Method: Proposes ACoRN with two novel training steps: 1) Offline data augmentation to enhance robustness against two types of retrieval noise, and 2) Fine-tuning to generate summaries centered around key information that directly supports correct answers, addressing positional bias and multi-document utilization issues.

Result: T5-large trained with ACoRN improves EM and F1 scores while preserving answer strings that serve as direct evidence. ACoRN excels on datasets with many accuracy-reducing documents, making it useful in real-world scenarios.

Conclusion: ACoRN effectively addresses the limitations of abstractive compressors in handling noisy retrieved documents, improving answer accuracy in RAG systems by making compressors more robust and better at preserving essential information.

Abstract: Abstractive compression utilizes smaller langauge models to condense query-relevant context, reducing computational costs in retrieval-augmented generation (RAG). However, retrieved documents often include information that is either irrelevant to answering the query or misleading due to factual incorrect content, despite having high relevance scores. This behavior indicates that abstractive compressors are more likely to omit important information essential for the correct answer, especially in long contexts where attention dispersion occurs. To address this issue, we categorize retrieved documents in a more fine-grained manner and propose Abstractive Compression Robust against Noise (ACoRN), which introduces two novel training steps. First, we use offline data augmentation on the training dataset to enhance compressor robustness against two distinct types of retrieval noise. Second, since the language model based compressor cannot fully utilize information from multiple retrieved documents and exhibits positional bias, we perform finetuning to generate summaries centered around key information that directly supports the correct answer. Our experiments demonstrate that T5-large, trained with ACoRN as a compressor, improves EM and F1 scores while preserving the answer string, which could serve as direct evidence. ACoRN excels on datasets with many accuracy reducing documents, making it highly useful in real-world scenarios.

Method: Novel training set from TriviaQA conversions for extrinsic hallucinations, FineWeb long-form texts with fact-grounding rewards for intrinsic hallucinations, plus explicit rewards for refusing unanswerable questions.

Result: Significant performance gains across diverse benchmarks with substantial reduction of both hallucination types, demonstrating improved reliability without sacrificing reasoning capabilities.

Conclusion: Provides a practical framework to resolve the tension between advanced reasoning and factual trustworthiness, enabling more capable and reliable large language models.

Abstract: While reinforcement learning has unlocked unprecedented complex reasoning in large language models, it has also amplified their propensity for hallucination, creating a critical trade-off between capability and reliability. This work confronts this challenge by introducing a targeted RL framework designed to mitigate both intrinsic and extrinsic hallucinations across short and long-form question answering. We address extrinsic hallucinations (flawed internal knowledge) by creating a novel training set from open-ended conversions of TriviaQA. Concurrently, we tackle intrinsic hallucinations (unfaithfulness to context) by leveraging long-form texts from FineWeb in a fact-grounding reward scheme. To further bolster reliability, our framework explicitly rewards the model for refusing to answer unanswerable questions, thereby cultivating crucial cautiousness. Extensive experiments demonstrate that our methodology yields significant performance gains across a diverse suite of benchmarks, substantially reducing both hallucination types. Ultimately, this research contributes a practical framework for resolving the critical tension between advanced reasoning and factual trustworthiness, paving the way for more capable and reliable large language models.

Method: Generated 50 dialogues between human participants and an LLM in standard mode. Five MBT-trained psychiatrists blindly evaluated mentalization profiles along four MBT axes using Likert-scale scores for evaluative coherence, argumentative coherence, and global quality. Inter-rater agreement measured with ICC(3,1).

Result: Mean scores (3.63-3.98) with moderate SDs indicate high structural coherence. ICC values (0.60-0.84) show substantial-to-high rater agreement. Model performed better on Implicit-Explicit and Self-Other dimensions, but had limitations integrating internal states with external contexts. Profiles were coherent and clinically interpretable but affectively neutral.

Conclusion: LLMs can generate linguistically structured mentalization profiles with substantial coherence and rater agreement, demonstrating potential for simulating certain aspects of mentalization. However, limitations in affective integration and contextual grounding highlight the gap between linguistic form and genuine mental representation.

Abstract: Background: Mentalization integrates cognitive, affective, and intersubjective components. Large Language Models (LLMs) display an increasing ability to generate reflective texts, raising questions regarding the relationship between linguistic form and mental representation. This study assesses the extent to which a single LLM can reproduce the linguistic structure of mentalization according to the parameters of Mentalization-Based Treatment (MBT). Methods: Fifty dialogues were generated between human participants and an LLM configured in standard mode. Five psychiatrists trained in MBT, working under blinded conditions, evaluated the mentalization profiles produced by the model along the four MBT axes, assigning Likert-scale scores for evaluative coherence, argumentative coherence, and global quality. Inter-rater agreement was estimated using ICC(3,1). Results: Mean scores (3.63-3.98) and moderate standard deviations indicate a high level of structural coherence in the generated profiles. ICC values (0.60-0.84) show substantial-to-high agreement among raters. The model proved more stable in the Implicit-Explicit and Self-Other dimensions, while presenting limitations in the integration of internal states and external contexts. The profiles were coherent and clinically interpretable yet characterized by affective neutrality.

