures outperform Llama 3.2, reducing validation loss at 350M-3B parameter scales and improving downstream task accuracy by up to 8.3% (average 1.1-3.1%) while enhancing both training and inference efficiency.

Conclusion: Composer provides an effective automated approach for discovering optimal hybrid model architectures that outperform existing state-of-the-art models across multiple metrics.

Abstract: Hybrid model architectures that combine computational primitives (e.g., Attention, MLP) in different ratios have shown promising performance beyond Transformers. Some studies have shown that different interleavings of primitives can affect model quality as well. However, prior works explore the hybrid model architecture design space manually. Due to the large design space and training costs, discovering hybrid models that combine key computational primitives for pre-training is challenging. In this work, we take a principled approach in designing a modular hybrid model architecture search framework – Composer. Composer explores model architectures at a small scale and extrapolates the top-performing model architectures to a larger scale using our proposed scaling strategies. Using Composer, we discover new hybrid LLM architectures that outperform Llama 3.2. Compared to Llama 3.2 and previous state-of-the-art baselines, the new model architectures consistently reduce validation loss at parameter scales of 350M-3B and improve evaluation accuracy on the downstream tasks by up to 2.8-8.3% (1.1-3.1% on average) while improving both training and inference efficiency.

Method: A conceptually simple approach for learning probabilistic tensor networks efficiently while maintaining numerical stability.

Result: Achieved 10x reduction in latency for MNIST generative modeling and enabled learning distributions with 10x more variables on density estimation benchmarks.

Conclusion: The proposed method provides significant improvements in time and space complexity for probabilistic tensor network learning while maintaining numerical stability.

Abstract: Tensor networks (TNs) enable compact representations of large tensors through shared parameters. Their use in probabilistic modeling is particularly appealing, as probabilistic tensor networks (PTNs) allow for tractable computation of marginals. However, existing approaches for learning parameters of PTNs are either computationally demanding and not fully compatible with automatic differentiation frameworks, or numerically unstable. In this work, we propose a conceptually simple approach for learning PTNs efficiently, that is numerically stable. We show our method provides significant improvements in time and space complexity, achieving 10x reduction in latency for generative modeling on the MNIST dataset. Furthermore, our approach enables learning of distributions with 10x more variables than previous approaches when applied to a variety of density estimation benchmarks. Our code is publicly available at github.com/marawangamal/ptn.

Method: Places a GP prior on the Hamiltonian surface and encodes variable-step multistep integrator constraints as finite linear functionals, enabling closed-form conditioning of both vector field and Hamiltonian surface without latent states.

Result: Achieves improved vector-field recovery and well-calibrated Hamiltonian uncertainty on mass-spring, Van der Pol, and Duffing benchmarks, with stated finite-sample vector-field bounds.

Conclusion: MS-PHS GP effectively learns physically consistent dynamics with Hamiltonian uncertainty quantification while maintaining energy balance and passivity properties.

Abstract: We propose the multistep port-Hamiltonian Gaussian process (MS-PHS GP) to learn physically consistent continuous-time dynamics and a posterior over the Hamiltonian from noisy, irregularly-sampled trajectories. By placing a GP prior on the Hamiltonian surface $H$ and encoding variable-step multistep integrator constraints as finite linear functionals, MS-PHS GP enables closed-form conditioning of both the vector field and the Hamiltonian surface without latent states, while enforcing energy balance and passivity by design. We state a finite-sample vector-field bound that separates the estimation and variable-step discretization terms. Lastly, we demonstrate improved vector-field recovery and well-calibrated Hamiltonian uncertainty on mass-spring, Van der Pol, and Duffing benchmarks.

Method: Invert the usual train-validation roles: perform inference on the training pool before and after fine-tuning on the validation set, then select samples whose predictions change the most. The key insight is that samples most affected by fine-tuning on validation data are most beneficial for target distribution.

Result: Experiments on instruction tuning and named entity recognition show the method achieves lower test log-loss than state-of-the-art approaches in most cases.

Conclusion: The proposed inverted role approach provides a simpler and faster data selection method that effectively identifies beneficial training samples for fine-tuning, supported by both empirical results and theoretical analysis.

Abstract: State-of-the-art machine learning often follows a two-stage process: $(i)$~pre-training on large, general-purpose datasets; $(ii)$~fine-tuning on task-specific data. In fine-tuning, selecting training examples that closely reflect the target distribution is crucial. However, it is often the case that only a few samples are available from the target distribution. Existing data selection methods treat these target samples as a validation set and estimate the effect of adding or removing a single sample from the training pool by performing inference on the validation set. We propose a simpler and faster alternative that inverts the usual role of train and validation: we perform inference on the training pool before and after fine-tuning on the validation set. We then select samples whose predictions change the most. Our key insight is that the training samples most affected by fine-tuning on a small validation set tend to be the most beneficial for reducing test loss on the target distribution. Experiments on instruction tuning and named entity recognition tasks show that, in most cases, our method achieves lower test log-loss than state-of-the-art approaches. We support our findings with theoretical analysis.

Method: Developed Dynamic Latent Variable Model (DLVM) for cross-task inference and Dynamic Adaptive Learning Engine (DALE) for optimal adaptive sampling.

Result: DLVM consistently outperformed IMLE, especially with smaller data amounts, and converged faster. DALE’s adaptive sampling outperformed random sampling and fixed test batteries, particularly within the first 80 trials.

Conclusion: Combining DLVMs for cross-task inference with DALE’s adaptive sampling provides a principled basis for more efficient cognitive assessments.

Abstract: Estimation (IMLE). DLVM integrates observations across multiple executive function tasks and individuals, allowing parameter estimation even under sparse or incomplete data conditions. DLVM consistently outperformed IMLE, especially under with smaller amounts of data, and converges faster to highly accurate estimates of the true distributions. In a second set of analyses, DALE adaptively guided sampling to maximize information gain, outperforming random sampling and fixed test batteries, particularly within the first 80 trials. These findings establish the advantages of combining DLVMs cross-task inference with DALEs optimal adaptive sampling, providing a principled basis for more efficient cognitive assessments.

Method: G2R represents nodes as closed regions in embedding space, capturing adjacency patterns and node features. Graph embeddings summarize regions where shape reflects structure and volume reflects size. Overlap approximates MCS, while disjoint parts serve as proxy for GED.

Result: G2R achieves up to 60.0% relative accuracy improvement over state-of-the-art methods in MCS similarity learning while maintaining training and inference efficiency. It can predict both MCS and GED similarities simultaneously.

Conclusion: G2R provides an efficient and effective geometric approach for graph similarity computation that outperforms existing methods and offers holistic assessment of both MCS and GED similarities.

Abstract: Graph similarity is critical in graph-related tasks such as graph retrieval, where metrics like maximum common subgraph (MCS) and graph edit distance (GED) are commonly used. However, exact computations of these metrics are known to be NP-Hard. Recent neural network-based approaches approximate the similarity score in embedding spaces to alleviate the computational burden, but they either involve expensive pairwise node comparisons or fail to effectively utilize structural and scale information of graphs. To tackle these issues, we propose a novel geometric-based graph embedding method called Graph2Region (G2R). G2R represents nodes as closed regions and recovers their adjacency patterns within graphs in the embedding space. By incorporating the node features and adjacency patterns of graphs, G2R summarizes graph regions, i.e., graph embeddings, where the shape captures the underlying graph structures and the volume reflects the graph size. Consequently, the overlap between graph regions can serve as an approximation of MCS, signifying similar node regions and adjacency patterns. We further analyze the relationship between MCS and GED and propose using disjoint parts as a proxy for GED similarity. This analysis enables concurrent computation of MCS and GED, incorporating local and global structural information. Experimental evaluation highlights G2R’s competitive performance in graph similarity computation. It achieves up to a 60.0% relative accuracy improvement over state-of-the-art methods in MCS similarity learning, while maintaining efficiency in both training and inference. Moreover, G2R showcases remarkable capability in predicting both MCS and GED similarities simultaneously, providing a holistic assessment of graph similarity. Code available at https://github.com/liuzhouyang/Graph2Region.

Method: Theoretical analysis of one-layer Mamba model consisting of linear attention followed by nonlinear gating layer. Comparison with linear Transformers under same setting, examining training dynamics and generalization on binary classification tasks with additive outliers.

Result: Mamba leverages linear attention to select informative context examples and uses nonlinear gating to suppress outlier influence. While requiring more training iterations, Mamba maintains accurate predictions even when outlier proportion exceeds linear Transformer’s tolerance threshold.

Conclusion: Mamba’s nonlinear gating mechanism provides superior robustness to outliers compared to linear Transformers, making it more reliable for in-context learning tasks with noisy data, though at the cost of slower convergence.

Abstract: The Mamba model has gained significant attention for its computational advantages over Transformer-based models, while achieving comparable performance across a wide range of language tasks. Like Transformers, Mamba exhibits in-context learning (ICL) capabilities, i.e., making predictions for new tasks based on a prompt containing input-label pairs and a query, without requiring fine-tuning. Despite its empirical success, the theoretical understanding of Mamba remains limited, largely due to the nonlinearity introduced by its gating mechanism. To the best of our knowledge, this paper presents the first theoretical analysis of the training dynamics of a one-layer Mamba model, which consists of a linear attention component followed by a nonlinear gating layer, and its ICL generalization on unseen binary classification tasks, even when the prompt includes additive outliers. Our analysis shows that Mamba leverages the linear attention layer to select informative context examples and uses the nonlinear gating layer to suppress the influence of outliers. By establishing and comparing to the analysis of linear Transformers under the same setting, we show that although Mamba may require more training iterations to converge, it maintains accurate predictions even when the proportion of outliers exceeds the threshold that a linear Transformer can tolerate. These theoretical findings are supported by empirical experiments.

Method: NC-Iso preserves relative feature positions by building hierarchical dependencies between adjacent echelons within node-rooted subtrees, and introduces a similarity dominance ratio-enhanced measure to quantify similarity dominance over dissimilarity between graph pairs.

Result: Empirical results on nine datasets validate NC-Iso’s effectiveness, generalization ability, scalability, and transferability while maintaining time efficiency.

Conclusion: NC-Iso offers a more discriminative neural subgraph matching solution for subgraph retrieval by addressing containment constraint issues and enhancing ranking capabilities.

Abstract: Subgraph matching is challenging as it necessitates time-consuming combinatorial searches. Recent Graph Neural Network (GNN)-based approaches address this issue by employing GNN encoders to extract graph information and hinge distance measures to ensure containment constraints in the embedding space. These methods significantly shorten the response time, making them promising solutions for subgraph retrieval. However, they suffer from scale differences between graph pairs during encoding, as they focus on feature counts but overlook the relative positions of features within node-rooted subtrees, leading to disturbed containment constraints and false predictions. Additionally, their hinge distance measures lack discriminative power for matched graph pairs, hindering ranking applications. We propose NC-Iso, a novel GNN architecture for neural subgraph matching. NC-Iso preserves the relative positions of features by building the hierarchical dependencies between adjacent echelons within node-rooted subtrees, ensuring matched graph pairs maintain consistent hierarchies while complying with containment constraints in feature counts. To enhance the ranking ability for matched pairs, we introduce a novel similarity dominance ratio-enhanced measure, which quantifies the dominance of similarity over dissimilarity between graph pairs. Empirical results on nine datasets validate the effectiveness, generalization ability, scalability, and transferability of NC-Iso while maintaining time efficiency, offering a more discriminative neural subgraph matching solution for subgraph retrieval. Code available at https://github.com/liuzhouyang/NC-Iso.

Method: The authors unroll the proximal gradient method for sparse coding to derive SAE variants, and propose AbsTopK SAE which applies hard thresholding over largest-magnitude activations to preserve both positive and negative activations.

Result: AbsTopK improves reconstruction fidelity, enhances interpretability, enables single features to encode contrasting concepts, and matches or surpasses supervised Difference-in-Mean method across four LLMs and seven tasks.

Conclusion: The AbsTopK SAE framework addresses fundamental limitations of existing SAEs by enabling bidirectional concept representation through magnitude-based sparsity constraints, achieving superior performance without requiring labeled data.

Abstract: Sparse autoencoders (SAEs) have emerged as powerful techniques for interpretability of large language models (LLMs), aiming to decompose hidden states into meaningful semantic features. While several SAE variants have been proposed, there remains no principled framework to derive SAEs from the original dictionary learning formulation. In this work, we introduce such a framework by unrolling the proximal gradient method for sparse coding. We show that a single-step update naturally recovers common SAE variants, including ReLU, JumpReLU, and TopK. Through this lens, we reveal a fundamental limitation of existing SAEs: their sparsity-inducing regularizers enforce non-negativity, preventing a single feature from representing bidirectional concepts (e.g., male vs. female). This structural constraint fragments semantic axes into separate, redundant features, limiting representational completeness. To address this issue, we propose AbsTopK SAE, a new variant derived from the $\ell_0$ sparsity constraint that applies hard thresholding over the largest-magnitude activations. By preserving both positive and negative activations, AbsTopK uncovers richer, bidirectional conceptual representations. Comprehensive experiments across four LLMs and seven probing and steering tasks show that AbsTopK improves reconstruction fidelity, enhances interpretability, and enables single features to encode contrasting concepts. Remarkably, AbsTopK matches or even surpasses the Difference-in-Mean method, a supervised approach that requires labeled data for each concept and has been shown in prior work to outperform SAEs.

Method: Proposes ZO Fine-tuner with a compact and memory-efficient design that learns perturbation strategies through learning-to-learn (L2L) approach, supporting one-time training per LLM with minimal overhead.

Result: Experiments on 4 LLMs and 7 datasets show ZO Fine-tuner outperforms prior zeroth-order baselines in 82.1% of task-model combinations, demonstrating strong performance and scalability.

Conclusion: ZO Fine-tuner successfully scales learning-to-learn to the foundation-model era, providing an efficient zeroth-order optimization solution for LLM fine-tuning with better performance than existing methods.

Abstract: Zeroth-order optimizers have recently emerged as a practical approach for fine-tuning large language models (LLMs), significantly reducing GPU memory consumption compared to traditional first-order methods. Yet, existing zeroth-order methods rely on hand-crafted, static sampling strategies that are not adaptable to model-specific structures. To address this, we propose ZO Fine-tuner, a learning-based zeroth-order optimizer for LLMs that automatically learns efficient perturbation strategies through a compact and memory-efficient design. Crucially, our approach is motivated by the observation that only a small number of foundation models and their derivatives are widely adopted in practice. Therefore, learning the optimizer once for a given LLM and reusing it across diverse downstream tasks is both feasible and highly desirable. Accordingly, ZO Fine-tuner is designed to scale learning to learn (L2L) to the foundation-model era by supporting one-time training per LLM with minimal overhead. Experiments on 4 LLMs and 7 datasets show that ZO Fine-tuner outperforms prior zeroth-order baselines in 82.1% of task-model combinations, thereby demonstrating strong performance and scalability for efficient LLM fine-tuning. Our code is available at https://github.com/ASTRAL-Group/ZO_Fine_tuner.git.

Method: Curated C-SRRG dataset integrating comprehensive clinical context including multi-view X-ray images, clinical indication, imaging techniques, and prior studies with comparisons based on patient histories. Used state-of-the-art multimodal large language models.

Result: Incorporating clinical context with C-SRRG significantly improves report generation quality compared to existing methods.

Conclusion: Clinical context is essential for clinically-aligned automated radiology report generation, and the proposed C-SRRG framework effectively addresses limitations of existing systems.

Abstract: Automated structured radiology report generation (SRRG) from chest X-ray images offers significant potential to reduce workload of radiologists by generating reports in structured formats that ensure clarity, consistency, and adherence to clinical reporting standards. While radiologists effectively utilize available clinical contexts in their diagnostic reasoning, existing SRRG systems overlook these essential elements. This fundamental gap leads to critical problems including temporal hallucinations when referencing non-existent clinical contexts. To address these limitations, we propose contextualized SRRG (C-SRRG) that comprehensively incorporates rich clinical context for SRRG. We curate C-SRRG dataset by integrating comprehensive clinical context encompassing 1) multi-view X-ray images, 2) clinical indication, 3) imaging techniques, and 4) prior studies with corresponding comparisons based on patient histories. Through extensive benchmarking with state-of-the-art multimodal large language models, we demonstrate that incorporating clinical context with the proposed C-SRRG significantly improves report generation quality. We publicly release dataset, code, and checkpoints to facilitate future research for clinically-aligned automated RRG at https://github.com/vuno/contextualized-srrg.

Method: Train a multimodal LLM with RL to iteratively update prompts based on intermediate latent states of diffusion models, creating step-wise prompt refinement without modifying diffusion model weights.

Result: Achieves effective reward optimization, generalizes to unseen models, composes with existing alignment methods, and mitigates over-optimization and reward hacking across diverse reward functions and diffusion backbones.

Conclusion: PromptLoop successfully bridges diffusion RL and prompt-based alignment, offering a flexible and robust alternative to direct model fine-tuning while maintaining strong performance and generalization capabilities.

Abstract: Despite the recent progress, reinforcement learning (RL)-based fine-tuning of diffusion models often struggles with generalization, composability, and robustness against reward hacking. Recent studies have explored prompt refinement as a modular alternative, but most adopt a feed-forward approach that applies a single refined prompt throughout the entire sampling trajectory, thereby failing to fully leverage the sequential nature of reinforcement learning. To address this, here we introduce PromptLoop, a plug-and-play RL framework that incorporates latent feedback into step-wise prompt refinement. Rather than modifying diffusion model weights, a multimodal large language model (MLLM) is trained with RL to iteratively update prompts based on intermediate latent states of diffusion models. This design achieves a structural analogy to the Diffusion RL approach, while retaining the flexibility and generality of prompt-based alignment. Extensive experiments across diverse reward functions and diffusion backbones demonstrate that PromptLoop (i) achieves effective reward optimization, (ii) generalizes seamlessly to unseen models, (iii) composes orthogonally with existing alignment methods, and (iv) mitigates over-optimization and reward hacking.

Method: Estimates sample influence by projecting gradients onto accumulated model updates and computing temporal variance. Low-variance samples get stronger augmentations for diversity, while unstable samples get milder transformations.

Result: Consistent improvements across benchmarks: +7.3% on fine-grained tasks and +4.3% on long-tailed datasets. The method is lightweight and plug-and-play.

Conclusion: SADA effectively addresses the dynamic nature of training by adapting augmentations to sample influence, demonstrating practical effectiveness across various datasets and architectures.

Abstract: Data augmentation has been widely employed to improve the generalization of deep neural networks. Most existing methods apply fixed or random transformations. However, we find that sample difficulty evolves along with the model’s generalization capabilities in dynamic training environments. As a result, applying uniform or stochastic augmentations, without accounting for such dynamics, can lead to a mismatch between augmented data and the model’s evolving training needs, ultimately degrading training effectiveness. To address this, we introduce SADA, a Sample-Aware Dynamic Augmentation that performs on-the-fly adjustment of augmentation strengths based on each sample’s evolving influence on model optimization. Specifically, we estimate each sample’s influence by projecting its gradient onto the accumulated model update direction and computing the temporal variance within a local training window. Samples with low variance, indicating stable and consistent influence, are augmented more strongly to emphasize diversity, while unstable samples receive milder transformations to preserve semantic fidelity and stabilize learning. Our method is lightweight, which does not require auxiliary models or policy tuning. It can be seamlessly integrated into existing training pipelines as a plug-and-play module. Experiments across various benchmark datasets and model architectures show consistent improvements of SADA, including +7.3% on fine-grained tasks and +4.3% on long-tailed datasets, highlighting the method’s effectiveness and practicality.

Method: Applies Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to tune LRP hyperparameters based on quantitative interpretability metrics like faithfulness and sparseness.

Result: EVO-LRP outperforms traditional XAI approaches in both interpretability metric performance and visual coherence, with strong sensitivity to class-specific features.

Conclusion: Attribution quality can be systematically improved through principled, task-specific optimization rather than relying on heuristic rule sets.

Abstract: Explainable AI (XAI) methods help identify which image regions influence a model’s prediction, but often face a trade-off between detail and interpretability. Layer-wise Relevance Propagation (LRP) offers a model-aware alternative. However, LRP implementations commonly rely on heuristic rule sets that are not optimized for clarity or alignment with model behavior. We introduce EVO-LRP, a method that applies Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to tune LRP hyperparameters based on quantitative interpretability metrics, such as faithfulness or sparseness. EVO-LRP outperforms traditional XAI approaches in both interpretability metric performance and visual coherence, with strong sensitivity to class-specific features. These findings demonstrate that attribution quality can be systematically improved through principled, task-specific optimization.

Method: Adapts control-theoretic matrix sketching to neural network layer activations using three complementary sketch matrices maintained through exponential moving averages (EMA) with adaptive rank adjustment.

Result: Empirical evaluation on MNIST, CIFAR-10, and physics-informed neural networks demonstrates controllable accuracy-memory tradeoff and enables real-time gradient norm tracking with minimal memory overhead.

Conclusion: Sketched activation storage provides a viable path toward memory-efficient neural network training and analysis.

Abstract: Neural network training relies on gradient computation through backpropagation, yet memory requirements for storing layer activations present significant scalability challenges. We present the first adaptation of control-theoretic matrix sketching to neural network layer activations, enabling memory-efficient gradient reconstruction in backpropagation. This work builds on recent matrix sketching frameworks for dynamic optimization problems, where similar state trajectory storage challenges motivate sketching techniques. Our approach sketches layer activations using three complementary sketch matrices maintained through exponential moving averages (EMA) with adaptive rank adjustment, automatically balancing memory efficiency against approximation quality. Empirical evaluation on MNIST, CIFAR-10, and physics-informed neural networks demonstrates a controllable accuracy-memory tradeoff. We demonstrate a gradient monitoring application on MNIST showing how sketched activations enable real-time gradient norm tracking with minimal memory overhead. These results establish that sketched activation storage provides a viable path toward memory-efficient neural network training and analysis.

Method: Decomposes adversarial documents into Attention Attractors (optimized to direct attention) and Focus Regions (containing semantic baits or malicious instructions), targeting specific attention heads correlated with attack success.

Result: Across 18 RAG settings, attack success rates increased from 21.9% to 57.8% (2.6x improvement over prior work), with single optimized attractors transferring to unseen black box systems without retraining.

Conclusion: Modular, reusable components pose practical threats to AI systems, and attention concentration strongly correlates with model outputs, informing both security and interpretability research.

Abstract: Existing data poisoning attacks on retrieval-augmented generation (RAG) systems scale poorly because they require costly optimization of poisoned documents for each target phrase. We introduce Eyes-on-Me, a modular attack that decomposes an adversarial document into reusable Attention Attractors and Focus Regions. Attractors are optimized to direct attention to the Focus Region. Attackers can then insert semantic baits for the retriever or malicious instructions for the generator, adapting to new targets at near zero cost. This is achieved by steering a small subset of attention heads that we empirically identify as strongly correlated with attack success. Across 18 end-to-end RAG settings (3 datasets $\times$ 2 retrievers $\times$ 3 generators), Eyes-on-Me raises average attack success rates from 21.9 to 57.8 (+35.9 points, 2.6$\times$ over prior work). A single optimized attractor transfers to unseen black box retrievers and generators without retraining. Our findings establish a scalable paradigm for RAG data poisoning and show that modular, reusable components pose a practical threat to modern AI systems. They also reveal a strong link between attention concentration and model outputs, informing interpretability research.

Method: UrbanGraph integrates heterogeneous and dynamic spatio-temporal graphs that encode physical processes (vegetation evapotranspiration, shading, convective diffusion) while modeling complex spatial dependencies among diverse urban entities and their temporal evolution.

Result: UrbanGraph improves R² by up to 10.8% and reduces FLOPs by 17.0% over all baselines, with heterogeneous and dynamic graphs contributing 3.5% and 7.1% gains respectively. The UMC4/12 dataset provides the first high-resolution benchmark for spatio-temporal microclimate modeling.

Conclusion: UrbanGraph effectively addresses urban microclimate prediction challenges and extends to broader urban heterogeneous dynamic computing tasks, providing a comprehensive framework for spatio-temporal modeling.

Abstract: With rapid urbanization, predicting urban microclimates has become critical, as it affects building energy demand and public health risks. However, existing generative and homogeneous graph approaches fall short in capturing physical consistency, spatial dependencies, and temporal variability. To address this, we introduce UrbanGraph, a physics-informed framework integrating heterogeneous and dynamic spatio-temporal graphs. It encodes key physical processes – vegetation evapotranspiration, shading, and convective diffusion – while modeling complex spatial dependencies among diverse urban entities and their temporal evolution. We evaluate UrbanGraph on UMC4/12, a physics-based simulation dataset covering diverse urban configurations and climates. Results show that UrbanGraph improves $R^2$ by up to 10.8% and reduces FLOPs by 17.0% over all baselines, with heterogeneous and dynamic graphs contributing 3.5% and 7.1% gains. Our dataset provides the first high-resolution benchmark for spatio-temporal microclimate modeling, and our method extends to broader urban heterogeneous dynamic computing tasks.

Method: Formulates anomaly detection as regularized robust low-rank + sparse tensor decomposition with total variation constraints for spatiotemporal smoothness, plus statistical anomaly scoring framework.

Result: Evaluated on synthetic and real data, showing ability to capture spatiotemporal dependencies and provide statistical confidence.

Conclusion: Proposed unsupervised tensor-based method effectively handles spatiotemporal anomalies with statistical confidence measures.

Abstract: Anomaly detection in spatiotemporal data is a challenging problem encountered in a variety of applications, including video surveillance, medical imaging data, and urban traffic monitoring. Existing anomaly detection methods focus mainly on point anomalies and cannot deal with temporal and spatial dependencies that arise in spatio-temporal data. Tensor-based anomaly detection methods have been proposed to address this problem. Although existing methods can capture dependencies across different modes, they are primarily supervised and do not account for the specific structure of anomalies. Moreover, these methods focus mainly on extracting anomalous features without providing any statistical confidence. In this paper, we introduce an unsupervised tensor-based anomaly detection method that simultaneously considers the sparse and spatiotemporally smooth nature of anomalies. The anomaly detection problem is formulated as a regularized robust low-rank + sparse tensor decomposition where the total variation of the tensor with respect to the underlying spatial and temporal graphs quantifies the spatiotemporal smoothness of the anomalies. Once the anomalous features are extracted, we introduce a statistical anomaly scoring framework that accounts for local spatio-temporal dependencies. The proposed framework is evaluated on both synthetic and real data.

Method: Proposes static-dynamic decomposition: (1) time-invariant component captured by global embedding module learning persistent representations, (2) time-varying component processed by frequency-domain filtering inspired by full-spectrum analysis.

Result: Outperforms state-of-the-art baselines on real-world datasets with fewer computational resources. Comprehensive analyses verify efficacy of static-dynamic disentanglement.

Conclusion: TimeEmb effectively handles temporal non-stationarity through component decomposition and can be easily integrated to improve existing time-series forecasting methods.

Abstract: Temporal non-stationarity, the phenomenon that time series distributions change over time, poses fundamental challenges to reliable time series forecasting. Intuitively, the complex time series can be decomposed into two factors, \ie time-invariant and time-varying components, which indicate static and dynamic patterns, respectively. Nonetheless, existing methods often conflate the time-varying and time-invariant components, and jointly learn the combined long-term patterns and short-term fluctuations, leading to suboptimal performance facing distribution shifts. To address this issue, we initiatively propose a lightweight static-dynamic decomposition framework, TimeEmb, for time series forecasting. TimeEmb innovatively separates time series into two complementary components: (1) time-invariant component, captured by a novel global embedding module that learns persistent representations across time series, and (2) time-varying component, processed by an efficient frequency-domain filtering mechanism inspired by full-spectrum analysis in signal processing. Experiments on real-world datasets demonstrate that TimeEmb outperforms state-of-the-art baselines and requires fewer computational resources. We conduct comprehensive quantitative and qualitative analyses to verify the efficacy of static-dynamic disentanglement. This lightweight framework can also improve existing time-series forecasting methods with simple integration. To ease reproducibility, the code is available at https://github.com/showmeon/TimeEmb.

Method: Conducted systematic stability analysis of EAP-IG circuit discovery method using controlled perturbations including input resampling, prompt paraphrasing, hyperparameter variation, and injected noise.

Result: EAP-IG exhibits high structural variance and sensitivity to hyperparameters across diverse models and tasks, questioning the stability of its findings.

Conclusion: Advocates for routine reporting of stability metrics in interpretability research to promote more rigorous and statistically grounded science.

Abstract: The development of trustworthy artificial intelligence requires moving beyond black-box performance metrics toward an understanding of models’ internal computations. Mechanistic Interpretability (MI) aims to meet this need by identifying the algorithmic mechanisms underlying model behaviors. Yet, the scientific rigor of MI critically depends on the reliability of its findings. In this work, we argue that interpretability methods, such as circuit discovery, should be viewed as statistical estimators, subject to questions of variance and robustness. To illustrate this statistical framing, we present a systematic stability analysis of a state-of-the-art circuit discovery method: EAP-IG. We evaluate its variance and robustness through a comprehensive suite of controlled perturbations, including input resampling, prompt paraphrasing, hyperparameter variation, and injected noise within the causal analysis itself. Across a diverse set of models and tasks, our results demonstrate that EAP-IG exhibits high structural variance and sensitivity to hyperparameters, questioning the stability of its findings. Based on these results, we offer a set of best-practice recommendations for the field, advocating for the routine reporting of stability metrics to promote a more rigorous and statistically grounded science of interpretability.

Method: The proposed F2OCL method integrates prompt learning with an NCM (Nearest Class Mean) classifier to tackle catastrophic forgetting without storing samples and without using task boundaries or identities.

Result: Extensive experiments on two benchmarks demonstrate the effectiveness of the proposed method in addressing catastrophic forgetting in rehearsal-free and task-free online continual learning scenarios.

Conclusion: The study successfully addresses catastrophic forgetting in online continual learning without requiring rehearsal buffers or task information, providing a practical solution for scenarios with privacy concerns and unknown task boundaries.

Abstract: The main challenge of continual learning is \textit{catastrophic forgetting}. Because of processing data in one pass, online continual learning (OCL) is one of the most difficult continual learning scenarios. To address catastrophic forgetting in OCL, some existing studies use a rehearsal buffer to store samples and replay them in the later learning process, other studies do not store samples but assume a sequence of learning tasks so that the task identities can be explored. However, storing samples may raise data security or privacy concerns and it is not always possible to identify the boundaries between learning tasks in one pass of data processing. It motivates us to investigate rehearsal-free and task-free OCL (F2OCL). By integrating prompt learning with an NCM classifier, this study has effectively tackled catastrophic forgetting without storing samples and without usage of task boundaries or identities. The extensive experimental results on two benchmarks have demonstrated the effectiveness of the proposed method.

Method: Apply eigenanalysis to empirical neural tangent kernel (eNTK) across two toy models: Toy Models of Superposition (TMS) and 1-layer MLP on modular addition, analyzing spectral properties and layerwise eNTK.

Result: eNTK exhibits sharp spectral cliffs aligning with ground-truth features, recovers features in both sparse and dense regimes of TMS, identifies Fourier features in modular arithmetic, localizes features to specific layers, and detects grokking phase transitions.

Conclusion: eNTK analysis provides a practical approach for feature discovery and detecting phase changes in small neural network models.

Abstract: We provide evidence that eigenanalysis of the empirical neural tangent kernel (eNTK) can surface the features used by trained neural networks. Across two standard toy models for mechanistic interpretability, Toy Models of Superposition (TMS) and a 1-layer MLP trained on modular addition, we find that the eNTK exhibits sharp spectral cliffs whose top eigenspaces align with ground-truth features. In TMS, the eNTK recovers the ground-truth features in both the sparse (high superposition) and dense regimes. In modular arithmetic, the eNTK can be used to recover Fourier feature families. Moreover, we provide evidence that a layerwise eNTK localizes features to specific layers and that the evolution of the eNTK eigenspectrum can be used to diagnose the grokking phase transition. These results suggest that eNTK analysis may provide a practical handle for feature discovery and for detecting phase changes in small models.

Method: Analyzed CQF by training binary classifiers to distinguish pretraining data from high-quality sets, comparing with models trained on synthetic data with increasing quality via random token permutations.

Result: CQF improves downstream tasks but not language modeling on high-quality data. Models trained with CQF show different trends than those trained on synthetic quality data.

Conclusion: CQF may not capture a meaningful notion of data quality, challenging current understanding of quality filtering methods.

Abstract: Large-scale models are pretrained on massive web-crawled datasets containing documents of mixed quality, making data filtering essential. A popular method is Classifier-based Quality Filtering (CQF), which trains a binary classifier to distinguish between pretraining data and a small, high-quality set. It assigns each pretraining document a quality score defined as the classifier’s score and retains only the top-scoring ones. We provide an in-depth analysis of CQF. We show that while CQF improves downstream task performance, it does not necessarily enhance language modeling on the high-quality dataset. We explain this paradox by the fact that CQF implicitly filters the high-quality dataset as well. We further compare the behavior of models trained with CQF to those trained on synthetic data of increasing quality, obtained via random token permutations, and find starkly different trends. Our results challenge the view that CQF captures a meaningful notion of data quality.

Method: Proposed Einstellung Rigidity Index (ERI) with three components: Adaptation Delay (AD), Performance Deficit (PD), and Relative Suboptimal Feature Reliance (SFR_rel). Evaluated on CIFAR-100 CL benchmark with spurious magenta patch in Phase 2, testing methods including SGD, EWC_on, DER++, GPM, and DGR.

Result: CL methods reached accuracy thresholds faster than Scratch-T2 baseline (negative AD) but achieved lower final accuracy on shortcut classes (positive PD). Masking the patch improved CL method accuracy while slightly reducing Scratch-T2 accuracy, yielding negative SFR_rel, indicating the patch acted as a distractor rather than helpful shortcut.

Conclusion: The ERI diagnostic successfully disentangles genuine transfer from shortcut-inflated performance, revealing that shortcut features can create rigidity in continual learning systems that hinders adaptation to new tasks.

Abstract: Deep neural networks frequently exploit shortcut features, defined as incidental correlations between inputs and labels without causal meaning. Shortcut features undermine robustness and reduce reliability under distribution shifts. In continual learning (CL), the consequences of shortcut exploitation can persist and intensify: weights inherited from earlier tasks bias representation reuse toward whatever features most easily satisfied prior labels, mirroring the cognitive Einstellung effect, a phenomenon where past habits block optimal solutions. Whereas catastrophic forgetting erodes past skills, shortcut-induced rigidity throttles the acquisition of new ones. We introduce the Einstellung Rigidity Index (ERI), a compact diagnostic that disentangles genuine transfer from cue-inflated performance using three interpretable facets: (i) Adaptation Delay (AD), (ii) Performance Deficit (PD), and (iii) Relative Suboptimal Feature Reliance (SFR_rel). On a two-phase CIFAR-100 CL benchmark with a deliberately spurious magenta patch in Phase 2, we evaluate Naive fine-tuning (SGD), online Elastic Weight Consolidation (EWC_on), Dark Experience Replay (DER++), Gradient Projection Memory (GPM), and Deep Generative Replay (DGR). Across these continual learning methods, we observe that CL methods reach accuracy thresholds earlier than a Scratch-T2 baseline (negative AD) but achieve slightly lower final accuracy on patched shortcut classes (positive PD). Masking the patch improves accuracy for CL methods while slightly reducing Scratch-T2, yielding negative SFR_rel. This pattern indicates the patch acted as a distractor for CL models in this setting rather than a helpful shortcut.

Method: Encodes source feature label information into latent vicinities using Gaussian priors over k-nearest neighbors, trains diffusion network to denoise samples back to label-consistent representations, and aligns target encoder to generated source-like cues using InfoNCE loss.

Result: Outperforms state-of-the-art methods on standard SFDA benchmarks, enhances source classifier accuracy on in-domain data, and boosts performance in supervised classification and domain generalization.

Conclusion: DVD reinterprets LDMs as practical, privacy-preserving bridges for explicit knowledge transfer, solving a core challenge in source-free domain adaptation that prior methods couldn’t address.

Abstract: Recent work on latent diffusion models (LDMs) has focused almost exclusively on generative tasks, leaving their potential for discriminative transfer largely unexplored. We introduce Discriminative Vicinity Diffusion (DVD), a novel LDM-based framework for a more practical variant of source-free domain adaptation (SFDA): the source provider may share not only a pre-trained classifier but also an auxiliary latent diffusion module, trained once on the source data and never exposing raw source samples. DVD encodes each source feature’s label information into its latent vicinity by fitting a Gaussian prior over its k-nearest neighbors and training the diffusion network to drift noisy samples back to label-consistent representations. During adaptation, we sample from each target feature’s latent vicinity, apply the frozen diffusion module to generate source-like cues, and use a simple InfoNCE loss to align the target encoder to these cues, explicitly transferring decision boundaries without source access. Across standard SFDA benchmarks, DVD outperforms state-of-the-art methods. We further show that the same latent diffusion module enhances the source classifier’s accuracy on in-domain data and boosts performance in supervised classification and domain generalization experiments. DVD thus reinterprets LDMs as practical, privacy-preserving bridges for explicit knowledge transfer, addressing a core challenge in source-free domain adaptation that prior methods have yet to solve.

Method: Reframe GRPO as contrastive learning, connect it to DPO, and investigate the two-rollout case (2-GRPO) with theoretical analysis and empirical validation.

Result: 2-GRPO achieves performance comparable to 16-GRPO using only 1/8 of the rollouts and reducing training time by over 70%.

Conclusion: Large group sizes are not necessary for GRPO; minimal two-rollout configuration is both feasible and efficient.

Abstract: Group Relative Policy Optimization (GRPO) is a prominent reinforcement learning algorithm for post-training Large Language Models (LLMs). It is commonly believed that GRPO necessitates a large group size to ensure stable training via precise statistical estimation, which incurs substantial computational overhead. In this work, we challenge this assumption by reframing GRPO as a form of contrastive learning, which reveals a fundamental connection to Direct Preference Optimization (DPO). Motivated by DPO’s empirical success, we investigate the minimal two-rollout case (2-GRPO), a configuration previously deemed infeasible. We provide a rigorous theoretical analysis to validate 2-GRPO and demonstrate empirically that it achieves performance on par with 16-GRPO, despite using only 1/8 of the rollouts and reducing training time by over 70%.

Method: Dual-branch network structure with unique prompts per branch, reconstruction learning at prompt and input levels, built upon time-series foundation model to handle spatio-temporal dynamics.

Result: CPFM achieves improved results with noticeable margins over competitors in three time-series datasets across different application domains.

Conclusion: CPFM effectively addresses black-box time-series domain adaptation by leveraging foundation models and cross-prompt learning, demonstrating superior performance across multiple domains.