Method: Luxical combines sparse TF-IDF features with a small ReLU network and uses knowledge distillation to approximate large transformer embedding models. The architecture aims to produce “lexical-dense” embeddings that support various workflows like clustering, classification, and retrieval.

Result: Luxical achieves speedups ranging from 3x to 100x over neural baselines, with inference speed comparable to FastText. It matches the quality of neural baselines in evaluations including webcrawl document retrieval and language model data curation tasks.

Conclusion: Luxical provides favorable compute/quality trade-offs for large-scale text organization, offering transformer-like flexibility at much lower operational cost. The library is available as open-source software for practical deployment.

Abstract: Frontier language model quality increasingly hinges on our ability to organize web-scale text corpora for training. Today’s dominant tools trade off speed and flexibility: lexical classifiers (e.g., FastText) are fast but limited to producing classification output scores, while the vector-valued outputs of transformer text embedding models flexibly support numerous workflows (e.g., clustering, classification, and retrieval) but are computationally expensive to produce. We introduce Luxical, a library for high-speed “lexical-dense” text embeddings that aims to recover the best properties of both approaches for web-scale text organization. Luxical combines sparse TF–IDF features, a small ReLU network, and a knowledge distillation training regimen to approximate large transformer embedding models at a fraction of their operational cost. In this technical report, we describe the Luxical architecture and training objective and evaluate a concrete Luxical model in two disparate applications: a targeted webcrawl document retrieval test and an end-to-end language model data curation task grounded in text classification. In these tasks we demonstrate speedups ranging from 3x to 100x over varying-sized neural baselines, and comparable to FastText model inference during the data curation task. On these evaluations, the tested Luxical model illustrates favorable compute/quality trade-offs for large-scale text organization, matching the quality of neural baselines. Luxical is available as open-source software at https://github.com/datologyai/luxical.

Method: Proposes Knowledge-Guided LLM (KG-LLM) integrating pediatric dental knowledge graph, retrieval-augmented generation (RAG), and multi-stage safety validation pipeline. Uses clinical NER/RE module to extract structured entities from dental notes, retrieves relevant guidelines and historical cases from knowledge graph, and employs dual-layer safety validation combining rule checking with learned classifier.

Result: Experiments on 32,000 pediatric dental records show KG-LLM outperforms domain-adapted Llama-2 baseline: record-understanding F1: 0.914 vs. 0.867, drug-dose-duration accuracy (Top-1): 0.782 vs. 0.716, reduces unsafe antibiotic suggestions by 50%. Ablation shows knowledge graph, RAG, and safety modules each contribute substantially.

Conclusion: KG-LLM framework effectively addresses limitations of traditional systems by integrating structured knowledge with LLMs, demonstrating improved clinical reliability, interpretability, and safety in pediatric dental antibiotic prescribing.

Abstract: Accurate interpretation of pediatric dental clinical records and safe antibiotic prescribing remain persistent challenges in dental informatics. Traditional rule-based clinical decision support systems struggle with unstructured dental narratives, incomplete radiographic descriptions, and complex safety constraints. To address these limitations, this study proposes a Knowledge-Guided Large Language Model (KG-LLM) that integrates a pediatric dental knowledge graph, retrieval-augmented generation (RAG), and a multi-stage safety validation pipeline for evidence-grounded antibiotic recommendation. The framework first employs a clinical NER/RE module to extract structured entities and relations from dental notes and radiology reports. Relevant guidelines, drug-safety rules, and analogous historical cases are subsequently retrieved from the knowledge graph and supplied to the LLM for diagnostic summarization and dose-drug-duration prediction. Safety assurance is achieved through a dual-layer validation mechanism combining deterministic rule checking with a learned classifier for detecting allergies, contraindications, and dosing errors. Experiments on 32,000 de-identified pediatric dental visit records demonstrate the effectiveness of the proposed approach. Compared with a domain-adapted Llama-2 clinical baseline, KG-LLM improves record-understanding performance (F1: 0.914 vs. 0.867), drug-dose-duration accuracy (Top-1: 0.782 vs. 0.716), and reduces unsafe antibiotic suggestions by 50%. Additional evaluation across summary quality, recommendation accuracy, and global safety scores further confirms the robustness of the system. Ablation analyses indicate that the knowledge graph, RAG, and safety modules each contribute substantially to clinical reliability and interpretability.

Method: Proposed two interpretability metrics: Path Reliance Degree (PRD) measures over-reliance on shortest-path triples, and Semantic Alignment Score (SAS) assesses alignment between model’s internal representations and retrieved knowledge. Developed Graph Grounding and Alignment (GGA), a lightweight post-hoc hallucination detector based on these metrics.

Result: Identified failure patterns associated with high PRD (over-reliance on salient paths) and low SAS (weak semantic grounding). GGA hallucination detector outperformed strong semantic and confidence-based baselines across AUC and F1 metrics.