Abstract: The black-box domain adaptation (BBDA) topic is developed to address the privacy and security issues where only an application programming interface (API) of the source model is available for domain adaptations. Although the BBDA topic has attracted growing research attentions, existing works mostly target the vision applications and are not directly applicable to the time-series applications possessing unique spatio-temporal characteristics. In addition, none of existing approaches have explored the strength of foundation model for black box time-series domain adaptation (BBTSDA). This paper proposes a concept of Cross-Prompt Foundation Model (CPFM) for the BBTSDA problems. CPFM is constructed under a dual branch network structure where each branch is equipped with a unique prompt to capture different characteristics of data distributions. In the domain adaptation phase, the reconstruction learning phase in the prompt and input levels is developed. All of which are built upon a time-series foundation model to overcome the spatio-temporal dynamic. Our rigorous experiments substantiate the advantage of CPFM achieving improved results with noticeable margins from its competitors in three time-series datasets of different application domains.

Method: Tested dual-architecture latent reasoning with Base-Coprocessor communication, comparing two hypotheses: increasing channel capacity (H1) and joint finetuning (H2) against a unified soft-embedding baseline with matched latent-token budgets on GPT-2 and Qwen-3.

Result: H2 (joint finetuning) was consistently strongest while H1 yielded modest gains. The unified baseline nearly matched H2 and surpassed H1. Scaling latent-token budgets beyond small values failed to improve robustness across GSM8K, ProsQA, and Countdown tests. Latent analyses showed overlapping subspaces with limited specialization.

Conclusion: Dual-model latent reasoning remains promising in principle but likely requires objectives and communication mechanisms that explicitly shape latent spaces for algorithmic planning, as current dual designs mostly add compute without qualitative reasoning improvements.

Abstract: Should LLM reasoning live in a separate module, or within a single model’s forward pass and representational space? We study dual-architecture latent reasoning, where a fluent Base exchanges latent messages with a Coprocessor, and test two hypotheses aimed at improving latent communication over Liu et al. (2024): (H1) increase channel capacity; (H2) learn communication via joint finetuning. Under matched latent-token budgets on GPT-2 and Qwen-3, H2 is consistently strongest while H1 yields modest gains. A unified soft-embedding baseline, a single model with the same forward pass and shared representations, using the same latent-token budget, nearly matches H2 and surpasses H1, suggesting current dual designs mostly add compute rather than qualitatively improving reasoning. Across GSM8K, ProsQA, and a Countdown stress test with increasing branching factor, scaling the latent-token budget beyond small values fails to improve robustness. Latent analyses show overlapping subspaces with limited specialization, consistent with weak reasoning gains. We conclude dual-model latent reasoning remains promising in principle, but likely requires objectives and communication mechanisms that explicitly shape latent spaces for algorithmic planning.

Method: DAV formulates diffusion alignment as variational EM with two alternating phases: E-step uses test-time search to generate diverse reward-aligned samples, M-step refines the diffusion model using these samples.

Result: DAV successfully optimizes rewards while preserving diversity in both continuous (text-to-image synthesis) and discrete (DNA sequence design) tasks.

Conclusion: The variational EM framework provides an effective approach for diffusion alignment that avoids reward over-optimization and mode collapse while maintaining sample diversity.

Abstract: Diffusion alignment aims to optimize diffusion models for the downstream objective. While existing methods based on reinforcement learning or direct backpropagation achieve considerable success in maximizing rewards, they often suffer from reward over-optimization and mode collapse. We introduce Diffusion Alignment as Variational Expectation-Maximization (DAV), a framework that formulates diffusion alignment as an iterative process alternating between two complementary phases: the E-step and the M-step. In the E-step, we employ test-time search to generate diverse and reward-aligned samples. In the M-step, we refine the diffusion model using samples discovered by the E-step. We demonstrate that DAV can optimize reward while preserving diversity for both continuous and discrete tasks: text-to-image synthesis and DNA sequence design.

Method: GEM provides a standardized framework with asynchronous vectorized execution, flexible wrappers, diverse environments, integrated tools, and example scripts for five popular RL frameworks. It also includes baselines using REINFORCE with Return Batch Normalization.

Result: GEM enables apple-to-apple benchmarking of PPO, GRPO and REINFORCE in single- and multi-turn settings, and functions as both training environment and evaluation toolkit.

Conclusion: GEM aims to accelerate future agentic LLM research by providing a comprehensive framework for experience-based learning and standardized benchmarking.

Abstract: The training paradigm for large language models (LLMs) is moving from static datasets to experience-based learning, where agents acquire skills via interacting with complex environments. To facilitate this transition we introduce GEM (General Experience Maker), an open-source environment simulator designed for the age of LLMs. Analogous to OpenAI-Gym for traditional reinforcement learning (RL), GEM provides a standardized framework for the environment-agent interface, including asynchronous vectorized execution for high throughput, and flexible wrappers for easy extensibility. GEM also features a diverse suite of environments, robust integrated tools, and single-file example scripts demonstrating using GEM with five popular RL training frameworks. Along with this, we also provide a set of baselines across 24 environments using REINFORCE with Return Batch Normalization (ReBN), which – unlike GRPO – is compatible with the full RL setting of dense per-turn rewards and offers better credit assignment. We further conduct apple-to-apple benchmarking of PPO, GRPO and REINFORCE in both single- and multi-turn settings using GEM to shed light on the algorithmic designs. Lastly, GEM also functions as a convenient evaluation toolkit besides a training environment. We hope this framework can help accelerate future agentic LLM research.

Method: Theoretical analysis of negative gradient alignment in DA, empirical validation on ViT/DiffViT and pretrained CLIP/DiffCLIP models across five datasets, and depth-dependent experiments examining noise cancellation effects.

Result: DA shows higher attack success rates, frequent gradient opposition, and stronger local sensitivity compared to standard attention. Depth-dependent experiments reveal robustness crossover where DA layers attenuate small perturbations but protection fades under larger attacks.

Conclusion: There is a fundamental trade-off: DA improves discriminative focus on clean inputs but increases adversarial vulnerability, highlighting the need to jointly design for both selectivity and robustness in future attention mechanisms.

Abstract: Differential Attention (DA) has been proposed as a refinement to standard attention, suppressing redundant or noisy context through a subtractive structure and thereby reducing contextual hallucination. While this design sharpens task-relevant focus, we show that it also introduces a structural fragility under adversarial perturbations. Our theoretical analysis identifies negative gradient alignment-a configuration encouraged by DA’s subtraction-as the key driver of sensitivity amplification, leading to increased gradient norms and elevated local Lipschitz constants. We empirically validate this Fragile Principle through systematic experiments on ViT/DiffViT and evaluations of pretrained CLIP/DiffCLIP, spanning five datasets in total. These results demonstrate higher attack success rates, frequent gradient opposition, and stronger local sensitivity compared to standard attention. Furthermore, depth-dependent experiments reveal a robustness crossover: stacking DA layers attenuates small perturbations via depth-dependent noise cancellation, though this protection fades under larger attack budgets. Overall, our findings uncover a fundamental trade-off: DA improves discriminative focus on clean inputs but increases adversarial vulnerability, underscoring the need to jointly design for selectivity and robustness in future attention mechanisms.

Method: Breaks down design space into three pillars (environment, reward, policy), empirically tests on TextWorld, ALFWorld, and SWE-Gym domains, analyzing task complexity, reward sparsity, and policy gradient methods including PPO, GRPO, and RLOO.

Result: Found that simple environments provide signal for generalization to complex tasks; dense rewards accelerate training but performance depends on RL algorithm choice; identified optimal SFT-to-RL training ratios and interplay between reward sparsity and policy methods.

Conclusion: Developed a training recipe that guides co-design across environment, reward, and policy pillars to facilitate multi-turn agentic RL research and practical applications.

Abstract: We study what actually works and what doesn’t for training large language models as agents via multi-turn reinforcement learning. Despite rapid progress, existing frameworks and definitions are fragmented, and there is no systematic formulation or analysis of which design choices matter across tasks. We address this gap by first breaking down the design space into three inter-related pillars – environment, reward, and policy – and empirically derive a recipe for training LLM agents in situated textual domains. In particular, we test TextWorld and ALFWorld, popular domains for testing situated embodied reasoning, as well as SWE-Gym for more software engineering style tasks. (i) For the environment, we analyze the impacts of task complexity in terms of sizes of the state and action spaces as well as optimal solution length, finding that even simple environments within a domain can provide signal on how well an agent can generalize to more complex tasks. (ii) For the reward, we ablate relative reward sparsity, observing that while dense turn-level rewards accelerate training, performance and stability is highly dependent on the choice of RL algorithm. (iii) And for the agent’s policy, we explore the interplay between reward sparsity and biased (PPO, GRPO) and unbiased (RLOO) policy gradient methods in addition to showing how to find the optimal Supervised Fine-tuning (SFT) to RL training ratio given a fixed budget. We distill these findings into a training recipe that guides co-design across the three pillars, facilitating research and practical efforts in multi-turn agentic RL. Code: https://github.com/pearls-lab/meow-tea-taro

Method: Developed a lightweight diagnostic suite including Hard Rank (participation ratio), Soft Rank (Shannon rank), Spectral Concentration, and Spectral Utilization Index to analyze spectral utilization across LLaMA, GPT-2, and nGPT families.

Result: Found asymmetric spectral scaling law: soft rank follows power law with FFN width, while hard rank grows sublinearly with high variance, indicating widening mostly adds low-energy tail directions while dominant-mode subspaces saturate early.

Conclusion: FFN width selection involves a trade-off between tail capacity and dominant-mode capacity, providing guidance for inference-efficient LLM design.

Abstract: As large language models (LLMs) scale, the question is not only how large they become, but how much of their capacity is effectively utilized. Existing scaling laws relate model size to loss, yet overlook how components exploit their latent space. We study feed-forward networks (FFNs) and recast width selection as a spectral utilization problem. Using a lightweight diagnostic suite – Hard Rank (participation ratio), Soft Rank (Shannon rank), Spectral Concentration, and the composite Spectral Utilization Index (SUI) – we quantify how many latent directions are meaningfully activated across LLaMA, GPT-2, and nGPT families. Our key finding is an asymmetric spectral scaling law: soft rank follows an almost perfect power law with FFN width, while hard rank grows only sublinearly and with high variance. This asymmetry suggests that widening FFNs mostly adds low-energy tail directions, while dominant-mode subspaces saturate early. Moreover, at larger widths, variance further collapses into a narrow subspace, leaving much of the latent space under-utilized. These results recast FFN width selection as a principled trade-off between tail capacity and dominant-mode capacity, offering concrete guidance for inference-efficient LLM design.

Method: PCL uses a concurrently updated value model to identify prompts of intermediate difficulty in an on-policy manner, focusing on prompts with high effective ratios without costly rollouts.

Result: PCL achieves 12.1× and 16.9× faster prompt identification on MATH and DeepScaleR datasets respectively, and either reaches highest performance or requires significantly less time for comparable performance.

Conclusion: PCL provides an improved tradeoff between performance and efficiency for reasoning-focused RL by focusing on progressively challenging prompts through accurate difficulty prediction.

Abstract: We introduce Prompt Curriculum Learning (PCL), a lightweight reinforcement learning (RL) algorithm that selects intermediate-difficulty prompts using a learned value model to post-train language models. Since post-training LLMs via RL remains sensitive to batching and prompt selection strategies, we first conduct a series of systematic experiments where we (1) determine the optimal training batch size that balances generation efficiency and gradient quality and (2) establish the importance of focusing on prompts of intermediate difficulty for the policy. We build upon these results to design PCL, which identifies prompts of intermediate difficulty for the current policy in an on-policy manner by using a value model that is concurrently updated based on the current policy. By focusing on informative prompts that yield high effective ratios, PCL achieves either the highest performance or requires significantly less time to reach comparable performance to its counterparts. Compared to rollout-based filtering methods, PCL avoids costly rollouts and achieves $12.1\times$ and $16.9\times$ faster speed on identifying intermediate-difficulty prompts when training on MATH and DeepScaleR, respectively. We further demonstrate that our value model accurately predicts prompt difficulty and allows PCL to focus on progressively more challenging prompts during RL. Our results present a new methodology that delivers improved tradeoff between upper-bound performance and efficiency for reasoning-focused RL.

Method: Used three ML models (LR, RDT, RF) with WHO/UN dataset, extensive preprocessing for missing values, and performance evaluation using R², MAE, and RMSE metrics.

Result: Random Forest achieved highest predictive accuracy (R²=0.9423), significantly outperforming Linear Regression and Regression Decision Tree. Key predictors identified were immunization rates (diphtheria, measles) and demographic attributes (HIV/AIDS, adult mortality).

Conclusion: Ensemble methods like Random Forest combined with interpretability techniques provide effective solutions for life expectancy prediction, with future research directions including advanced imputation, neural networks, and updated data for improved accuracy.

Abstract: Life expectancy is a fundamental indicator of population health and socio-economic well-being, yet accurately forecasting it remains challenging due to the interplay of demographic, environmental, and healthcare factors. This study evaluates three machine learning models – Linear Regression (LR), Regression Decision Tree (RDT), and Random Forest (RF), using a real-world dataset drawn from World Health Organization (WHO) and United Nations (UN) sources. After extensive preprocessing to address missing values and inconsistencies, each model’s performance was assessed with $R^2$, Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE). Results show that RF achieves the highest predictive accuracy ($R^2 = 0.9423$), significantly outperforming LR and RDT. Interpretability was prioritized through p-values for LR and feature importance metrics for the tree-based models, revealing immunization rates (diphtheria, measles) and demographic attributes (HIV/AIDS, adult mortality) as critical drivers of life-expectancy predictions. These insights underscore the synergy between ensemble methods and transparency in addressing public-health challenges. Future research should explore advanced imputation strategies, alternative algorithms (e.g., neural networks), and updated data to further refine predictive accuracy and support evidence-based policymaking in global health contexts.

Method: Standardizes process reward model training across verifiable/non-verifiable settings, performs multi-objective alignment using MAH-DPO with vectorized rewards, and enables fine-grained inference-time user control.

Result: Experiments show improved performance across multiple objectives simultaneously while minimizing cross-objective trade-offs and enabling flexible inference-time control.

Conclusion: The proposed framework successfully addresses multi-objective alignment challenges across diverse domains while maintaining user control.

Abstract: Aligning large language models to human preferences is inherently multidimensional, yet most pipelines collapse heterogeneous signals into a single optimizeable objective. We seek to answer what it would take to simultaneously align a model across various domains spanning those with: verifiable rewards (mathematical accuracy), non-verifiable subjective preferences (human values), and complex interactive scenarios (multi-turn AI tutoring dialogues). Such multi-objective reinforcement learning setups are often plagued by the individual objectives being at odds with each other, resulting in inefficient training and little user control during inference. We propose a unified framework that: (i) standardizes {process reward model} (PRM) training across both verifiable and non-verifiable settings to better supervise models’ chain-of-thought reasoning; (ii) performs {multi-objective alignment} by training the LLM with our $\textbf{M}$ulti-$\textbf{A}$ction-$\textbf{H}$ead $\textbf{DPO}$ (MAH-DPO) and a vectorized reward where the dimensions of the vector correspond to the various objectives instead of a single scalar; and (iii) demonstrates how such a system provides fine-grained inference-time user control. Experiments across math reasoning, value alignment, and multi-turn dialogue show that our framework improves performance across multiple objectives simultaneously, while minimizing cross-objective trade-offs and enabling flexible inference time user control. The code can be found at https://github.com/pearls-lab/multiobj-align.

Method: Identified two key properties: Rank-1 Dominance (top singular subspace determines reasoning improvements) and Rank-1 Linear Dynamics (dominant subspace evolves linearly). Proposed AlphaRL framework that extrapolates final parameter updates from early training.

Result: Validated across 8 LLMs and 7 algorithms, achieving up to 2.5× speedup while retaining >96% reasoning performance without extra modules or hyperparameter tuning.

Conclusion: The findings provide a versatile tool for large-scale RL, enabling principled, interpretable, and efficient training paradigm for LLMs.

Abstract: Recent advances in reasoning capabilities of large language models (LLMs) are largely driven by reinforcement learning (RL), yet the underlying parameter dynamics during RL training remain poorly understood. This work identifies two fundamental properties of RL-induced parameter updates in LLMs: (1) Rank-1 Dominance, where the top singular subspace of the parameter update matrix nearly fully determines reasoning improvements, recovering over 99% of performance gains; and (2) Rank-1 Linear Dynamics, where this dominant subspace evolves linearly throughout training, enabling accurate prediction from early checkpoints. Extensive experiments across 8 LLMs and 7 algorithms validate the generalizability of these properties. More importantly, based on these findings, we propose AlphaRL, a plug-in acceleration framework that extrapolates the final parameter update using a short early training window, achieving up to 2.5 speedup while retaining \textgreater 96% of reasoning performance without extra modules or hyperparameter tuning. This positions our finding as a versatile and practical tool for large-scale RL, opening a path toward principled, interpretable, and efficient training paradigm for LLMs.

Method: Synthesized trajectories from real MCP environments using 5 query generation models, quality filtering, and 3 teacher models with 2 agentic frameworks, plus extension mechanisms for task diversification and multi-turn conversations.

Result: Models fine-tuned on Toucan outperform larger closed-source counterparts on BFCL V3 benchmark and advance the Pareto frontier on MCP-Universe Bench.

Conclusion: Toucan successfully addresses the data gap for tool-agentic training and enables open-source LLM agents to achieve state-of-the-art performance.

Abstract: Large Language Model (LLM) agents are rapidly emerging as powerful systems for automating tasks across domains. Yet progress in the open-source community is constrained by the lack of high quality permissively licensed tool-agentic training data. Existing datasets are often limited in diversity, realism, and complexity, particularly regarding multi-tool and multi-turn interactions. To address this gap, we introduce Toucan, the largest publicly available tool-agentic dataset to date, containing 1.5 million trajectories synthesized from nearly 500 real-world Model Context Protocols (MCPs). Unlike prior work, Toucan leverages authentic MCP environments to generate diverse, realistic, and challenging tasks with trajectories involving real tool execution. Our pipeline first produces a broad spectrum of tool-use queries using five distinct models, applies model-based quality filtering, and then generates agentic trajectories with three teacher models using two agentic frameworks. Rigorous rule-based and model-based validation ensures high-quality outputs. We also introduce three extension mechanisms to further diversify tasks and simulate multi-turn conversations. Models fine-tuned on Toucan outperform larger closed-source counterparts on the BFCL V3 benchmark and push the Pareto frontier forward on MCP-Universe Bench.

Method: Theoretical analysis of SSM memory capacity by examining how input time series are stored in states, revealing tradeoffs and equivalence between S4 and simplified diagonal S4. Proposed training strategy with fixed recurrent weights.

Result: Analysis showed importance of initial parameters - successful learning requires longest possible initial memory structure. Experiments confirmed extending memory is difficult. Fixed recurrent weights achieved comparable or higher performance with faster convergence.

Conclusion: Provides new theoretical foundation for SSMs and offers novel optimization strategy through fixed recurrent weights, explaining SSM learning dynamics and improving training efficiency.

Abstract: State space models (SSMs) have gained attention by showing potential to outperform Transformers. However, previous studies have not sufficiently addressed the mechanisms underlying their high performance owing to a lack of theoretical explanation of SSMs’ learning dynamics. In this study, we provide such an explanation and propose an improved training strategy. The memory capacity of SSMs can be evaluated by examining how input time series are stored in their current state. Such an examination reveals a tradeoff between memory accuracy and length, as well as the theoretical equivalence between the structured state space sequence model (S4) and a simplified S4 with diagonal recurrent weights. This theoretical foundation allows us to elucidate the learning dynamics, proving the importance of initial parameters. Our analytical results suggest that successful learning requires the initial memory structure to be the longest possible even if memory accuracy may deteriorate or the gradient lose the teacher information. Experiments on tasks requiring long memory confirmed that extending memory is difficult, emphasizing the importance of initialization. Furthermore, we found that fixing recurrent weights can be more advantageous than adapting them because it achieves comparable or even higher performance with faster convergence. Our results provide a new theoretical foundation for SSMs and potentially offer a novel optimization strategy.

Method: Uses mass balance equation analysis to characterize probability mass changes during RL. Increases rollouts per example to hundreds for exhaustive exploration, reducing unsampled token effects and ensuring correct-mass expansion.

Result: BroRL revives models saturated after 3K ProRL steps and achieves continuous improvement. Achieves state-of-the-art results for 1.5B model across diverse benchmarks.

Conclusion: Increasing rollouts per example provides an effective complementary scaling paradigm to training steps, enabling continuous performance gains beyond ProRL saturation points.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a key ingredient for unlocking complex reasoning capabilities in large language models. Recent work ProRL has shown promise in scaling RL by increasing the number of training steps. However, performance plateaus after thousands of steps, with clear diminishing returns from allocating more computation to additional training. In this work, we investigate a complementary paradigm for scaling RL, BroR-Lincreasing the number of rollouts per example to hundreds to exhaustively Broaden exploration, which yields continuous performance gains beyond the saturation point observed in ProRL when scaling the number of training steps. Our approach is motivated by a mass balance equation analysis allowing us to characterize the rate of change in probability mass for correct and incorrect tokens during the reinforcement process. We show that under a one-step RL assumption, sampled rollout tokens always contribute to correct-mass expansion, while unsampled tokens outside rollouts may lead to gains or losses depending on their distribution and the net reward balance. Importantly, as the number of rollouts per example N increases, the effect of unsampled terms diminishes, ensuring overall correct-mass expansion. To validate our theoretical analysis, we conduct simulations under more relaxed conditions and find that a sufficiently large rollout size N-corresponding to ample exploration-guarantees an increase in the probability mass of all correct tokens. Empirically, BroRL revives models saturated after 3K ProRL training steps and demonstrates robust, continuous improvement, achieving state-of-the-art results for the 1.5B model across diverse benchmarks.

Method: Uses data-adaptive learned orthogonal transforms that compact over 90% of signal energy into the first half of dimensions, enabling early candidate pruning with partial distance computations. Integrated into existing ANNS methods without index modification using level-major memory layouts, SIMD-vectorized partial distance computations, and cache-aware access patterns.

Result: Achieves 2-30× end-to-end speedup across diverse datasets (CIFAR-10, GIST, OpenAI’s Ada 2 and Large 3) with no recall loss.

Conclusion: PANORAMA effectively tackles the ANNS verification bottleneck through learned transforms and optimization techniques, providing significant performance improvements without compromising accuracy.

Abstract: Approximate Nearest-Neighbor Search (ANNS) efficiently finds data items whose embeddings are close to that of a given query in a high-dimensional space, aiming to balance accuracy with speed. Used in recommendation systems, image and video retrieval, natural language processing, and retrieval-augmented generation (RAG), ANNS algorithms such as IVFPQ, HNSW graphs, Annoy, and MRPT utilize graph, tree, clustering, and quantization techniques to navigate large vector spaces. Despite this progress, ANNS systems spend up to 99% of query time to compute distances in their final refinement phase. In this paper, we present PANORAMA, a machine learning-driven approach that tackles the ANNS verification bottleneck through data-adaptive learned orthogonal transforms that facilitate the accretive refinement of distance bounds. Such transforms compact over 90% of signal energy into the first half of dimensions, enabling early candidate pruning with partial distance computations. We integrate PANORAMA into state-of-the-art ANNS methods, namely IVFPQ/Flat, HNSW, MRPT, and Annoy, without index modification, using level-major memory layouts, SIMD-vectorized partial distance computations, and cache-aware access patterns. Experiments across diverse datasets – from image-based CIFAR-10 and GIST to modern embedding spaces including OpenAI’s Ada 2 and Large 3 – demonstrate that PANORAMA affords a 2–30$\times$ end-to-end speedup with no recall loss.

Method: Developed new algorithms for private online learning that maintain differential privacy while achieving improved performance bounds. The approach works for concept classes with finite Littlestone dimension.

Result: Achieved optimal O_d(log T) mistake bound against adaptive adversaries in the realizable setting, closing the gap from prior work. Also obtained Õ_d(√T) regret bound in the agnostic setting.

Conclusion: Littlestone classes are privately online learnable with optimal bounds against adaptive adversaries in the realizable setting and sublinear regret in the agnostic setting, demonstrating strong private learnability guarantees.

Abstract: We revisit the problem of private online learning, in which a learner receives a sequence of $T$ data points and has to respond at each time-step a hypothesis. It is required that the entire stream of output hypotheses should satisfy differential privacy. Prior work of Golowich and Livni [2021] established that every concept class $\mathcal{H}$ with finite Littlestone dimension $d$ is privately online learnable in the realizable setting. In particular, they proposed an algorithm that achieves an $O_{d}(\log T)$ mistake bound against an oblivious adversary. However, their approach yields a suboptimal $\tilde{O}{d}(\sqrt{T})$ bound against an adaptive adversary. In this work, we present a new algorithm with a mistake bound of $O{d}(\log T)$ against an adaptive adversary, closing this gap. We further investigate the problem in the agnostic setting, which is more general than the realizable setting as it does not impose any assumptions on the data. We give an algorithm that obtains a sublinear regret of $\tilde{O}_d(\sqrt{T})$ for generic Littlestone classes, demonstrating that they are also privately online learnable in the agnostic setting.

Method: Introduces the expected thresholded calibration error (ETCE) metric and extends computer vision post-processing techniques to forecasting, specifically using selective scaling with lead time conditioning.

Result: The proposed approach reduces model miscalibration without reducing forecast quality, showing improved calibration performance compared to standard methods.

Conclusion: ETCE better captures miscalibration in ordered precipitation classes, and selective scaling with lead time conditioning effectively improves forecast calibration for neural weather models.

Abstract: Reliable precipitation nowcasting is critical for weather-sensitive decision-making, yet neural weather models (NWMs) can produce poorly calibrated probabilistic forecasts. Standard calibration metrics such as the expected calibration error (ECE) fail to capture miscalibration across precipitation thresholds. We introduce the expected thresholded calibration error (ETCE), a new metric that better captures miscalibration in ordered classes like precipitation amounts. We extend post-processing techniques from computer vision to the forecasting domain. Our results show that selective scaling with lead time conditioning reduces model miscalibration without reducing the forecast quality.

Method: Introduces structural causal optimal transport that incorporates both causal graph and structural equations into ambiguity sets. Also proposes a relaxed version with regularization replacing complex causal constraints, enabling efficient solution via difference-of-convex programming.

Result: The approach creates more realistic distributions in ambiguity sets, provides finite sample guarantees when structural information is estimated, and achieves faster shrinkage with dimension-free order, overcoming the curse of dimensionality.

Conclusion: Incorporating structural equations significantly enhances distributionally robust optimization by creating more realistic ambiguity sets, enabling efficient computation, and providing robustness against dimensionality issues while maintaining theoretical guarantees.

Abstract: Distributionally robust optimization tackles out-of-sample issues like overfitting and distribution shifts by adopting an adversarial approach over a range of possible data distributions, known as the ambiguity set. To balance conservatism and accuracy, these sets must include realistic probability distributions by leveraging information from the nominal distribution. Assuming that nominal distributions arise from a structural causal model with a directed acyclic graph $\mathcal{G}$ and structural equations, previous methods such as adapted and $\mathcal{G}$-causal optimal transport have only utilized causal graph information in designing ambiguity sets. In this work, we propose incorporating structural equations, which include causal graph information, to enhance ambiguity sets, resulting in more realistic distributions. We introduce structural causal optimal transport and its associated ambiguity set, demonstrating their advantages and connections to previous methods. A key benefit of our approach is a relaxed version, where a regularization term replaces the complex causal constraints, enabling an efficient algorithm via difference-of-convex programming to solve structural causal optimal transport. We also show that when structural information is absent and must be estimated, our approach remains effective and provides finite sample guarantees. Lastly, we address the radius of ambiguity sets, illustrating how our method overcomes the curse of dimensionality in optimal transport problems, achieving faster shrinkage with dimension-free order.

Method: Error: OutputParser failed

Result: Error: OutputParser failed

Conclusion: Error: OutputParser failed

Abstract: In many real-world applications such as recommendation systems, multiple learning agents must balance exploration and exploitation while maintaining safety guarantees to avoid catastrophic failures. We study the stochastic linear bandit problem in a multi-agent networked setting where agents must satisfy stage-wise conservative constraints. A network of $N$ agents collaboratively maximizes cumulative reward while ensuring that the expected reward at every round is no less than $(1-\alpha)$ times that of a baseline policy. Each agent observes local rewards with unknown parameters, but the network optimizes for the global parameter (average of local parameters). Agents communicate only with immediate neighbors, and each communication round incurs additional regret. We propose MA-SCLUCB (Multi-Agent Stage-wise Conservative Linear UCB), an episodic algorithm alternating between action selection and consensus-building phases. We prove that MA-SCLUCB achieves regret $\tilde{O}\left(\frac{d}{\sqrt{N}}\sqrt{T}\cdot\frac{\log(NT)}{\sqrt{\log(1/|\lambda_2|)}}\right)$ with high probability, where $d$ is the dimension, $T$ is the horizon, and $|\lambda_2|$ is the network’s second largest eigenvalue magnitude. Our analysis shows: (i) collaboration yields $\frac{1}{\sqrt{N}}$ improvement despite local communication, (ii) communication overhead grows only logarithmically for well-connected networks, and (iii) stage-wise safety adds only lower-order regret. Thus, distributed learning with safety guarantees achieves near-optimal performance in reasonably connected networks.

Method: FAME uses bidirectional neural controlled differential equations coupled with mixture-of-experts vector fields to capture intra-functional continuity, and multi-head cross attention to model inter-functional dependencies.

Result: FAME achieves state-of-the-art accuracy on functional regression benchmarks and shows strong robustness to arbitrarily sampled discrete observations of functions.

Conclusion: FAME provides an end-to-end, fully data-driven framework that effectively handles the continuous nature of functional data while achieving superior performance in function-on-function regression tasks.

Abstract: Functional data play a pivotal role across science and engineering, yet their infinite-dimensional nature makes representation learning challenging. Conventional statistical models depend on pre-chosen basis expansions or kernels, limiting the flexibility of data-driven discovery, while many deep-learning pipelines treat functions as fixed-grid vectors, ignoring inherent continuity. In this paper, we introduce Functional Attention with a Mixture-of-Experts (FAME), an end-to-end, fully data-driven framework for function-on-function regression. FAME forms continuous attention by coupling a bidirectional neural controlled differential equation with MoE-driven vector fields to capture intra-functional continuity, and further fuses change to inter-functional dependencies via multi-head cross attention. Extensive experiments on synthetic and real-world functional-regression benchmarks show that FAME achieves state-of-the-art accuracy, strong robustness to arbitrarily sampled discrete observations of functions.

Method: EF21-Muon extends error feedback to non-Euclidean settings, supports stochastic gradients, momentum, and bidirectional compression, and recovers existing methods when compression is disabled.

Result: The method achieves up to 7× communication savings with no accuracy degradation in NanoGPT experiments, matching best-known Euclidean rates and enabling faster convergence under suitable norms.

Conclusion: EF21-Muon provides the first efficient distributed implementation of non-Euclidean LMO-based optimizers with rigorous theory covering various smoothness regimes and practical communication benefits.

Abstract: Recent optimizers like Muon, Scion, and Gluon have pushed the frontier of large-scale deep learning by exploiting layer-wise linear minimization oracles (LMOs) over non-Euclidean norm balls, capturing neural network structure in ways traditional algorithms cannot. Yet, no principled distributed framework exists for these methods, and communication bottlenecks remain unaddressed. The very few distributed variants are heuristic, with no convergence guarantees in sight. We introduce EF21-Muon, the first communication-efficient, non-Euclidean LMO-based optimizer with rigorous convergence guarantees. EF21-Muon supports stochastic gradients, momentum, and bidirectional compression with error feedback-marking the first extension of error feedback beyond the Euclidean setting. It recovers Muon/Scion/Gluon when compression is off and specific norms are chosen, providing the first efficient distributed implementation of this powerful family. Our theory covers non-Euclidean smooth and the more general $(L^0, L^1)$-smooth setting, matching best-known Euclidean rates and enabling faster convergence under suitable norm choices. We further extend the analysis to layer-wise (generalized) smoothness regimes, capturing the anisotropic structure of deep networks. Experiments on NanoGPT benchmarking EF21-Muon against uncompressed Muon/Scion/Gluon demonstrate up to $7\times$ communication savings with no accuracy degradation.

Method: Models pile as Euler-Bernoulli beam and soil as Pasternak foundation, formulates soil-pile interaction as fourth-order ODE (physics component), incorporates measured data (data-driven component), and trains ELM network using least squares method within 1 second.

Result: Validated against BEM and FDM methods; parametric studies show optimal data monitoring locations are at positions with significant pile deflection gradients (pile tip/top and near tunneling zones); approach enables efficient real-time analysis.

Conclusion: PIELM framework successfully combines physics and data for soil-pile interaction analysis, showing great potential for real-time monitoring, safety assessment of pile foundations, and intelligent early-warning systems in geotechnical engineering.

Abstract: Physics-informed machine learning has been a promising data-driven and physics-informed approach in geotechnical engineering. This study proposes a physics-informed extreme learning machine (PIELM) framework for analyzing tunneling-induced soil-pile interactions. The pile foundation is modeled as an Euler-Bernoulli beam, and the surrounding soil is modeled as a Pasternak foundation. The soil-pile interaction is formulated into a fourth-order ordinary differential equation (ODE) that constitutes the physics-informed component, while measured data are incorporated into PIELM as the data-driven component. Combining physics and data yields a loss vector of the extreme learning machine (ELM) network, which is trained within 1 second by the least squares method. After validating the PIELM approach by the boundary element method (BEM) and finite difference method (FDM), parametric studies are carried out to examine the effects of ELM network architecture, data monitoring locations and numbers on the performance of PIELM. The results indicate that monitored data should be placed at positions where the gradients of pile deflections are significant, such as at the pile tip/top and near tunneling zones. Two application examples highlight the critical role of physics-informed and data-driven approach for tunnelling-induced soil-pile interactions. The proposed approach shows great potential for real-time monitoring and safety assessment of pile foundations, and benefits for intelligent early-warning systems in geotechnical engineering.

Method: Used multiple ML algorithms (Decision Trees, k-NN, SVM, AdaBoost, SGD, Naive Bayes, Logistic Regression) with oversampling for class imbalance. Employed One vs Rest approach instead of single multiclass model, evaluated using accuracy, precision, recall, and F1-score.

Result: The study found that data quantity alone doesn’t determine performance - class similarity and dissimilarity are also crucial factors in classification effectiveness.

Conclusion: The One vs Rest combined model approach can optimize classification performance, and successful automated classification depends on both data volume and the inherent characteristics of class distributions.

Abstract: Automated document classification is a trending topic in Natural Language Processing (NLP) due to the extensive growth in digital databases. However, a model that fits well for a specific classification task might perform weakly for another dataset due to differences in the context. Thus, training and evaluating several models is necessary to optimise the results. This study employs a publicly available document database on worldwide digital development interventions categorised under twelve areas. Since digital interventions are still emerging, utilising NLP in the field is relatively new. Given the exponential growth of digital interventions, this research has a vast scope for improving how digital-development-oriented organisations report their work. The paper examines the classification performance of Machine Learning (ML) algorithms, including Decision Trees, k-Nearest Neighbors, Support Vector Machine, AdaBoost, Stochastic Gradient Descent, Naive Bayes, and Logistic Regression. Accuracy, precision, recall and F1-score are utilised to evaluate the performance of these models, while oversampling is used to address the class-imbalanced nature of the dataset. Deviating from the traditional approach of fitting a single model for multiclass classification, this paper investigates the One vs Rest approach to build a combined model that optimises the performance. The study concludes that the amount of data is not the sole factor affecting the performance; features like similarity within classes and dissimilarity among classes are also crucial.

Method: Uses neural networks to map input variables to physical parameters (initial condition, drift, diffusion) of FHT processes like Brownian motion, yielding closed-form survival and hazard functions.

Result: Achieves predictive accuracy comparable to state-of-the-art approaches while providing physics-based interpretable parameterization that elucidates feature-risk relationships.

Conclusion: The combination of stochastic process theory and deep learning offers a principled approach for modeling survival phenomena in complex systems with both accuracy and interpretability.

Abstract: We introduce DeepFHT, a survival-analysis framework that couples deep neural networks with first hitting time (FHT) distributions from stochastic process theory. Time to event is represented as the first passage of a latent diffusion process to an absorbing boundary. A neural network maps input variables to physically meaningful parameters including initial condition, drift, and diffusion, within a chosen FHT process such as Brownian motion, both with drift and driftless. This yields closed-form survival and hazard functions and captures time-varying risk without assuming proportional-hazards. We compare DeepFHT with Cox regression and other existing parametric survival models, using synthetic and real-world datasets. The method achieves predictive accuracy on par with state-of-the-art approaches, while maintaining a physics-based interpretable parameterization that elucidates the relation between input features and risk. This combination of stochastic process theory and deep learning provides a principled avenue for modeling survival phenomena in complex systems.

Method: TD-JEPA trains state and task encoders, a policy-conditioned multi-step predictor, and parameterized policies in latent space using temporal difference learning on offline, reward-free transitions.

Result: Theoretically, TD-JEPA avoids collapse and learns encoders that capture low-rank factorization of long-term policy dynamics. Empirically, it matches or outperforms state-of-the-art baselines on locomotion, navigation, and manipulation tasks across 13 datasets, especially in zero-shot RL from pixels.

Conclusion: TD-JEPA demonstrates that TD-based latent-predictive representations enable effective unsupervised RL and zero-shot reward optimization, advancing the capabilities of latent prediction methods in reinforcement learning.

Abstract: Latent prediction–where agents learn by predicting their own latents–has emerged as a powerful paradigm for training general representations in machine learning. In reinforcement learning (RL), this approach has been explored to define auxiliary losses for a variety of settings, including reward-based and unsupervised RL, behavior cloning, and world modeling. While existing methods are typically limited to single-task learning, one-step prediction, or on-policy trajectory data, we show that temporal difference (TD) learning enables learning representations predictive of long-term latent dynamics across multiple policies from offline, reward-free transitions. Building on this, we introduce TD-JEPA, which leverages TD-based latent-predictive representations into unsupervised RL. TD-JEPA trains explicit state and task encoders, a policy-conditioned multi-step predictor, and a set of parameterized policies directly in latent space. This enables zero-shot optimization of any reward function at test time. Theoretically, we show that an idealized variant of TD-JEPA avoids collapse with proper initialization, and learns encoders that capture a low-rank factorization of long-term policy dynamics, while the predictor recovers their successor features in latent space. Empirically, TD-JEPA matches or outperforms state-of-the-art baselines on locomotion, navigation, and manipulation tasks across 13 datasets in ExoRL and OGBench, especially in the challenging setting of zero-shot RL from pixels.