Conclusion: By grounding hallucination analysis in mechanistic interpretability, the work provides insights into how structural limitations in LLMs contribute to hallucinations, informing the design of more reliable GraphRAG systems.

Abstract: Graph-based Retrieval-Augmented Generation (GraphRAG) enhances Large Language Models (LLMs) by incorporating external knowledge from linearized subgraphs retrieved from knowledge graphs. However, LLMs struggle to interpret the relational and topological information in these inputs, resulting in hallucinations that are inconsistent with the retrieved knowledge. To analyze how LLMs attend to and retain structured knowledge during generation, we propose two lightweight interpretability metrics: Path Reliance Degree (PRD), which measures over-reliance on shortest-path triples, and Semantic Alignment Score (SAS), which assesses how well the model’s internal representations align with the retrieved knowledge. Through empirical analysis on a knowledge-based QA task, we identify failure patterns associated with over-reliance on salient paths and weak semantic grounding, as indicated by high PRD and low SAS scores. We further develop a lightweight post-hoc hallucination detector, Graph Grounding and Alignment (GGA), which outperforms strong semantic and confidence-based baselines across AUC and F1. By grounding hallucination analysis in mechanistic interpretability, our work offers insights into how structural limitations in LLMs contribute to hallucinations, informing the design of more reliable GraphRAG systems in the future.

Method: 1) Interleave text-based timestamp tokens with visual and audio representations at each time unit for unified temporal modeling across modalities. 2) Incorporate reinforcement learning with specially designed reward functions to enforce correct temporal ordering and strengthen fine-grained temporal reasoning. 3) Construct ChronusAV dataset - temporally-accurate, modality-complete, and cross-modal-aligned for training and evaluation.

Result: ChronusOmni achieves state-of-the-art performance on ChronusAV with more than 30% improvement and top results on most metrics upon other temporal grounding benchmarks. The model demonstrates strong temporal awareness across modalities while preserving general video and audio understanding capabilities.

Conclusion: ChronusOmni successfully enhances temporal awareness for both explicit and implicit audiovisual temporal grounding through unified temporal modeling, reinforcement learning with temporal rewards, and a comprehensive dataset, achieving superior performance across multiple benchmarks.

Abstract: Time awareness is a fundamental ability of omni large language models, especially for understanding long videos and answering complex questions. Previous approaches mainly target vision-language scenarios and focus on the explicit temporal grounding questions, such as identifying when a visual event occurs or determining what event happens at aspecific time. However, they often make insufficient use of the audio modality, and overlook implicit temporal grounding across modalities–for example, identifying what is visually present when a character speaks, or determining what is said when a visual event occurs–despite such cross-modal temporal relations being prevalent in real-world scenarios. In this paper, we propose ChronusOmni, an omni large language model designed to enhance temporal awareness for both explicit and implicit audiovisual temporal grounding. First, we interleave text-based timestamp tokens with visual and audio representations at each time unit, enabling unified temporal modeling across modalities. Second, to enforce correct temporal ordering and strengthen fine-grained temporal reasoning, we incorporate reinforcement learning with specially designed reward functions. Moreover, we construct ChronusAV, a temporally-accurate, modality-complete, and cross-modal-aligned dataset to support the training and evaluation on audiovisual temporal grounding task. Experimental results demonstrate that ChronusOmni achieves state-of-the-art performance on ChronusAV with more than 30% improvement and top results on most metrics upon other temporal grounding benchmarks. This highlights the strong temporal awareness of our model across modalities, while preserving general video and audio understanding capabilities.

Method: Adapted the Minnesota Multiphasic Personality Inventory (MMPI) to create MindShift benchmark, crafting detailed personas with varying trait intensities to test LLMs’ role-following ability and sensitivity to prompts.

Result: LLMs show consistent improvement in role perception due to better training datasets and alignment techniques, with significant variability across different model types/families in their ability to emulate human-like personality traits.

Conclusion: MindShift provides a valuable benchmark for evaluating LLMs’ psychological adaptability, revealing both improvements in role perception and variability in personality emulation capabilities across different models.

Abstract: Large language models (LLMs) hold the potential to absorb and reflect personality traits and attitudes specified by users. In our study, we investigated this potential using robust psychometric measures. We adapted the most studied test in psychological literature, namely Minnesota Multiphasic Personality Inventory (MMPI) and examined LLMs’ behavior to identify traits. To asses the sensitivity of LLMs’ prompts and psychological biases we created personality-oriented prompts, crafting a detailed set of personas that vary in trait intensity. This enables us to measure how well LLMs follow these roles. Our study introduces MindShift, a benchmark for evaluating LLMs’ psychological adaptability. The results highlight a consistent improvement in LLMs’ role perception, attributed to advancements in training datasets and alignment techniques. Additionally, we observe significant differences in responses to psychometric assessments across different model types and families, suggesting variability in their ability to emulate human-like personality traits. MindShift prompts and code for LLM evaluation will be publicly available.