Method: Used forecasting as the downstream task to evaluate time series foundation models’ zero-shot capabilities and fine-tuning performance compared to smaller dedicated models.

Result: Zero-shot capabilities are domain-dependent, and fine-tuned foundation models don’t consistently provide substantially better results than smaller task-specific models, relative to their parameter count and memory requirements.

Conclusion: Time series data’s diversity makes foundation models less effective than for language/vision, with limited generalization and questionable efficiency compared to dedicated models.

Abstract: Foundation Models are designed to serve as versatile embedding machines, with strong zero shot capabilities and superior generalization performance when fine-tuned on diverse downstream tasks. While this is largely true for language and vision foundation models, we argue that the inherent diversity of time series data makes them less suited for building effective foundation models. We demonstrate this using forecasting as our downstream task. We show that the zero-shot capabilities of a time series foundation model are significantly influenced and tied to the specific domains it has been pretrained on. Furthermore, when applied to unseen real-world time series data, fine-tuned foundation models do not consistently yield substantially better results, relative to their increased parameter count and memory footprint, than smaller, dedicated models tailored to the specific forecasting task at hand.

Method: Combines differentiable approximation of Euler Characteristic Transform (DECT) and its local variant (ℓ-ECT) to create LEAP, an end-to-end trainable local structural positional encoding for graphs.

Result: Evaluated on multiple real-world datasets and a synthetic task designed to test topological feature extraction capability, showing promising performance.

Conclusion: LEAP-based encodings demonstrate potential as a powerful component for graph representation learning pipelines, addressing limitations of standard MPNNs through geometric-topological invariants.

Abstract: Graph neural networks (GNNs) largely rely on the message-passing paradigm, where nodes iteratively aggregate information from their neighbors. Yet, standard message passing neural networks (MPNNs) face well-documented theoretical and practical limitations. Graph positional encoding (PE) has emerged as a promising direction to address these limitations. The Euler Characteristic Transform (ECT) is an efficiently computable geometric-topological invariant that characterizes shapes and graphs. In this work, we combine the differentiable approximation of the ECT (DECT) and its local variant ($\ell$-ECT) to propose LEAP, a new end-to-end trainable local structural PE for graphs. We evaluate our approach on multiple real-world datasets as well as on a synthetic task designed to test its ability to extract topological features. Our results underline the potential of LEAP-based encodings as a powerful component for graph representation learning pipelines.

Method: Investigate optimizer grade (zeroth-order to second-order) impact on unlearning robustness, finding that downgrading optimizers improves resilience. Propose a hybrid optimizer combining first-order and zeroth-order updates.

Result: Extensive experiments on MUSE and WMDP benchmarks show the proposed hybrid optimizer achieves more resilient forgetting across multiple LLM unlearning algorithms without sacrificing unlearning quality.

Conclusion: Optimizer choice is crucial for unlearning robustness, with lower-grade optimizers providing natural advantages. The hybrid optimizer approach effectively balances unlearning efficacy and robustness.

Abstract: Large language model (LLM) unlearning aims to surgically remove the influence of undesired data or knowledge from an existing model while preserving its utility on unrelated tasks. This paradigm has shown promise in addressing privacy and safety concerns. However, recent findings reveal that unlearning effects are often fragile: post-unlearning manipulations such as weight quantization or fine-tuning can quickly neutralize the intended forgetting. Prior efforts to improve robustness primarily reformulate unlearning objectives by explicitly assuming the role of vulnerability sources. In this work, we take a different perspective by investigating the role of the optimizer, independent of unlearning objectives and formulations, in shaping unlearning robustness. We show that the ‘grade’ of the optimizer, defined by the level of information it exploits, ranging from zeroth-order (gradient-free) to first-order (gradient-based) to second-order (Hessian-based), is tightly linked to the resilience of unlearning. Surprisingly, we find that downgrading the optimizer, such as using zeroth-order methods or compressed-gradient variants (e.g., gradient sign-based optimizers), often leads to stronger robustness. While these optimizers produce noisier and less precise updates, they encourage convergence to harder-to-disturb basins in the loss landscape, thereby resisting post-training perturbations. By connecting zeroth-order methods with randomized smoothing, we further highlight their natural advantage for robust unlearning. Motivated by these insights, we propose a hybrid optimizer that combines first-order and zeroth-order updates, preserving unlearning efficacy while enhancing robustness. Extensive experiments on the MUSE and WMDP benchmarks, across multiple LLM unlearning algorithms, validate that our approach achieves more resilient forgetting without sacrificing unlearning quality.

Method: In-place feedback paradigm where users directly edit the LLM’s previous response, and the model conditions on this modified response to generate revisions, rather than relying on new messages.

Result: Empirical evaluations show in-place feedback achieves better performance than conventional multi-turn feedback while using 79.1% fewer tokens. It resolves the core limitation of multi-turn feedback by enabling precise application of feedback to erroneous parts.

Conclusion: In-place feedback offers a more natural and effective mechanism for guiding LLMs in reasoning-intensive tasks by allowing direct editing of responses rather than relying on new messages.

Abstract: Large language models (LLMs) are increasingly studied in the context of multi-turn reasoning, where models iteratively refine their outputs based on user-provided feedback. Such settings are crucial for tasks that require complex reasoning, yet existing feedback paradigms often rely on issuing new messages. LLMs struggle to integrate these reliably, leading to inconsistent improvements. In this work, we introduce in-place feedback, a novel interaction paradigm in which users directly edit an LLM’s previous response, and the model conditions on this modified response to generate its revision. Empirical evaluations on diverse reasoning-intensive benchmarks reveal that in-place feedback achieves better performance than conventional multi-turn feedback while using $79.1%$ fewer tokens. Complementary analyses on controlled environments further demonstrate that in-place feedback resolves a core limitation of multi-turn feedback: models often fail to apply feedback precisely to erroneous parts of the response, leaving errors uncorrected and sometimes introducing new mistakes into previously correct content. These findings suggest that in-place feedback offers a more natural and effective mechanism for guiding LLMs in reasoning-intensive tasks.

Method: Provides design choices and implementation for sample-efficient exploration using explicit, system-agnostic constraints to define regions of interest, enabling constrained diversity maximization.

Result: The approach allows interactive exploration of complex systems that maximizes pattern diversity within user-defined constraints while preserving global diversity.

Conclusion: The proposed method successfully addresses the challenges of exploring complex systems with user expectations, enabling efficient discovery of diverse patterns in constrained regions of interest.

Abstract: The diversity of patterns that emerge from complex systems motivates their use for scientific or artistic purposes. When exploring these systems, the challenges faced are the size of the parameter space and the strongly non-linear mapping between parameters and emerging patterns. In addition, artists and scientists who explore complex systems do so with an expectation of particular patterns. Taking these expectations into account adds a new set of challenges, which the exploration process must address. We provide design choices and their implementation to address these challenges; enabling the maximization of the diversity of patterns discovered in the user’s region of interest – which we call the constrained diversity – in a sample-efficient manner. The region of interest is expressed in the form of explicit constraints. These constraints are formulated by the user in a system-agnostic way, and their addition enables interactive system exploration leading to constrained diversity, while maintaining global diversity.

Method: Extends EvoMol evolutionary algorithm with reinforcement learning, using Extended Connectivity Fingerprints (ECFPs) to learn context-aware mutation policies that prioritize chemically plausible transformations.

Result: Significantly improves generation of valid and realistic molecules, reduces structural artifacts, enhances optimization performance, and consistently outperforms baseline in molecular pre-filtering realism.

Conclusion: Combining reinforcement learning with molecular fingerprints effectively generates chemically relevant molecular structures.

Abstract: The efficient exploration of chemical space remains a central challenge, as many generative models still produce unstable or non-synthesizable compounds. To address these limitations, we present EvoMol-RL, a significant extension of the EvoMol evolutionary algorithm that integrates reinforcement learning to guide molecular mutations based on local structural context. By leveraging Extended Connectivity Fingerprints (ECFPs), EvoMol-RL learns context-aware mutation policies that prioritize chemically plausible transformations. This approach significantly improves the generation of valid and realistic molecules, reducing the frequency of structural artifacts and enhancing optimization performance. The results demonstrate that EvoMol-RL consistently outperforms its baseline in molecular pre-filtering realism. These results emphasize the effectiveness of combining reinforcement learning with molecular fingerprints to generate chemically relevant molecular structures.

Method: Proposes a two-stage algorithm based on low-rank matrix bandits: first performs subspace estimation to identify low-dimensional structure, then uses linear bandit algorithm within the compact representation. Only observes scalar feedback of overall polarization and disagreement after interventions.

Result: The algorithm achieves O(√T) cumulative regret over time horizon T. Empirical results show it significantly outperforms linear bandit baseline in both cumulative regret and running time.

Conclusion: Successfully connects algorithmic interventions on social media platforms with multi-armed bandit theory, providing an effective solution for online polarization minimization under incomplete information.

Abstract: We study the problem of minimizing polarization and disagreement in the Friedkin-Johnsen opinion dynamics model under incomplete information. Unlike prior work that assumes a static setting with full knowledge of users’ innate opinions, we address the more realistic online setting where innate opinions are unknown and must be learned through sequential observations. This novel setting, which naturally mirrors periodic interventions on social media platforms, is formulated as a regret minimization problem, establishing a key connection between algorithmic interventions on social media platforms and theory of multi-armed bandits. In our formulation, a learner observes only a scalar feedback of the overall polarization and disagreement after an intervention. For this novel bandit problem, we propose a two-stage algorithm based on low-rank matrix bandits. The algorithm first performs subspace estimation to identify an underlying low-dimensional structure, and then employs a linear bandit algorithm within the compact dimensional representation derived from the estimated subspace. We prove that our algorithm achieves an $ \widetilde{O}(\sqrt{T}) $ cumulative regret over any time horizon $T$. Empirical results validate that our algorithm significantly outperforms a linear bandit baseline in terms of both cumulative regret and running time.

Method: Integration of enhanced MCTS into GFlowNets sampling process, using MCTS-based policy evaluation to guide generation toward high-reward trajectories and Polynomial Upper Confidence Trees (PUCT) to balance exploration and exploitation adaptively, with a controllable mechanism to regulate greediness.

Result: The method accelerates discovery of high-reward regions and continuously generates high-reward samples while preserving the diversity of the generative distribution.

Conclusion: The proposed approach enhances exploitation without sacrificing diversity by dynamically balancing exploration and reward-driven guidance in GFlowNets.

Abstract: Generative Flow Networks (GFlowNets) have emerged as a powerful tool for generating diverse and high-reward structured objects by learning to sample from a distribution proportional to a given reward function. Unlike conventional reinforcement learning (RL) approaches that prioritize optimization of a single trajectory, GFlowNets seek to balance diversity and reward by modeling the entire trajectory distribution. This capability makes them especially suitable for domains such as molecular design and combinatorial optimization. However, existing GFlowNets sampling strategies tend to overexplore and struggle to consistently generate high-reward samples, particularly in large search spaces with sparse high-reward regions. Therefore, improving the probability of generating high-reward samples without sacrificing diversity remains a key challenge under this premise. In this work, we integrate an enhanced Monte Carlo Tree Search (MCTS) into the GFlowNets sampling process, using MCTS-based policy evaluation to guide the generation toward high-reward trajectories and Polynomial Upper Confidence Trees (PUCT) to balance exploration and exploitation adaptively, and we introduce a controllable mechanism to regulate the degree of greediness. Our method enhances exploitation without sacrificing diversity by dynamically balancing exploration and reward-driven guidance. The experimental results show that our method can not only accelerate the speed of discovering high-reward regions but also continuously generate high-reward samples, while preserving the diversity of the generative distribution. All implementations are available at https://github.com/ZRNB/MG2FlowNet.

Method: Using synthetic datasets with varying periodic structures, measuring trade-off between adaptation to new data and retention of prior knowledge through sequential fine-tuning.

Result: Fine-tuning improves performance on new tasks but causes significant degradation on previously learned ones, demonstrating catastrophic forgetting.

Conclusion: TSFMs face a fundamental stability-plasticity dilemma when adapted sequentially, highlighting the need for better continual learning approaches.

Abstract: Time Series Foundation Models (TSFMs) have shown promising zero-shot generalization across diverse forecasting tasks. However, their robustness to continual adaptation remains underexplored. In this work, we investigate the extent to which TSFMs suffer from catastrophic forgetting when fine-tuned sequentially on multiple datasets. Using synthetic datasets designed with varying degrees of periodic structure, we measure the trade-off between adaptation to new data and retention of prior knowledge. Our experiments reveal that, while fine-tuning improves performance on new tasks, it often causes significant degradation on previously learned ones, illustrating a fundamental stability-plasticity dilemma.

Method: Learn continuous guidance weights ω_c,(s,t) as functions of conditioning c, current denoising time t, and target time s. Minimize distributional mismatch between noised samples from true conditional distribution and guided diffusion process. Extend to reward-guided sampling using reward functions R(x_0,c) on clean data.

Result: Demonstrated effectiveness on low-dimensional toy examples and high-dimensional image generation. Improved Fréchet inception distance (FID) for image generation. In text-to-image applications, CLIP score-based reward functions improved image-prompt alignment.

Conclusion: Dynamic, learned guidance weights outperform static CFG by better balancing perceptual quality and distributional alignment, with additional benefits from reward-guided sampling for improved conditional generation.

Abstract: Classifier-free guidance (CFG) is a widely used technique for improving the perceptual quality of samples from conditional diffusion models. It operates by linearly combining conditional and unconditional score estimates using a guidance weight $\omega$. While a large, static weight can markedly improve visual results, this often comes at the cost of poorer distributional alignment. In order to better approximate the target conditional distribution, we instead learn guidance weights $\omega_{c,(s,t)}$, which are continuous functions of the conditioning $c$, the time $t$ from which we denoise, and the time $s$ towards which we denoise. We achieve this by minimizing the distributional mismatch between noised samples from the true conditional distribution and samples from the guided diffusion process. We extend our framework to reward guided sampling, enabling the model to target distributions tilted by a reward function $R(x_0,c)$, defined on clean data and a conditioning $c$. We demonstrate the effectiveness of our methodology on low-dimensional toy examples and high-dimensional image settings, where we observe improvements in Fr'echet inception distance (FID) for image generation. In text-to-image applications, we observe that employing a reward function given by the CLIP score leads to guidance weights that improve image-prompt alignment.

Method: Formalizes policy gradients with second-order geometry, tracks curvature information during updates, and uses data selection to mask samples causing unstable updates.

Result: CAPO achieves stable updates under aggressive learning regimes where baselines fail, with only 8% token rejection rate and up to 30x sample efficiency improvement over GRPO.

Conclusion: Curvature-aware optimization enables more stable and sample-efficient RL training for LLMs, unlocking scalable post-training with minimal intervention.

Abstract: Reinforcement Learning, particularly through policy gradient methods, has played a central role in enabling reasoning capabilities of Large Language Models. However, the optimization stability of policy gradients in this setting remains understudied. As a result, existing implementations often resort to conservative hyperparameter choices to ensure stability, which requires more training samples and increases computational costs. Hence, developing models for reliably tracking the underlying optimization dynamics and leveraging them into training enables more sample-efficient regimes and further unleashes scalable post-training. We address this gap by formalizing the stochastic optimization problem of policy gradients with explicit consideration of second-order geometry. We propose a tractable computational framework that tracks and leverages curvature information during policy updates. We further employ this framework to design interventions in the optimization process through data selection. The resultant algorithm, Curvature-Aware Policy Optimization (CAPO), identifies samples that contribute to unstable updates and masks them out. Theoretically, we establish monotonic improvement guarantees under realistic assumptions. On standard math reasoning benchmarks, we empirically show that CAPO ensures stable updates under aggressive learning regimes where baselines catastrophically fail. With minimal intervention (rejecting fewer than 8% of tokens), CAPO achieves up to 30x improvement in sample efficiency over standard GRPO for LLM reasoning.

Method: Formulates routing as contextual dueling bandits using pairwise preference feedback. Introduces CCFT with contrastive fine-tuning and categorical weighting to derive model embeddings, enabling FGTS.CDB algorithm with four variants integrating model quality and cost.

Result: Empirical evaluation on RouterBench and MixInstruct datasets shows lower cumulative regret, faster convergence, better robustness, and superior performance-cost balance compared to baselines using general-purpose OpenAI embeddings.

Conclusion: The proposed methods effectively solve LLM routing by combining theoretical grounding with practical implementation, achieving superior performance in model selection while balancing quality and cost considerations.

Abstract: We study LLM routing, the problem of selecting the best model for each query while balancing user satisfaction, model expertise, and inference cost. We formulate routing as contextual dueling bandits, learning from pairwise preference feedback rather than absolute scores, thereby yielding label-efficient and dynamic adaptation. Building on this formulation, we introduce Category-Calibrated Fine-Tuning (CCFT), a representation-learning method that derives model embeddings from offline data using contrastive fine-tuning with categorical weighting. These embeddings enable the practical instantiation of Feel-Good Thompson Sampling for Contextual Dueling Bandits (FGTS.CDB), a theoretically grounded posterior-sampling algorithm. We propose four variants of the categorical weighting that explicitly integrate model quality and cost, and we empirically evaluate the proposed methods on the RouterBench and MixInstruct datasets. Across both benchmarks, our methods achieve lower cumulative regret and faster convergence, with better robustness and performance-cost balance than strong baselines built with a general-purpose OpenAI embedding model.

Method: Uses Wasserstein GAN with mask matrices to represent missing values, training on incomplete microsamples and comparing with models trained on complete data.

Result: Successfully generates synthetic populations that closely resemble both models trained with complete data and the actual population, validated using Swedish national travel survey.

Conclusion: Provides robust solution for population synthesis with incomplete data, demonstrating potential of deep generative models in advancing population synthesis capabilities.

Abstract: This paper presents a population synthesis model that utilizes the Wasserstein Generative-Adversarial Network (WGAN) for training on incomplete microsamples. By using a mask matrix to represent missing values, the study proposes a WGAN training algorithm that lets the model learn from a training dataset that has some missing information. The proposed method aims to address the challenge of missing information in microsamples on one or more attributes due to privacy concerns or data collection constraints. The paper contrasts WGAN models trained on incomplete microsamples with those trained on complete microsamples, creating a synthetic population. We conducted a series of evaluations of the proposed method using a Swedish national travel survey. We validate the efficacy of the proposed method by generating synthetic populations from all the models and comparing them to the actual population dataset. The results from the experiments showed that the proposed methodology successfully generates synthetic data that closely resembles a model trained with complete data as well as the actual population. The paper contributes to the field by providing a robust solution for population synthesis with incomplete data, opening avenues for future research, and highlighting the potential of deep generative models in advancing population synthesis capabilities.

Method: Used Conditional Tabular Generative Adversarial Network (CT-GAN) to create target populations directly from marginal constraints or through hybrid CT-GAN+FBS-CO approach combining deep generative models with traditional optimization.

Result: Stand-alone CT-GAN performed best overall, generating realistic single-variable distributions but struggling with multi-variable relationships. Hybrid model improved FBS-CO performance by using CT-GAN to generate descriptive base populations refined by FBS-CO.

Conclusion: CT-GAN is effective for target population synthesis, and deep generative models can be successfully integrated with conventional techniques to enhance performance in transportation planning applications.

Abstract: Agent-based models used in scenario planning for transportation and urban planning usually require detailed population information from the base as well as target scenarios. These populations are usually provided by synthesizing fake agents through deterministic population synthesis methods. However, these deterministic population synthesis methods face several challenges, such as handling high-dimensional data, scalability, and zero-cell issues, particularly when generating populations for target scenarios. This research looks into how a deep generative model called Conditional Tabular Generative Adversarial Network (CT-GAN) can be used to create target populations either directly from a collection of marginal constraints or through a hybrid method that combines CT-GAN with Fitness-based Synthesis Combinatorial Optimization (FBS-CO). The research evaluates the proposed population synthesis models against travel survey and zonal-level aggregated population data. Results indicate that the stand-alone CT-GAN model performs the best when compared with FBS-CO and the hybrid model. CT-GAN by itself can create realistic-looking groups that match single-variable distributions, but it struggles to maintain relationships between multiple variables. However, the hybrid model demonstrates improved performance compared to FBS-CO by leveraging CT-GAN ability to generate a descriptive base population, which is then refined using FBS-CO to align with target-year marginals. This study demonstrates that CT-GAN represents an effective methodology for target populations and highlights how deep generative models can be successfully integrated with conventional synthesis techniques to enhance their performance.

Method: Interactive web-based dashboard using Python visualization techniques including time series plots, heatmaps, box plots, histograms, correlation matrices, and anomaly-sensitive KPIs (skewness, modified z-scores) for human-in-the-loop data quality assessment.

Result: Demonstrated on real-world Danish district heating dataset (4+ years, 7000 meters), showing visual analytics can uncover systemic data issues and guide future data cleaning strategies to enhance LSTM and GRU model performance for heat demand forecasting.

Conclusion: Provides a scalable, generalizable framework for visual data inspection and emphasizes the critical role of data quality in AI-driven energy management systems.

Abstract: High-quality data is a prerequisite for training reliable Artificial Intelligence (AI) models in the energy domain. In district heating networks, sensor and metering data often suffer from noise, missing values, and temporal inconsistencies, which can significantly degrade model performance. This paper presents a systematic approach for evaluating and improving data quality using visual diagnostics, implemented through an interactive web-based dashboard. The dashboard employs Python-based visualization techniques, including time series plots, heatmaps, box plots, histograms, correlation matrices, and anomaly-sensitive KPIs such as skewness and anomaly detection based on the modified z-scores. These tools al-low human experts to inspect and interpret data anomalies, enabling a human-in-the-loop strategy for data quality assessment. The methodology is demonstrated on a real-world dataset from a Danish district heating provider, covering over four years of hourly data from nearly 7000 meters. The findings show how visual analytics can uncover systemic data issues and, in the future, guide data cleaning strategies that enhance the accuracy, stability, and generalizability of Long Short-Term Memory and Gated Recurrent Unit models for heat demand forecasting. The study contributes to a scalable, generalizable framework for visual data inspection and underlines the critical role of data quality in AI-driven energy management systems.

Method: Training a pre-existing autoencoder using cosmic event data with optimized architecture and parameters.

Result: Significant increase in signal-to-noise ratio of processed signals.

Conclusion: Autoencoders show strong potential for analyzing small signals with multiple interference sources.

Abstract: This brief study focuses on the application of autoencoders to improve the quality of low-amplitude signals, such as gravitational events. A pre-existing autoencoder was trained using cosmic event data, optimizing its architecture and parameters. The results show a significant increase in the signal-to-noise ratio of the processed signals, demonstrating the potential of autoencoders in the analysis of small signals with multiple sources of interference.

Method: Separates structural knowledge (fixed once stabilized) from quantitative knowledge (continuously optimized), reformulating MLPs as path combinations while maintaining universal approximation capabilities.

Result: Reduces training time by ~40% on average while matching or exceeding MLP accuracy with equivalent parameters; applicable across diverse setups including supervised learning, self-supervised learning, and few-shot classification.

Conclusion: GLAI establishes a new design principle for efficient neural network training that can potentially be integrated into large-scale architectures like Transformers where MLPs dominate computation.

Abstract: In this work we introduce GreenLightningAI (GLAI), a new architectural block designed as an alternative to conventional MLPs. The central idea is to separate two types of knowledge that are usually entangled during training: (i) structural knowledge, encoded by the stable activation patterns induced by ReLU activations; and (ii) quantitative knowledge, carried by the numerical weights and biases. By fixing the structure once stabilized, GLAI reformulates the MLP as a combination of paths, where only the quantitative component is optimized. This reformulation retains the universal approximation capabilities of MLPs, yet achieves a more efficient training process, reducing training time by ~40% on average across the cases examined in this study. Crucially, GLAI is not just another classifier, but a generic block that can replace MLPs wherever they are used, from supervised heads with frozen backbones to projection layers in self-supervised learning or few-shot classifiers. Across diverse experimental setups, GLAI consistently matches or exceeds the accuracy of MLPs with an equivalent number of parameters, while converging faster. Overall, GLAI establishes a new design principle that opens a direction for future integration into large-scale architectures such as Transformers, where MLP blocks dominate the computational footprint.

Method: Theoretical analysis comparing mean absolute error vs mean squared error for policy suboptimality control, and empirical evaluation of different loss functions (binary/categorical cross-entropy vs squared loss) in linear reinforcement learning.

Result: Mean absolute error is theoretically superior to mean squared error for controlling learned policy’s suboptimality gap. Cross-entropy losses outperform squared loss in empirical linear RL experiments.

Conclusion: The choice of loss function significantly impacts RL performance, with cross-entropy losses aligned with MAE being more effective than squared loss aligned with MSE for value-based methods.

Abstract: This paper investigates the impact of the loss function in value-based methods for reinforcement learning through an analysis of underlying prediction objectives. We theoretically show that mean absolute error is a better prediction objective than the traditional mean squared error for controlling the learned policy’s suboptimality gap. Furthermore, we present results that different loss functions are better aligned with these different regression objectives: binary and categorical cross-entropy losses with the mean absolute error and squared loss with the mean squared error. We then provide empirical evidence that algorithms minimizing these cross-entropy losses can outperform those based on the squared loss in linear reinforcement learning.

Method: Proposes BoMGene, a hybrid approach integrating Boruta and mRMR feature selection techniques, tested on 25 gene expression datasets using SVM, Random Forest, XGBoost, and GBM classifiers.

Result: The Boruta-mRMR combination reduces feature count compared to mRMR alone, speeds up training time, and maintains or improves classification accuracy over individual methods.

Conclusion: The proposed approach shows clear advantages in accuracy, stability, and practical applicability for multi-class gene expression data analysis.

Abstract: Feature selection is a crucial step in analyzing gene expression data, enhancing classification performance, and reducing computational costs for high-dimensional datasets. This paper proposes BoMGene, a hybrid feature selection method that effectively integrates two popular techniques: Boruta and Minimum Redundancy Maximum Relevance (mRMR). The method aims to optimize the feature space and enhance classification accuracy. Experiments were conducted on 25 publicly available gene expression datasets, employing widely used classifiers such as Support Vector Machine (SVM), Random Forest, XGBoost (XGB), and Gradient Boosting Machine (GBM). The results show that using the Boruta-mRMR combination cuts down the number of features chosen compared to just using mRMR, which helps to speed up training time while keeping or even improving classification accuracy compared to using individual feature selection methods. The proposed approach demonstrates clear advantages in accuracy, stability, and practical applicability for multi-class gene expression data analysis

Method: Proposes Risk-based Policy Optimization (RiskPO) with Mixed Value-at-Risk objectives that integrate weighted attention over multiple reward distribution regions, plus a bundling scheme that aggregates multiple questions for richer feedback.

Result: Achieves consistent and significant improvements in mathematical reasoning, multi-modal reasoning, and code generation benchmarks, surpassing GRPO and variants on both Pass@1 and Pass@k metrics.

Conclusion: Risk-based optimization provides a rigorous and effective paradigm for enhancing LLM reasoning capabilities by preventing entropy collapse and promoting exploration.

Abstract: Reinforcement learning with verifiable reward has recently emerged as a central paradigm for post-training large language models (LLMs); however, prevailing mean-based methods, such as Group Relative Policy Optimization (GRPO), suffer from entropy collapse and limited reasoning gains. We argue that these issues stem from overemphasizing high-probability output sequences while neglecting rare but informative reasoning paths. To address these challenges, we propose Risk-based Policy Optimization (RiskPO), which substitutes classical mean-based objectives with principled risk measures. Specifically, we introduce a Mixed Value-at-Risk objective that integrates weighted attention over multiple regions of the reward distribution, thereby amplifying gradient signals on challenging instances and preventing overconfident convergence. We further design a bundling scheme that aggregates multiple questions into bundles, thus enriching the feedback signal and yielding more stable and informative training dynamics. Theoretically, we prove that the risk-averse update alleviates entropy collapse and promotes exploration. Numerically, RiskPO achieves consistent and significant improvements in mathematical reasoning, multi-modal reasoning, and code generation benchmarks, surpassing GRPO and its variants on both Pass@1 and Pass@k metrics. Our results demonstrate that risk-based optimization provides a rigorous and effective paradigm for enhancing LLM reasoning capabilities.

Method: Model verifier as stochastic reward channel with asymmetric noise rates, then derive two correction algorithms: backward correction (de-biases observed binary reward) and forward correction (reweights score-function terms using only FN rate). Implemented in GRPO-based RLVR pipeline.

Result: Both corrections improve over uncorrected training across models and datasets; forward variant converges faster and remains stable under heavier noise. Lightweight LLM verifier estimating FN rate online outperforms other state-of-the-art methods.

Conclusion: The proposed correction algorithms effectively mitigate verifier unreliability in RLVR systems, with forward correction showing particular advantages in convergence speed and noise stability.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) trains policies against automated verifiers to avoid costly human labeling. To reduce vulnerability to verifier hacking, many RLVR systems collapse rewards to binary ${0,1}$ during training. This choice carries a cost: it introduces \textit{false negatives} (rejecting correct answers, FNs) and \textit{false positives} (accepting incorrect ones, FPs). For instance, a rule-based checker may mark the correct fraction $\frac{12}{36}$ as wrong when compared against the canonical $\frac{1}{3}$ due to brittle parsing/equivalence rules (FN), while a large language model (LLM) judges can be gamed by superficial cues or even a single adversarial token, yielding inflated correctness for wrong solutions (FP). We formalize verifier unreliability by modeling the verifier as a stochastic reward channel with asymmetric noise rates. From this abstraction, we derive two correction algorithms for verifier errors. The first is a \textit{backward} correction that de-biases the observed binary reward to recover an \textit{unbiased} estimator of the clean policy gradient. The second is a \textit{forward} correction that reweights score-function terms so that the expected update direction aligns with the \textit{clean gradient}; notably, it requires only the FN rate. We implement both as lightweight hooks in a group relative policy optimization (GRPO)-based RLVR pipeline and evaluate them on math-reasoning models and benchmarks. Across models and datasets, both corrections improve over uncorrected training; the forward variant converges faster and remains stable under heavier noise. Finally, we show a practical appeal mechanism in which a lightweight LLM verifier estimates the FN rate online by rechecking rule-based negatives, obtaining outperformance compared with other state-of-the-art contenders.

Method: Uses reinforcement learning from human feedback (RLHF) with synthetically generated counter-aligned CoT prefills and standard prompts, requiring no additional training cost or modifications.

Result: Substantially improves safety and jailbreak robustness, reduces overrefusal, preserves core reasoning capability, and maintains inference token budget.

Conclusion: RECAP-trained models engage in more self-reflection and remain robust under adaptive attacks while preserving safety even after repeated override attempts.

Abstract: Large reasoning models (LRMs) “think” by generating structured chain-of-thought (CoT) before producing a final answer, yet they still lack the ability to reason critically about safety alignment and are easily biased when a flawed premise is injected into their thought process. We propose RECAP (Robust Safety Alignment via Counter-Aligned Prefilling), a principled reinforcement learning (RL) method for post-training that explicitly teaches models to override flawed reasoning trajectories and reroute to safe and helpful responses. RECAP trains on a mixture of synthetically generated counter-aligned CoT prefills and standard prompts, requires no additional training cost or modifications beyond vanilla reinforcement learning from human feedback (RLHF), and substantially improves safety and jailbreak robustness, reduces overrefusal, and preserves core reasoning capability – all while maintaining inference token budget. Extensive analysis shows that RECAP-trained models engage in self-reflection more frequently and remain robust under adaptive attacks, preserving safety even after repeated attempts to override their reasoning.

Method: Use covariant derivative and exponential-map parameterization to derive closed-form training objectives, with two variants: Riemannian consistency distillation (RCD) and Riemannian consistency training (RCT).

Result: Superior generative quality demonstrated on various non-Euclidean manifolds including flat-tori, spheres, and SO(3) rotation group.

Conclusion: RCM successfully enables few-step consistency modeling on Riemannian manifolds while respecting intrinsic geometric constraints, with theoretical insights from kinematics perspective.

Abstract: Consistency models are a class of generative models that enable few-step generation for diffusion and flow matching models. While consistency models have achieved promising results on Euclidean domains like images, their applications to Riemannian manifolds remain challenging due to the curved geometry. In this work, we propose the Riemannian Consistency Model (RCM), which, for the first time, enables few-step consistency modeling while respecting the intrinsic manifold constraint imposed by the Riemannian geometry. Leveraging the covariant derivative and exponential-map-based parameterization, we derive the closed-form solutions for both discrete- and continuous-time training objectives for RCM. We then demonstrate theoretical equivalence between the two variants of RCM: Riemannian consistency distillation (RCD) that relies on a teacher model to approximate the marginal vector field, and Riemannian consistency training (RCT) that utilizes the conditional vector field for training. We further propose a simplified training objective that eliminates the need for the complicated differential calculation. Finally, we provide a unique kinematics perspective for interpreting the RCM objective, offering new theoretical angles. Through extensive experiments, we manifest the superior generative quality of RCM in few-step generation on various non-Euclidean manifolds, including flat-tori, spheres, and the 3D rotation group SO(3).

Method: Adapts Random Feature Methods from ANNs to Spike Response Model SNNs, creating S-SWIM algorithm for end-to-end training without spike function gradient approximation.

Result: S-SWIM achieves high accuracy on time series forecasting and serves as effective initialization before gradient-based training, outperforming random weight sampling.

Conclusion: The proposed S-SWIM method provides a fast, high-performance, and interpretable approach for SNN training that addresses gradient propagation difficulties.

Abstract: Spiking Neural Networks (SNNs) as Machine Learning (ML) models have recently received a lot of attention as a potentially more energy-efficient alternative to conventional Artificial Neural Networks. The non-differentiability and sparsity of the spiking mechanism can make these models very difficult to train with algorithms based on propagating gradients through the spiking non-linearity. We address this problem by adapting the paradigm of Random Feature Methods (RFMs) from Artificial Neural Networks (ANNs) to Spike Response Model (SRM) SNNs. This approach allows training of SNNs without approximation of the spike function gradient. Concretely, we propose a novel data-driven, fast, high-performance, and interpretable algorithm for end-to-end training of SNNs inspired by the SWIM algorithm for RFM-ANNs, which we coin S-SWIM. We provide a thorough theoretical discussion and supplementary numerical experiments showing that S-SWIM can reach high accuracies on time series forecasting as a standalone strategy and serve as an effective initialisation strategy before gradient-based training. Additional ablation studies show that our proposed method performs better than random sampling of network weights.

Method: Develops an algorithm that interleaves label-collection and distribution estimation to estimate both the logistic model parameters θ* and context distribution P, using a conservative data-driven threshold on logistic scores to decide when testing is necessary.

Result: The algorithm guarantees with high probability that misclassification rates don’t exceed the target error tolerance α, while requiring only O(√T) excess tests compared to an oracle that knows both θ* and the patient feature distribution.

Conclusion: Establishes the first no-regret guarantees for error-constrained logistic testing, providing a practical solution for cost-sensitive medical screening that maintains safety while minimizing testing costs.

Abstract: We study the problem of sequentially testing individuals for a binary disease outcome whose true risk is governed by an unknown logistic model. At each round, a patient arrives with feature vector $x_t$, and the decision maker may either pay to administer a (noiseless) diagnostic test–revealing the true label–or skip testing and predict the patient’s disease status based on their feature vector and prior history. Our goal is to minimize the total number of costly tests required while guaranteeing that the fraction of misclassifications does not exceed a prespecified error tolerance $\alpha$, with probability at least $1-\delta$. To address this, we develop a novel algorithm that interleaves label-collection and distribution estimation to estimate both $\theta^{}$ and the context distribution $P$, and computes a conservative, data-driven threshold $\tau_t$ on the logistic score $|x_t^\top\theta|$ to decide when testing is necessary. We prove that, with probability at least $1-\delta$, our procedure does not exceed the target misclassification rate, and requires only $O(\sqrt{T})$ excess tests compared to the oracle baseline that knows both $\theta^{}$ and the patient feature distribution $P$. This establishes the first no-regret guarantees for error-constrained logistic testing, with direct applications to cost-sensitive medical screening. Simulations corroborate our theoretical guarantees, showing that in practice our procedure efficiently estimates $\theta^{*}$ while retaining safety guarantees, and does not require too many excess tests.

Method: A novel geometric scattering transform designed for graphs with scalar and vector node features, incorporated into a geometric GNN framework with SE(3)-equivariance properties.

Result: Empirical results show the equivariant scattering-based GNN achieves performance comparable to other equivariant message-passing GNNs while using significantly fewer parameters.

Conclusion: The proposed geometric scattering transform provides an effective and parameter-efficient alternative to traditional equivariant message-passing GNNs for geometric graph learning tasks.

Abstract: We introduce a novel version of the geometric scattering transform for geometric graphs containing scalar and vector node features. This new scattering transform has desirable symmetries with respect to rigid-body roto-translations (i.e., $SE(3)$-equivariance) and may be incorporated into a geometric GNN framework. We empirically show that our equivariant scattering-based GNN achieves comparable performance to other equivariant message-passing-based GNNs at a fraction of the parameter count.

Method: Approaches the problem from RL gradient optimization perspective, identifies key factors (prompt selection and rollout allocation), and proposes CurES with Bayesian posterior estimation to minimize computational overhead.

Result: Outperforms GRPO by +3.30 points with 1.5B model and +4.82 points with 7B model, while achieving faster convergence.

Conclusion: CurES provides a systematic and theoretically grounded approach to improve LLM training efficiency through optimized prompt curriculum design.

Abstract: Curriculum learning plays a crucial role in enhancing the training efficiency of large language models (LLMs) on reasoning tasks. However, existing methods often fail to adequately account for variations in prompt difficulty or rely on simplistic filtering mechanisms to select prompt datasets within a narrow criterion range, resulting in significant computational waste. In this work, we approach the problem from the perspective of reinforcement learning gradient optimization, offering a systematic and theoretical investigation into how to improve the training efficiency of LLMs. We identify two key factors influencing training efficiency: the selection of training prompts and the allocation of rollout quantities across different prompts. Our theoretical analysis reveals that the sampling distribution of prompts dictates the convergence rate of gradient descent, while the allocation of the rollout quantity influences the consistency and stability of overall gradient updates. Based on these insights, we propose CurES, an efficient training method that accelerates convergence and employs Bayesian posterior estimation to minimize computational overhead. Experiments demonstrate that our CurES outperforms Group Relative Policy Optimization (GRPO) by \textbf{+3.30} points and \textbf{+4.82} points with 1.5B and 7B models, respectively. Additionally, CurES exhibits faster convergence compared to baselines, including GRPO.