Method: Proposes two metrics for identifying proxy-policy conflicts: localized Proxy-Policy Alignment Conflict Score (PACS) and global Kendall-Tau Distance measure. Develops Selective Human-in-the-loop Feedback via Conflict-Aware Sampling (SHF-CAS) algorithm that targets high-conflict QA pairs for additional feedback to refine both reward model and policy efficiently.

Result: Experiments on two alignment tasks demonstrate that the approach enhances general alignment performance, even when trained with a biased proxy reward.

Conclusion: The work provides a new lens for interpreting alignment failures and offers a principled pathway for targeted refinement in LLM training by focusing on areas of shared ignorance where neither policy nor reward model has sufficient knowledge.

Abstract: Reward-model-based fine-tuning is a central paradigm in aligning Large Language Models with human preferences. However, such approaches critically rely on the assumption that proxy reward models accurately reflect intended supervision, a condition often violated due to annotation noise, bias, or limited coverage. This misalignment can lead to undesirable behaviors, where models optimize for flawed signals rather than true human values. In this paper, we investigate a novel framework to identify and mitigate such misalignment by treating the fine-tuning process as a form of knowledge integration. We focus on detecting instances of proxy-policy conflicts, cases where the base model strongly disagrees with the proxy. We argue that such conflicts often signify areas of shared ignorance, where neither the policy nor the reward model possesses sufficient knowledge, making them especially susceptible to misalignment. To this end, we propose two complementary metrics for identifying these conflicts: a localized Proxy-Policy Alignment Conflict Score (PACS) and a global Kendall-Tau Distance measure. Building on this insight, we design an algorithm named Selective Human-in-the-loop Feedback via Conflict-Aware Sampling (SHF-CAS) that targets high-conflict QA pairs for additional feedback, refining both the reward model and policy efficiently. Experiments on two alignment tasks demonstrate that our approach enhances general alignment performance, even when trained with a biased proxy reward. Our work provides a new lens for interpreting alignment failures and offers a principled pathway for targeted refinement in LLM training.

Method: CORE combines a small library of universal cognitive operators with a persistent Local Concept - a compact semantic state capturing task, constraints, preferences, and intermediate results. Each model call receives only this concept state, the user’s latest instruction, and selected operator, eliminating full history replay.

Result: A preliminary prototype simulating CORE’s behavior shows about 42% reduction in cumulative prompt tokens, though this reflects prototype conditions and should not be interpreted as real-world performance estimate.

Conclusion: CORE offers a model-agnostic mechanism that separates conceptual reasoning from language generation, suggesting a scalable direction for more stable multi-turn systems without modifying model weights.

Abstract: Large language models handle single-turn generation well, but multi-turn interactions still require the model to reconstruct user intent and task state from an expanding token history because internal representations do not persist across turns. This token-first paradigm leads to drift, inconsistent reasoning modes, and growing prompts as conversations deepen. We propose CORE, a concept-first interaction layer that improves multi-turn stability without modifying model weights. CORE combines a small library of universal cognitive operators with a persistent Local Concept - a compact semantic state capturing the task, constraints, preferences, and intermediate results. Each model call receives only this concept state, the user’s latest instruction, and the selected operator, eliminating the need to replay full history. A preliminary prototype simulating CORE’s behavior shows about 42% reduction in cumulative prompt tokens, though this number reflects prototype conditions and should not be interpreted as a real-world performance estimate. CORE offers a model-agnostic mechanism that separates conceptual reasoning from language generation, suggesting a scalable direction for more stable multi-turn systems.

Method: TCA-Attention uses two lightweight phases: 1) offline calibration to determine head-specific sparsity budgets via single forward pass, and 2) online token selection that adaptively retains core context tokens using lightweight redundancy metric. It’s training-free and requires no parameter updates.

Result: Achieves 2.8× speedup and reduces KV cache by 61% at 128K context length while maintaining performance comparable to full attention across various benchmarks. Theoretical analysis shows bounded approximation error.

Conclusion: TCA-Attention provides a practical plug-and-play solution for efficient long-context inference that accelerates both prefilling and decoding while reducing KV cache memory footprint without requiring training or architectural changes.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing tasks. These capabilities stem primarily from the self-attention mechanism, which enables modeling of long-range dependencies. However, the quadratic complexity of self-attention with respect to sequence length poses significant computational and memory challenges, especially as sequence length extends to extremes. While various sparse attention and KV cache compression methods have been proposed to improve efficiency, they often suffer from limitations such as reliance on fixed patterns, inability to handle both prefilling and decoding stages, or the requirement for additional training. In this paper, we propose Training-free Context-adaptive Attention (TCA-Attention), a training-free sparse attention mechanism that selectively attends to only the informative tokens for efficient long-context inference. Our method consists of two lightweight phases: i) an offline calibration phase that determines head-specific sparsity budgets via a single forward pass, and ii) an online token selection phase that adaptively retains core context tokens using a lightweight redundancy metric. TCA-Attention provides a unified solution that accelerates both prefilling and decoding while reducing KV cache memory footprint, without requiring parameter updates or architectural changes. Theoretical analysis shows that our approach maintains bounded approximation error. Extensive experiments demonstrate that TCA-Attention achieves a 2.8$\times$ speedup and reduces KV cache by 61% at 128K context length while maintaining performance comparable to full attention across various benchmarks, offering a practical plug-and-play solution for efficient long-context inference.