Method: Proposes Activation Redistribution Module (ARM) - an inference-time technique that identifies near-zero activations after non-linear functions and shifts them outward, redistributing activation patterns without changing input sequences.

Result: Extensive experiments show ARM consistently improves LLM performance on reasoning tasks across diverse benchmarks and model architectures, requiring only minimal code implementation.

Conclusion: The work provides both a mechanistic explanation for meaningless token benefits and a simple, effective technique that harnesses activation redistribution to improve LLM reasoning performance.

Abstract: Motivated by the puzzling observation that inserting long sequences of meaningless tokens before the query prompt can consistently enhance LLM reasoning performance, this work analyzes the underlying mechanism driving this phenomenon and based on these insights proposes a more principled method that allows for similar performance gains. First, we find that the improvements arise from a redistribution of activations in the LLM’s MLP layers, where near zero activations become less frequent while large magnitude activations increase. This redistribution enhances the model’s representational capacity by suppressing weak signals and promoting stronger, more informative ones. Building on this insight, we propose the Activation Redistribution Module (ARM), a lightweight inference-time technique that modifies activations directly without altering the input sequence. ARM adaptively identifies near-zero activations after the non-linear function and shifts them outward, implicitly reproducing the beneficial effects of meaningless tokens in a controlled manner. Extensive experiments across diverse benchmarks and model architectures clearly show that ARM consistently improves LLM performance on reasoning tasks while requiring only a few lines of simple code to implement. Our findings deliver both a clear mechanistic explanation for the unexpected benefits of meaningless tokens and a simple yet effective technique that harnesses activation redistribution to further improve LLM performance.

Method: Proposes Gated X-TFC with soft, learned domain decomposition using differentiable logistic gates that dynamically adapt RBF kernel widths across the domain, eliminating interface penalties. Includes operator-conditioned meta-learning for fast warm-starting.

Result: On 1D convection-diffusion benchmark: order-of-magnitude lower error than standard X-TFC, 80% fewer collocation points, 66% reduction in training time. Scalable to multiple subdomains and higher dimensions (2D Poisson with sharp Gaussian source).

Conclusion: Gated X-TFC provides a simple, accurate, and computationally efficient alternative to PINNs for challenging boundary-layer regimes, with future work focusing on nonlinear problems.

Abstract: Physics-informed neural networks (PINNs) and related methods struggle to resolve sharp gradients in singularly perturbed boundary value problems without resorting to some form of domain decomposition, which often introduce complex interface penalties. While the Extreme Theory of Functional Connections (X-TFC) avoids multi-objective optimization by employing exact boundary condition enforcement, it remains computationally inefficient for boundary layers and incompatible with decomposition. We propose Gated X-TFC, a novel framework for both forward and inverse problems, that overcomes these limitations through a soft, learned domain decomposition. Our method replaces hard interfaces with a differentiable logistic gate that dynamically adapts radial basis function (RBF) kernel widths across the domain, eliminating the need for interface penalties. This approach yields not only superior accuracy but also dramatic improvements in computational efficiency: on a benchmark one dimensional (1D) convection-diffusion, Gated X-TFC achieves an order-of-magnitude lower error than standard X-TFC while using 80 percent fewer collocation points and reducing training time by 66 percent. In addition, we introduce an operator-conditioned meta-learning layer that learns a probabilistic mapping from PDE parameters to optimal gate configurations, enabling fast, uncertainty-aware warm-starting for new problem instances. We further demonstrate scalability to multiple subdomains and higher dimensions by solving a twin boundary-layer equation and a 2D Poisson problem with a sharp Gaussian source. Overall, Gated X-TFC delivers a simple alternative alternative to PINNs that is both accurate and computationally efficient for challenging boundar-layer regimes. Future work will focus on nonlinear problems.

Method: Proposes Parallel-Distill-Refine (PDR): (i) generate diverse drafts in parallel, (ii) distill them into a bounded textual workspace, (iii) refine conditioned on this workspace. Also explores Sequential Refinement (SR) as a subcase. Trains an 8B model with RL to be consistent with PDR inference.

Result: PDR achieves better accuracy than long CoT with lower latency. On math tasks, iterative pipelines surpass single-pass baselines, with PDR delivering largest gains (+11% on AIME 2024, +9% on AIME 2025). Sequential Refinement also outperforms long CoT.

Conclusion: Model orchestrations like PDR can shift the Pareto frontier of reasoning performance, offering better accuracy with lower computational costs. Training models to be consistent with such inference methods further enhances performance.

Abstract: Reasoning training incentivizes LLMs to produce long chains of thought (long CoT), which among other things, allows them to explore solution strategies with self-checking. This results in higher accuracy, but inflates context length, token/compute cost, and answer latency. We ask: Can current models leverage their metacognition to provide other combinations on this Pareto frontier, e.g., better accuracy with lower context length and/or latency? Abstractly, we view the model as an improvement operator on its own “thoughts” with a continuum of possible strategies. We identify an interesting inference family Parallel-Distill-Refine (PDR), which performs the following: (i) generate diverse drafts in parallel; (ii) distill them into a bounded, textual workspace; and (iii) refine conditioned on this workspace, producing an output that seeds the next round. Importantly, context length (hence compute cost) is controllable via degree of parallelism, and is no longer conflated with the total number of generated tokens. We report PDR instantiations of current models that give better accuracy than long CoT while incurring lower latency. Setting degree of parallelism to 1 yields an interesting subcase, Sequential Refinement (SR) (iteratively improve a single candidate answer) which provides performance superior to long CoT. Success of such model orchestrations raises the question whether further training could shift the Pareto frontier. To this end, we train an 8B thinking model with Reinforcement Learning (RL) to make it consistent with PDR as the inference method. On math tasks with verifiable answers, iterative pipelines surpass single-pass baselines at matched sequential budgets, with PDR delivering the largest gains (e.g., +11% on AIME 2024 and +9% on AIME 2025).

Method: Trained three families of LLMs with specific hidden knowledge, then designed and tested black-box (prefill attacks) and white-box (logit lens, sparse autoencoders) techniques to elicit the secret knowledge. Created a benchmark for evaluating secret elicitation methods.

Result: Many techniques improved on simple baselines. Prefill attacks were most effective in 2/3 settings, while white-box techniques (logit lens and SAEs) worked best in the remaining setting. The authors released models and code as a public benchmark.

Conclusion: Secret elicitation is feasible with both black-box and white-box approaches, with prefill attacks being generally effective and white-box methods showing promise in specific scenarios. The work establishes a foundation for evaluating methods to detect hidden knowledge in AI systems.

Abstract: We study secret elicitation: discovering knowledge that an AI possesses but does not explicitly verbalize. As a testbed, we train three families of large language models (LLMs) to possess specific knowledge that they apply downstream but deny knowing when asked directly. For example, in one setting, we train an LLM to generate replies that are consistent with knowing the user is female, while denying this knowledge when asked directly. We then design various black-box and white-box secret elicitation techniques and evaluate them based on whether they can help an LLM auditor successfully guess the secret knowledge. Many of our techniques improve on simple baselines. Our most effective techniques (performing best in 2/3 settings) are based on prefill attacks, a black-box technique where the LLM reveals secret knowledge when generating a completion from a predefined prefix. In our remaining setting, white-box techniques based on logit lens and sparse autoencoders (SAEs) are most effective. We release our models and code, establishing a public benchmark for evaluating secret elicitation methods.

Method: Uses auto-decoder based Implicit Neural Representation with learnable row and feature embeddings to model tabular data as neural functions, enabling instance adaptive imputations without model modification.

Result: Consistently strong imputation accuracy across 12 real-world datasets and multiple missingness mechanisms, mostly matching or outperforming classical and deep learning models, with clearest gains on high-dimensional datasets.

Conclusion: TabINR provides an effective framework for tabular data imputation that handles discrete structure well and delivers robust performance across various dataset sizes and missingness scenarios.

Abstract: Tabular data builds the basis for a wide range of applications, yet real-world datasets are frequently incomplete due to collection errors, privacy restrictions, or sensor failures. As missing values degrade the performance or hinder the applicability of downstream models, and while simple imputing strategies tend to introduce bias or distort the underlying data distribution, we require imputers that provide high-quality imputations, are robust across dataset sizes and yield fast inference. We therefore introduce TabINR, an auto-decoder based Implicit Neural Representation (INR) framework that models tables as neural functions. Building on recent advances in generalizable INRs, we introduce learnable row and feature embeddings that effectively deal with the discrete structure of tabular data and can be inferred from partial observations, enabling instance adaptive imputations without modifying the trained model. We evaluate our framework across a diverse range of twelve real-world datasets and multiple missingness mechanisms, demonstrating consistently strong imputation accuracy, mostly matching or outperforming classical (KNN, MICE, MissForest) and deep learning based models (GAIN, ReMasker), with the clearest gains on high-dimensional datasets.

Method: Two-level ensemble model: first stage uses hyperparameter-tuned base models (Linear SVC, Random Forest, Gradient Boosting, XGBoost) with internal stacking; second stage aggregates predictions using Random Forest as meta-learner.

Result: Achieved Accuracy 0.9433, F1 Score 0.9425, Recall 0.9207, Precision 0.9653, ROC-AUC 0.9844, and AUPRC 0.9875, surpassing one-level stacking and FCN baselines.

Conclusion: The hierarchical stacking strategy provides accurate and interpretable diabetic retinopathy risk prediction suitable for clinical settings, with better generalization and computational efficiency than deep learning approaches.

Abstract: Preprint Note: This is the author preprint version of a paper accepted for presentation at the IISE Annual Conference & Expo 2025. The final version will appear in the official proceedings. Diabetic retinopathy (DR) is a leading cause of blindness in working-age adults, and current diagnostic methods rely on resource-intensive eye exams and specialized equipment. Image-based AI tools have shown limitations in early-stage detection, motivating the need for alternative approaches. We propose a non-image-based, two-level ensemble model for DR prediction using routine laboratory test results. In the first stage, base models (Linear SVC, Random Forest, Gradient Boosting, and XGBoost) are hyperparameter tuned and internally stacked across different configurations to optimize metrics such as accuracy, recall, and precision. In the second stage, predictions are aggregated using Random Forest as a meta-learner. This hierarchical stacking strategy improves generalization, balances performance across multiple metrics, and remains computationally efficient compared to deep learning approaches. The model achieved Accuracy 0.9433, F1 Score 0.9425, Recall 0.9207, Precision 0.9653, ROC-AUC 0.9844, and AUPRC 0.9875, surpassing one-level stacking and FCN baselines. These results highlight the model potential for accurate and interpretable DR risk prediction in clinical settings.

Method: M2PO (Second-Moment Trust Policy Optimization) constrains the second moment of importance weights to suppress only extreme outliers while preserving informative updates from stale data.

Result: M2PO reduces clipped tokens from 1.22% to 0.06% under high staleness, maintains stable optimization, and matches on-policy performance across six models (1.7B to 32B) and eight benchmarks.

Conclusion: M2PO enables stable off-policy training with data stale by at least 256 model updates, achieving on-policy performance while improving training efficiency and scalability.

Abstract: Reinforcement learning has been central to recent advances in large language model reasoning, but most algorithms rely on on-policy training that demands fresh rollouts at every update, limiting efficiency and scalability. Asynchronous RL systems alleviate this by decoupling rollout generation from training, yet their effectiveness hinges on tolerating large staleness in rollout data, a setting where existing methods either degrade in performance or collapse. We revisit this challenge and uncover a prosperity-before-collapse phenomenon: stale data can be as informative as on-policy data if exploited properly. Building on this insight, we introduce M2PO (Second-Moment Trust Policy Optimization), which constrains the second moment of importance weights to suppress only extreme outliers while preserving informative updates. Notably, M2PO sharply reduces the fraction of clipped tokens under high staleness (from 1.22% to 0.06% over training), precisely masking high-variance tokens while maintaining stable optimization. Extensive evaluation across six models (from 1.7B to 32B) and eight benchmarks shows that M2PO delivers stable off-policy training even with data stale by at least 256 model updates and matches on-policy performance.

Method: Uses multi-actor multi-critic architecture with three key components: non-dominated sorting for actor selection based on skill and creativity factors, quantile-based ensemble strategy for actor/critic evaluation, and exploitation of best skill actors. Proves learning stability and bounded estimation bias theoretically.

Result: Outperforms state-of-the-art deep deterministic reinforcement learning methods on MuJoCo benchmark. Experimental analysis confirms effectiveness of proposed components and shows benefits on complicated problems.

Conclusion: The MAMC framework provides a robust solution for reinforcement learning in complex environments by combining multiple actors and critics, with proven theoretical guarantees and superior empirical performance compared to existing methods.

Abstract: Reinforcement learning has gathered much attention in recent years due to its rapid development and rich applications, especially on control systems and robotics. When tackling real-world applications with reinforcement learning method, the corresponded Markov decision process may have huge discrete or even continuous state/action space. Deep reinforcement learning has been studied for handling these issues through deep learning for years, and one promising branch is the actor-critic architecture. Many past studies leveraged multiple critics to enhance the accuracy of evaluation of a policy for addressing the overestimation and underestimation issues. However, few studies have considered the architecture with multiple actors together with multiple critics. This study proposes a novel multi-actor multi-critic (MAMC) deep deterministic reinforcement learning method. The proposed method has three main features, including selection of actors based on non-dominated sorting for exploration with respect to skill and creativity factors, evaluation for actors and critics using a quantile-based ensemble strategy, and exploiting actors with best skill factor. Theoretical analysis proves the learning stability and bounded estimation bias for the MAMC. The present study examines the performance on a well-known reinforcement learning benchmark MuJoCo. Experimental results show that the proposed framework outperforms state-of-the-art deep deterministic based reinforcement learning methods. Experimental analysis also indicates the proposed components are effective. Empirical analysis further investigates the validity of the proposed method, and shows its benefit on complicated problems. The source code can be found at https://github.com/AndyWu101/MAMC.

Method: A set of novel techniques for time series data generation called Fiaingen, evaluated across three criteria: (a) overlap in reduced dimensionality space, (b) downstream ML task performance, and (c) runtime performance.

Result: Fiaingen achieves state-of-the-art performance across all three criteria. Synthetic data closely mirrors original time series while keeping generation time near seconds, ensuring scalability. Models trained on synthetic data achieve performance close to those trained with real data.

Conclusion: Fiaingen provides an effective solution for generating synthetic financial time series data that addresses data scarcity issues while maintaining high quality and computational efficiency.

Abstract: Data is vital in enabling machine learning models to advance research and practical applications in finance, where accurate and robust models are essential for investment and trading decision-making. However, real-world data is limited despite its quantity, quality, and variety. The data shortage of various financial assets directly hinders the performance of machine learning models designed to trade and invest in these assets. Generative methods can mitigate this shortage. In this paper, we introduce a set of novel techniques for time series data generation (we name them Fiaingen) and assess their performance across three criteria: (a) overlap of real-world and synthetic data on a reduced dimensionality space, (b) performance on downstream machine learning tasks, and (c) runtime performance. Our experiments demonstrate that the methods achieve state-of-the-art performance across the three criteria listed above. Synthetic data generated with Fiaingen methods more closely mirrors the original time series data while keeping data generation time close to seconds - ensuring the scalability of the proposed approach. Furthermore, models trained on it achieve performance close to those trained with real-world data.

Method: Uses variational autoencoders for dynamical systems to estimate both system state trajectories and noise time series from data, incorporating teacher forcing strategy and supporting multi-step system evolution.

Result: Demonstrated performance on six test problems including simulated and experimental data, showing the effects of teacher forcing interval on internal dynamics and comparing favorably to deterministic models with equivalent architecture.

Conclusion: The proposed approach effectively addresses limitations of autoencoder-based methods for stochastic systems by jointly estimating states and noise, with teacher forcing improving performance and enabling better multi-step predictions.

Abstract: Learning stochastic models of dynamical systems underlying observed data is of interest in many scientific fields. Here we propose a novel method for this task, based on the framework of variational autoencoders for dynamical systems. The method estimates from the data both the system state trajectories and noise time series. This approach allows to perform multi-step system evolution and supports a teacher forcing strategy, alleviating limitations of autoencoder-based approaches for stochastic systems. We demonstrate the performance of the proposed approach on six test problems, covering simulated and experimental data. We further show the effects of the teacher forcing interval on the nature of the internal dynamics, and compare it to the deterministic models with equivalent architecture.

Method: Formulate exemplar selection as multi-objective optimization problem targeting accuracy maximization and calibration error minimization. Solve using COM-BOM algorithm to find Pareto front that optimally trades off both objectives.

Result: COM-BOM beats or matches baselines at jointly optimizing accuracy and calibration on multiple MMLU-Pro benchmark tasks, while requiring minimal LLM API calls.

Conclusion: Multi-objective optimization approach for exemplar selection effectively balances accuracy and calibration, providing more trustworthy in-context learning with efficient computation.

Abstract: Selecting an optimal set of exemplars is critical for good performance of in-context learning. However, prior exemplar search methods narrowly optimize for predictive accuracy, critically neglecting model calibration–a key determinant of trustworthiness and safe deployment. In this paper, we formulate exemplar selection as a multi-objective optimization problem, explicitly targeting both the maximization of predictive accuracy and the minimization of expected calibration error. We solve this problem with a sample-efficient Combinatorial Bayesian Optimization algorithm (COM-BOM) to find the Pareto front that optimally trades off the two objectives of accuracy and calibration. We evaluate COM-BOM on multiple tasks from unsaturated MMLU-Pro benchmark and find that COM-BOM beats or matches the baselines at jointly optimizing the two objectives, while requiring a minimal number of LLM API calls.

Method: Analyze models through intrinsic dimension estimation, comparing ID_H (last-layer features) with ID_Y (regression targets) across control tasks and synthetic datasets.

Result: Collapsed models show ID_H < ID_Y, leading to poor generalization. Non-collapsed models typically maintain ID_H > ID_Y, with performance depending on data quantity and noise levels. Two regimes identified: over-compressed (ID_H < ID_Y) and under-compressed.

Conclusion: The study provides geometric insights into neural regression and suggests practical strategies for improving generalization by managing the relationship between feature and target intrinsic dimensions.

Abstract: Neural multivariate regression underpins a wide range of domains such as control, robotics, and finance, yet the geometry of its learned representations remains poorly characterized. While neural collapse has been shown to benefit generalization in classification, we find that analogous collapse in regression consistently degrades performance. To explain this contrast, we analyze models through the lens of intrinsic dimension. Across control tasks and synthetic datasets, we estimate the intrinsic dimension of last-layer features (ID_H) and compare it with that of the regression targets (ID_Y). Collapsed models exhibit ID_H < ID_Y, leading to over-compression and poor generalization, whereas non-collapsed models typically maintain ID_H > ID_Y. For the non-collapsed models, performance with respect to ID_H depends on the data quantity and noise levels. From these observations, we identify two regimes (over-compressed and under-compressed) that determine when expanding or reducing feature dimensionality improves performance. Our results provide new geometric insights into neural regression and suggest practical strategies for enhancing generalization.

Method: Augment an LLM with specialized time series perception using a patch-based encoder-decoder architecture, trained on over 2 million interleaved time series and text examples covering forecasting, QA, pattern explanation, classification, and report generation.

Result: TsLLM demonstrates strong performance on tasks requiring integration of time series analysis with natural language, though not designed to surpass specialized models on traditional benchmarks.

Conclusion: This work establishes a new paradigm for time series analysis that bridges numerical computation and natural language understanding, democratizing access to sophisticated temporal reasoning through natural language interaction.

Abstract: Time series data is fundamental to decision-making in many crucial domains including healthcare, finance, and environmental science. However, analyzing this data often requires incorporating unstructured contextual information, answering domain-specific questions, and generating natural language explanations – capabilities that traditional time series models lack due to their inability to process text. While Large Language Models (LLMs) excel at contextual reasoning and knowledge integration, they struggle with numerical time series due to inefficient text-based representations and limited exposure to temporal data during pretraining. We address this gap by augmenting an LLM with specialized time series perception through a patch-based encoder-decoder architecture. We train this Time Series-augmented LLM (TsLLM) on a large corpus of over 2 million interleaved time series and text examples spanning diverse analysis tasks: forecasting with contextual information, time series question-answering, pattern explanation, classification with natural language outputs, and report generation. This training enables TsLLM to leverage both its language understanding and newly acquired temporal reasoning capabilities. While not designed to surpass specialized models on traditional benchmarks, TsLLM demonstrates strong performance on tasks requiring the integration of time series analysis with natural language – capabilities that existing approaches cannot provide. Our work establishes a new paradigm for time series analysis that bridges numerical computation and natural language understanding, democratizing access to sophisticated temporal reasoning through natural language interaction.

Method: Attention mechanism at central server weights local model updates based on significance, using Siamese-CNN for feature extraction and differential privacy with secure update protocols.

Result: A3-FL achieved 0.8413 accuracy vs 0.8164 for FedAvg, 0.7664 for Local-only, and 0.7997 for Centralized. Maintained 0.8330 accuracy with differential privacy.

Conclusion: A3-FL provides a scalable, privacy-preserving biometric system with superior accuracy, convergence speed, and robustness for distributed environments.

Abstract: Because biometric data is sensitive, centralized training poses a privacy risk, even though biometric recognition is essential for contemporary applications. Federated learning (FL), which permits decentralized training, provides a privacy-preserving substitute. Conventional FL, however, has trouble with interpretability and heterogeneous data (non-IID). In order to handle non-IID biometric data, this framework adds an attention mechanism at the central server that weights local model updates according to their significance. Differential privacy and secure update protocols safeguard data while preserving accuracy. The A3-FL framework is evaluated in this study using FVC2004 fingerprint data, with each client’s features extracted using a Siamese Convolutional Neural Network (Siamese-CNN). By dynamically modifying client contributions, the attention mechanism increases the accuracy of the global model.The accuracy, convergence speed, and robustness of the A3-FL framework are superior to those of standard FL (FedAvg) and static baselines, according to experimental evaluations using fingerprint data (FVC2004). The accuracy of the attention-based approach was 0.8413, while FedAvg, Local-only, and Centralized approaches were 0.8164, 0.7664, and 0.7997, respectively. Accuracy stayed high at 0.8330 even with differential privacy. A scalable and privacy-sensitive biometric system for secure and effective recognition in dispersed environments is presented in this work.

Method: Uses Residual Vector-Quantized VAE to create discrete tokens for time series data, followed by two-stage training: supervised fine-tuning on time series tasks, then Group Relative Policy Optimization training with prompting strategies that encourage explicit reasoning steps.

Result: The RL-driven approach with intermediate CoT reasoning significantly enhances LLM performance across various time series analysis tasks.

Conclusion: COUNTS opens new possibilities for complex temporal data reasoning by enabling explicit multi-step reasoning in time series analysis.

Abstract: Complex numerical time series analysis often demands multi-step reasoning capabilities beyond current models’ reach. Tasks like medical diagnosis and weather forecasting require sequential reasoning processes – including counterfactual analysis, logical deduction, knowledge application, and multi-modal contextual integration – that existing time series models cannot explicitly perform. While recent research has shown large language models (LLMs) can achieve sophisticated Chain-of-Thought (CoT) reasoning through reinforcement learning (RL), these advances have primarily focused on mathematical and coding domains, with LLMs still demonstrating poor performance on time series tasks. We introduce Chain Of thought for Understanding Numerical Time Series (COUNTS), the first framework that trains LLMs to perform CoT reasoning across diverse time series tasks using RL with verifiable rewards. Our approach employs a Residual Vector-Quantized VAE to create high-fidelity discrete tokens that seamlessly integrate into a pre-trained LLM’s vocabulary. COUNTS undergoes a two-stage training process: first, supervised fine-tuning on time series analysis tasks to master our novel representations, followed by Group Relative Policy Optimization training on verifiable problems using prompting strategies that encourage explicit reasoning steps before producing final answers. Our experiments demonstrate that this RL-driven approach with intermediate CoT reasoning significantly enhances LLM performance across various time series analysis tasks, opening new possibilities for complex temporal data reasoning.

Method: Transform sequence feature representations into hyperboloid space, compute kernel matrix from hyperboloid features using inner products to measure sequence similarities.

Result: Experimental evaluation shows the approach effectively captures important sequence correlations and improves classification accuracy.

Conclusion: Hyperboloid space transformation enables better representation of biological sequence structures and relationships compared to traditional Euclidean approaches.

Abstract: Genomic sequence analysis plays a crucial role in various scientific and medical domains. Traditional machine-learning approaches often struggle to capture the complex relationships and hierarchical structures of sequence data when working in high-dimensional Euclidean spaces. This limitation hinders accurate sequence classification and similarity measurement. To address these challenges, this research proposes a method to transform the feature representation of biological sequences into the hyperboloid space. By applying a transformation, the sequences are mapped onto the hyperboloid, preserving their inherent structural information. Once the sequences are represented in the hyperboloid space, a kernel matrix is computed based on the hyperboloid features. The kernel matrix captures the pairwise similarities between sequences, enabling more effective analysis of biological sequence relationships. This approach leverages the inner product of the hyperboloid feature vectors to measure the similarity between pairs of sequences. The experimental evaluation of the proposed approach demonstrates its efficacy in capturing important sequence correlations and improving classification accuracy.

Method: Uses Kronecker products to compress feed-forward layer matrices, introduces modified Van Loan decomposition for initialization, and proposes pruning-based initialization technique.

Result: Compressed 124M-parameter GPT-2 to 80M-96M models; 81M variant outperforms DistilGPT2 on next-token prediction across all standard language modeling datasets.

Conclusion: Krony-PT provides effective compression of GPT-2 with competitive performance compared to larger Kronecker-based compressions, demonstrating efficient model size reduction.

Abstract: We introduce Krony-PT, a compression technique for GPT-2 based on Kronecker products. We specifically target the feed-forward weights of each transformer block, and systematically compress the feed-forward layer matrices to various degrees. We introduce a modified Van Loan decomposition to initialize new Kronecker factors, and also propose a new pruning-based initialization technique. Our method compresses the original 124M-parameter GPT-2 to various smaller models, ranging from 80M to 96M. Our 81M model variant outperforms DistilGPT2 on next-token prediction across all standard language modeling datasets, and shows competitive or comparable performance with significantly larger Kronecker-based compressions of GPT-2.

Method: Proposed a post-processing algorithm leveraging random matrix theory to denoise gradients and restore their low-rank structure while maintaining privacy.

Result: Applied to DP-SGD fine-tuning of RoBERTa on GLUE tasks, the method improves sample efficiency and reduces training time compared to state-of-the-art approaches.

Conclusion: Matrix recovery techniques can enhance utility in private language model training without compromising privacy guarantees.

Abstract: We address the challenge of sample efficiency in differentially private fine-tuning of large language models (LLMs) using DP-SGD. While DP-SGD provides strong privacy guarantees, the added noise significantly increases the entropy of gradient matrices, disrupting their low-rank structure and slowing optimization. We propose a post-processing algorithm that leverages random matrix theory to denoise gradients, restore low-rank structure, and improve alignment with the original signal. Applied to DP-SGD fine-tuning of RoBERTa on GLUE tasks, our method improves sample efficiency compared to state-of-the-art approaches, substantially reducing training time when optimal performance is not required. This work demonstrates that matrix recovery techniques can enhance the utility of private language model training without compromising privacy guarantees.

Method: Introduces TDBench with 2000 curated questions per rotation, RotationalEval (RE) to measure consistency across four rotated views, and a reliability framework to separate genuine knowledge from chance.

Result: The benchmark provides comprehensive evaluation of VLMs in top-down perception and offers new insights into model trustworthiness through rigorous reliability metrics.

Conclusion: TDBench not only benchmarks VLMs in top-down perception but also provides a framework for developing more robust and grounded AI systems by focusing on trustworthiness and reliability.

Abstract: Top-down images play an important role in safety-critical settings such as autonomous navigation and aerial surveillance, where they provide holistic spatial information that front-view images cannot capture. Despite this, Vision Language Models (VLMs) are mostly trained and evaluated on front-view benchmarks, leaving their performance in the top-down setting poorly understood. Existing evaluations also overlook a unique property of top-down images: their physical meaning is preserved under rotation. In addition, conventional accuracy metrics can be misleading, since they are often inflated by hallucinations or “lucky guesses”, which obscures a model’s true reliability and its grounding in visual evidence. To address these issues, we introduce TDBench, a benchmark for top-down image understanding that includes 2000 curated questions for each rotation. We further propose RotationalEval (RE), which measures whether models provide consistent answers across four rotated views of the same scene, and we develop a reliability framework that separates genuine knowledge from chance. Finally, we conduct four case studies targeting underexplored real-world challenges. By combining rigorous evaluation with reliability metrics, TDBench not only benchmarks VLMs in top-down perception but also provides a new perspective on trustworthiness, guiding the development of more robust and grounded AI systems. Project homepage: https://github.com/Columbia-ICSL/TDBench

Method: Uses Hamilton-Jacobi equation viscosity solution to characterize OT map, applies method of characteristics to derive closed-form bidirectional transport maps, trains single neural network with characteristic-derived loss function in pure minimization framework.

Result: Achieves accurate, scalable, and efficient optimal transport across diverse datasets, eliminates need for numerical integration and adversarial training, reduces computational complexity while maintaining optimality.

Conclusion: The framework provides a principled and versatile tool for OT applications with provable optimality, supporting various cost functions and class-conditional transport while being computationally efficient.

Abstract: We present a novel framework for solving optimal transport (OT) problems based on the Hamilton–Jacobi (HJ) equation, whose viscosity solution uniquely characterizes the OT map. By leveraging the method of characteristics, we derive closed-form, bidirectional transport maps, thereby eliminating the need for numerical integration. The proposed method adopts a pure minimization framework: a single neural network is trained with a loss function derived from the method of characteristics of the HJ equation. This design guarantees convergence to the optimal map while eliminating adversarial training stages, thereby substantially reducing computational complexity. Furthermore, the framework naturally extends to a wide class of cost functions and supports class-conditional transport. Extensive experiments on diverse datasets demonstrate the accuracy, scalability, and efficiency of the proposed method, establishing it as a principled and versatile tool for OT applications with provable optimality.

Method: PODS decouples rollout generation from policy updates by training only on a strategically selected subset of rollouts using max-variance down-sampling criterion that maximizes reward diversity, with efficient O(n log n) implementation.

Result: Group Relative Policy Optimization (GRPO) with PODS achieves the peak test accuracy of vanilla GRPO at least 1.7x faster across different reasoning benchmarks and hardware configurations.

Conclusion: PODS provides an effective solution to the compute/memory asymmetry in RLVR by enabling efficient training through strategic subset selection while maintaining learning quality.

Abstract: Reinforcement learning with verifiable rewards (RLVR) has emerged as the leading approach for enhancing reasoning capabilities in large language models. However, it faces a fundamental compute and memory asymmetry: rollout generation is embarrassingly parallel and memory-light, whereas policy updates are communication-heavy and memory-intensive. To address this, we introduce PODS (Policy Optimization with Down-Sampling), which decouples rollout generation from policy updates by training only on a strategically selected subset of rollouts, maintaining learning quality while dramatically reducing update costs. We propose a principled subset selection criterion, max-variance down-sampling, that maximizes reward diversity, and provide an efficient $O(n\log n)$ implementation. Empirically, Group Relative Policy Optimization (GRPO) with PODS achieves the peak test accuracy of vanilla GRPO at least $\mathbf{1.7\times}$ faster across the different reasoning benchmarks and hardware configurations we tested.

Method: Uses GAN-inspired adversarial loss to fit neurally parametrised interpolant curves between source and target points, ensuring marginal distributions at intermediate time points match observed distributions. The interpolants are then marginalised by flow matching to train a vector field for the underlying dynamics.

Result: Outperforms existing baselines on spatial transcriptomics and cell tracking datasets, while performing on par with them on single-cell trajectory prediction. The method produces smooth, unique trajectories under mild assumptions.

Conclusion: ALI-CFM provides a versatile and scalable approach for trajectory inference from discrete time observations, demonstrating superior performance on complex biological datasets while maintaining competitive performance on standard tasks.

Abstract: Learning the dynamics of a process given sampled observations at several time points is an important but difficult task in many scientific applications. When no ground-truth trajectories are available, but one has only snapshots of data taken at discrete time steps, the problem of modelling the dynamics, and thus inferring the underlying trajectories, can be solved by multi-marginal generalisations of flow matching algorithms. This paper proposes a novel flow matching method that overcomes the limitations of existing multi-marginal trajectory inference algorithms. Our proposed method, ALI-CFM, uses a GAN-inspired adversarial loss to fit neurally parametrised interpolant curves between source and target points such that the marginal distributions at intermediate time points are close to the observed distributions. The resulting interpolants are smooth trajectories that, as we show, are unique under mild assumptions. These interpolants are subsequently marginalised by a flow matching algorithm, yielding a trained vector field for the underlying dynamics. We showcase the versatility and scalability of our method by outperforming the existing baselines on spatial transcriptomics and cell tracking datasets, while performing on par with them on single-cell trajectory prediction. Code: https://github.com/mmacosha/adversarially-learned-interpolants.

Method: Introduces stepwise guided policy optimization (SGPO) that incorporates response diversity within groups using a step-wise judge model, which can be trained or adapted from existing LLMs.

Result: SGPO consistently outperforms GRPO across 7B, 14B, and 32B models on 9 benchmarks, showing particular gains in early and mid-training stages where all-negative-sample groups are prevalent.

Conclusion: SGPO effectively mitigates the all-negative-sample issue without requiring judge models to generate correct answers, differentiating it from knowledge distillation methods and bridging a key gap between artificial and human intelligence.

Abstract: Reinforcement learning (RL) has proven effective in strengthening the reasoning capabilities of large language models (LLMs). A widely adopted method, Group Relative Policy Optimization (GRPO), has shown strong empirical results in training DeepSeek-R1. However, GRPO fails to update the policy when all responses within a group are incorrect (i.e., \emph{all-negative-sample} groups). This limitation underscores a key gap between artificial and human intelligence: unlike humans, who can learn from mistakes, GRPO discards these signals. Our first contribution is to introduce a simple framework that mitigates the all-negative-sample issue by incorporating response diversity within groups using a \textit{step-wise} judge model, which can be either directly trained or adapted from existing LLMs. We prove that this diversification can accelerate GRPO’s learning dynamics in a simplified setting. We also empirically validate the proposed stepwise guided policy optimization (SGPO) method, demonstrating consistent gains across model sizes (7B, 14B, 32B) in offline and online training on 9 benchmarks, including base and distilled variants. Our results highlight two advantages: (i) SGPO surpasses GRPO, especially in the early and mid-training stages where all-negative-sample groups are prevalent; and (ii) SGPO does not require judge models to generate correct answers, differentiating it from knowledge distillation methods.

Method: Developed a theoretical framework unifying task selection and generalization, extending Bayesian posterior consistency to heavy-tailed priors and dependent sequences. Empirically studied ICL performance on challenging numerical tasks like stochastic differential equations.

Result: Showed that distributional properties (tail behavior, coverage) govern sample efficiency, task retrieval, and robustness in ICL. Found that controlling statistical properties of pretraining data is crucial for ICL capability.

Conclusion: Controlling key statistical properties of the pretraining distribution is essential for building reliable and capable in-context learning systems in large language models.

Abstract: The emergence of in-context learning (ICL) in large language models (LLMs) remains poorly understood despite its consistent effectiveness, enabling models to adapt to new tasks from only a handful of examples. To clarify and improve these capabilities, we characterize how the statistical properties of the pretraining distribution (e.g., tail behavior, coverage) shape ICL on numerical tasks. We develop a theoretical framework that unifies task selection and generalization, extending and sharpening earlier results, and show how distributional properties govern sample efficiency, task retrieval, and robustness. To this end, we generalize Bayesian posterior consistency and concentration results to heavy-tailed priors and dependent sequences, better reflecting the structure of LLM pretraining data. We then empirically study how ICL performance varies with the pretraining distribution on challenging tasks such as stochastic differential equations and stochastic processes with memory. Together, these findings suggest that controlling key statistical properties of the pretraining distribution is essential for building ICL-capable and reliable LLMs.

Method: Applied (generalized) weight normalization with Riemannian optimization to the overparameterized matrix sensing problem.

Result: Proved that weight normalization achieves linear convergence, providing exponential speedup over standard methods without WN, with both iteration and sample complexity improving polynomially as overparameterization increases.

Conclusion: This work provides the first characterization of how weight normalization leverages overparameterization for faster convergence in matrix sensing, establishing theoretical benefits of normalization techniques.

Abstract: While normalization techniques are widely used in deep learning, their theoretical understanding remains relatively limited. In this work, we establish the benefits of (generalized) weight normalization (WN) applied to the overparameterized matrix sensing problem. We prove that WN with Riemannian optimization achieves linear convergence, yielding an exponential speedup over standard methods that do not use WN. Our analysis further demonstrates that both iteration and sample complexity improve polynomially as the level of overparameterization increases. To the best of our knowledge, this work provides the first characterization of how WN leverages overparameterization for faster convergence in matrix sensing.

Method: Uses an energy-based framework to rank CoT solutions, learning to distinguish correct from incorrect reasoning using only simple outcome labels without expensive annotations.

Result: Boosts Llama 3 8B accuracy to 90.7% on GSM8k and 63.7% on MATH, matching or exceeding resource-intensive Best-of-N sampling while being highly efficient.

Conclusion: EORM generalizes effectively to out-of-distribution problems and unseen models, learning fundamental reasoning principles, making it a practical tool for deploying dependable LLMs in real-world applications.

Abstract: Large Language Models (LLMs) struggle with reliable mathematical reasoning, and current verification methods are often computationally expensive. This paper introduces the Energy Outcome Reward Model (EORM), a highly efficient, lightweight post-hoc verifier designed to address this challenge. EORM uses an energy-based framework to rank Chain-of-Thought (CoT) solutions, learning to distinguish correct from incorrect reasoning using only simple outcome labels, thus eliminating the need for expensive annotations. With only 55M parameters, over 127 times smaller than typical reward models, EORM boosts the accuracy of Llama 3 8B to 90.7% on GSM8k and 63.7% on MATH. This performance is achieved by efficiently selecting the optimal reasoning path from a pool of candidates, allowing it to match or exceed the accuracy of far more resource-intensive Best-of-N sampling techniques. Crucially, our experiments show that EORM generalizes effectively to out-of-distribution problems and unseen models, indicating it learns fundamental principles of valid reasoning. This robustness, combined with its efficiency, establishes EORM as a practical tool for deploying more dependable LLMs in complex, real-world applications.

Method: Rescaling learned score functions of noisy data distributions to control local sampling temperature, compatible with existing models and samplers.

Result: Validated on 2D data and applied to five tasks; sharper distributions improved depth prediction while flatter distributions benefited image generation.

Conclusion: Score function rescaling is an effective, training-free approach for controlling sampling diversity across diverse generative modeling tasks.

Abstract: We present a mechanism to steer the sampling diversity of denoising diffusion and flow matching models, allowing users to sample from a sharper or broader distribution than the training distribution. We build on the observation that these models leverage (learned) score functions of noisy data distributions for sampling and show that rescaling these allows one to effectively control a `local’ sampling temperature. Notably, this approach does not require any finetuning or alterations to training strategy, and can be applied to any off-the-shelf model and is compatible with both deterministic and stochastic samplers. We first validate our framework on toy 2D data, and then demonstrate its application for diffusion models trained across five disparate tasks – image generation, pose estimation, depth prediction, robot manipulation, and protein design. We find that across these tasks, our approach allows sampling from sharper (or flatter) distributions, yielding performance gains e.g., depth prediction models benefit from sampling more likely depth estimates, whereas image generation models perform better when sampling a slightly flatter distribution. Project page: https://temporalscorerescaling.github.io

Method: Introduces Dirichlet-Prior Shaping Loss (DPSL), which directly shapes routing probability distributions by matching expert assignments to a target Dirichlet prior, enabling fine-grained control over expert balance and specialization.

Result: DPSL consistently outperforms existing upcycling strategies and regularization techniques across standard vision-language benchmarks with various LLM backbones (Qwen2, Phi3, Llama3.2).

Conclusion: DPSL effectively addresses poor specialization in upcycled MoEs, fosters more adaptive and higher-performing models, and is a general tool applicable to any module outputting categorical probability distributions.

Abstract: Upcycling pre-trained dense models into sparse Mixture-of-Experts (MoEs) efficiently increases model capacity but often suffers from poor expert specialization due to naive weight replication. Our analysis reveals that upcycled MoEs, even with conventional regularization, exhibit low-confidence, weakly differentiated routing, hindering performance. We introduce Dirichlet-Prior Shaping Loss (DPSL), a novel router regularization technique that directly shapes routing probability distributions by matching expert assignments to a target Dirichlet prior. DPSL offers fine-grained control over expert balance and specialization, and enables encoding of inductive biases such as encouraging experts to focus on specific modalities or tasks, without requiring manual intervention; notably, DPSL is a general tool applicable to any module that outputs categorical probability distributions, extending its utility beyond MoE training. Experiments on upcycled MoE vision-language models (with Qwen2, Phi3, Llama3.2 LLM backbones) show DPSL consistently outperforms upcycling strategies and regularization techniques across standard vision-language benchmarks, addressing the critical issue of poor specialization and fostering more adaptive, higher-performing models.

Method: Introduced Seamless Spurious Token Injection (SSTI) to measure vulnerability, where models exclusively focus on single spurious tokens correlated with labels during LoRA finetuning.

Result: Experiments across model families and datasets show that SSTI allows on-demand manipulation of model predictions at test-time, and existing checkers/preprocessors cannot sanitize datasets.

Conclusion: LoRA finetuning creates significant security vulnerabilities that raise new concerns for data quality and AI safety, with current mitigation methods being ineffective.

Abstract: Large Language Models (LLMs) are commonly finetuned for a variety of use cases and domains. A common approach is to leverage Low-Rank Adaptation (LoRA) – known to provide strong performance at low resource costs. In this study, we demonstrate that LoRA actually opens the door to short-cut vulnerabilities – and the more resource efficient is the LoRA setup, the more vulnerable will be the finetuned model to aggressive attacks. To measure that vulnerability, we introduce Seamless Spurious Token Injection (SSTI), where we find that LoRA exclusively focuses on even just a single token that is spuriously correlated with downstream labels. In short, injection of that spurious token during finetuning ensure that the model’s prediction at test-time can be manipulated on-demand. We conducted experiments across model families and datasets to evaluate the impact of SSTI during LoRA finetuning while providing possible mitigations. Our experiments conclude that none of the existing checkers and preprocessors can sanitize a dataset raising new concerns for data quality and AI safety.

Method: Developed a fully Riemannian framework with three components: (1) Riemannion optimizer on fixed-rank manifold, (2) Riemannian gradient-informed LoRA initialization, and (3) efficient implementation using automatic differentiation and numerical linear algebra best practices.

Result: Comprehensive experiments on LLM and diffusion models show consistent and noticeable improvements in convergence speed and final task performance over standard LoRA and state-of-the-art modifications.

Conclusion: The Riemannian framework provides a geometrically principled approach to LoRA that outperforms existing methods by directly optimizing on the fixed-rank manifold.

Abstract: This work presents a novel, fully Riemannian framework for Low-Rank Adaptation (LoRA) that geometrically treats low-rank adapters by optimizing them directly on the fixed-rank manifold. This formulation eliminates the parametrization ambiguity present in standard Euclidean optimizers. Our framework integrates three key components to achieve this: (1) we derive Riemannion, a new Riemannian optimizer on the fixed-rank matrix manifold that generalizes the recently proposed Muon optimizer; (2) we develop a Riemannian gradient-informed LoRA initialization, and (3) we provide an efficient implementation without prominent overhead that uses automatic differentiation to compute arising geometric operations while adhering to best practices in numerical linear algebra. Comprehensive experimental results on both LLM and diffusion model architectures demonstrate that our approach yields consistent and noticeable improvements in convergence speed and final task performance over both standard LoRA and its state-of-the-art modifications.

Method: Two innovations: 1) Row-sparse update formulation using softmax-based top-k classification for sparse state updates, 2) Sparse State Expansion (SSE) that partitions contextual state to decouple parameter size from state capacity.

Result: SSE achieves strong retrieval performance and scales well with state size. A 2B SSE-H model achieves SOTA mathematical reasoning scores (64.5 on AIME24, 50.2 on AIME25), significantly outperforming similarly sized Transformers.

Conclusion: SSE is a promising and efficient architecture for long-context modeling, effectively balancing performance and efficiency.

Abstract: The Transformer architecture, despite its widespread success, struggles with long-context scenarios due to quadratic computation and linear memory growth. While various linear attention variants mitigate these efficiency constraints by compressing context into fixed-size states, they often degrade performance in tasks such as in-context retrieval and reasoning. To address this limitation and achieve more effective context compression, we propose two key innovations. First, we introduce a row-sparse update formulation for linear attention by conceptualizing state updating as information classification. This enables sparse state updates via softmax-based top-$k$ hard classification, thereby extending receptive fields and reducing inter-class interference. Second, we present Sparse State Expansion (SSE) within the sparse framework, which expands the contextual state into multiple partitions, effectively decoupling parameter size from state capacity while maintaining the sparse classification paradigm. Supported by efficient parallelized implementations, our design achieves effective classification and highly discriminative state representations. We extensively validate SSE in both pure linear and hybrid (SSE-H) architectures across language modeling, in-context retrieval, and mathematical reasoning benchmarks. SSE demonstrates strong retrieval performance and scales favorably with state size. Moreover, after reinforcement learning (RL) training, our 2B SSE-H model achieves state-of-the-art mathematical reasoning performance among small reasoning models, scoring 64.5 on AIME24 and 50.2 on AIME25, significantly outperforming similarly sized open-source Transformers. These results highlight SSE as a promising and efficient architecture for long-context modeling.

Method: EVOL-RL mirrors evolutionary principles by retaining majority-voted answers as anchors for stability while adding novelty-aware rewards that score solutions based on how different their reasoning is from other concurrently generated responses.

Result: EVOL-RL consistently outperforms majority-only baselines, improving Qwen3-4B-Base AIME25 pass@1 from 4.6% to 16.4% and pass@16 from 18.5% to 37.9%. It also improves out-of-domain generalization on tasks like GPQA, MMLU-Pro, and BBEH.

Conclusion: EVOL-RL effectively prevents diversity collapse while improving both in-domain performance and out-of-domain generalization through its variation-selection approach to self-improvement.

Abstract: Large language models (LLMs) are increasingly trained with reinforcement learning from verifiable rewards (RLVR), yet real-world deployment demands models that can self-improve without labels or external judges. Existing self-improvement approaches primarily rely on self-confirmation signals (e.g., confidence, entropy, or consistency) to generate rewards. This reliance drives models toward over-confident, majority-favored solutions, causing an entropy collapse that degrades pass@n and reasoning complexity. To address this, we propose EVOL-RL, a label-free framework that mirrors the evolutionary principle of balancing selection with variation. Concretely, EVOL-RL retains the majority-voted answer as an anchor for stability, but adds a novelty-aware reward that scores each sampled solution by how different its reasoning is from other concurrently generated responses. This majority-for-stability + novelty-for-exploration rule mirrors the variation-selection principle: selection prevents drift, while novelty prevents collapse. Evaluation results show that EVOL-RL consistently outperforms the majority-only baseline; e.g., training on label-free AIME24 lifts Qwen3-4B-Base AIME25 pass@1 from baseline’s 4.6% to 16.4%, and pass@16 from 18.5% to 37.9%. EVOL-RL not only prevents in-domain diversity collapse but also improves out-of-domain generalization (from math reasoning to broader tasks, e.g., GPQA, MMLU-Pro, and BBEH). The code is available at: https://github.com/YujunZhou/EVOL-RL.

Method: Shift analysis to hidden-state space using Effective Rank (ER) and propose novel derivatives (ERV and ERA). Develop VERL method that shapes RL advantage function using ERA as predictive meta-controller to create dual-channel incentive structure.

Result: Experiments show consistent gains across diverse LLMs and reasoning benchmarks, including up to 21.4% absolute accuracy improvement on Gaokao 2024 dataset.

Conclusion: Exploration and exploitation can be decoupled at hidden-state level, enabling simultaneous enhancement through synergistic approaches like VERL that avoid forcing traditional trade-offs.

Abstract: A prevailing view in Reinforcement Learning for Verifiable Rewards (RLVR) interprets recent progress through the lens of an exploration-exploitation trade-off, a perspective largely shaped by token-level metrics. We re-examine this perspective, proposing that this perceived trade-off may not be a fundamental constraint but rather an artifact of the measurement level. To investigate this, we shift the analysis to the semantically rich hidden-state space, adopting Effective Rank (ER) to quantify exploration and proposing its novel first- and second-order derivatives, named Effective Rank Velocity (ERV) and Effective Rank Acceleration (ERA), to capture exploitation dynamics. Our analysis reveals that at the hidden-state level, exploration and exploitation could be decoupled (Sec. 4). This finding reveals an opportunity to enhance both capacities simultaneously. This insight motivates our method, Velocity-Exploiting Rank-Learning (VERL), the first to operationalize the principle of synergistic exploration-exploitation enhancement by directly shaping the RL advantage function. The key innovation is leveraging the theoretically stable ERA as a predictive meta-controller to create a synergistic, dual-channel incentive structure. Instead of forcing a trade-off, VERL prospectively amplifies rewards for exploration to preempt overconfidence and reinforces exploitative gains to consolidate reasoning. Experiments across diverse LLMs and reasoning benchmarks show consistent gains, including up to 21.4% absolute accuracy improvement on the challenging Gaokao 2024 dataset.

Method: Uses neural controlled differential equations with a graph-enhanced neural network vector field and time-varying graph path as control signal, enabling continuous-time modeling without piecewise integration.

Result: Empirical evaluation shows effectiveness in capturing complex dynamics of dynamic graphs, with capabilities for trajectory calibration and robustness to missing observations.

Conclusion: GN-CDEs provide an effective continuous-time framework for dynamic graph representation learning that can handle evolving graphs and missing data while capturing complex temporal dynamics.

Abstract: This paper focuses on representation learning for dynamic graphs with temporal interactions. A fundamental issue is that both the graph structure and the nodes own their own dynamics, and their blending induces intractable complexity in the temporal evolution over graphs. Drawing inspiration from the recent progress of physical dynamic models in deep neural networks, we propose Graph Neural Controlled Differential Equations (GN-CDEs), a continuous-time framework that jointly models node embeddings and structural dynamics by incorporating a graph enhanced neural network vector field with a time-varying graph path as the control signal. Our framework exhibits several desirable characteristics, including the ability to express dynamics on evolving graphs without piecewise integration, the capability to calibrate trajectories with subsequent data, and robustness to missing observations. Empirical evaluation on a range of dynamic graph representation learning tasks demonstrates the effectiveness of our proposed approach in capturing the complex dynamics of dynamic graphs.

Method: The authors use information theory to rigorously analyze distributed DNN robustness, then conduct extensive experiments with 6 DNN architectures, 6 distributed approaches, and 10 adversarial attacks on ImageNet-1K.

Result: Theoretical analysis proves that compressed latent dimensions improve robustness but affect task performance, and deeper splitting points enhance robustness but increase computational burden.

Conclusion: The identified trade-offs provide a novel perspective for designing robust distributed DNNs, addressing both security and performance considerations in edge computing scenarios.

Abstract: Distributed deep neural networks (DNNs) have been shown to reduce the computational burden of mobile devices and decrease the end-to-end inference latency in edge computing scenarios. While distributed DNNs have been studied, to the best of our knowledge, the resilience of distributed DNNs to adversarial action remains an open problem. In this paper, we fill the existing research gap by rigorously analyzing the robustness of distributed DNNs against adversarial action. We cast this problem in the context of information theory and rigorously proved that (i) the compressed latent dimension improves the robustness but also affect task-oriented performance; and (ii) the deeper splitting point enhances the robustness but also increases the computational burden. These two trade-offs provide a novel perspective to design robust distributed DNN. To test our theoretical findings, we perform extensive experimental analysis by considering 6 different DNN architectures, 6 different approaches for distributed DNN and 10 different adversarial attacks using the ImageNet-1K dataset.

Method: Hot PATE introduces a diversity-preserving ensemble sampler that efficiently transfers diversity without additional privacy cost, requiring only API access to proprietary models as a drop-in replacement for existing Cold PATE samplers.

Result: Empirical evaluations show significant improvements in privacy-utility trade-off on in-context learning tasks, with better preservation of diversity and more relevant responses.

Conclusion: Hot PATE effectively addresses the diversity-privacy trade-off in generative settings, providing a practical solution that maintains output quality while preserving privacy.

Abstract: The Private Aggregation of Teacher Ensembles (PATE) framework enables privacy-preserving machine learning by aggregating responses from disjoint subsets of sensitive data. Adaptations of PATE to tasks with inherent output diversity such as text generation, where the desired output is a sample from a distribution, face a core tension: as diversity increases, samples from different teachers are less likely to agree, but lower agreement results in reduced utility for the same privacy requirements. Yet suppressing diversity to artificially increase agreement is undesirable, as it distorts the output of the underlying model, and thus reduces output quality. We propose Hot PATE, a variant of PATE designed for diverse generative settings. We formalize the notion of a diversity-preserving ensemble sampler and introduce an efficient sampler that provably transfers diversity without incurring additional privacy cost. Hot PATE requires only API access to proprietary models and can be used as a drop-in replacement for existing Cold PATE samplers. Our empirical evaluations corroborate and quantify the benefits, showing significant improvements in the privacy utility trade-off on evaluated in-context learning tasks, both in preserving diversity and in returning relevant responses.

Method: Propose SMaRt (Score Matching Regularity) that improves GAN optimization by incorporating score matching, which persistently pushes generated data points towards the real data manifold.

Result: Consistently boosts synthesis performance of various state-of-the-art GANs on real-world datasets. Improved FID from 8.87 to 7.11 on ImageNet 64×64 with Aurora, matching one-step consistency model performance.

Conclusion: Score matching serves as a promising solution to address GANs’ limitations in learning from highly diverse data, and SMaRt effectively enhances GAN training through score matching regularization.

Abstract: Generative adversarial networks (GANs) usually struggle in learning from highly diverse data, whose underlying manifold is complex. In this work, we revisit the mathematical foundations of GANs, and theoretically reveal that the native adversarial loss for GAN training is insufficient to fix the problem of \textit{subsets with positive Lebesgue measure of the generated data manifold lying out of the real data manifold}. Instead, we find that score matching serves as a promising solution to this issue thanks to its capability of persistently pushing the generated data points towards the real data manifold. We thereby propose to improve the optimization of GANs with score matching regularity (SMaRt). Regarding the empirical evidences, we first design a toy example to show that training GANs by the aid of a ground-truth score function can help reproduce the real data distribution more accurately, and then confirm that our approach can consistently boost the synthesis performance of various state-of-the-art GANs on real-world datasets with pre-trained diffusion models acting as the approximate score function. For instance, when training Aurora on the ImageNet $64\times64$ dataset, we manage to improve FID from 8.87 to 7.11, on par with the performance of one-step consistency model. Code is available at \href{https://github.com/thuxmf/SMaRt}{https://github.com/thuxmf/SMaRt}.

Method: Uses angles between feature representations and decision boundaries viewed from the mean of in-distribution features as discriminative factor. Also enables ensemble strategy via simple score summation due to scale-invariant nature.

Result: Achieves lowest FPR in 5 out of 9 ImageNet models, best average FPR overall, and consistently ranks among top 3 across all evaluated models. Shows strong performance with contrastive representations (ResNet SCL and CLIP).

Conclusion: Angle-based metric relative to in-distribution structure is an effective approach for OOD detection, outperforming existing methods and working well with contrastive representations.

Abstract: Deep learning systems deployed in real-world applications often encounter data that is different from their in-distribution (ID). A reliable model should ideally abstain from making decisions in this out-of-distribution (OOD) setting. Existing state-of-the-art methods primarily focus on feature distances, such as k-th nearest neighbors and distances to decision boundaries, either overlooking or ineffectively using in-distribution statistics. In this work, we propose a novel angle-based metric for OOD detection that is computed relative to the in-distribution structure. We demonstrate that the angles between feature representations and decision boundaries, viewed from the mean of in-distribution features, serve as an effective discriminative factor between ID and OOD data. We evaluate our method on nine ImageNet-pretrained models. Our approach achieves the lowest FPR in 5 out of 9 ImageNet models, obtains the best average FPR overall, and consistently ranking among the top 3 across all evaluated models. Furthermore, we highlight the benefits of contrastive representations by showing strong performance with ResNet SCL and CLIP architectures. Finally, we demonstrate that the scale-invariant nature of our score enables an ensemble strategy via simple score summation. Code is available at https://github.com/berkerdemirel/ORA-OOD-Detection-with-Relative-Angles.

Method: Uses FHE with three innovations: FHE-friendly architectural modifications (polynomial approximations, low-rank adapters), two-stage split learning (offline knowledge distillation + online encrypted inference), and privacy-boosting schemes (sample permutations, stochastic block sampling).

Result: Empirical results on four image classification datasets demonstrate practical feasibility, though with high communication costs and large computational complexity for the service provider.

Conclusion: BlindFed provides a viable framework for privacy-preserving collaborative FM adaptation that protects both data owners and service providers, despite computational and communication overheads.

Abstract: Foundation models (FMs) excel in zero-shot tasks but benefit from task-specific adaptation. However, privacy concerns prevent data sharing among multiple data owners, and proprietary restrictions prevent the learning service provider (LSP) from sharing the FM. In this work, we propose BlindFed, a framework enabling collaborative FM adaptation while protecting both parties: data owners do not access the FM or each other’s data, and the LSP does not see sensitive task data. BlindFed relies on fully homomorphic encryption (FHE) and consists of three key innovations: (i) FHE-friendly architectural modifications via polynomial approximations and low-rank adapters, (ii) a two-stage split learning approach combining offline knowledge distillation and online encrypted inference for adapter training without backpropagation through the FM, and (iii) a privacy-boosting scheme using sample permutations and stochastic block sampling to mitigate model extraction attacks. Empirical results on four image classification datasets demonstrate the practical feasibility of the BlindFed framework, albeit at a high communication cost and large computational complexity for the LSP.

Method: Proposes a 3D geometric pre-training strategy that constructs a 3D virtual interaction environment to train 2D MRL models to learn both global and local 3D geometric information of molecular interactions.

Result: Extensive experiments on real-world datasets show 3DMRL achieves up to 24.93% performance improvement across 40 tasks, including challenging out-of-distribution and extrapolation scenarios.

Conclusion: 3DMRL successfully bridges the gap between 2D topological and 3D geometric molecular representations, providing an effective and computationally feasible approach for molecular relational learning with significant performance gains.

Abstract: Molecular Relational Learning (MRL) is a rapidly growing field that focuses on understanding the interaction dynamics between molecules, which is crucial for applications ranging from catalyst engineering to drug discovery. Despite recent progress, earlier MRL approaches are limited to using only the 2D topological structure of molecules, as obtaining the 3D interaction geometry remains prohibitively expensive. This paper introduces a novel 3D geometric pre-training strategy for MRL (3DMRL) that incorporates a 3D virtual interaction environment, overcoming the limitations of costly traditional quantum mechanical calculation methods. With the constructed 3D virtual interaction environment, 3DMRL trains 2D MRL model to learn the global and local 3D geometric information of molecular interaction. Extensive experiments on various tasks using real-world datasets, including out-of-distribution and extrapolation scenarios, demonstrate the effectiveness of 3DMRL, showing up to a 24.93% improvement in performance across 40 tasks. Our code is publicly available at https://github.com/Namkyeong/3DMRL.

Method: Dynamic Connection Masking (DCM) mechanism that adaptively masks less important edges during training by evaluating their information-carrying capacity. Can be integrated with various noise-robust training methods.

Result: Extensive experiments show the method consistently outperforms state-of-the-art approaches on both synthetic and real-world benchmarks. Also reveals KANs’ superior noise robustness over MLPs in real-world noisy scenarios.

Conclusion: DCM provides an effective regularization approach for noisy label robustness that can be seamlessly integrated with existing methods. KANs show promise as robust classifiers against noisy labels.

Abstract: Noisy labels are inevitable in real-world scenarios. Due to the strong capacity of deep neural networks to memorize corrupted labels, these noisy labels can cause significant performance degradation. Existing research on mitigating the negative effects of noisy labels has mainly focused on robust loss functions and sample selection, with comparatively limited exploration of regularization in model architecture. Inspired by the sparsity regularization used in Kolmogorov-Arnold Networks (KANs), we propose a Dynamic Connection Masking (DCM) mechanism for both Multi-Layer Perceptron Networks (MLPs) and KANs to enhance the robustness of classifiers against noisy labels. The mechanism can adaptively mask less important edges during training by evaluating their information-carrying capacity. Through theoretical analysis, we demonstrate its efficiency in reducing gradient error. Our approach can be seamlessly integrated into various noise-robust training methods to build more robust deep networks, including robust loss functions, sample selection strategies, and regularization techniques. Extensive experiments on both synthetic and real-world benchmarks demonstrate that our method consistently outperforms state-of-the-art (SOTA) approaches. Furthermore, we are also the first to investigate KANs as classifiers against noisy labels, revealing their superior noise robustness over MLPs in real-world noisy scenarios. Our code will soon be publicly available.

Method: Offline: Use cloud model predictions to set divergence threshold. Runtime: Construct credal sets via thresholding divergence in probability simplex to guarantee inclusion of cloud model predictions.

Result: CD-CI significantly improves calibration performance compared to low-complexity Bayesian methods like Laplace approximation on visual and language tasks.

Conclusion: CD-CI provides practical and efficient uncertainty quantification for edge AI deployments by bridging cloud and edge model capabilities.

Abstract: Deploying artificial intelligence (AI) models on edge devices involves a delicate balance between meeting stringent complexity constraints, such as limited memory and energy resources, and ensuring reliable performance in sensitive decision-making tasks. One way to enhance reliability is through uncertainty quantification via Bayesian inference. This approach, however, typically necessitates maintaining and running multiple models in an ensemble, which may exceed the computational limits of edge devices. This paper introduces a low-complexity methodology to address this challenge by distilling calibration information from a more complex model. In an offline phase, predictive probabilities generated by a high-complexity cloud-based model are leveraged to determine a threshold based on the typical divergence between the cloud and edge models. At run time, this threshold is used to construct credal sets – ranges of predictive probabilities that are guaranteed, with a user-selected confidence level, to include the predictions of the cloud model. The credal sets are obtained through thresholding of a divergence measure in the simplex of predictive probabilities. Experiments on visual and language tasks demonstrate that the proposed approach, termed Conformalized Distillation for Credal Inference (CD-CI), significantly improves calibration performance compared to low-complexity Bayesian methods, such as Laplace approximation, making it a practical and efficient solution for edge AI deployments.

Method: The authors conduct extensive benchmarking of existing continuous GWOT approaches across different scenarios, carefully analyze results, and propose a new continuous GWOT method that doesn’t rely on discrete techniques.

Result: Experimental findings show that current continuous GWOT solvers are unreliable and the community lacks a robust solution. The proposed new method partially solves some competitor problems.

Conclusion: Further research is needed for reliable continuous GWOT solvers. The proposed method represents a first step in this direction by avoiding discrete techniques.

Abstract: Recently, the Gromov-Wasserstein Optimal Transport (GWOT) problem has attracted the special attention of the ML community. In this problem, given two distributions supported on two (possibly different) spaces, one has to find the most isometric map between them. In the discrete variant of GWOT, the task is to learn an assignment between given discrete sets of points. In the more advanced continuous formulation, one aims at recovering a parametric mapping between unknown continuous distributions based on i.i.d. samples derived from them. The clear geometrical intuition behind the GWOT makes it a natural choice for several practical use cases, giving rise to a number of proposed solvers. Some of them claim to solve the continuous version of the problem. At the same time, GWOT is notoriously hard, both theoretically and numerically. Moreover, all existing continuous GWOT solvers still heavily rely on discrete techniques. Natural questions arise: to what extent do existing methods unravel the GWOT problem, what difficulties do they encounter, and under which conditions they are successful? Our benchmark paper is an attempt to answer these questions. We specifically focus on the continuous GWOT as the most interesting and debatable setup. We crash-test existing continuous GWOT approaches on different scenarios, carefully record and analyze the obtained results, and identify issues. Our findings experimentally testify that the scientific community is still missing a reliable continuous GWOT solver, which necessitates further research efforts. As the first step in this direction, we propose a new continuous GWOT method which does not rely on discrete techniques and partially solves some of the problems of the competitors.

Method: I²PFL incorporates both intra-domain and inter-domain prototypes. For intra-domain, it uses feature alignment with MixUp-based augmented prototypes to capture local diversity. For inter-domain, it introduces a reweighting mechanism to generate generalized prototypes that reduce domain shift.

Result: Extensive experiments on Digits, Office-10, and PACS datasets show superior performance compared to other baselines.

Conclusion: The proposed I²PFL method effectively mitigates domain shift from both intra-domain and inter-domain perspectives, learning a generalized global model across multiple domains in federated learning.

Abstract: Federated Learning (FL) has emerged as a decentralized machine learning technique, allowing clients to train a global model collaboratively without sharing private data. However, most FL studies ignore the crucial challenge of heterogeneous domains where each client has a distinct feature distribution, which is popular in real-world scenarios. Prototype learning, which leverages the mean feature vectors within the same classes, has become a prominent solution for federated learning under domain shift. However, existing federated prototype learning methods focus soley on inter-domain prototypes and neglect intra-domain perspectives. In this work, we introduce a novel federated prototype learning method, namely I$^2$PFL, which incorporates $\textbf{I}$ntra-domain and $\textbf{I}$nter-domain $\textbf{P}$rototypes, to mitigate domain shift from both perspectives and learn a generalized global model across multiple domains in federated learning. To construct intra-domain prototypes, we propose feature alignment with MixUp-based augmented prototypes to capture the diversity within local domains and enhance the generalization of local features. Additionally, we introduce a reweighting mechanism for inter-domain prototypes to generate generalized prototypes that reduce domain shift while providing inter-domain knowledge across multiple clients. Extensive experiments on the Digits, Office-10, and PACS datasets illustrate the superior performance of our method compared to other baselines.

Method: Analysis of the Fisher-Rao (natural) gradient of the ELBO gap, derivation of conditions for equivalence between ELBO maximization and KL minimization under model constraints.

Result: The gap between evidence and ELBO has vanishing natural gradient, making ELBO maximization equivalent to KL divergence minimization. This equivalence persists under constrained optimization when models satisfy cylindrical model conditions.

Conclusion: The study provides geometric characterization through cylindrical models for when ELBO optimization remains equivalent to KL divergence minimization, offering insights for constrained learning in generative models.

Abstract: This article studies the Fisher-Rao gradient, also referred to as the natural gradient, of the evidence lower bound (ELBO) which plays a central role in generative machine learning. It reveals that the gap between the evidence and its lower bound, the ELBO, has essentially a vanishing natural gradient within unconstrained optimization. As a result, maximization of the ELBO is equivalent to minimization of the Kullback-Leibler divergence from a target distribution, the primary objective function of learning. Building on this insight, we derive a condition under which this equivalence persists even when optimization is constrained to a model. This condition yields a geometric characterization, which we formalize through the notion of a cylindrical model.

Method: Replaces conventional dot-product and Softmax-based attention with an alternative mechanism using only addition and ReLU activation, maintaining core functionality while being more computationally efficient.

Result: Training experiments on four benchmark tasks show comparable test set prediction scores to conventional Transformers with dot-product attention, with significant computational savings in both plaintext and encrypted scenarios.

Conclusion: The ReLU and addition-based attention mechanism enables efficient execution of quantized Transformers, potentially enabling privacy-preserving AI applications under homomorphic encryption by avoiding costly multiplication of encrypted variables.

Abstract: To enhance the computational efficiency of quantized Transformers, we replace the dot-product and Softmax-based attention with an alternative mechanism involving addition and ReLU activation only. This side-steps the expansion to double precision often required by matrix multiplication and avoids costly Softmax evaluations but maintains much of the core functionality of conventional dot-product attention. It can enable more efficient execution and support larger quantized Transformer models on resource-constrained hardware or alternative arithmetic systems like homomorphic encryption. Training experiments on four common benchmark tasks show test set prediction scores comparable to those of conventional Transformers with dot-product attention. Our scaling experiments also suggest significant computational savings, both in plaintext and under encryption. In particular, we believe that the ReLU and addition-based attention mechanism examined in this paper may enable privacy-preserving AI applications operating under homomorphic encryption by avoiding the costly multiplication of encrypted variables.

Method: Analyze distributions of rotated/unrotated MNIST digits in low-order PCA space, introduce local PCA entropy by binning PCA space and calculating class occurrence entropy, and implement data balancing through oversampling high-entropy regions.

Result: Digits with salient geometric features map to restricted PCA regions and achieve higher prediction accuracy, while ambiguous digits occupy overlapping PCA volumes. Local PCA entropy successfully identifies regions with high prediction confusion.

Conclusion: The geometric structure in PCA space correlates with prediction accuracy, and local PCA entropy provides an effective metric for locating confusing data regions, enabling improved data balancing strategies.

Abstract: This paper examines in detail the geometric structure of principal component analysis (PCA) by considering in detail the distributions of both unrotated and rotated MNIST digits in the space defined by the lowest order PCA components. Since digits possessing salient geometric features are mapped to restricted regions far from the origin, they are predicted by neural networks with a greater accuracy than digits that are mapped to broad, diffuse and overlapping volumes of the low order PCA space. Motivated by these results, a new quantity, the local PCA entropy, obtained by dividing the spatial region spanned by the low order principal components into histogram bins and evaluating the entropy associated with the number of occurrences of each input class within a bin, is introduced. The metric locates the input data records that yield the largest confusion in prediction accuracy within reduced coordinate volumes that optimally discriminate among geometric features. As an example of the potential utility of the local PCA entropy, a simple data balancing procedure is realized by oversampling the data records in regions of large local entropy.

Method: Adds a higher-order penalty on learning rate based on velocity, creating automatic slowdown during large weight updates. Combines velocity-based regularization with Adam’s per-parameter scaling for global damping.

Result: VRAdam exceeds performance of standard optimizers including AdamW across various tasks: image classification, language modeling, and generative modeling using CNNs, Transformers, and GFlowNets.

Conclusion: VRAdam provides theoretical convergence guarantees with rate O(ln(N)/√N) for non-convex objectives and demonstrates practical improvements over existing optimizers through velocity-based stabilization.

Abstract: We introduce Velocity-Regularized Adam (VRAdam), a physics-inspired optimizer for training deep neural networks that draws on ideas from quartic terms for kinetic energy with its stabilizing effects on various system dynamics. Previous algorithms, including the ubiquitous Adam, operate at the so-called adaptive edge of stability regime during training, leading to rapid oscillations and slowed convergence of loss. However, VRAdam adds a higher order penalty on the learning rate based on the velocity such that the algorithm automatically slows down whenever weight updates become large. In practice, we observe that the effective dynamic learning rate shrinks in high-velocity regimes, and damping oscillations. By combining this velocity-based regularizer for global damping with per-parameter scaling of Adam, we create a powerful hybrid optimizer. For this optimizer, we provide rigorous theoretical analysis of operation at the edge of stability from a physical and control perspective for the momentum. Furthermore, we derive convergence bounds with the rate $\mathcal{O}(\ln(N)/\sqrt{N})$ for a stochastic non convex objective under mild assumptions. We demonstrate that VRAdam exceeds the performance against standard optimizers including AdamW. We benchmark various tasks such as image classification, language modeling, and generative modeling using diverse architectures and training methodologies including Convolutional Neural Networks (CNNs), Transformers, and GFlowNets.

Method: Provides systematic survey and fine-grained taxonomy of existing state-of-the-art methods for addressing spurious correlations in ML models.

Result: Summarizes existing datasets, benchmarks, and metrics to facilitate future research on spurious correlation mitigation.

Conclusion: Discusses broader impacts, recent advancements, and future challenges in generative AI era, providing valuable insights for ML researchers.

Abstract: Back in the early 20th century, a horse named Hans appeared to perform arithmetic and other intellectual tasks during exhibitions in Germany, while it actually relied solely on involuntary cues in the body language from the human trainer. Modern machine learning models are no different. These models are known to be sensitive to spurious correlations between non-essential features of the inputs (e.g., background, texture, and secondary objects) and the corresponding labels. Such features and their correlations with the labels are known as “spurious” because they tend to change with shifts in real-world data distributions, which can negatively impact the model’s generalization and robustness. In this paper, we provide a comprehensive survey of this emerging issue, along with a fine-grained taxonomy of existing state-of-the-art methods for addressing spurious correlations in machine learning models. Additionally, we summarize existing datasets, benchmarks, and metrics to facilitate future research. The paper concludes with a discussion of the broader impacts, the recent advancements, and future challenges in the era of generative AI, aiming to provide valuable insights for researchers in the related domains of the machine learning community.

Method: DDPN outputs parameters of Double Poisson distribution, enabling flexible aleatoric uncertainty representation and improved epistemic uncertainty estimation when ensembled, with learnable loss attenuation.

Result: DDPN outperforms current baselines in accuracy, calibration, and out-of-distribution detection across diverse datasets.

Conclusion: DDPN establishes a new state-of-the-art in deep count regression with effective uncertainty representation.

Abstract: Neural networks capable of accurate, input-conditional uncertainty representation are essential for real-world AI systems. Deep ensembles of Gaussian networks have proven highly effective for continuous regression due to their ability to flexibly represent aleatoric uncertainty via unrestricted heteroscedastic variance, which in turn enables accurate epistemic uncertainty estimation. However, no analogous approach exists for count regression, despite many important applications. To address this gap, we propose the Deep Double Poisson Network (DDPN), a novel neural discrete count regression model that outputs the parameters of the Double Poisson distribution, enabling arbitrarily high or low predictive aleatoric uncertainty for count data and improving epistemic uncertainty estimation when ensembled. We formalize and prove that DDPN exhibits robust regression properties similar to heteroscedastic Gaussian models via learnable loss attenuation, and introduce a simple loss modification to control this behavior. Experiments on diverse datasets demonstrate that DDPN outperforms current baselines in accuracy, calibration, and out-of-distribution detection, establishing a new state-of-the-art in deep count regression.

Method: TFPS employs a dual-domain encoder for time and frequency features, uses subspace clustering to identify distinct patterns across data patches, and applies pattern-specific experts to model unique patterns for tailored predictions.

Result: Extensive experiments show TFPS outperforms state-of-the-art methods, particularly in long-term forecasting, through its dynamic and pattern-aware learning approach.

Conclusion: TFPS achieves significantly improved forecasting accuracy by explicitly learning and adapting to evolving patterns in time series data.

Abstract: Time series forecasting, which aims to predict future values based on historical data, has garnered significant attention due to its broad range of applications. However, real-world time series often exhibit complex non-uniform distribution with varying patterns across segments, such as season, operating condition, or semantic meaning, making accurate forecasting challenging. Existing approaches, which typically train a single model to capture all these diverse patterns, often struggle with the pattern drifts between patches and may lead to poor generalization. To address these challenges, we propose TFPS, a novel architecture that leverages pattern-specific experts for more accurate and adaptable time series forecasting. TFPS employs a dual-domain encoder to capture both time-domain and frequency-domain features, enabling a more comprehensive understanding of temporal dynamics. It then uses subspace clustering to dynamically identify distinct patterns across data patches. Finally, pattern-specific experts model these unique patterns, delivering tailored predictions for each patch. By explicitly learning and adapting to evolving patterns, TFPS achieves significantly improved forecasting accuracy. Extensive experiments on real-world datasets demonstrate that TFPS outperforms state-of-the-art methods, particularly in long-term forecasting, through its dynamic and pattern-aware learning approach. The data and codes are available: https://github.com/syrGitHub/TFPS.

Method: CYCle protocol uses novel reputation scoring based on gradient alignment between local cross-entropy and distillation losses, enabling decentralized operation and extending to gossip-based algorithms like Gossip-SGD.

Result: Theoretical analysis shows CYCle outperforms FedAvg in two-client mean estimation under high heterogeneity. Empirical results demonstrate effectiveness in ensuring positive and fair collaboration gain for all participants, even with highly skewed data distributions.

Conclusion: CYCle protocol successfully addresses fairness in collaborative learning by maximizing mean collaboration gain while minimizing gain spread, working effectively in decentralized settings with heterogeneous data distributions.

Abstract: Collaborative learning (CL) enables multiple participants to jointly train machine learning (ML) models on decentralized data sources without raw data sharing. While the primary goal of CL is to maximize the expected accuracy gain for each participant, it is also important to ensure that the gains are fairly distributed: no client should be negatively impacted, and gains should reflect contributions. Most existing CL methods require central coordination and focus only on gain maximization, overlooking fairness. In this work, we first show that the existing measure of collaborative fairness based on the correlation between accuracy values without and with collaboration has drawbacks because it does not account for negative collaboration gain. We argue that maximizing mean collaboration gain (MCG) while simultaneously minimizing the collaboration gain spread (CGS) is a fairer alternative. Next, we propose the CYCle protocol that enables individual participants in a private decentralized learning (PDL) framework to achieve this objective through a novel reputation scoring method based on gradient alignment between the local cross-entropy and distillation losses. We further extend the CYCle protocol to operate on top of gossip-based decentralized algorithms such as Gossip-SGD. We also theoretically show that CYCle performs better than standard FedAvg in a two-client mean estimation setting under high heterogeneity. Empirical experiments demonstrate the effectiveness of the CYCle protocol to ensure positive and fair collaboration gain for all participants, even in cases where the data distributions of participants are highly skewed.