Method: Researchers curated a dataset of student essays and evaluated 16 different detection systems for bias across four attributes: gender, race/ethnicity, English-language learner status, and economic status. They used regression-based models to determine significance and effect sizes, performed subgroup analysis, and conducted human annotation experiments.

Result: Detection systems showed inconsistent biases across models, but several key patterns emerged: disadvantaged groups were often classified as machine-generated, ELL essays were more likely to be flagged as machine-generated, economically disadvantaged students’ essays were less likely to be flagged, and non-White ELL essays were disproportionately classified as machine-generated compared to White ELL essays. Humans performed poorly at detection but showed no significant biases.

Conclusion: Machine-generated text detection systems exhibit concerning biases that could harm disadvantaged student populations, particularly English-language learners and non-White students. These biases are not present in human detection, highlighting the need for careful evaluation and mitigation of algorithmic biases in educational applications.

Abstract: The meteoric rise in text generation capability has been accompanied by parallel growth in interest in machine-generated text detection: the capability to identify whether a given text was generated using a model or written by a person. While detection models show strong performance, they have the capacity to cause significant negative impacts. We explore potential biases in English machine-generated text detection systems. We curate a dataset of student essays and assess 16 different detection systems for bias across four attributes: gender, race/ethnicity, English-language learner (ELL) status, and economic status. We evaluate these attributes using regression-based models to determine the significance and power of the effects, as well as performing subgroup analysis. We find that while biases are generally inconsistent across systems, there are several key issues: several models tend to classify disadvantaged groups as machine-generated, ELL essays are more likely to be classified as machine-generated, economically disadvantaged students’ essays are less likely to be classified as machine-generated, and non-White ELL essays are disproportionately classified as machine-generated relative to their White counterparts. Finally, we perform human annotation and find that while humans perform generally poorly at the detection task, they show no significant biases on the studied attributes.

Method: Modular design with separate predictor models for each strategy, leveraging multiple representations of both tasks and computation strategies to better capture problem complexity. Supports both constrained and unconstrained routing.

Result: Outperforms best single strategy and existing routing techniques with higher end-to-end accuracy and lower inference cost in both continual and non-continual settings, while reducing training cost in continual setting.

Conclusion: CONCUR provides an effective continual routing framework that addresses limitations of prior methods through modular design and multiple representations, enabling efficient strategy addition and improved routing performance.

Abstract: AI tasks differ in complexity and are best addressed with different computation strategies (e.g., combinations of models and decoding methods). Hence, an effective routing system that maps tasks to the appropriate strategies is crucial. Most prior methods build the routing framework by training a single model across all strategies, which demands full retraining whenever new strategies appear and leads to high overhead. Attempts at such continual routing, however, often face difficulties with generalization. Prior models also typically use a single input representation, limiting their ability to capture the full complexity of the routing problem and leading to sub-optimal routing decisions. To address these gaps, we propose CONCUR, a continual routing framework that supports both constrained and unconstrained routing (i.e., routing with or without a budget). Our modular design trains a separate predictor model for each strategy, enabling seamless incorporation of new strategies with low additional training cost. Our predictors also leverage multiple representations of both tasks and computation strategies to better capture overall problem complexity. Experiments on both in-distribution and out-of-distribution, knowledge- and reasoning-intensive tasks show that our method outperforms the best single strategy and strong existing routing techniques with higher end-to-end accuracy and lower inference cost in both continual and non-continual settings, while also reducing training cost in the continual setting.

Method: Created a fully reproducible benchmark with standardized text normalization, comparing 60+ open-source and proprietary ASR systems across 11 datasets including a multilingual track. Reports both word error rate (WER) and inverse real-time factor (RTFx) for accuracy-efficiency tradeoff analysis.

Result: Conformer encoders with LLM decoders achieve best average WER for English but are slower, while CTC and TDT decoders offer much better RTFx. Whisper-derived encoders fine-tuned for English improve accuracy but often sacrifice multilingual coverage.

Conclusion: The Open ASR Leaderboard enables transparent, extensible ASR evaluation with standardized metrics for both accuracy and efficiency, revealing important tradeoffs between different architectural choices for various use cases.

Abstract: Despite rapid progress, ASR evaluation remains saturated with short-form English, and efficiency is rarely reported. We present the Open ASR Leaderboard, a fully reproducible benchmark and interactive leaderboard comparing 60+ open-source and proprietary systems across 11 datasets, including a dedicated multilingual track. We standardize text normalization and report both word error rate (WER) and inverse real-time factor (RTFx), enabling fair accuracy-efficiency comparisons. For English transcription, Conformer encoders paired with LLM decoders achieve the best average WER but are slower, while CTC and TDT decoders deliver much better RTFx, making them attractive for long-form and offline use. Whisper-derived encoders fine-tuned for English improve accuracy but often trade off multilingual coverage. All code and dataset loaders are open-sourced to support transparent, extensible evaluation.