Method: Freeze all base model weights and train only a single d-dimensional steering vector per layer using reinforcement learning.

Result: Matches accuracy of fully RL-tuned models across various base models and math reasoning benchmarks, with minimal parameter overhead.

Conclusion: Millions of adapter weights are unnecessary for high-level reasoning; minimal steering vectors suffice while reducing optimizer memory and inter-GPU communication costs.

Abstract: We show that training a single $d$-dimensional steering vector per layer with reinforcement learning, while freezing all base weights, matches the accuracy of fully RL-tuned reasoning models on mathematical-reasoning tasks. On an 8 billion-parameter model this adds only $\approx 0.0016%$ additional parameters and reproduces performance across a range of base models and mathematical-reasoning benchmarks. These results tighten the upper bound on the parameter budget required for high-level chain-of-thought reasoning, indicating that millions of adapter weights are unnecessary. The minimal trainable footprint reduces optimizer memory and inter-GPU communication, lowering the overall cost of fine-tuning. Moreover, a logit-lens analysis shows that the learned vectors amplify coherent token directions, providing clearer insight into the model’s internal computations.

Method: Proposes a structured bandit architecture operating in trajectory prompt space with linear scaling, using pre-trained PDT as feature extractor for efficient reward modeling.

Result: The method consistently enhances performance across various tasks, high-dimensional environments, and out-of-distribution scenarios, outperforming existing prompt tuning baselines.

Conclusion: Bandit-based prompt tuning with structured architecture and pre-trained feature extraction provides superior performance for adapting Decision Transformers in offline RL.

Abstract: Prompt tuning has emerged as a key technique for adapting large pre-trained Decision Transformers (DTs) in offline Reinforcement Learning (RL), particularly in multi-task and few-shot settings. The Prompting Decision Transformer (PDT) enables task generalization via trajectory prompts sampled uniformly from expert demonstrations – without accounting for prompt informativeness. In this work, we propose a bandit-based prompt-tuning method that learns to construct optimal trajectory prompts from demonstration data at inference time. We devise a structured bandit architecture operating in the trajectory prompt space, achieving linear rather than combinatorial scaling with prompt size. Additionally, we show that the pre-trained PDT itself can serve as a powerful feature extractor for the bandit, enabling efficient reward modeling across various environments. We theoretically establish regret bounds and demonstrate empirically that our method consistently enhances performance across a wide range of tasks, high-dimensional environments, and out-of-distribution scenarios, outperforming existing baselines in prompt tuning.

Method: Object-Centric Concept Bottlenecks (OCB) framework that integrates concept-based models with pre-trained object-centric foundation models, using strategies for aggregating object-concept encodings.

Result: OCB outperforms traditional CBMs on complex image datasets and enables interpretable decisions for complex visual tasks.

Conclusion: OCB successfully combines the strengths of CBMs and object-centric models to achieve both high performance and interpretability in complex vision tasks.

Abstract: Developing high-performing, yet interpretable models remains a critical challenge in modern AI. Concept-based models (CBMs) attempt to address this by extracting human-understandable concepts from a global encoding (e.g., image encoding) and then applying a linear classifier on the resulting concept activations, enabling transparent decision-making. However, their reliance on holistic image encodings limits their expressiveness in object-centric real-world settings and thus hinders their ability to solve complex vision tasks beyond single-label classification. To tackle these challenges, we introduce Object-Centric Concept Bottlenecks (OCB), a framework that combines the strengths of CBMs and pre-trained object-centric foundation models, boosting performance and interpretability. We evaluate OCB on complex image datasets and conduct a comprehensive ablation study to analyze key components of the framework, such as strategies for aggregating object-concept encodings. The results show that OCB outperforms traditional CBMs and allows one to make interpretable decisions for complex visual tasks.

Method: Designed key data types for fundamental molecular features, proposed a module integrating complementary information from different molecular encoders to leverage distinct advantages of molecular representations.

Result: Experimental results show Mol-LLaMA can comprehend general molecular features and provide informative responses, demonstrating potential as a general-purpose molecular analysis assistant.

Conclusion: Mol-LLaMA represents a significant advancement in molecular language models by incorporating reasoning capabilities and comprehensive molecular knowledge, making it suitable for drug discovery and molecular analysis tasks.

Abstract: Understanding molecules is key to understanding organisms and driving advances in drug discovery, requiring interdisciplinary knowledge across chemistry and biology. Although large molecular language models have achieved notable success in task transfer, they often struggle to accurately analyze molecular features due to limited knowledge and reasoning capabilities. To address this issue, we present Mol-LLaMA, a large molecular language model that grasps the general knowledge centered on molecules and exhibits explainability and reasoning ability. To this end, we design key data types that encompass the fundamental molecular features, taking into account the essential abilities for molecular reasoning. Further, to improve molecular understanding, we propose a module that integrates complementary information from different molecular encoders, leveraging the distinct advantages of molecular representations. Our experimental results demonstrate that Mol-LLaMA is capable of comprehending the general features of molecules and providing informative responses, implying its potential as a general-purpose assistant for molecular analysis. Our project page is at https://mol-llama.github.io/.

Method: Proposes model-based risk-sensitive Q-value iteration (MB-RS-QVI) algorithm and provides both upper bounds through PAC analysis and matching lower bounds.

Result: Shows that both upper and lower bounds have exponential dependence on |β|/(1-γ), proving this dependence is unavoidable and that polynomial dependence on all model parameters is impossible in risk-sensitive settings.

Conclusion: Risk-sensitive RL fundamentally requires exponential sample complexity in the risk parameter and effective horizon, unlike classical risk-neutral RL where polynomial dependence is possible.

Abstract: In this paper, we analyze the sample complexities of learning the optimal state-action value function $Q^$ and an optimal policy $\pi^$ in a finite discounted Markov decision process (MDP) where the agent has recursive entropic risk-preferences with risk-parameter $\beta\neq 0$ and where a generative model of the MDP is available. We provide and analyze a simple model based approach which we call model-based risk-sensitive $Q$-value-iteration (MB-RS-QVI) which leads to $(\varepsilon,\delta)$-PAC-bounds on $|Q^-Q^k|$, and $|V^-V^{\pi_k}|$ where $Q_k$ is the output of MB-RS-QVI after k iterations and $\pi_k$ is the greedy policy with respect to $Q_k$. Both PAC-bounds have exponential dependence on the effective horizon $\frac{1}{1-\gamma}$ and the strength of this dependence grows with the learners risk-sensitivity $|\beta|$. We also provide two lower bounds which shows that exponential dependence on $|\beta|\frac{1}{1-\gamma}$ is unavoidable in both cases. The lower bounds reveal that the PAC-bounds are tight in the parameters $S,A,\delta,\varepsilon$ and that unlike in the classical setting it is not possible to have polynomial dependence in all model parameters.

Method: Replaced GRU with LSTM in DSTGCRN models for better long-term dependency capture, and integrated Client-Side Validation mechanism to validate server-aggregated parameters before local model updates.

Result: Substantial improvements over conventional methods in multimodal transport demand forecasting and OD matrix forecasting, with enhanced ability to capture complex spatiotemporal dependencies while maintaining data privacy.

Conclusion: The framework provides scalable, privacy-preserving solution for real-time region-specific forecasting and demonstrates the potential of leveraging distributed data sources in federated learning contexts.

Abstract: This paper presents an advanced Federated Learning (FL) framework for forecasting complex spatiotemporal data, improving upon recent state-of-the-art models. In the proposed approach, the original Gated Recurrent Unit (GRU) module within previous Dynamic Spatial–Temporal Graph Convolutional Recurrent Network (DSTGCRN) modeling is first replaced with a Long Short-Term Memory (LSTM) network, enabling the resulting model to more effectively capture long-term dependencies inherent to time series data. The resulting architecture significantly improves the model’s capacity to handle complex temporal patterns in diverse forecasting applications. Furthermore, the proposed FL framework integrates a novel Client-Side Validation (CSV) mechanism, introducing a critical validation step at the client level before incorporating aggregated parameters from the central server into local models, ensuring only the most effective updates are retained and improving both the robustness and accuracy of the forecasting model across clients. The efficiency of our approach is demonstrated through extensive experiments on real-world applications, including public datasets for multimodal transport demand forecasting and private datasets for Origin-Destination (OD) matrix forecasting in urban areas. The results demonstrate substantial improvements over conventional methods, highlighting the framework’s ability to capture complex spatiotemporal dependencies while preserving data privacy. This work not only provides a scalable and privacy-preserving solution for real-time, region-specific forecasting and management but also underscores the potential of leveraging distributed data sources in a FL context. We provide our algorithms as open-source on GitHub

Method: Uses structured prediction bias of VLMs as inductive bias for teacher models. Proposes Probabilistic CoreSet (PCoreSet) selection strategy that selects probabilistically diverse unlabeled samples to maximize coverage in probability space rather than feature space.

Result: ActiveKD consistently improves performance across selection methods (+29.07% on ImageNet averaged over methods). PCoreSet ranks first in 64/73 settings (87.7%) across 5 student and 3 teacher networks, achieving best performance except for first 2 AL rounds.

Conclusion: ActiveKD effectively integrates AL with KD using VLMs, with PCoreSet demonstrating superior performance in selecting diverse samples for efficient knowledge transfer under limited annotation budgets.

Abstract: Knowledge distillation (KD) is a widely used framework for training compact, task-specific models by transferring the knowledge from teacher models. However, its application to active learning (AL), which aims to minimize annotation costs through iterative sample selection, remains underexplored. This gap stems from the fact that KD typically assumes access to sufficient labeled data, whereas AL operates in data-scarce scenarios where task-specific teacher models are often unavailable. In this paper, we first introduce ActiveKD, a framework that integrates AL with KD by leveraging the zero- and few-shot capabilities of large vision-language models (VLMs). A key aspect of ActiveKD is the structured prediction bias of VLMs-i.e., their predictions form clusters in the probability space. We regard this structure as an inductive bias of the teacher model, capturing generalizable output patterns beneficial to student learning. To exploit this bias, we propose Probabilistic CoreSet (PCoreSet), a selection strategy that maximizes coverage in the probability space rather than the feature space. PCoreSet strategically selects probabilistically diverse unlabeled samples, facilitating more efficient transfer of teacher knowledge under limited annotation budgets. Extensive evaluations on 11 datasets show that ActiveKD consistently improves performance across selection methods (e.g., +29.07% on ImageNet, averaged over methods). Under ActiveKD, PCoreSet ranks first in 64/73 settings (approximately 87.7%) across 5 student and 3 teacher networks, always achieving the best performance except for first 2 AL rounds. Our code is available at https://github.com/erjui/PCoreSet.

Method: Introduces Damped Linear Oscillatory State-Space models (D-LinOSS) that learn to dissipate latent state energy on arbitrary time scales through flexible parameterization while maintaining stable dynamics.

Result: D-LinOSS consistently outperforms previous LinOSS methods on long-range learning tasks, achieves faster convergence, and reduces the hyperparameter search space by 50% without adding complexity.

Conclusion: D-LinOSS provides a more general and effective class of oscillatory SSMs with flexible forgetting mechanisms that improve long-range reasoning capabilities while maintaining computational efficiency.

Abstract: State-space models (SSMs) are a class of networks for sequence learning that benefit from fixed state size and linear complexity with respect to sequence length, contrasting the quadratic scaling of typical attention mechanisms. Inspired from observations in neuroscience, Linear Oscillatory State-Space models (LinOSS) are a recently proposed class of SSMs constructed from layers of discretized forced harmonic oscillators. Although these models perform competitively, leveraging fast parallel scans over diagonal recurrent matrices and achieving state-of-the-art performance on tasks with sequence length up to 50k, LinOSS models rely on rigid energy dissipation (“forgetting”) mechanisms that are inherently coupled to the time scale of state evolution. As forgetting is a crucial mechanism for long-range reasoning, we demonstrate the representational limitations of these models and introduce Damped Linear Oscillatory State-Space models (D-LinOSS), a more general class of oscillatory SSMs that learn to dissipate latent state energy on arbitrary time scales. We analyze the spectral distribution of the model’s recurrent matrices and prove that the SSM layers exhibit stable dynamics under a simple, flexible parameterization. Without introducing additional complexity, D-LinOSS consistently outperforms previous LinOSS methods on long-range learning tasks, achieves faster convergence, and reduces the hyperparameter search space by 50%.

Method: Iterative training procedure using direct preference optimization where speakers learn to compose explanations that maximize communicative utility for listeners, requiring only pairwise preference data.

Result: Method successfully aligns speakers with simulated listener preferences on image classification across three datasets, and pragmatic explanations improve classification accuracy in user studies.

Conclusion: Listener-adaptive explanations grounded in pragmatic reasoning effectively address the communicative context gap in explanation methods and improve human understanding.

Abstract: Concept-based explanation methods aim at making machine learning models more transparent by finding the most important semantic features of an input (e.g., colors, patterns, shapes) for a given prediction task. However, these methods generally ignore the communicative context of explanations, such as the preferences of a listener. For example, medical doctors understand explanations in terms of clinical markers, but patients may not, needing a different vocabulary to rationalize the same diagnosis. We address this gap with listener-adaptive explanations grounded in principles of pragmatic reasoning and the rational speech act. We introduce an iterative training procedure based on direct preference optimization where a speaker learns to compose explanations that maximize communicative utility for a listener. Our approach only needs access to pairwise preferences, which can be collected from human feedback, making it particularly relevant in real-world scenarios where a model of the listener may not be available. We demonstrate that our method is able to align speakers with the preferences of simulated listeners on image classification across three datasets, and further validate that pragmatic explanations generated with our method improve the classification accuracy of participants in a user study.

Method: Recycles few-shot examples to evaluate candidate outputs using two combined scores motivated by Bayes’ rule, selecting outputs through additional LLM inferences based on confidence and contextual coherence.

Result: Significantly improves LLM accuracy across seven diverse tasks with three different LLMs, achieving average gain of 4.8% without additional training.

Conclusion: ReFeri effectively enhances LLM performance through response selection by leveraging existing few-shot examples for verification, providing a cost-effective alternative to existing methods.

Abstract: Although LLMs have achieved remarkable performance, the inherent stochasticity of their reasoning process and varying conclusions present significant challenges. Majority voting or Best-of-N with external verification models has been explored to find the most promising solution among multiple LLM outputs. However, these approaches have certain limitations, such as limited applicability or the cost of an additional training step. To address this problem, we propose a novel and effective framework that Recycles Few-shot examples to verify LLM outputs (ReFeri). Our key idea is to additionally utilize the given few-shot examples to evaluate the candidate outputs of the target query, not only using them to generate outputs as the conventional few-shot prompting setup. Specifically, ReFeri evaluates the generated outputs by combining two different scores, designed motivated from Bayes’ rule, and subsequently selects the candidate that is both confidently determined and contextually coherent through a few additional LLM inferences. Experiments with three different LLMs and across seven diverse tasks demonstrate that our framework significantly improves the accuracy of LLMs-achieving an average gain of 4.8%-through effective response selection, without additional training.

Method: Generalizes gradient path kernels to realistic settings like AdamW, derives novel parameter- and step-wise influence scores from kernel feature maps, and jointly interprets model components and data over the training process.

Result: Validated that CNNs and Transformers are accurately replicated by the kernel reformulation; influence scores show comparable effectiveness for parameter pruning; Grokking analysis supports proposed stages while refining the final phase as alignment of embeddings and final layers.

Conclusion: ExPLAIND provides a theoretically grounded, unified framework to interpret model behavior and training dynamics, overcoming limitations of isolated interpretability approaches.

Abstract: Post-hoc interpretability methods typically attribute a model’s behavior to its components, data, or training trajectory in isolation. This leads to explanations that lack a unified view and may miss key interactions. While combining existing methods or applying them at different training stages offers broader insights, such approaches usually lack theoretical support. In this work, we present ExPLAIND, a unified framework that integrates all these perspectives. First, we generalize recent work on gradient path kernels, which reformulate models trained by gradient descent as a kernel machine, to realistic settings like AdamW. We empirically validate that a CNN and a Transformer are accurately replicated by this reformulation. Second, we derive novel parameter- and step-wise influence scores from the kernel feature maps. Their effectiveness for parameter pruning is comparable to existing methods, demonstrating their value for model component attribution. Finally, jointly interpreting model components and data over the training process, we leverage ExPLAIND to analyze a Transformer that exhibits Grokking. Our findings support previously proposed stages of Grokking, while refining the final phase as one of alignment of input embeddings and final layers around a representation pipeline learned after the memorization phase. Overall, ExPLAIND provides a theoretically grounded, unified framework to interpret model behavior and training dynamics.

Method: Proposes a multi-scale time series reshape module for cross-period patch interactions and variable dependencies, combined with a context-aware dynamic deformable convolution mechanism to handle uneven temporal feature distributions.

Result: MS-DFTVNet significantly outperforms strong baselines and achieves an average 7.5% improvement across six public datasets, setting new state-of-the-art results.

Conclusion: The proposed multi-scale deformable convolutional framework effectively addresses long-term time series forecasting challenges and demonstrates superior performance over existing approaches.

Abstract: Research on long-term time series prediction has primarily relied on Transformer and MLP models, while the potential of convolutional networks in this domain remains underexplored. To address this, we propose a novel multi-scale time series reshape module that effectively captures cross-period patch interactions and variable dependencies. Building on this, we develop MS-DFTVNet, the multi-scale 3D deformable convolutional framework tailored for long-term forecasting. Moreover, to handle the inherently uneven distribution of temporal features, we introduce a context-aware dynamic deformable convolution mechanism, which further enhances the model’s ability to capture complex temporal patterns. Extensive experiments demonstrate that MS-DFTVNet not only significantly outperforms strong baselines but also achieves an average improvement of about 7.5% across six public datasets, setting new state-of-the-art results.

Method: Combines soft best-of-n test-time scaling with reward model r(x,y) and speculative samples from a small auxiliary model π_S(y|x) to approximate the optimal tilted policy π_β,B(y|x) ∝ π_B(y|x)exp(β r(x,y)).

Result: Achieves higher accuracy than standard soft best-of-n with π_S and reward-guided speculative decoding, and in certain settings outperforms soft best-of-n with π_B on reasoning benchmarks (MATH500, OlympiadBench, Minerva Math, MMLU-STEM, GSM8K).

Conclusion: GSI provides an effective approach for efficient reward-guided decoding that can approximate optimal policies and achieve improved performance on reasoning tasks compared to existing methods.

Abstract: We propose Guided Speculative Inference (GSI), a novel algorithm for efficient reward-guided decoding in large language models. GSI combines soft best-of-$n$ test-time scaling with a reward model $r(x,y)$ and speculative samples from a small auxiliary model $\pi_S(y\mid x)$. We provably approximate both the optimal tilted policy $\pi_{\beta,B}(y\mid x) \propto \pi_B(y\mid x)\exp(\beta,r(x,y))$ of soft best-of-$n$ under the base model $\pi_B$, as well as the expected reward under the optimal policy. In experiments on reasoning benchmarks (MATH500, OlympiadBench, Minerva Math, MMLU-STEM, GSM8K), our method achieves higher accuracy than standard soft best-of-$n$ with $\pi_S$ and reward-guided speculative decoding (Liao et al., 2025), and in certain settings even outperforms soft best-of-$n$ with $\pi_B$. The code is available at https://github.com/j-geuter/GSI .

Method: EFRame augments GRPO with three components: additional rollouts for targeted exploration, online filtering to remove low-quality samples, and experience replay to amplify rare informative trajectories.

Result: EFRame achieves consistent gains across diverse reasoning benchmarks, including a 37.9% relative improvement on Geometry3K over GRPO, while supporting fine-grained sample categorization and precise entropy control.

Conclusion: EFRame establishes a principled training cycle that balances exploration, efficiency, and stability, serving as a robust solution for advancing deeper reasoning in LLMs.

Abstract: Recent advances in reinforcement learning (RL) have significantly enhanced the reasoning capabilities of large language models (LLMs). Group Relative Policy Optimization (GRPO), a lightweight variant of Proximal Policy Optimization (PPO), improves efficiency but suffers from limited exploration and training instability, limiting its effectiveness on complex reasoning tasks. To address these challenges, we introduce EFRame, an Exploration-Filter-Replay framework that augments GRPO across three dimensions: additional rollouts enable deeper and more targeted exploration, online filtering removes low-quality samples to stabilize gradients and accelerate training, and experience replay amplifies rare yet informative trajectories for stable convergence. This unified framework establishes a principled training cycle that balances exploration, efficiency, and stability. Experiments on diverse reasoning benchmarks demonstrate that EFRame achieves consistent gains, including a 37.9% relative improvement on Geometry3K over GRPO. EFRame further supports fine-grained sample categorization and precise entropy control, highlighting it as a robust solution for advancing deeper reasoning in LLMs. Our code is available at https://github.com/597358816/EFRame.

Method: Replace iterative optimization with random feature-based parameter construction for Hamiltonian Graph Networks, enabling fast training while maintaining permutation, rotation, and translation invariances.

Result: Achieves up to 600x faster training with comparable accuracy, robust performance in diverse N-body systems (up to 10,000 particles), and zero-shot generalization from 8-node to 4096-node systems without retraining.

Conclusion: The work challenges the dominance of iterative gradient-descent optimization for training neural network models in physical systems, demonstrating superior efficiency through random feature-based parameter construction.

Abstract: Learning dynamical systems that respect physical symmetries and constraints remains a fundamental challenge in data-driven modeling. Integrating physical laws with graph neural networks facilitates principled modeling of complex N-body dynamics and yields accurate and permutation-invariant models. However, training graph neural networks with iterative, gradient-based optimization algorithms (e.g., Adam, RMSProp, LBFGS) often leads to slow training, especially for large, complex systems. In comparison to 15 different optimizers, we demonstrate that Hamiltonian Graph Networks (HGN) can be trained up to 600x faster–but with comparable accuracy–by replacing iterative optimization with random feature-based parameter construction. We show robust performance in diverse simulations, including N-body mass-spring and molecular systems in up to 3 dimensions and 10,000 particles with different geometries, while retaining essential physical invariances with respect to permutation, rotation, and translation. Our proposed approach is benchmarked using a NeurIPS 2022 Datasets and Benchmarks Track publication to further demonstrate its versatility. We reveal that even when trained on minimal 8-node systems, the model can generalize in a zero-shot manner to systems as large as 4096 nodes without retraining. Our work challenges the dominance of iterative gradient-descent-based optimization algorithms for training neural network models for physical systems.

Method: Two-stage fine-tuning method that trains reasoning models to explicitly generate progress estimates (0-100%) during reasoning, using linear probes on internal representations.

Result: Simple linear probes achieved 30% accuracy over 10 progress classes with MAE of 1.75; fine-tuned model predictions were on average 10% from true label for sequences under 16K tokens.

Conclusion: Reasoning progress in LLMs can be quantified and explicitly estimated, enabling better progress tracking for long reasoning tasks.

Abstract: Reasoning models that produce long, hidden chains of thought, have emerged as powerful tools for reasoning-intensive and agentic tasks. However, as the time horizons at which these models can operate grow exponentially, it becomes increasingly difficult to know how much progress the model is making on a task, making it challenging for users to set appropriate expectations about completion time. By probing the internal representations of Large Language Models (LLMs), we find evidence that their reasoning progress can be quantified, with simple linear probes achieving 30% accuracy over 10 progress classes and Mean Absolute Error (MAE) of 1.75. Rooted in this insight, we propose a two-stage fine-tuning method that trains existing reasoning models to explicitly generate progress estimates (0-100%) during their reasoning process. We find that the predictions of our best fine-tuned language model for sequences below 16K tokens are on average 10% from the true label.

Method: Training small, structured subnetworks on separate workers using two strategies: stochastic block dropping and width-wise subnetwork construction, with focus on uniform parameter representation across distributed setup.

Result: Stochastic block dropping consistently outperforms width-wise construction, achieving 20-40% reduction in memory usage without performance loss, attributed to stronger gradient alignment in subnetworks retaining blocks with skip connections.

Conclusion: The approach shows promise for distributed training by significantly reducing memory requirements while maintaining comparable communication bandwidth to standard data parallel schemes.

Abstract: Distributed pre-training of large models at scale often imposes heavy memory demands on individual nodes and incurs significant intra-node communication costs. We propose a novel alternative approach that reduces the memory requirements by training small, structured subnetworks of the model on separate workers. Unlike pipelining, our method avoids inter-node activation communication and maintains bandwidth requirements that are comparable to or lower than standard data parallel communication schemes based on all-reduce. We evaluate two subnetwork construction strategies guided by the principle of ensuring uniform representation of each parameter across the distributed training setup. Our results show that the stochastic block dropping technique consistently outperforms the width-wise subnetwork construction previously explored in federated learning. We empirically attribute this superior performance to stronger gradient alignment in subnetworks that retain blocks having skip connections. Preliminary experiments highlight the promise of our approach, achieving a 20-40% reduction in memory usage without any loss in performance.

Method: Develops a fair CCA method for representation learning that enforces independence between projected features and sensitive attributes, validated on synthetic data and real-world Alzheimer’s Disease Neuroimaging Initiative (ADNI) data.

Result: The method maintains high correlation analysis performance while improving fairness in classification tasks, demonstrating effectiveness in neuroimaging studies where unbiased analysis is essential.

Conclusion: The proposed fair CCA approach enables fair machine learning in neuroimaging studies by ensuring fairness without sacrificing accuracy, with code publicly available for implementation.

Abstract: Canonical correlation analysis (CCA) is a technique for finding correlations between different data modalities and learning low-dimensional representations. As fairness becomes crucial in machine learning, fair CCA has gained attention. However, previous approaches often overlook the impact on downstream classification tasks, limiting applicability. We propose a novel fair CCA method for fair representation learning, ensuring the projected features are independent of sensitive attributes, thus enhancing fairness without compromising accuracy. We validate our method on synthetic data and real-world data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), demonstrating its ability to maintain high correlation analysis performance while improving fairness in classification tasks. Our work enables fair machine learning in neuroimaging studies where unbiased analysis is essential. Code is available in https://github.com/ZhanliangAaronWang/FR-CCA-ADNI.

Method: The attack exploits in-vehicular networks to induce delays across sensor streams (camera and LiDAR), creating temporal misalignments. It was validated using automotive Ethernet testbeds and Autoware stack simulations.

Result: Single-frame LiDAR delay reduced car detection mAP by up to 88.5%, while three-frame camera delay dropped multiple object tracking accuracy (MOTA) for cars by 73%. The attack caused collisions and phantom braking in validation scenarios.

Conclusion: Multimodal fusion systems have task-specific sensor dependencies that can be exploited through temporal misalignment attacks, highlighting critical security vulnerabilities in autonomous driving perception systems.

Abstract: Multimodal fusion (MMF) plays a critical role in the perception of autonomous driving, which primarily fuses camera and LiDAR streams for a comprehensive and efficient scene understanding. However, its strict reliance on precise temporal synchronization exposes it to new vulnerabilities. In this paper, we introduce DejaVu, an attack that exploits the in-vehicular network and induces delays across sensor streams to create subtle temporal misalignments, severely degrading downstream MMF-based perception tasks. Our comprehensive attack analysis across different models and datasets reveals the sensors’ task-specific imbalanced sensitivities: object detection is overly dependent on LiDAR inputs, while object tracking is highly reliant on the camera inputs. Consequently, with a single-frame LiDAR delay, an attacker can reduce the car detection mAP by up to 88.5%, while with a three-frame camera delay, multiple object tracking accuracy (MOTA) for car drops by 73%. We further demonstrated two attack scenarios using an automotive Ethernet testbed for hardware-in-the-loop validation and the Autoware stack for end-to-end AD simulation, demonstrating the feasibility of the DejaVu attack and its severe impact, such as collisions and phantom braking.

Method: GRID incorporates a decoding mechanism with representative inputs, automatic task identification, and constrained decoding for backward transfer, plus gradient-guided prompt selection to compress less informative prompts into aggregated representations.

Result: Extensive experiments show GRID improves average accuracy and backward transfer, achieves competitive forward transfer, and substantially reduces prompt memory usage on long-sequence and negative transfer benchmarks.

Conclusion: GRID provides a scalable and memory-efficient solution for prompt-based continual learning that addresses key limitations of existing methods while maintaining strong performance across tasks.

Abstract: Prompt-based continual learning (CL) provides a parameter-efficient approach for adapting large language models (LLMs) across task sequences. However, most existing methods rely on task-aware inference and maintain a growing set of task-specific prompts, which introduces two major challenges: (1) severe performance degradation on earlier tasks under task-agnostic inference, and (2) limited scalability due to prompt memory accumulation as task sequences grow. In this paper, we present GRID, a unified framework designed to address these challenges. GRID incorporates a decoding mechanism that enhances backward transfer by leveraging representative inputs, automatic task identification, and constrained decoding. Furthermore, it employs a gradient-guided prompt selection strategy to compress less informative prompts into a single aggregated representation, ensuring scalable and memory-efficient continual learning. Extensive experiments on long-sequence and negative transfer benchmarks show that GRID improves average accuracy and backward transfer, achieves competitive forward transfer, and substantially reduces prompt memory usage.

Method: Mathematical analysis showing GPTQ’s equivalence to Babai’s algorithm, then designing clipping-free quantization methods based on the derived error bounds.

Result: Developed improved quantization methods that outperform original GPTQ, with efficient GPU inference kernels for the resulting representation.

Conclusion: GPTQ is placed on firm theoretical footing, enabling future quantization algorithm design by importing decades of lattice algorithm progress.

Abstract: Quantizing the weights of large language models (LLMs) from 16-bit to lower bitwidth is the de facto approach to deploy massive transformers onto more affordable accelerators. While GPTQ emerged as one of the standard methods for one-shot post-training quantization at LLM scale, its inner workings are described as a sequence of ad-hoc algebraic updates that obscure geometric meaning or worst-case guarantees. In this work, we show that, when executed back-to-front (from the last to first dimension) for a linear layer, GPTQ is mathematically identical to Babai’s nearest plane algorithm for the classical closest vector problem (CVP) on a lattice defined by the Hessian matrix of the layer’s inputs. This equivalence is based on a sophisticated mathematical argument, and has two analytical consequences: first, the GPTQ error propagation step gains an intuitive geometric interpretation; second, GPTQ inherits the error upper bound of Babai’s algorithm under the assumption that no weights are clipped. Leveraging this bound, we design post-training quantization methods that avoid clipping, and outperform the original GPTQ. In addition, we provide efficient GPU inference kernels for the resulting representation. Taken together, these results place GPTQ on a firm theoretical footing and open the door to importing decades of progress in lattice algorithms towards the design of future quantization algorithms for billion-parameter models.

Method: Leveraged the RAGTruth corpus with token-level annotations and reproduced logits to analyze how hallucination signals depend on a token’s position within hallucinated spans.

Result: The first hallucinated token carries a stronger signal and is more detectable than conditional tokens within hallucinated spans.

Conclusion: The analysis provides improved understanding of token-level hallucination, with findings that can help improve detection methods, and the framework and code are publicly released for further research.

Abstract: Hallucination, the generation of untruthful content, is one of the major concerns regarding foundational models. Detecting hallucinations at the token level is vital for real-time filtering and targeted correction, yet the variation of hallucination signals within token sequences is not fully understood. Leveraging the RAGTruth corpus with token-level annotations and reproduced logits, we analyse how these signals depend on a token’s position within hallucinated spans, contributing to an improved understanding of token-level hallucination. Our results show that the first hallucinated token carries a stronger signal and is more detectable than conditional tokens. We release our analysis framework, along with code for logit reproduction and metric computation at https://github.com/jakobsnl/RAGTruth_Xtended.

Method: Analyze stage-wise dynamics of attention during training, introducing new proof techniques to characterize how general non-degenerate L-Lipschitz task functions affect attention weights, with focus on the Lipschitz constant L as a key governing factor.

Result: Attention scores between query tokens and target features grow rapidly early then converge to one, while attention to irrelevant features decays slowly with oscillations. For different L regimes (below/above threshold), different time bounds guarantee near-zero prediction error.

Conclusion: Despite convergence time depending on underlying task functions, transformers consistently attend to prompt tokens with highly relevant features at convergence, demonstrating ICL capability for unseen functions.

Abstract: The transformer architecture, which processes sequences of input tokens to produce outputs for query tokens, has revolutionized numerous areas of machine learning. A defining feature of transformers is their ability to perform previously unseen tasks using task specific prompts without updating parameters, a phenomenon known as in-context learning (ICL). Recent research has actively explored the training dynamics behind ICL, with much of the focus on relatively simple tasks such as linear regression and binary classification. To advance the theoretical understanding of ICL, this paper investigates more complex nonlinear regression tasks, aiming to uncover how transformers acquire in-context learning capabilities in these settings. We analyze the stage-wise dynamics of attention during training: attention scores between a query token and its target features grow rapidly in the early phase, then gradually converge to one, while attention to irrelevant features decays more slowly and exhibits oscillatory behavior. Our analysis introduces new proof techniques that explicitly characterize how the nature of general non-degenerate $L$-Lipschitz task functions affects attention weights. Specifically, we identify that the Lipschitz constant $L$ of nonlinear function classes as a key factor governing the convergence dynamics of transformers in ICL. Leveraging these insights, for two distinct regimes depending on whether $L$ is below or above a threshold, we derive different time bounds to guarantee near-zero prediction error. Notably, despite the convergence time depending on the underlying task functions, we prove that query tokens consistently attend to prompt tokens with highly relevant features at convergence, demonstrating the ICL capability of transformers for unseen functions.

Method: Models Q-function as leader (updates less frequently) and perception network as follower (updates more frequently) in a hierarchical game, solved via bi-level optimization approximated by two-timescale algorithm.

Result: Extensive experiments on DQN variants show improved performance from algorithmic insight rather than model complexity, demonstrating stable co-adaptation and reduced bias.

Conclusion: SCORER’s asymmetric update strategy between representation and value learning provides a principled solution to instability in deep Q-learning through hierarchical game-theoretic formulation.

Abstract: Deep Q-learning jointly learns representations and values within monolithic networks, promising beneficial co-adaptation between features and value estimates. Although this architecture has attained substantial success, the coupling between representation and value learning creates instability as representations must constantly adapt to non-stationary value targets, while value estimates depend on these shifting representations. This is compounded by high variance in bootstrapped targets, which causes bias in value estimation in off-policy methods. We introduce Stackelberg Coupled Representation and Reinforcement Learning (SCORER), a framework for value-based RL that views representation and Q-learning as two strategic agents in a hierarchical game. SCORER models the Q-function as the leader, which commits to its strategy by updating less frequently, while the perception network (encoder) acts as the follower, adapting more frequently to learn representations that minimize Bellman error variance given the leader’s committed strategy. Through this division of labor, the Q-function minimizes MSBE while perception minimizes its variance, thereby reducing bias accordingly, with asymmetric updates allowing stable co-adaptation, unlike simultaneous parameter updates in monolithic solutions. Our proposed SCORER framework leads to a bi-level optimization problem whose solution is approximated by a two-timescale algorithm that creates an asymmetric learning dynamic between the two players. Extensive experiments on DQN and its variants demonstrate that gains stem from algorithmic insight rather than model complexity.

Method: Proposed conformal criteria to quantify exclusion of forgotten samples from prediction sets, developed empirical metrics for unlearning effectiveness, and created a practical method optimizing these conformal metrics.

Result: Extensive experiments across various forgetting scenarios, datasets and models show the approach effectively removes targeted data with statistical guarantees.

Conclusion: The conformal prediction-based approach provides statistically sound, uncertainty-aware guarantees for machine unlearning without relying on naive retraining concepts.

Abstract: The increasing demand for data privacy has made machine unlearning (MU) essential for removing the influence of specific training samples from machine learning models while preserving performance on retained data. However, most existing MU methods lack rigorous statistical guarantees or rely on heuristic metrics such as accuracy. To overcome these limitations, we introduce a new definition for MU based on conformal prediction (CP), providing statistically sound, uncertainty-aware guarantees without the need for the concept of naive retraining. We formalize the proposed conformal criteria that quantify how often forgotten samples are excluded from CP sets, and propose empirical metrics to measure the effectiveness of unlearning. We further present a practical unlearning method designed to optimize these conformal metrics. Extensive experiments across diverse forgetting scenarios, datasets and models demonstrate the efficacy of our approach in removing targeted data.

Method: Generalizes Neural Logic Networks with NOT operations and biases to account for unobserved data. Proposes a rigorous logical and probabilistic modeling framework using concept combinations, along with a novel factorized IF-THEN rule structure and modified learning algorithm.

Result: The method improves state-of-the-art in Boolean networks discovery and can learn relevant, interpretable rules in tabular classification, particularly in medical and industrial applications where interpretability is crucial.

Conclusion: The proposed Neural Logic Networks with enhanced logical operations provide both interpretable structure and improved performance, making them valuable for domains requiring transparent decision-making processes.

Abstract: Traditional neural networks have an impressive classification performance, but what they learn cannot be inspected, verified or extracted. Neural Logic Networks on the other hand have an interpretable structure that enables them to learn a logical mechanism relating the inputs and outputs with AND and OR operations. We generalize these networks with NOT operations and biases that take into account unobserved data and develop a rigorous logical and probabilistic modeling in terms of concept combinations to motivate their use. We also propose a novel factorized IF-THEN rule structure for the model as well as a modified learning algorithm. Our method improves the state-of-the-art in Boolean networks discovery and is able to learn relevant, interpretable rules in tabular classification, notably on examples from the medical and industrial fields where interpretability has tangible value.

Method: OFQL reformulates DQL policy within Flow Matching paradigm but learns an average velocity field that directly supports accurate one-step action generation, eliminating multi-step denoising and backpropagation-through-time updates.

Result: OFQL significantly reduces computation during training and inference while outperforming multi-step DQL by a large margin. It achieves state-of-the-art performance on D4RL benchmark.

Conclusion: OFQL demonstrates that effective one-step action generation is achievable without auxiliary modules or distillation, providing faster, more robust learning while maintaining superior performance compared to multi-step approaches.