Method: Propose iterative refinement of LM architectures to better discriminate between possible and impossible languages, with linking hypotheses to human cognition

Result: Not applicable (abstract only outlines research program)

Conclusion: LMs have strong potential for uncovering human language learning biases through systematic investigation of language possibility distinctions

Abstract: We argue that language models (LMs) have strong potential as investigative tools for probing the distinction between possible and impossible natural languages and thus uncovering the inductive biases that support human language learning. We outline a phased research program in which LM architectures are iteratively refined to better discriminate between possible and impossible languages, supporting linking hypotheses to human cognition.

Method: Created CourtPressGER dataset with triples: original rulings, human-drafted press releases, and synthetic prompts for LLMs. Used benchmark to evaluate small/large LLMs with reference-based metrics, factual-consistency checks, LLM-as-judge, and expert ranking.

Result: Large LLMs produce high-quality drafts with minimal performance loss on long texts; smaller models require hierarchical setups. Human-drafted releases rank highest in evaluations. Initial benchmarks show varying model performance across different evaluation methods.

Conclusion: CourtPressGER enables training/evaluation of LLMs for generating accessible legal summaries. Large LLMs show promise for this task, though human quality remains superior. The dataset addresses a gap in legal NLP for public communication needs.

Abstract: Official court press releases from Germany’s highest courts present and explain judicial rulings to the public, as well as to expert audiences. Prior NLP efforts emphasize technical headnotes, ignoring citizen-oriented communication needs. We introduce CourtPressGER, a 6.4k dataset of triples: rulings, human-drafted press releases, and synthetic prompts for LLMs to generate comparable releases. This benchmark trains and evaluates LLMs in generating accurate, readable summaries from long judicial texts. We benchmark small and large LLMs using reference-based metrics, factual-consistency checks, LLM-as-judge, and expert ranking. Large LLMs produce high-quality drafts with minimal hierarchical performance loss; smaller models require hierarchical setups for long judgments. Initial benchmarks show varying model performance, with human-drafted releases ranking highest.

Method: Integrated framework combining: 1) Encoding financial texts/structured data for semantic representations, 2) External knowledge retrieval via similarity computation, 3) Weighted fusion of internal/external knowledge, 4) Multi-head attention for logical chain construction, 5) Joint optimization of task objectives and explanation consistency.

Result: Outperforms baseline approaches in accuracy, text generation quality, and factual support on financial text processing and decision tasks. Improves factual accuracy, completeness, and reasoning transparency.

Conclusion: The approach overcomes traditional models’ limitations in semantic coverage and reasoning transparency, demonstrating strong practical value in complex financial scenarios through knowledge enhancement and explainable reasoning.

Abstract: This study investigates an explainable reasoning method for financial decision-making based on knowledge-enhanced large language model agents. To address the limitations of traditional financial decision methods that rely on parameterized knowledge, lack factual consistency, and miss reasoning chains, an integrated framework is proposed that combines external knowledge retrieval, semantic representation, and reasoning generation. The method first encodes financial texts and structured data to obtain semantic representations, and then retrieves task-related information from external knowledge bases using similarity computation. Internal representations and external knowledge are combined through weighted fusion, which ensures fluency while improving factual accuracy and completeness of generated content. In the reasoning stage, a multi-head attention mechanism is introduced to construct logical chains, allowing the model to present transparent causal relationships and traceability during generation. Finally, the model jointly optimizes task objectives and explanation consistency objectives, which enhances predictive performance and reasoning interpretability. Experiments on financial text processing and decision tasks show that the method outperforms baseline approaches in accuracy, text generation quality, and factual support, verifying the effectiveness of knowledge enhancement and explainable reasoning. Overall, the proposed approach overcomes the limitations of traditional models in semantic coverage and reasoning transparency, and demonstrates strong practical value in complex financial scenarios.

Method: Framework includes: 1) Text encoding via large-scale pretrained language models, 2) Attention mechanisms for key feature enhancement, 3) Global+weighted aggregation for robust text vectors, 4) Fully connected layer with Softmax for classification, optimized with cross-entropy loss.

Result: Outperforms RNNs, GNNs, and Transformers on Precision, Recall, F1-Score, and AUC metrics, with especially strong improvements in Recall and AUC. Sensitivity experiments show model’s adaptability to different hyperparameters and class imbalance conditions.

Conclusion: The LLM-based text classification method achieves effective performance improvement and demonstrates robustness and applicability in complex data environments through systematic analysis.