Abstract: Diffusion Q-Learning (DQL) has established diffusion policies as a high-performing paradigm for offline reinforcement learning, but its reliance on multi-step denoising for action generation renders both training and inference slow and fragile. Existing efforts to accelerate DQL toward one-step denoising typically rely on auxiliary modules or policy distillation, sacrificing either simplicity or performance. It remains unclear whether a one-step policy can be trained directly without such trade-offs. To this end, we introduce One-Step Flow Q-Learning (OFQL), a novel framework that enables effective one-step action generation during both training and inference, without auxiliary modules or distillation. OFQL reformulates the DQL policy within the Flow Matching (FM) paradigm but departs from conventional FM by learning an average velocity field that directly supports accurate one-step action generation. This design removes the need for multi-step denoising and backpropagation-through-time updates, resulting in substantially faster and more robust learning. Extensive experiments on the D4RL benchmark show that OFQL, despite generating actions in a single step, not only significantly reduces computation during both training and inference but also outperforms multi-step DQL by a large margin. Furthermore, OFQL surpasses all other baselines, achieving state-of-the-art performance in D4RL.

Method: Proposed MissionHD framework that encodes graphs with constrained graph-neural operations, aligns them with downstream task loss, and decodes refined structures using hyperdimensional computing.

Result: Experiments on VAD/VAR benchmarks show that MissionHD-refined graphs consistently improve performance in video anomaly tasks.

Conclusion: HDC-GSR establishes an effective pre-processing step for structured reasoning in video anomaly detection and recognition tasks.

Abstract: LLM-generated reasoning graphs, referred to as mission-specific graphs (MSGs), are increasingly used for video anomaly detection (VAD) and recognition (VAR). These MSGs are novel artifacts: they often exhibit skewed connectivity and lack large-scale datasets for pre-training, which makes existing graph structure refinement (GSR) methods ineffective. To address this challenge, we propose HDC-constrained Graph Structure Refinement (HDC-GSR), a paradigm that leverages hyperdimensional computing (HDC) to optimize decodable graph representations without relying on structural-distribution learning. Building on this paradigm, we introduce MissionHD, an HDC framework that encodes graphs with constrained graph-neural operations, aligns them directly with downstream task loss, and decodes refined structures. Experiments on VAD/VAR benchmarks demonstrate that MissionHD-refined graphs consistently improve performance, establishing HDC-GSR as an effective pre-processing step for structured reasoning in video anomaly tasks.

Method: Established connection between task vectors and gradients, proved equivalence under gradient descent, bounded error terms for multi-epoch finetuning, and validated with empirical analysis across seven vision benchmarks.

Result: Task vectors from one epoch of finetuning are exactly equivalent to negative gradients scaled by learning rate; multi-epoch vectors approximate this with bounded error. First-epoch gradient dominates finetuning trajectory.

Conclusion: Task arithmetic is a form of approximate multitask learning, with early training dynamics being critical for effective model merging. Single-epoch finetuning often suffices for comparable performance to fully converged models.

Abstract: Task arithmetic has emerged as a simple yet powerful technique for model merging, enabling the combination of multiple finetuned models into one. Despite its empirical success, a clear theoretical explanation of why and when it works is lacking. This paper provides a rigorous theoretical foundation for task arithmetic by establishing a connection between task vectors and gradients of the task losses. We show that under standard gradient descent, a task vector generated from one epoch of finetuning is exactly equivalent to the negative gradient of the loss, scaled by the learning rate. For the practical multi-epoch setting, we prove that this equivalence holds approximately, with a second-order error term that we explicitly bound for feed-forward networks. Our empirical analysis across seven vision benchmarks corroborates our theory, demonstrating that the first-epoch gradient dominates the finetuning trajectory in both norm and direction. A key implication is that merging models finetuned for only a single epoch often yields performance comparable to merging fully converged models. These findings reframe task arithmetic as a form of approximate multitask learning, providing a clear rationale for its effectiveness and highlighting the critical role of early training dynamics in model merging.

Method: The authors analyze minimax rates for mean estimation in multivariate bounded distributions, univariate Gaussian distributions, multivariate Gaussian mean estimation, and linear regression under heterogeneous corruption patterns.

Result: For mean estimation in multivariate bounded and univariate Gaussian distributions, tight minimax rates are established for all heterogeneous corruption patterns. For multivariate Gaussian mean estimation and linear regression, the minimax rate for squared error is established up to a factor of √d.

Conclusion: The optimal estimators may discard samples beyond a certain corruption threshold, which is determined by the empirical distribution of the given corruption rates.

Abstract: We study the problem of robust estimation under heterogeneous corruption rates, where each sample may be independently corrupted with a known but non-identical probability. This setting arises naturally in distributed and federated learning, crowdsourcing, and sensor networks, yet existing robust estimators typically assume uniform or worst-case corruption, ignoring structural heterogeneity. For mean estimation for multivariate bounded distributions and univariate gaussian distributions, we give tight minimax rates for all heterogeneous corruption patterns. For multivariate gaussian mean estimation and linear regression, we establish the minimax rate for squared error up to a factor of $\sqrt{d}$, where $d$ is the dimension. Roughly, our findings suggest that samples beyond a certain corruption threshold may be discarded by the optimal estimators – this threshold is determined by the empirical distribution of the corruption rates given.

Method: RankRefine uses inverse variance weighting to combine base regressor outputs with rank-based estimates from pairwise comparisons in a reference set. It’s model-agnostic and plug-and-play, requiring no retraining.

Result: In molecular property prediction, RankRefine achieves up to 10% relative reduction in mean absolute error using only 20 pairwise comparisons from general-purpose LLMs without finetuning.

Conclusion: RankRefine provides a practical and broadly applicable solution for improving regression accuracy in low-data settings using rankings from human experts or general-purpose LLMs.

Abstract: Accurate prediction of continuous properties is essential to many scientific and engineering tasks. Although deep-learning regressors excel with abundant labels, their accuracy deteriorates in data-scarce regimes. We introduce RankRefine, a model-agnostic, plug-and-play post hoc method that refines regression with expert knowledge coming from pairwise rankings. Given a query item and a small reference set with known properties, RankRefine combines the base regressor’s output with a rank-based estimate via inverse variance weighting, requiring no retraining. In molecular property prediction task, RankRefine achieves up to 10% relative reduction in mean absolute error using only 20 pairwise comparisons obtained through a general-purpose large language model (LLM) with no finetuning. As rankings provided by human experts or general-purpose LLMs are sufficient for improving regression across diverse domains, RankRefine offers practicality and broad applicability, especially in low-data settings.

Method: Chain of Merges (CoM) - a sequential layer-wise merging procedure that merges weights across layers while updating activation statistics to account for inter-layer interactions.

Result: Experiments on standard benchmarks show CoM achieves state-of-the-art performance in model merging.

Conclusion: CoM effectively mitigates covariate shift in model merging by explicitly handling inter-layer dependencies through sequential weight merging and activation statistics updates.

Abstract: Fine-tuning pretrained models has become a standard pathway to achieve state-of-the-art performance across a wide range of domains, leading to a proliferation of task-specific model variants. As the number of such specialized models increases, merging them into a unified model without retraining has become a critical challenge. Existing merging techniques operate at the level of individual layers, thereby overlooking the inter-layer dependencies inherent in deep networks. We show that this simplification leads to distributional mismatches, particularly in methods that rely on intermediate activations, as changes in early layers are not properly propagated to downstream layers during merging. We identify these mismatches as a form of internal covariate shift, comparable to the phenomenon encountered in the initial phases of neural networks training. To address this, we propose Chain of Merges (CoM), a layer-wise merging procedure that sequentially merges weights across layers while sequentially updating activation statistics. By explicitly accounting for inter-layer interactions, CoM mitigates covariate shift and produces a coherent merged model through a series of conditionally optimal updates. Experiments on standard benchmarks demonstrate that CoM achieves state-of-the-art performance.

Method: Uses adaptive Top-k logit selection to dynamically sparsify logits based on communication conditions, adaptive logits aggregation to handle dimensional inconsistency, and incorporates LoRA-adapted hidden-layer projection into distillation loss.

Result: Achieves superior performance compared to baseline methods while reducing communication overhead by approximately 50%.

Conclusion: The proposed scheme effectively addresses communication bottlenecks in federated LLM distillation through adaptive logit processing and enhanced distillation techniques.

Abstract: Federated learning (FL) for large language models (LLMs) offers a privacy-preserving scheme, enabling clients to collaboratively fine-tune locally deployed LLMs or smaller language models (SLMs) without exchanging raw data. While parameter-sharing methods in traditional FL models solves number of technical challenges, they still incur high communication overhead and struggle with adapting to heterogeneous model architectures. Federated distillation, a framework for mutual knowledge transfer via shared logits, typically offers lower communication overhead than parameter-sharing methods. However, transmitting logits from LLMs remains challenging for bandwidth-limited clients due to their high dimensionality. In this work, we focus on a federated LLM distillation with efficient communication overhead. To achieve this, we first propose an adaptive Top-k logit selection mechanism, dynamically sparsifying logits according to real-time communication conditions. Then to tackle the dimensional inconsistency introduced by the adaptive sparsification, we design an adaptive logits aggregation scheme, effectively alleviating the artificial and uninformative inputs introduced by conventional zero-padding methods. Finally, to enhance the distillation effect, we incorporate LoRA-adapted hidden-layer projection from LLM into the distillation loss, reducing the communication overhead further while providing richer representation. Experimental results demonstrate that our scheme achieves superior performance compared to baseline methods while effectively reducing communication overhead by approximately 50%.

Method: Use steering vectors inserted into residual stream trained with RL objective, analyzed via logit-lens readouts and path-patching on two models. Study SAE features and vector transferability across models.

Result: Steering vectors transfer to other models, combine across layers when trained in isolation, and concentrate magnitude on meaningful prompt segments. Different layers operate through distinct mechanisms: token bias, MLP/unembedding pathways, and language filtering.

Conclusion: Results deepen understanding of how trained steering vectors shape computation and inform future work in activation engineering and reasoning model studies.

Abstract: The mechanisms by which reasoning training reshapes LLMs’ internal computations remain unclear. We study lightweight steering vectors inserted into the base model’s residual stream and trained with a reinforcement-learning objective. These vectors match full fine-tuning performance while preserving the interpretability of small, additive interventions. Using logit-lens readouts and path-patching analyses on two models, we find that (i) the last-layer steering vector acts like a token-substitution bias concentrated on the first generated token, consistently boosting tokens such as “To” and “Step”; (ii) the penultimate-layer vector leaves attention patterns largely intact and instead operates through the MLP and unembedding, preferentially up-weighting process words and structure symbols; and (iii) middle layers de-emphasize non-English tokens. Next, we show that a SAE isolates features associated with correct generations. We also show that steering vectors (i) transfer to other models, (ii) combine across layers when trained in isolation, and (iii) concentrate magnitude on meaningful prompt segments under adaptive token-wise scaling. Taken together, these results deepen understanding of how trained steering vectors shape computation and should inform future work in activation engineering and the study of reasoning models.

Method: Progressive layer replacement from student to teacher model, with training that aligns intermediate feature representations and improves student output performance.

Result: Models maintain competitive distillation performance and gradually improve accuracy to match full teacher model accuracy, while preserving fast initial inference speed.

Conclusion: PWL is well-suited for dynamic, resource-constrained deployments requiring both responsiveness and high performance.

Abstract: Deep learning models have become increasingly large and complex, resulting in higher memory consumption and computational demands. Consequently, model loading times and initial inference latency have increased, posing significant challenges in mobile and latency-sensitive environments where frequent model loading and unloading are required, which directly impacts user experience. While Knowledge Distillation (KD) offers a solution by compressing large teacher models into smaller student ones, it often comes at the cost of reduced performance. To address this trade-off, we propose Progressive Weight Loading (PWL), a novel technique that enables fast initial inference by first deploying a lightweight student model, then incrementally replacing its layers with those of a pre-trained teacher model. To support seamless layer substitution, we introduce a training method that not only aligns intermediate feature representations between student and teacher layers, but also improves the overall output performance of the student model. Our experiments on VGG, ResNet, and ViT architectures demonstrate that models trained with PWL maintain competitive distillation performance and gradually improve accuracy as teacher layers are loaded-matching the final accuracy of the full teacher model without compromising initial inference speed. This makes PWL particularly suited for dynamic, resource-constrained deployments where both responsiveness and performance are critical.

Method: The authors first develop efficient algorithms for replicable random design regression and uncentered covariance estimation. They then leverage these tools to create replicable RL algorithms for linear Markov decision processes in both generative model and episodic settings.

Result: The paper provides the first provably efficient replicable RL algorithms for linear function approximation. Experimental evaluation shows these algorithms can inspire more consistent neural policies.

Conclusion: This work makes significant progress in extending replicability guarantees from tabular RL to practical function approximation settings, particularly linear Markov decision processes.

Abstract: Replication of experimental results has been a challenge faced by many scientific disciplines, including the field of machine learning. Recent work on the theory of machine learning has formalized replicability as the demand that an algorithm produce identical outcomes when executed twice on different samples from the same distribution. Provably replicable algorithms are especially interesting for reinforcement learning (RL), where algorithms are known to be unstable in practice. While replicable algorithms exist for tabular RL settings, extending these guarantees to more practical function approximation settings has remained an open problem. In this work, we make progress by developing replicable methods for linear function approximation in RL. We first introduce two efficient algorithms for replicable random design regression and uncentered covariance estimation, each of independent interest. We then leverage these tools to provide the first provably efficient replicable RL algorithms for linear Markov decision processes in both the generative model and episodic settings. Finally, we evaluate our algorithms experimentally and show how they can inspire more consistent neural policies.

Method: The authors propose a policy gradient optimization method that leverages dual representation of coherent risk measures and extends the envelope theorem to continuous cases, with extensions to episodic settings.

Result: The algorithm achieves a convergence rate of O(T^{-1/2} + r^{-1/2}) where T is policy gradient iterations and r is gradient estimator sample size. In episodic settings, it achieves O(ε) error bounds with established iteration complexity.

Conclusion: The proposed policy gradient approach successfully solves MDPs with coherent risk measures where traditional dynamic programming fails, with proven convergence guarantees in both stationary and episodic settings.

Abstract: Motivated by many application problems, we consider Markov decision processes (MDPs) with a general loss function and unknown parameters. To mitigate the epistemic uncertainty associated with unknown parameters, we take a Bayesian approach to estimate the parameters from data and impose a coherent risk functional (with respect to the Bayesian posterior distribution) on the loss. Since this formulation usually does not satisfy the interchangeability principle, it does not admit Bellman equations and cannot be solved by approaches based on dynamic programming. Therefore, We propose a policy gradient optimization method, leveraging the dual representation of coherent risk measures and extending the envelope theorem to continuous cases. We then show the stationary analysis of the algorithm with a convergence rate of $\mathcal{O}(T^{-1/2}+r^{-1/2})$, where $T$ is the number of policy gradient iterations and $r$ is the sample size of the gradient estimator. We further extend our algorithm to an episodic setting, and establish the global convergence of the extended algorithm and provide bounds on the number of iterations needed to achieve an error bound $\mathcal{O}(\epsilon)$ in each episode.

Method: The authors propose an entropy-based data selection strategy that monitors and selects synthetic training data based on entropy levels to prevent the transition from generalization to memorization.

Result: Empirical results show the approach significantly enhances visual quality and diversity in recursive generation, effectively preventing model collapse.

Conclusion: The transition from generalization to memorization driven by declining entropy in synthetic data is a key mechanism of model collapse, and entropy-based data selection effectively mitigates this issue.

Abstract: The widespread use of diffusion models has led to an abundance of AI-generated data, raising concerns about model collapse – a phenomenon in which recursive iterations of training on synthetic data lead to performance degradation. Prior work primarily characterizes this collapse via variance shrinkage or distribution shift, but these perspectives miss practical manifestations of model collapse. This paper identifies a transition from generalization to memorization during model collapse in diffusion models, where models increasingly replicate training data instead of generating novel content during iterative training on synthetic samples. This transition is directly driven by the declining entropy of the synthetic training data produced in each training cycle, which serves as a clear indicator of model degradation. Motivated by this insight, we propose an entropy-based data selection strategy to mitigate the transition from generalization to memorization and alleviate model collapse. Empirical results show that our approach significantly enhances visual quality and diversity in recursive generation, effectively preventing collapse.

Method: Builds on particle filters using pre-trained diffusion models for dynamical system emulation, requiring no further training.

Result: Proposed methodology successfully applied to GenCast, a diffusion-based global ensemble weather forecast model.

Conclusion: The approach provides a general and efficient framework for data assimilation using pre-trained diffusion models.

Abstract: Data assimilation is widely used in many disciplines such as meteorology, oceanography, and robotics to estimate the state of a dynamical system from noisy observations. In this work, we propose a lightweight and general method to perform data assimilation using diffusion models pre-trained for emulating dynamical systems. Our method builds on particle filters, a class of data assimilation algorithms, and does not require any further training. As a guiding example throughout this work, we illustrate our methodology on GenCast, a diffusion-based model that generates global ensemble weather forecasts.

Method: DBVI initiates reverse diffusion from a learnable, data-dependent initial distribution parameterized via an amortized neural network operating on inducing inputs, using Girsanov-based ELBOs and reverse-time SDEs with Doob-bridged diffusion.

Result: DBVI consistently outperforms DDVI and other variational baselines across regression, classification, and image reconstruction tasks in predictive accuracy, convergence speed, and posterior quality.

Conclusion: DBVI provides a principled extension to DDVI that bridges the posterior gap through learnable initialization, enabling more efficient and scalable inference for deep Gaussian processes.

Abstract: Deep Gaussian processes (DGPs) enable expressive hierarchical Bayesian modeling but pose substantial challenges for posterior inference, especially over inducing variables. Denoising diffusion variational inference (DDVI) addresses this by modeling the posterior as a time-reversed diffusion from a simple Gaussian prior. However, DDVI’s fixed unconditional starting distribution remains far from the complex true posterior, resulting in inefficient inference trajectories and slow convergence. In this work, we propose Diffusion Bridge Variational Inference (DBVI), a principled extension of DDVI that initiates the reverse diffusion from a learnable, data-dependent initial distribution. This initialization is parameterized via an amortized neural network and progressively adapted using gradients from the ELBO objective, reducing the posterior gap and improving sample efficiency. To enable scalable amortization, we design the network to operate on the inducing inputs, which serve as structured, low-dimensional summaries of the dataset and naturally align with the inducing variables’ shape. DBVI retains the mathematical elegance of DDVI, including Girsanov-based ELBOs and reverse-time SDEs,while reinterpreting the prior via a Doob-bridged diffusion process. We derive a tractable training objective under this formulation and implement DBVI for scalable inference in large-scale DGPs. Across regression, classification, and image reconstruction tasks, DBVI consistently outperforms DDVI and other variational baselines in predictive accuracy, convergence speed, and posterior quality.

Method: Train discrete diffusion models as policies using policy mirror descent to define regularized target distributions, framing policy updates as distributional matching problems for stable learning.

Result: Achieves state-of-the-art results and superior sample efficiency across diverse combinatorial benchmarks including DNA sequence generation, RL with macro-actions, and multi-agent systems.

Conclusion: Diffusion policies trained with this decoupled approach provide stable and effective solutions for complex RL problems with large combinatorial action spaces, outperforming other baselines.

Abstract: Reinforcement learning (RL) struggles to scale to large, combinatorial action spaces common in many real-world problems. This paper introduces a novel framework for training discrete diffusion models as highly effective policies in these complex settings. Our key innovation is an efficient online training process that ensures stable and effective policy improvement. By leveraging policy mirror descent (PMD) to define an ideal, regularized target policy distribution, we frame the policy update as a distributional matching problem, training the expressive diffusion model to replicate this stable target. This decoupled approach stabilizes learning and significantly enhances training performance. Our method achieves state-of-the-art results and superior sample efficiency across a diverse set of challenging combinatorial benchmarks, including DNA sequence generation, RL with macro-actions, and multi-agent systems. Experiments demonstrate that our diffusion policies attain superior performance compared to other baselines.

Method: Relax interpolation to regression, use neural networks to approximate high-dimensional target functions, and introduce redundancy by predicting larger subspaces than needed.

Result: Theoretical analysis shows reduced complexity and smoother mappings, with empirical results demonstrating significant accuracy improvements when predicting larger subspaces.

Conclusion: Subspace regression with neural networks and redundancy is effective for various parametric problems including eigenproblems, PDEs, and optimal control.

Abstract: It is often possible to perform reduced order modelling by specifying linear subspace which accurately captures the dynamics of the system. This approach becomes especially appealing when linear subspace explicitly depends on parameters of the problem. A practical way to apply such a scheme is to compute subspaces for a selected set of parameters in the computationally demanding offline stage and in the online stage approximate subspace for unknown parameters by interpolation. For realistic problems the space of parameters is high dimensional, which renders classical interpolation strategies infeasible or unreliable. We propose to relax the interpolation problem to regression, introduce several loss functions suitable for subspace data, and use a neural network as an approximation to high-dimensional target function. To further simplify a learning problem we introduce redundancy: in place of predicting subspace of a given dimension we predict larger subspace. We show theoretically that this strategy decreases the complexity of the mapping for elliptic eigenproblems with constant coefficients and makes the mapping smoother for general smooth function on the Grassmann manifold. Empirical results also show that accuracy significantly improves when larger-than-needed subspaces are predicted. With the set of numerical illustrations we demonstrate that subspace regression can be useful for a range of tasks including parametric eigenproblems, deflation techniques, relaxation methods, optimal control and solution of parametric partial differential equations.

Method: Proposes Boosted Direct Output (BDO) strategy combining AR and DO advantages, with smooth parameter tracking for learning stabilization. Based on a Multiple Neural Forecasting Theorem.

Result: A simple MLP with these principled improvements achieves state-of-the-art performance, outperforming recent complex models in nearly all cases without domain-specific considerations.

Conclusion: The work establishes a dynamic performance bound, verifies the proposed theorem, and identifies promising directions for future research in LTSF.

Abstract: Neural Forecasters (NFs) are a cornerstone of Long-term Time Series Forecasting (LTSF). However, progress has been hampered by an overemphasis on architectural complexity at the expense of fundamental forecasting principles. In this work, we return to first principles to redesign the LTSF paradigm. We begin by introducing a Multiple Neural Forecasting Theorem that provides a theoretical basis for our approach. We propose Boosted Direct Output (BDO), a novel forecasting strategy that synergistically combines the advantages of both Auto-Regressive (AR) and Direct Output (DO). In addition, we stabilize the learning process by smoothly tracking the model’s parameters. Extensive experiments show that these principled improvements enable a simple MLP to achieve state-of-the-art performance, outperforming recent, complex models in nearly all cases, without any specific considerations in the area. Finally, we empirically verify our theorem, establishing a dynamic performance bound and identifying promising directions for future research. The code for review is available at: .

Method: Uses Gaussian Process framework to model Q(V) as stochastic process, enabling analytical computation of dQ/dV derivatives with calibrated uncertainty and noise-aware inference.

Result: Successfully detects plating peaks under low-temperature, high-rate charging conditions while correctly identifying no peaks in baseline cases. Validated across various C-rates (0.2C-1C) and temperatures (0-40°C).

Conclusion: Provides accurate and robust lithium plating detection with uncertainty quantification, establishing practical pathway for real-time detection in battery management systems.

Abstract: Lithium plating during fast charging is a critical degradation mechanism that accelerates capacity fade and can trigger catastrophic safety failures. Recent work has identified a distinctive dQ/dV peak above 4.0 V as a reliable signature of plating onset; however, conventional methods for computing dQ/dV rely on finite differencing with filtering, which amplifies sensor noise and introduces bias in peak location. In this paper, we propose a Gaussian Process (GP) framework for lithium plating detection by directly modeling the charge-voltage relationship Q(V) as a stochastic process with calibrated uncertainty. Leveraging the property that derivatives of GPs remain GPs, we infer dQ/dV analytically and probabilistically from the posterior, enabling robust detection without ad hoc smoothing. The framework provides three key benefits: (i) noise-aware inference with hyperparameters learned from data, (ii) closed-form derivatives with credible intervals for uncertainty quantification, and (iii) scalability to online variants suitable for embedded BMS. Experimental validation on Li-ion coin cells across a range of C-rates (0.2C-1C) and temperatures (0-40\deg C) demonstrates that the GP-based method reliably detects plating peaks under low-temperature, high-rate charging, while correctly reporting no peaks in baseline cases. The concurrence of GP-identified differential peaks, reduced charge throughput, and capacity fade measured via reference performance tests confirms the method’s accuracy and robustness, establishing a practical pathway for real-time lithium plating detection.

Method: Statistical analysis of confidence dynamics identifies a volatility band region during decoding. AdaBlock-dLLM is a training-free scheduler that adaptively adjusts block size during runtime to align block boundaries with semantic steps.

Result: Achieves up to 5.3% accuracy improvement under the same throughput budget across diverse benchmarks.

Conclusion: The semantics-aware adaptive scheduling approach and confidence-based analysis can inspire future training strategies for diffusion-based LLMs beyond inference-time optimization.

Abstract: Diffusion-based large language models (dLLMs) are gaining attention for their inherent capacity for parallel decoding, offering a compelling alternative to autoregressive LLMs. Among various decoding strategies, blockwise semi-autoregressive (semi-AR) approaches are widely adopted due to their natural support for KV caching and their favorable accuracy-speed trade-off. However, this paper identifies two fundamental limitations in the conventional semi-AR decoding approach that applies a fixed block size: i) late decoding overhead, where the unmasking of high-confidence tokens outside the current block is unnecessarily delayed, and ii) premature decoding error, where low-confidence tokens inside the current block are committed too early, leading to incorrect tokens. This paper presents the first systematic investigation challenging the fixed block size assumption in semi-AR decoding. Through a statistical analysis of confidence dynamics during the denoising process, we identify a volatility band (VB) region during dLLM decoding, which encodes local semantic structure and can be used to guide adaptive block sizing. Leveraging these insights, we introduce AdaBlock-dLLM, a training-free, plug-and-play scheduler that adaptively aligns block boundaries with semantic steps by adjusting block size during runtime. Extensive experiments across diverse benchmarks show that AdaBlock-dLLM achieves up to 5.3% accuracy improvement under the same throughput budget. Beyond inference-time optimization, we hope our semantics-aware adaptive scheduling approach and confidence-based analysis will inspire future training strategies for dLLMs.

Method: Uses a multi-agent system with head agent for high-level planning and worker agents implementing ReActEval (plan, reason, act, evaluate cycle) for low-level drone actions. Operates entirely in natural language.

Result: Evaluated in simulated environment with two worker agents, showing performance across varying task complexity levels and workflow efficiency. Drones can handle tasks from simple navigation to complex industrial inspections.

Conclusion: The framework offers a novel, flexible, and user-accessible alternative to traditional drone solutions, enabling autonomous problem-solving for industrial inspection without extensive user intervention through natural language processing.

Abstract: Autonomous inspection systems are essential for ensuring the performance and longevity of industrial assets. Recently, agentic frameworks have demonstrated significant potential for automating inspection workflows but have been limited to digital tasks. Their application to physical assets in real-world environments, however, remains underexplored. In this work, our contributions are two-fold: first, we propose a hierarchical agentic framework for autonomous drone control, and second, a reasoning methodology for individual function executions which we refer to as ReActEval. Our framework focuses on visual inspection tasks in indoor industrial settings, such as interpreting industrial readouts or inspecting equipment. It employs a multi-agent system comprising a head agent and multiple worker agents, each controlling a single drone. The head agent performs high-level planning and evaluates outcomes, while worker agents implement ReActEval to reason over and execute low-level actions. Operating entirely in natural language, ReActEval follows a plan, reason, act, evaluate cycle, enabling drones to handle tasks ranging from simple navigation (e.g., flying forward 10 meters and land) to complex high-level tasks (e.g., locating and reading a pressure gauge). The evaluation phase serves as a feedback and/or replanning stage, ensuring actions align with user objectives while preventing undesirable outcomes. We evaluate the framework in a simulated environment with two worker agents, assessing performance qualitatively and quantitatively based on task completion across varying complexity levels and workflow efficiency. By leveraging natural language processing for agent communication, our approach offers a novel, flexible, and user-accessible alternative to traditional drone-based solutions, enabling autonomous problem-solving for industrial inspection without extensive user intervention.

Method: Formalizes agent states using prompt templates, reasoning context vectors, and capability matrices. Uses a distributed architecture for dynamic task routing with proven convergence when step sizes satisfy α < 1/(2L) where L is the Lipschitz constant of state transitions.

Result: 42% reduction in reasoning latency, 23% improvement in logical consistency (ROUGE-L), 89% success rate for task completion without context loss. Performance degrades beyond 10 agent transitions and requires 76.5GB memory for 1,000 concurrent agents.

Conclusion: Establishes a new paradigm for scalable reasoning in multi-agent systems with theoretical foundations for understanding reasoning emergence across coordinated language models.

Abstract: The emergence of large language models has enabled sophisticated multi-agent systems, yet coordinating their reasoning capabilities through prompt engineering remains challenging. We present a theoretically-grounded framework for dynamic prompt orchestration that enhances reasoning across multiple specialized agents. This framework addresses three core challenges: logical consistency preservation during agent transitions, reasoning-aware prompt adaptation, and scalable coordination of distributed inference. Our approach formalizes agent states using prompt templates, reasoning context vectors, and capability matrices. We prove system convergence to stable coordination patterns when step sizes satisfy $\alpha < \frac{1}{2L}$ where $L$ is the Lipschitz constant of the state transition function. We implement this through a distributed architecture that dynamically routes reasoning tasks while maintaining semantic coherence. Experimental results on 1,000 synthetic multi-agent conversations demonstrate a 42% reduction in reasoning latency, a 23% improvement in logical consistency measured by ROUGE-L score, and an 89% success rate for task completion without context loss across agent transitions. Ablation studies identify the consensus mechanism as the primary performance driver, while revealing limitations: performance degrades beyond 10 agent transitions, and the system requires 76.5GB memory for 1,000 concurrent agents. These findings establish a new paradigm for scalable reasoning in multi-agent systems, providing theoretical foundations for understanding reasoning emergence across coordinated language models.

Method: Uses Conflict-Based Search (CBS) as a central planning protocol that requires only one specific API: finding collision-free paths satisfying space-time constraints. This allows integration of diverse single-agent planners including Heuristic Search, Sampling-Based Search, Optimization, Diffusion, and Reinforcement Learning.

Result: The CBS protocol successfully enables efficient collision-free movements between algorithmically heterogeneous agents, demonstrating multi-agent motion planning capability across a variety of single-agent planning approaches.

Conclusion: Conflict-Based Search provides an effective protocol for multi-agent motion planning in heterogeneous teams, allowing different robots with various motion planning algorithms to coordinate safely in shared environments through a standardized API requirement.

Abstract: Imagine the future construction site, hospital, office, or even sophisticated household with dozens of robots bought from different manufacturers. How can we enable these different systems to effectively move in a shared environment, given that each robot may have its own independent motion planning system? This work shows how we can get efficient collision-free movements between algorithmically heterogeneous agents by using Conflict-Based Search (Sharon et al. 2015) as a protocol. At its core, the CBS Protocol requires one specific single-agent motion planning API; finding a collision-free path that satisfies certain space-time constraints. Given such an API, CBS uses a central planner to find collision-free paths - independent of how the API is implemented. We show how this protocol enables multi-agent motion planning for a heterogeneous team of agents completing independent tasks with a variety of single-agent planners including: Heuristic Search (e.g., A*), Sampling Based Search (e.g., RRT), Optimization (e.g., Direct Collocation), Diffusion, and Reinforcement Learning.

Method: Agents independently respond to queries and assess peer contributions using Shapley value approximations. A directed acyclic graph (DAG) is dynamically constructed to regulate response propagation from high-contributing agents to others, updated based on previous rounds.

Result: SelfOrg demonstrates robust performance with both strong and weak LLM backends, showing significant gains in the weak regime where prior methods fail. Experiments confirm improved collaboration efficiency.

Conclusion: The framework enables effective self-organization of multi-agent systems without additional supervision, theoretically ensuring correct responses dominate information flow and increasing overall system correctness.

Abstract: Multi-agent systems (MAS) based on Large Language Models (LLMs) have the potential to solve tasks that are beyond the reach of any single LLM. However, this potential can only be realized when the collaboration mechanism between agents is optimized. Specifically, optimizing the communication structure between agents is critical for fruitful collaboration. Most existing approaches rely on fixed topologies, pretrained graph generators, optimization over edges, or employ external LLM judges, thereby adding to the complexity. In this work, we introduce a response-conditioned framework that adapts communication on-the-fly. Agents independently generate responses to the user query and assess peer contributions using an approximation of the Shapley value. A directed acyclic graph (DAG) is then constructed to regulate the propagation of the responses among agents, which ensures stable and efficient message transmission from high-contributing agents to others. This graph is dynamically updated based on the agent responses from the previous collaboration round. Since the proposed framework enables the self-organization of agents without additional supervision or training, we refer to it as SelfOrg. The SelfOrg framework goes beyond task- and query-level optimization and takes into account the stochastic nature of agent responses. Experiments with both strong and weak LLM backends demonstrate robust performance, with significant gains in the weak regime where prior methods collapse. We also theoretically show that multiple agents increase the chance of correctness and that the correct responses naturally dominate the information flow.

Method: Proposed the Bootstrap Percolation and Mean Subsequence Reduced (BP-MSR) algorithm, a distributed approach that establishes sufficient conditions for individual followers to achieve consensus in arbitrary time-varying graphs.

Result: Simulations validate that the BP-MSR algorithm guarantees partial leader-follower consensus, enabling a subset of non-adversarial followers to track the leader’s state even when standard resilient consensus algorithms fail.

Conclusion: The BP-MSR algorithm successfully achieves partial leader-follower consensus in adversarial environments with insufficient robustness, providing resilience guarantees where traditional methods would fail.

Abstract: This work studies resilient leader-follower consensus with a bounded number of adversaries. Existing approaches typically require robustness conditions of the entire network to guarantee resilient consensus. However, the behavior of such systems when these conditions are not fully met remains unexplored. To address this gap, we introduce the notion of partial leader-follower consensus, in which a subset of non-adversarial followers successfully tracks the leader’s reference state despite insufficient robustness. We propose a novel distributed algorithm - the Bootstrap Percolation and Mean Subsequence Reduced (BP-MSR) algorithm - and establish sufficient conditions for individual followers to achieve consensus via the BP-MSR algorithm in arbitrary time-varying graphs. We validate our findings through simulations, demonstrating that our method guarantees partial leader-follower consensus, even when standard resilient consensus algorithms fail.

Method: Uses inversion-regeneration paradigm with word-to-sounding-object aligned audio generation model and holistically-optimized editing algorithm for structural preservation.

Result: Achieved advanced results in audio-video object-level editing tasks with fine semantic alignment, and the audio generation model also performed well.

Conclusion: Object-AVEdit successfully enables object-level audio-visual editing with good structural preservation and semantic alignment across modalities.

Abstract: There is a high demand for audio-visual editing in video post-production and the film making field. While numerous models have explored audio and video editing, they struggle with object-level audio-visual operations. Specifically, object-level audio-visual editing requires the ability to perform object addition, replacement, and removal across both audio and visual modalities, while preserving the structural information of the source instances during the editing process. In this paper, we present \textbf{Object-AVEdit}, achieving the object-level audio-visual editing based on the inversion-regeneration paradigm. To achieve the object-level controllability during editing, we develop a word-to-sounding-object well-aligned audio generation model, bridging the gap in object-controllability between audio and current video generation models. Meanwhile, to achieve the better structural information preservation and object-level editing effect, we propose an inversion-regeneration holistically-optimized editing algorithm, ensuring both information retention during the inversion and better regeneration effect. Extensive experiments demonstrate that our editing model achieved advanced results in both audio-video object-level editing tasks with fine audio-visual semantic alignment. In addition, our developed audio generation model also achieved advanced performance. More results on our project page: https://gewu-lab.github.io/Object_AVEdit-website/.

Method: Proposes DiffAU, a cascaded Ambisonics upscaling method that leverages diffusion models with novel adaptation to spatial audio to generate 3rd order Ambisonics from FOA input.

Result: Experiments in anechoic conditions with multiple speakers show strong objective and perceptual performance, demonstrating reliable reproduction of high-order Ambisonics in various settings.

Conclusion: DiffAU provides a principled approach that rapidly and reliably reproduces high-order Ambisonics, addressing the spatial resolution limitations of first-order Ambisonics while leveraging its hardware efficiency.

Abstract: Spatial audio enhances immersion by reproducing 3D sound fields, with Ambisonics offering a scalable format for this purpose. While first-order Ambisonics (FOA) notably facilitates hardware-efficient acquisition and storage of sound fields as compared to high-order Ambisonics (HOA), its low spatial resolution limits realism, highlighting the need for Ambisonics upscaling (AU) as an approach for increasing the order of Ambisonics signals. In this work we propose DiffAU, a cascaded AU method that leverages recent developments in diffusion models combined with novel adaptation to spatial audio to generate 3rd order Ambisonics from FOA. By learning data distributions, DiffAU provides a principled approach that rapidly and reliably reproduces HOA in various settings. Experiments in anechoic conditions with multiple speakers, show strong objective and perceptual performance.

Method: Collected 45-minute recordings from 36 participants reading newspaper articles under controlled conditions using five different quality microphones.

Result: Created 20 deepfake voices and compiled a dataset with anonymized speaker demographics, enabling more realistic synthetic voice generation.

Conclusion: ELAD SVDSR is expected to advance audio forensics, biometric security, and voice authentication systems by providing challenging deepfake examples for detection training.

Abstract: This paper introduces the Extended Length Audio Dataset for Synthetic Voice Detection and Speaker Recognition (ELAD SVDSR), a resource specifically designed to facilitate the creation of high quality deepfakes and support the development of detection systems trained against them. The dataset comprises 45 minute audio recordings from 36 participants, each reading various newspaper articles recorded under controlled conditions and captured via five microphones of differing quality. By focusing on extended duration audio, ELAD SVDSR captures a richer range of speech attributes such as pitch contours, intonation patterns, and nuanced delivery enabling models to generate more realistic and coherent synthetic voices. In turn, this approach allows for the creation of robust deepfakes that can serve as challenging examples in datasets used to train and evaluate synthetic voice detection methods. As part of this effort, 20 deepfake voices have already been created and added to the dataset to showcase its potential. Anonymized metadata accompanies the dataset on speaker demographics. ELAD SVDSR is expected to spur significant advancements in audio forensics, biometric security, and voice authentication systems.

Method: Directly optimize FDN parameters through differentiable programming-based optimization to match target RIR acoustic and psychoacoustic metrics like clarity and definition.

Result: Enables dynamic real-time adjustments for listener/source movement and produces renderings with quality similar to convolution with long BRIR filters but at much lower computational cost.

Conclusion: The method provides efficient real-time RIR rendering that can be combined with HRIR representations for complete auditory object rendering.