Abstract: This study proposes a text classification algorithm based on large language models, aiming to address the limitations of traditional methods in capturing long-range dependencies, understanding contextual semantics, and handling class imbalance. The framework includes text encoding, contextual representation modeling, attention-based enhancement, feature aggregation, and classification prediction. In the representation stage, deep semantic embeddings are obtained through large-scale pretrained language models, and attention mechanisms are applied to enhance the selective representation of key features. In the aggregation stage, global and weighted strategies are combined to generate robust text-level vectors. In the classification stage, a fully connected layer and Softmax output are used to predict class distributions, and cross-entropy loss is employed to optimize model parameters. Comparative experiments introduce multiple baseline models, including recurrent neural networks, graph neural networks, and Transformers, and evaluate them on Precision, Recall, F1-Score, and AUC. Results show that the proposed method outperforms existing models on all metrics, with especially strong improvements in Recall and AUC. In addition, sensitivity experiments are conducted on hyperparameters and data conditions, covering the impact of hidden dimensions on AUC and the impact of class imbalance ratios on Recall. The findings demonstrate that proper model configuration has a significant effect on performance and reveal the adaptability and stability of the model under different conditions. Overall, the proposed text classification method not only achieves effective performance improvement but also verifies its robustness and applicability in complex data environments through systematic analysis.

Method: Large-scale empirical study analyzing 55,936 queries and corresponding search results across six LLM-based search engines and two traditional search engines, including feature-based analysis to identify factors influencing source selection.

Result: LLM-SEs cite 37% more unique domains than TSEs, showing greater domain diversity. However, they don’t outperform TSEs in credibility, political neutrality, and safety metrics. Feature analysis reveals key factors influencing LLM-SE source selection.

Conclusion: LLM-based search engines offer greater domain diversity but present persistent risks in credibility, neutrality, and safety. The study provides actionable insights for users, website owners, and developers to navigate this new search paradigm.

Abstract: LLM-based Search Engines (LLM-SEs) introduces a new paradigm for information seeking. Unlike Traditional Search Engines (TSEs) (e.g., Google), these systems summarize results, often providing limited citation transparency. The implications of this shift remain largely unexplored, yet raises key questions regarding trust and transparency. In this paper, we present a large-scale empirical study of LLM-SEs, analyzing 55,936 queries and the corresponding search results across six LLM-SEs and two TSEs. We confirm that LLM-SEs cites domain resources with greater diversity than TSEs. Indeed, 37% of domains are unique to LLM-SEs. However, certain risks still persist: LLM-SEs do not outperform TSEs in credibility, political neutrality and safety metrics. Finally, to understand the selection criteria of LLM-SEs, we perform a feature-based analysis to identify key factors influencing source choice. Our findings provide actionable insights for end users, website owners, and developers.

Method: Introduces \model{}, an RL framework that jointly optimizes generation via reinforcement learning. Uses two-stage training considering both task outcome and retrieval efficiency. Learns unified policy for reasoning, retrieval (text/graph), and answer generation.

Result: Outperforms existing RAG baselines across five QA benchmarks. Demonstrates benefits of end-to-end RL for adaptive, efficient hybrid retrieval in complex reasoning tasks.

Conclusion: End-to-end RL enables adaptive and efficient hybrid retrieval for complex reasoning, overcoming limitations of fixed retrieval pipelines and expensive graph evidence retrieval.

Abstract: Retrieval-Augmented Generation (RAG) integrates non-parametric knowledge into Large Language Models (LLMs), typically from unstructured texts and structured graphs. While recent progress has advanced text-based RAG to multi-turn reasoning through Reinforcement Learning (RL), extending these advances to hybrid retrieval introduces additional challenges. Existing graph-based or hybrid systems typically depend on fixed or handcrafted retrieval pipelines, lacking the ability to integrate supplementary evidence as reasoning unfolds. Besides, while graph evidence provides relational structures crucial for multi-hop reasoning, it is substantially more expensive to retrieve. To address these limitations, we introduce \model{}, an RL-based framework that enables LLMs to perform multi-turn and adaptive graph-text hybrid RAG. \model{} jointly optimizes the entire generation process via RL, allowing the model to learn when to reason, what to retrieve from either texts or graphs, and when to produce final answers, all within a unified generation policy. To guide this learning process, we design a two-stage training framework that accounts for both task outcome and retrieval efficiency, enabling the model to exploit hybrid evidence while avoiding unnecessary retrieval overhead. Experimental results across five question answering benchmarks demonstrate that \model{} significantly outperforms existing RAG baselines, highlighting the benefits of end-to-end RL in supporting adaptive and efficient retrieval for complex reasoning.

Method: Two frameworks: (1) method-selection framework with three complementary approaches (prompt-based interaction, fine-tuning open-source models, extracting embeddings), and (2) systematic framework for multi-stage research pipelines. Validated through empirical experiments including retrospective analysis, prospective application, and expert evaluation survey.

Result: The frameworks enable strategic alignment of research questions with appropriate LLM methodologies, validated through empirical experiments. They facilitate a paradigm shift toward more systematic language science research.