Abstract: We introduce a computationally efficient and tunable feedback delay network (FDN) architecture for real-time room impulse response (RIR) rendering that addresses the computational and latency challenges inherent in traditional convolution and Fourier transform based methods. Our approach directly optimizes FDN parameters to match target RIR acoustic and psychoacoustic metrics such as clarity and definition through novel differentiable programming-based optimization. Our method enables dynamic, real-time adjustments of room impulse responses that accommodates listener and source movement. When combined with previous work on representation of head-related impulse responses via infinite impulse responses, an efficient rendering of auditory objects is possible when the HRIR and RIR are known. Our method produces renderings with quality similar to convolution with long binaural room impulse response (BRIR) filters, but at a fraction of the computational cost.

Method: Proposed personalizing OVR systems through data augmentation and fine-tuning, compared to generic counterparts. Evaluated using objective metrics and subjective listening tests under various conditions.

Result: Personalized OVR provides benefits over generic OVR for some talkers only. Objective metrics don’t always accurately predict system performance comparisons, with certain disturbances leading to consistent overestimation of quality compared to subjective ratings.

Conclusion: Personalization shows limited benefits and objective metrics have limitations in predicting actual subjective quality, particularly overestimating quality for certain types of disturbances.

Abstract: Own voice pickup technology for hearable devices facilitates communication in noisy environments. Own voice reconstruction (OVR) systems enhance the quality and intelligibility of the recorded noisy own voice signals. Since disturbances affecting the recorded own voice signals depend on individual factors, personalized OVR systems have the potential to outperform generic OVR systems. In this paper, we propose personalizing OVR systems through data augmentation and fine-tuning, comparing them to their generic counterparts. We investigate the influence of personalization on speech quality assessed by objective metrics and conduct a subjective listening test to evaluate quality under various conditions. In addition, we assess the prediction accuracy of the objective metrics by comparing predicted quality with subjectively measured quality. Our findings suggest that personalized OVR provides benefits over generic OVR for some talkers only. Our results also indicate that performance comparisons between systems are not always accurately predicted by objective metrics. In particular, certain disturbances lead to a consistent overestimation of quality compared to actual subjective ratings.

Method: The authors analyze static and dynamic post-training quantization schemes for audio DiTs, using two extensions: (1) denoising-timestep-aware smoothing that adapts quantization scales per input channel and timestep, and (2) lightweight LoRA-based branches from SVD to compensate for weight errors. They benchmark W8A8 and W4A8 configurations using Stable Audio Open.

Result: Dynamic quantization preserves fidelity even at lower precision, while static methods remain competitive with lower latency. Low-precision DiTs can retain high-fidelity generation while reducing memory usage by up to 79%.

Conclusion: Post-training quantization enables practical deployment of audio DiTs by significantly reducing computational and storage requirements while maintaining audio quality, with dynamic quantization showing better fidelity preservation at lower precision levels.

Abstract: Diffusion Transformers (DiTs) enable high-quality audio synthesis but are often computationally intensive and require substantial storage, which limits their practical deployment. In this paper, we present a comprehensive evaluation of post-training quantization (PTQ) techniques for audio DiTs, analyzing the trade-offs between static and dynamic quantization schemes. We explore two practical extensions (1) a denoising-timestep-aware smoothing method that adapts quantization scales per-input-channel and timestep to mitigate activation outliers, and (2) a lightweight low-rank adapter (LoRA)-based branch derived from singular value decomposition (SVD) to compensate for residual weight errors. Using Stable Audio Open we benchmark W8A8 and W4A8 configurations across objective metrics and human perceptual ratings. Our results show that dynamic quantization preserves fidelity even at lower precision, while static methods remain competitive with lower latency. Overall, our findings show that low-precision DiTs can retain high-fidelity generation while reducing memory usage by up to 79%.

Method: Proposed DR-BioL framework that uses contrastive learning to promote cohesion within same species and mitigate inter-domain discrepancies, combined with species-conditional distribution alignment to enhance cross-domain species representation.

Result: Experiments on multi-domain mosquito bioacoustic dataset from diverse environments show DR-BioL improves accuracy and robustness of baselines.

Conclusion: DR-BioL demonstrates potential for reliable cross-domain mosquito species classification in real-world applications by addressing domain shift issues.

Abstract: Mosquito Species Classification (MSC) is crucial for vector surveillance and disease control. The collection of mosquito bioacoustic data is often limited by mosquito activity seasons and fieldwork. Mosquito recordings across regions, habitats, and laboratories often show non-biological variations from the recording environment, which we refer to as domain features. This study finds that models directly trained on audio recordings with domain features tend to rely on domain information rather than the species’ acoustic cues for identification, resulting in illusory good performance while actually performing poor cross-domain generalization. To this end, we propose a Domain-Robust Bioacoustic Learning (DR-BioL) framework that combines contrastive learning with distribution alignment. Contrastive learning aims to promote cohesion within the same species and mitigate inter-domain discrepancies, and species-conditional distribution alignment further enhances cross-domain species representation. Experiments on a multi-domain mosquito bioacoustic dataset from diverse environments show that the DR-BioL improves the accuracy and robustness of baselines, highlighting its potential for reliable cross-domain MSC in the real world.

Method: Uses flow matching generative model to capture conditional distribution of complex-valued spectral coefficients and directly reconstructs waveforms through inverse Short-Time Fourier Transform.

Result: Achieves state-of-the-art performance, consistently producing high-fidelity 48 kHz audio across diverse upsampling factors on both speech and general audio datasets.

Conclusion: The vocoder-free approach simplifies end-to-end optimization and overcomes vocoder performance bottlenecks, enabling superior audio super-resolution.

Abstract: In this paper, we present a vocoder-free framework for audio super-resolution that employs a flow matching generative model to capture the conditional distribution of complex-valued spectral coefficients. Unlike conventional two-stage diffusion-based approaches that predict a mel-spectrogram and then rely on a pre-trained neural vocoder to synthesize waveforms, our method directly reconstructs waveforms via the inverse Short-Time Fourier Transform (iSTFT), thereby eliminating the dependence on a separate vocoder. This design not only simplifies end-to-end optimization but also overcomes a critical bottleneck of two-stage pipelines, where the final audio quality is fundamentally constrained by vocoder performance. Experiments show that our model consistently produces high-fidelity 48 kHz audio across diverse upsampling factors, achieving state-of-the-art performance on both speech and general audio datasets.

Method: Used bidirectional LSTM models on 3+ hours of realtime dynamic MRI speech data, with denoised speech signals and automatically segmented articulator contours, testing individual vs simultaneous articulator inversion.

Result: Achieved average RMSE precision of 1.65 mm on test set, comparable to pixel size of 1.62 mm, successfully inverting entire vocal tract from glottis to lips.

Conclusion: This represents the first complete inversion of the entire vocal tract, demonstrating feasibility of comprehensive acoustic-to-articulatory mapping using MRI data and bidirectional LSTM approaches.

Abstract: Acoustic to articulatory inversion has often been limited to a small part of the vocal tract because the data are generally EMA (ElectroMagnetic Articulography) data requiring sensors to be glued to easily accessible articulators. The presented acoustic to articulation model focuses on the inversion of the entire vocal tract from the glottis, the complete tongue, the velum, to the lips. It relies on a realtime dynamic MRI database of more than 3 hours of speech. The data are the denoised speech signal and the automatically segmented articulator contours. Several bidirectional LSTM-based approaches have been used, either inverting each articulator individually or inverting all articulators simultaneously. To our knowledge, this is the first complete inversion of the vocal tract. The average RMSE precision on the test set is 1.65 mm to be compared with the pixel size which is 1.62 mm.

Method: Used X-vector system from Kaldi for audio trials, visual modality models for audio-visual trials. Employed pretrained models on VoxBlink2 and VoxCeleb2 datasets, and trained X-vector models from scratch using CTS superset dataset.

Result: Submitted results for both closed-set and open-set conditions to the competition website.

Conclusion: The paper reports on the performance of the proposed speaker recognition systems evaluated on the SRE24 benchmark.

Abstract: The CL-UZH team submitted one system each for the fixed and open conditions of the NIST SRE 2024 challenge. For the closed-set condition, results for the audio-only trials were achieved using the X-vector system developed with Kaldi. For the audio-visual results we used only models developed for the visual modality. Two sets of results were submitted for the open-set and closed-set conditions, one based on a pretrained model using the VoxBlink2 and VoxCeleb2 datasets. An Xvector-based model was trained from scratch using the CTS superset dataset for the closed set. In addition to the submission of the results of the SRE24 evaluation to the competition website, we talked about the performance of the proposed systems on the SRE24 evaluation in this report.

Method: Proposed Spiralformer encoder that combines layer dropping and early exiting for block processing. It skips layer computation cyclically and shifts computed layers spirally across blocks, completing all layer computations over the block processing.

Result: Achieved 21.6% reduction in averaged token emission delay in Librispeech and 7.0% in CSJ compared to baseline, while maintaining similar computational cost and word error rates.

Conclusion: Spiralformer effectively reduces encoding latency in streaming speech recognition through efficient layer computation strategies, demonstrating significant improvements in token emission delay without sacrificing accuracy or computational efficiency.

Abstract: For streaming speech recognition, a Transformer-based encoder has been widely used with block processing. Although many studies addressed improving emission latency of transducers, little work has been explored for improving encoding latency of the block processing. We seek to reduce latency by frequently emitting a chunk with a small shift rather than scarce large-chunk emissions, resulting in higher computational costs. To efficiently compute with the small chunk shift, we propose a new encoder, Spiralformer, tailored for block processing by combining layer dropping and early exiting. We skip layer computation in a cyclic manner and shift the computed layer in each block spirally, which completes computation for all the layers over the block processing. Experimentally, we observed that our method achieved 21.6% reduction in the averaged token emission delay in Librispeech, and 7.0% in CSJ, compared with the baseline with similar computational cost and word error rates.

Method: Uses graph shift operators to build learnable graph topology and defines learnable graph Fourier basis through singular value matrices using 1-D convolution neural layers, avoiding matrix inversion.

Result: Eliminates numerical errors and stability problems associated with matrix inversion in traditional GFT approaches.

Conclusion: The proposed learnable GFT-SVD framework provides a simple yet effective solution for speech enhancement with improved adaptability and numerical stability.

Abstract: The Graph Fourier Transform (GFT) has recently demonstrated promising results in speech enhancement. However, existing GFT-based speech enhancement approaches often employ fixed graph topologies to build the graph Fourier basis, whose the representation lacks the adaptively and flexibility. In addition, they suffer from the numerical errors and instability introduced by matrix inversion in GFT based on both Singular Value Decomposition (GFT-SVD) and Eigen Vector Decomposition (GFT-EVD). Motivated by these limitations, this paper propose a simple yet effective learnable GFT-SVD framework for speech enhancement. Specifically, we leverage graph shift operators to construct a learnable graph topology and define a learnable graph Fourier basis by the singular value matrices using 1-D convolution (Conv-1D) neural layer. This eliminates the need for matrix inversion, thereby avoiding the associated numerical errors and stability problem.

Method: Conducted comprehensive survey and proposed systematic taxonomy with four evaluation dimensions: General Auditory Awareness and Processing, Knowledge and Reasoning, Dialogue-oriented Ability, and Fairness, Safety, and Trustworthiness.

Result: Developed first survey specifically focused on LALM evaluations, providing detailed overviews within each category and highlighting field challenges.

Conclusion: Provides clear evaluation guidelines for the community and will maintain collection of surveyed papers to support ongoing advancements in large audio-language models.

Abstract: With advancements in large audio-language models (LALMs), which enhance large language models (LLMs) with auditory capabilities, these models are expected to demonstrate universal proficiency across various auditory tasks. While numerous benchmarks have emerged to assess LALMs’ performance, they remain fragmented and lack a structured taxonomy. To bridge this gap, we conduct a comprehensive survey and propose a systematic taxonomy for LALM evaluations, categorizing them into four dimensions based on their objectives: (1) General Auditory Awareness and Processing, (2) Knowledge and Reasoning, (3) Dialogue-oriented Ability, and (4) Fairness, Safety, and Trustworthiness. We provide detailed overviews within each category and highlight challenges in this field, offering insights into promising future directions. To the best of our knowledge, this is the first survey specifically focused on the evaluations of LALMs, providing clear guidelines for the community. We will release the collection of the surveyed papers and actively maintain it to support ongoing advancements in the field.

Method: Uses object-oriented processing strategy with dynamic temporal kernel-based feature extractor, transformer-based aggregator, and object separator. Implements temporal coherence matching within chain-of-inference for multi-task fusion and iterative refinement of object features.

Result: Achieves state-of-the-art performance across all evaluated tasks on spatial audio benchmark datasets (ASA2, MC-FUSS, STARSS23), demonstrating effectiveness in both source separation and auditory parameter estimation.

Conclusion: DeepASA provides a unified framework that successfully handles multiple auditory scene analysis tasks through object-centric representations and iterative refinement, resolving parameter association ambiguity and achieving robust performance in diverse spatial auditory scenes.

Abstract: We propose DeepASA, a multi-purpose model for auditory scene analysis that performs multi-input multi-output (MIMO) source separation, dereverberation, sound event detection (SED), audio classification, and direction-of-arrival estimation (DoAE) within a unified framework. DeepASA is designed for complex auditory scenes where multiple, often similar, sound sources overlap in time and move dynamically in space. To achieve robust and consistent inference across tasks, we introduce an object-oriented processing (OOP) strategy. This approach encapsulates diverse auditory features into object-centric representations and refines them through a chain-of-inference (CoI) mechanism. The pipeline comprises a dynamic temporal kernel-based feature extractor, a transformer-based aggregator, and an object separator that yields per-object features. These features feed into multiple task-specific decoders. Our object-centric representations naturally resolve the parameter association ambiguity inherent in traditional track-wise processing. However, early-stage object separation can lead to failure in downstream ASA tasks. To address this, we implement temporal coherence matching (TCM) within the chain-of-inference, enabling multi-task fusion and iterative refinement of object features using estimated auditory parameters. We evaluate DeepASA on representative spatial audio benchmark datasets, including ASA2, MC-FUSS, and STARSS23. Experimental results show that our model achieves state-of-the-art performance across all evaluated tasks, demonstrating its effectiveness in both source separation and auditory parameter estimation under diverse spatial auditory scenes.

Method: The study investigates uncertainty-aware deep learning models with rejection mechanisms, using Variational Bayesian models that reject low-confidence predictions in a deferred decision-making clinical context.

Result: Results show trade-offs between prediction coverage and reliability, with Variational Bayesian models adopting conservative strategies that reject uncertain predictions. Performance metrics include accuracy on accepted predictions, coverage, rejection-ratio, and Expected Calibration Error.

Conclusion: Uncertainty estimation and selective rejection improve model reliability in safety-critical diagnostic use cases, demonstrating a clear trade-off between accuracy and caution.

Abstract: Diabetic retinopathy (DR) is a major cause of visual impairment, and effective treatment options depend heavily on timely and accurate diagnosis. Deep learning models have demonstrated great success identifying DR from retinal images. However, relying only on predictions made by models, without any indication of model confidence, creates uncertainty and poses significant risk in clinical settings. This paper investigates an alternative in uncertainty-aware deep learning models, including a rejection mechanism to reject low-confidence predictions, contextualized by deferred decision-making in clinical practice. The results show there is a trade-off between prediction coverage and coverage reliability. The Variational Bayesian model adopted a more conservative strategy when predicting DR, subsequently rejecting the uncertain predictions. The model is evaluated by means of important performance metrics such as Accuracy on accepted predictions, the proportion of accepted cases (coverage), the rejection-ratio, and Expected Calibration Error (ECE). The findings also demonstrate a clear trade-off between accuracy and caution, establishing that the use of uncertainty estimation and selective rejection improves the model’s reliability in safety-critical diagnostic use cases.

Method: Incorporates N-gram context into the Swin Transformer architecture to expand the regions considered for pixel restoration and increase context awareness across neighboring windows.

Result: Achieves -5.86% improvement in BD-Rate over existing variable-rate learned image compression techniques and improves quality of regions of interest (ROI) in images.

Conclusion: The method is particularly beneficial for object-focused applications in manufacturing and industrial vision systems due to improved ROI quality and high-resolution reconstruction.

Abstract: This paper presents an N-gram context-based Swin Transformer for learned image compression. Our method achieves variable-rate compression with a single model. By incorporating N-gram context into the Swin Transformer, we overcome its limitation of neglecting larger regions during high-resolution image reconstruction due to its restricted receptive field. This enhancement expands the regions considered for pixel restoration, thereby improving the quality of high-resolution reconstructions. Our method increases context awareness across neighboring windows, leading to a -5.86% improvement in BD-Rate over existing variable-rate learned image compression techniques. Additionally, our model improves the quality of regions of interest (ROI) in images, making it particularly beneficial for object-focused applications in fields such as manufacturing and industrial vision systems.

Method: Uses CNN architecture with separable convolutions, batch normalization, and dropout regularization. Applies data augmentation and adaptive learning rate strategies on extensive chest X-ray dataset. Integrates medical ontologies with semantic technology.

Result: Achieved 91% accuracy with strong performance across precision, recall, and F1 score metrics. Enhanced generalization capabilities and diagnostic reliability.

Conclusion: The approach provides a scalable, efficient pneumonia detection solution that advances AI integration into clinical settings with more precise automated diagnostic methods delivering consistent medical imaging results.

Abstract: Deep learning integration into medical imaging systems has transformed disease detection and diagnosis processes with a focus on pneumonia identification. The study introduces an intricate deep learning system using Convolutional Neural Networks for automated pneumonia detection from chest Xray images which boosts diagnostic precision and speed. The proposed CNN architecture integrates sophisticated methods including separable convolutions along with batch normalization and dropout regularization to enhance feature extraction while reducing overfitting. Through the application of data augmentation techniques and adaptive learning rate strategies the model underwent training on an extensive collection of chest Xray images to enhance its generalization capabilities. A convoluted array of evaluation metrics such as accuracy, precision, recall, and F1 score collectively verify the model exceptional performance by recording an accuracy rate of 91. This study tackles critical clinical implementation obstacles such as data privacy protection, model interpretability, and integration with current healthcare systems beyond just model performance. This approach introduces a critical advancement by integrating medical ontologies with semantic technology to improve diagnostic accuracy. The study enhances AI diagnostic reliability by integrating machine learning outputs with structured medical knowledge frameworks to boost interpretability. The findings demonstrate AI powered healthcare tools as a scalable efficient pneumonia detection solution. This study advances AI integration into clinical settings by developing more precise automated diagnostic methods that deliver consistent medical imaging results.

Method: Uses gray and white matter MRI slices as inputs to three pretrained CNNs (ResNet50, NASNet, MobileNet) fine-tuned end-to-end, with stacked ensemble learning and weighted averaging to combine models.

Result: Achieves 99.21% accuracy for AD vs. MCI and 91.0% for MCI vs. Normal Controls on ADNI dataset, outperforming conventional transfer learning and baseline ensemble methods.

Conclusion: The framework shows potential for robust and scalable clinical decision support in neurodegenerative disease diagnostics, with enhanced interpretability through Explainable AI techniques.

Abstract: Early and accurate diagnosis of Alzheimer Disease is critical for effective clinical intervention, particularly in distinguishing it from Mild Cognitive Impairment, a prodromal stage marked by subtle structural changes. In this study, we propose a hybrid deep learning ensemble framework for Alzheimer Disease classification using structural magnetic resonance imaging. Gray and white matter slices are used as inputs to three pretrained convolutional neural networks such as ResNet50, NASNet, and MobileNet, each fine tuned through an end to end process. To further enhance performance, we incorporate a stacked ensemble learning strategy with a meta learner and weighted averaging to optimally combine the base models. Evaluated on the Alzheimer Disease Neuroimaging Initiative dataset, the proposed method achieves state of the art accuracy of 99.21% for Alzheimer Disease vs. Mild Cognitive Impairment and 91.0% for Mild Cognitive Impairment vs. Normal Controls, outperforming conventional transfer learning and baseline ensemble methods. To improve interpretability in image based diagnostics, we integrate Explainable AI techniques by Gradient weighted Class Activation, which generates heatmaps and attribution maps that highlight critical regions in gray and white matter slices, revealing structural biomarkers that influence model decisions. These results highlight the frameworks potential for robust and scalable clinical decision support in neurodegenerative disease diagnostics.

Method: System with RGB-D camera + wearable sensors, mobile app for guidance, and AI server using RAST-G@ model (ST-GCN + transformer attention) for movement assessment. Built NRC dataset with 10 ADL and 5 ROM activities annotated by physiotherapists.

Result: RAST-G@ outperforms baselines on KIMORE and NRC datasets in MAD, RMSE, and MAPE metrics. System provides patient-centered assessment and monitoring feedback.

Conclusion: The proposed system offers scalable, quantitative, and consistent home-based rehabilitation assessment for stroke patients.

Abstract: Effective stroke recovery requires continuous rehabilitation integrated with daily living. To support this need, we propose a home-based rehabilitation exercise and feedback system. The system consists of (1) hardware setup with RGB-D camera and wearable sensors to capture Stroke movements, (2) a mobile application for exercise guidance, and (3) an AI server for assessment and feedback. When Stroke user exercises following the application guidance, the system records skeleton sequences, which are then Assessed by the deep learning model, RAST-G@. The model employs a spatio-temporal graph convolutional network (ST-GCN) to extract skeletal features and integrates transformer-based temporal attention to figure out action quality. For system implementation, we constructed the NRC dataset, include 10 upper-limb activities of daily living (ADL) and 5 range-of-motion (ROM) collected from stroke and non-disabled participants, with Score annotations provided by licensed physiotherapists. Results on the KIMORE and NRC datasets show that RAST-G@ improves over baseline in terms of MAD, RMSE, and MAPE. Furthermore, the system provides user feedback that combines patient-centered assessment and monitoring. The results demonstrate that the proposed system offers a scalable approach for quantitative and consistent domiciliary rehabilitation assessment.

Method: Extends InfoVAE to explicitly maximize mutual information between 3D brain MRI images and latent variables, producing compact, structured embeddings that retain clinically meaningful content. Evaluated on two cohorts: healthy controls (n=6527) with age data and multiple sclerosis patients (n=904) with SDMT scores.

Result: Learned latents support accurate brain-age and SDMT regression, preserve key medical attributes, and form intuitive clusters for interpretation. Consistently outperforms other VAE variants across reconstruction and downstream prediction tasks, indicating stronger information capture in the embedding space.

Conclusion: InfoVAE-Med3D unites predictive performance with interpretability, offering a practical path toward MRI-based biomarkers and more transparent analysis of cognitive deterioration in neurological disease.

Abstract: We present InfoVAE-Med3D, a latent-representation learning approach for 3D brain MRI that targets interpretable biomarkers of cognitive decline. Standard statistical models and shallow machine learning often lack power, while most deep learning methods behave as black boxes. Our method extends InfoVAE to explicitly maximize mutual information between images and latent variables, producing compact, structured embeddings that retain clinically meaningful content. We evaluate on two cohorts: a large healthy-control dataset (n=6527) with chronological age, and a clinical multiple sclerosis dataset from Charles University in Prague (n=904) with age and Symbol Digit Modalities Test (SDMT) scores. The learned latents support accurate brain-age and SDMT regression, preserve key medical attributes, and form intuitive clusters that aid interpretation. Across reconstruction and downstream prediction tasks, InfoVAE-Med3D consistently outperforms other VAE variants, indicating stronger information capture in the embedding space. By uniting predictive performance with interpretability, InfoVAE-Med3D offers a practical path toward MRI-based biomarkers and more transparent analysis of cognitive deterioration in neurological disease.

Method: Proposes DPsurv - a dual-prototype whole-slide image evidential fusion network that uses patch prototype assignment maps, component prototypes, and component-wise relative risk aggregation to provide uncertainty-aware survival predictions.

Result: Achieves the highest mean concordance index and the lowest mean integrated Brier score across five publicly available datasets, demonstrating superior performance and reliability.

Conclusion: DPsurv provides transparent predictions at feature, reasoning, and decision levels, significantly enhancing the trustworthiness and interpretability of WSI-based survival analysis.

Abstract: Pathology whole-slide images (WSIs) are widely used for cancer survival analysis because of their comprehensive histopathological information at both cellular and tissue levels, enabling quantitative, large-scale, and prognostically rich tumor feature analysis. However, most existing methods in WSI survival analysis struggle with limited interpretability and often overlook predictive uncertainty in heterogeneous slide images. In this paper, we propose DPsurv, a dual-prototype whole-slide image evidential fusion network that outputs uncertainty-aware survival intervals, while enabling interpretation of predictions through patch prototype assignment maps, component prototypes, and component-wise relative risk aggregation. Experiments on five publicly available datasets achieve the highest mean concordance index and the lowest mean integrated Brier score, validating the effectiveness and reliability of DPsurv. The interpretation of prediction results provides transparency at the feature, reasoning, and decision levels, thereby enhancing the trustworthiness and interpretability of DPsurv.

Method: Evaluated SkinGPT-4 on SCIN dataset, developed fine-tuned models for custom classification tasks, and assessed fairness using demographic parity and equalized odds metrics across Fitzpatrick skin types.

Result: SkinGPT-4 showed 0.10 average demographic parity bias, with 0.10-0.15 differences between lightest and darkest tones. Custom models achieved 0.75 average F1 score and 0.75 demographic parity, with best model reaching 0.83-0.90 parity scores across Fitzpatrick types.

Conclusion: Large language models exhibit skin tone bias, but fine-tuning existing backbones can create accurate, fair models for skin disease classification across diverse populations.

Abstract: SkinGPT-4, a large vision-language model, leverages annotated skin disease images to augment clinical workflows in underserved communities. However, its training dataset predominantly represents lighter skin tones, limiting diagnostic accuracy for darker tones. Here, we evaluated performance biases in SkinGPT-4 across skin tones on common skin diseases, including eczema, allergic-contact dermatitis, and psoriasis using the open-sourced SCIN dataset. We leveraged the SkinGPT-4 backbone to develop finetuned models for custom skin disease classification tasks and explored bias mitigation strategies. Clinical evaluation by board-certified dermatologists on six relevant skin diseases from 300 SCIN cases assessed images for diagnostic accuracy, informativity, physician utility, and patient utility. Model fairness metrics, including demographic parity and equalized odds, were calculated across skin tones. SkinGPT-4 achieved an average demographic parity of 0.10 across Fitzpatrick types, with notable differences of 0.10-0.15 between lightest and darkest tones across evaluation metrics. Model hallucinations in artifacts and anatomy occurred at a rate of 17.8. Our customized models achieved average F1, precision, and AUROC of 0.75, 0.78, and 0.78 across visually similar disease pairs. Fairness analysis showed an average demographic parity of 0.75, with a maximum disparity of 0.21 across skin tones. The best model achieved parity scores of 0.83, 0.83, 0.76, 0.89, 0.90, and 0.90 for Fitzpatrick I-VI, indicating robust fairness. Large language models such as SkinGPT-4 showed weaker performance on darker tones. Model biases exist across evaluation criteria, and hallucinations may affect diagnostic efficacy. These findings demonstrate the efficacy of training accurate, fair models using existing backbones for custom skin disease classification.

Method: The authors use a tri-axial framework that couples imaging modalities (DXA, X-ray, CT, MRI) with clinical tasks and AI methodologies (classical ML, CNNs, transformers, self-supervised learning, explainable AI). They follow PRISMA-guided search strategy and introduce taxonomy via a roadmap figure.

Result: The survey synthesizes cross-study insights on key challenges including data scarcity, external validation, and interpretability. It identifies emerging trends and provides actionable research directions for the field.

Conclusion: This review provides AI scientists, medical imaging researchers, and musculoskeletal clinicians with a clear compass to accelerate rigorous, patient-centered innovation in osteoporosis care, addressing fragmentation in the literature through a unified framework.

Abstract: Osteoporosis silently erodes skeletal integrity worldwide; however, early detection through imaging can prevent most fragility fractures. Artificial intelligence (AI) methods now mine routine Dual-energy X-ray Absorptiometry (DXA), X-ray, Computed Tomography (CT), and Magnetic Resonance Imaging (MRI) scans for subtle, clinically actionable markers, but the literature is fragmented. This survey unifies the field through a tri-axial framework that couples imaging modalities with clinical tasks and AI methodologies (classical machine learning, convolutional neural networks (CNNs), transformers, self-supervised learning, and explainable AI). Following a concise clinical and technical primer, we detail our Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)-guided search strategy, introduce the taxonomy via a roadmap figure, and synthesize cross-study insights on data scarcity, external validation, and interpretability. By identifying emerging trends, open challenges, and actionable research directions, this review provides AI scientists, medical imaging researchers, and musculoskeletal clinicians with a clear compass to accelerate rigorous, patient-centered innovation in osteoporosis care. The project page of this survey can also be found on Github.

Method: V-info is introduced as a relaxation of TSI that can quantify image utility for specified families of sub-ideal observers. It’s evaluated in a magnetic resonance image restoration problem for signal detection and discrimination tasks.

Result: V-info correlates with ROC curve area for binary tasks and works for multi-class tasks where ROC analysis is challenging. It shows greater sensitivity than conventional metrics in saturation scenarios.

Conclusion: V-info represents a new objective image quality measure that complements conventional signal detection theory-based metrics, particularly for multi-class tasks and scenarios where traditional metrics saturate.

Abstract: Objective, task-based, measures of image quality (IQ) have been widely advocated for assessing and optimizing medical imaging technologies. Besides signal detection theory-based measures, information-theoretic quantities have been proposed to quantify task-based IQ. For example, task-specific information (TSI), defined as the mutual information between an image and task variable, represents an optimal measure of how informative an image is for performing a specified task. However, like the ideal observer from signal detection theory, TSI does not quantify the amount of task-relevant information in an image that can be exploited by a sub-ideal observer. A recently proposed relaxation of TSI, termed predictive V-information (V-info), removes this limitation and can quantify the utility of an image with consideration of a specified family of sub-ideal observers. In this study, for the first time, V-info is proposed and investigated as an objective, task-specific, IQ metric. To corroborate its usefulness, a stylized magnetic resonance image restoration problem is considered in which V-info is employed to quantify signal detection or discrimination performance. The presented results show that V-info correlates with area under the receiver operating characteristic (ROC) curve for binary tasks, while being readily applicable to multi-class (>2) tasks where ROC analysis is challenging. Notably, V-info exhibits greater sensitivity in scenarios where conventional metrics saturate. These findings demonstrate that V-info represents a new objective IQ measure that can complement conventional signal detection theory-based ones.

Method: Proposed deep learning framework that uses longitudinal information by incorporating prior full-dose MRI exams from the same patient to virtually enhance low-dose post-contrast T1-weighted MRI images.

Result: Longitudinal approach significantly improved image quality across multiple reconstruction metrics compared to non-longitudinal single session model, and showed robustness with varying simulated contrast doses.

Conclusion: Integration of prior imaging history into deep learning-based virtual contrast enhancement can reduce gadolinium contrast agent usage without compromising diagnostic utility, enabling safer longitudinal monitoring in clinical MRI practice.

Abstract: Gadolinium-based contrast agents (GBCAs) are widely used in magnetic resonance imaging (MRI) to enhance lesion detection and characterisation, particularly in the field of neuro-oncology. Nevertheless, concerns regarding gadolinium retention and accumulation in brain and body tissues, most notably for diseases that require close monitoring and frequent GBCA injection, have led to the need for strategies to reduce dosage. In this study, a deep learning framework is proposed for the virtual contrast enhancement of full-dose post-contrast T1-weighted MRI images from corresponding low-dose acquisitions. The contribution of the presented model is its utilisation of longitudinal information, which is achieved by incorporating a prior full-dose MRI examination from the same patient. A comparative evaluation against a non-longitudinal single session model demonstrated that the longitudinal approach significantly improves image quality across multiple reconstruction metrics. Furthermore, experiments with varying simulated contrast doses confirmed the robustness of the proposed method. These results emphasize the potential of integrating prior imaging history into deep learning-based virtual contrast enhancement pipelines to reduce GBCA usage without compromising diagnostic utility, thus paving the way for safer, more sustainable longitudinal monitoring in clinical MRI practice.

Method: Used U-Net with EfficientNet encoder for segmentation, implemented summed-area tables for fast voxel sum calculations enabling 3D full search, and designed user-controllable search metric prioritizing cubes over oblongs.

Result: Proposed computation (8 seconds) was 100-500 times faster than conventional method (11-40 minutes), and proposed metric achieved higher tumor fractions while preferring cube-shaped regions over oblongs.

Conclusion: The method is promising for fast and precise rectangular tumor region search in brain MRI diagnosis, significantly reducing processing time and improving region quality.

Abstract: Purpose: To develop a fast and precise method for searching rectangular regions in brain tumor images. Methods: The authors propose a new method for searching rectangular tumor regions in brain MR images. The proposed method consisted of a segmentation network and a fast search method with a user-controllable search metric. As the segmentation network, the U-Net whose encoder was replaced by the EfficientNet was used. In the fast search method, summed-area tables were used for accelerating sums of voxels in rectangular regions. Use of the summed-area tables enabled exhaustive search of the 3D offset (3D full search). The search metric was designed for giving priority to cubes over oblongs, and assigning better values for higher tumor fractions even if they exceeded target tumor fractions. The proposed computation and metric were compared with those used in a conventional method using the Brain Tumor Image Segmentation dataset. Results: When the 3D full search was used, the proposed computation (8 seconds) was 100-500 times faster than the conventional computation (11-40 minutes). When the user-controllable parts of the search metrics were changed variously, the tumor fractions of the proposed metric were higher than those of the conventional metric. In addition, the conventional metric preferred oblongs whereas the proposed metric preferred cubes. Conclusion: The proposed method is promising for implementing fast and precise search of rectangular tumor regions, which is useful for brain tumor diagnosis using MRI systems. The proposed computation reduced processing times of the 3D full search, and the proposed metric improved the quality of the assigned rectangular tumor regions.

Method: Proposes U-DFA with novel Local-Global Fusion Adapter (LGFA) that injects spatial features from CNN-based Spatial Pattern Adapter into frozen DINOv2 blocks at multiple stages.

Result: Achieves state-of-the-art performance on Synapse and ACDC datasets with only 33% of trainable model parameters.

Conclusion: U-DFA is a robust and scalable framework for medical image segmentation across multiple modalities.

Abstract: Accurate medical image segmentation plays a crucial role in overall diagnosis and is one of the most essential tasks in the diagnostic pipeline. CNN-based models, despite their extensive use, suffer from a local receptive field and fail to capture the global context. A common approach that combines CNNs with transformers attempts to bridge this gap but fails to effectively fuse the local and global features. With the recent emergence of VLMs and foundation models, they have been adapted for downstream medical imaging tasks; however, they suffer from an inherent domain gap and high computational cost. To this end, we propose U-DFA, a unified DINOv2-Unet encoder-decoder architecture that integrates a novel Local-Global Fusion Adapter (LGFA) to enhance segmentation performance. LGFA modules inject spatial features from a CNN-based Spatial Pattern Adapter (SPA) module into frozen DINOv2 blocks at multiple stages, enabling effective fusion of high-level semantic and spatial features. Our method achieves state-of-the-art performance on the Synapse and ACDC datasets with only 33% of the trainable model parameters. These results demonstrate that U-DFA is a robust and scalable framework for medical image segmentation across multiple modalities.

Method: Train CNN for damage localization, then combine MPCA and fine-tuning to adapt the model to new domains. MPCA extracts shared latent features from source and target domains, followed by fine-tuning without needing large datasets.

Result: Tested on 12 case studies with different composite materials and sensor arrays. Showed substantial reduction in localization error compared to standard TL techniques, with statistical metrics confirming improved domain alignment.

Conclusion: The MPCA-based TL framework is robust, data-efficient, and statistically effective for UGW-based structural health monitoring applications.

Abstract: Ultrasonic Guided Waves (UGWs) represent a promising diagnostic tool for Structural Health Monitoring (SHM) in thin-walled structures, and their integration with machine learning (ML) algorithms is increasingly being adopted to enable real-time monitoring capabilities. However, the large-scale deployment of UGW-based ML methods is constrained by data scarcity and limited generalisation across different materials and sensor configurations. To address these limitations, this work proposes a novel transfer learning (TL) framework based on Multilinear Principal Component Analysis (MPCA). First, a Convolutional Neural Network (CNN) for regression is trained to perform damage localisation for a plated structure. Then, MPCA and fine-tuning are combined to have the CNN work for a different plate. By jointly applying MPCA to the source and target domains, the method extracts shared latent features, enabling effective domain adaptation without requiring prior assumptions about dimensionality. Following MPCA, fine-tuning enables adapting the pre-trained CNN to a new domain without the need for a large training dataset. The proposed MPCA-based TL method was tested against 12 case studies involving different composite materials and sensor arrays. Statistical metrics were used to assess domains alignment both before and after MPCA, and the results demonstrate a substantial reduction in localisation error compared to standard TL techniques. Hence, the proposed approach emerges as a robust, data-efficient, and statistically based TL framework for UGW-based SHM.

Method: The framework adapts state-of-the-art 3D diffusion models with enhancements from Simple Diffusion and incorporates augmented conditioning to improve text alignment and image quality.

Result: The model successfully generates high-quality 3D counterfactual medical images for neurological MRI datasets, simulating lesion loads in Multiple Sclerosis and cognitive states in Alzheimer’s disease while preserving subject fidelity.

Conclusion: This work lays the groundwork for prompt-driven disease progression analysis in 3D medical imaging and represents the first language-guided native-3D diffusion model applied to neurological imaging.

Abstract: Vision-language models have demonstrated impressive capabilities in generating 2D images under various conditions; however, the success of these models is largely enabled by extensive, readily available pretrained foundation models. Critically, comparable pretrained models do not exist for 3D, significantly limiting progress. As a result, the potential of vision-language models to produce high-resolution 3D counterfactual medical images conditioned solely on natural language remains unexplored. Addressing this gap would enable powerful clinical and research applications, such as personalized counterfactual explanations, simulation of disease progression, and enhanced medical training by visualizing hypothetical conditions in realistic detail. Our work takes a step toward this challenge by introducing a framework capable of generating high-resolution 3D counterfactual medical images of synthesized patients guided by free-form language prompts. We adapt state-of-the-art 3D diffusion models with enhancements from Simple Diffusion and incorporate augmented conditioning to improve text alignment and image quality. To our knowledge, this is the first demonstration of a language-guided native-3D diffusion model applied to neurological imaging, where faithful three-dimensional modeling is essential. On two neurological MRI datasets, our framework simulates varying counterfactual lesion loads in Multiple Sclerosis and cognitive states in Alzheimer’s disease, generating high-quality images while preserving subject fidelity. Our results lay the groundwork for prompt-driven disease progression analysis in 3D medical imaging. Project link

• https://lesupermomo.github.io/imagining-alternatives/.