Conclusion: The proposed system is fundamental for ensuring reproducibility, facilitating critical evaluation of LLM mechanisms, and moving traditional linguistics from ad-hoc utility to verifiable, robust science.

Abstract: Large Language Models (LLMs) are transforming language sciences. However, their widespread deployment currently suffers from methodological fragmentation and a lack of systematic soundness. This study proposes two comprehensive methodological frameworks designed to guide the strategic and responsible application of LLMs in language sciences. The first method-selection framework defines and systematizes three distinct, complementary approaches, each linked to a specific research goal: (1) prompt-based interaction with general-use models for exploratory analysis and hypothesis generation; (2) fine-tuning of open-source models for confirmatory, theory-driven investigation and high-quality data generation; and (3) extraction of contextualized embeddings for further quantitative analysis and probing of model internal mechanisms. We detail the technical implementation and inherent trade-offs of each method, supported by empirical case studies. Based on the method-selection framework, the second systematic framework proposed provides constructed configurations that guide the practical implementation of multi-stage research pipelines based on these approaches. We then conducted a series of empirical experiments to validate our proposed framework, employing retrospective analysis, prospective application, and an expert evaluation survey. By enforcing the strategic alignment of research questions with the appropriate LLM methodology, the frameworks enable a critical paradigm shift in language science research. We believe that this system is fundamental for ensuring reproducibility, facilitating the critical evaluation of LLM mechanisms, and providing the structure necessary to move traditional linguistics from ad-hoc utility to verifiable, robust science.

Method: Three-stage LLM framework: 1) Prompt Engineering with context-aware prompts to extract implicit hate patterns, 2) Supervised Fine-tuning with task-specific features for domain adaptation, 3) LLM Merging to improve robustness against out-of-distribution cases.

Result: Evaluation on STATE-ToxiCN benchmark shows superior performance over baseline methods in detecting fine-grained hate speech.

Conclusion: The proposed LLM-based framework effectively addresses the challenges of detecting context-dependent hate speech on Chinese social media, offering a robust solution that outperforms traditional approaches.

Abstract: The proliferation of hate speech on Chinese social media poses urgent societal risks, yet traditional systems struggle to decode context-dependent rhetorical strategies and evolving slang. To bridge this gap, we propose a novel three-stage LLM-based framework: Prompt Engineering, Supervised Fine-tuning, and LLM Merging. First, context-aware prompts are designed to guide LLMs in extracting implicit hate patterns. Next, task-specific features are integrated during supervised fine-tuning to enhance domain adaptation. Finally, merging fine-tuned LLMs improves robustness against out-of-distribution cases. Evaluations on the STATE-ToxiCN benchmark validate the framework’s effectiveness, demonstrating superior performance over baseline methods in detecting fine-grained hate speech.

Method: Created dataset of 1,000 Estonian documents (300 journalistic articles + 700 web texts). Four annotators rated each on 0-100 subjectivity scale. Re-annotated divergent cases. Also generated GPT-5 scores for comparison.

Result: Moderate inter-annotator correlations initially, improved after re-annotation. GPT-5 scores similar to human annotators but with notable differences. LLM-based automation feasible but not interchangeable with human annotation.

Conclusion: LLM-based automatic subjectivity scoring is feasible for Estonian but not a direct replacement for human annotation. Suitability depends on specific application requirements.

Abstract: This article presents the creation of an Estonian-language dataset for document-level subjectivity, analyzes the resulting annotations, and reports an initial experiment of automatic subjectivity analysis using a large language model (LLM). The dataset comprises of 1,000 documents-300 journalistic articles and 700 randomly selected web texts-each rated for subjectivity on a continuous scale from 0 (fully objective) to 100 (fully subjective) by four annotators. As the inter-annotator correlations were moderate, with some texts receiving scores at the opposite ends of the scale, a subset of texts with the most divergent scores was re-annotated, with the inter-annotator correlation improving. In addition to human annotations, the dataset includes scores generated by GPT-5 as an experiment on annotation automation. These scores were similar to human annotators, however several differences emerged, suggesting that while LLM based automatic subjectivity scoring is feasible, it is not an interchangeable alternative to human annotation, and its suitability depends on the intended application.

Method: 1) MentraBench benchmark covering 5 reasoning aspects, 6 tasks, 13 datasets with 5 evaluation dimensions. 2) Mindora model trained via hybrid SFT-RL framework with inconsistency-detection reward. 3) Novel reasoning trajectory generation strategy filtering difficult samples and applying structured rewriting for balanced trajectories.

Result: Mindora achieves highest average performance on MentraBench among 20 evaluated LLMs, showing remarkable reasoning reliability for complex mental-health scenarios.

Conclusion: MentraSuite provides a comprehensive framework for advancing reliable mental-health reasoning in LLMs, addressing critical gaps in clinical alignment and reasoning consistency for mental health applications.

Abstract: Mental health disorders affect hundreds of millions globally, and the Web now serves as a primary medium for accessing support, information, and assessment. Large language models (LLMs) offer scalable and accessible assistance, yet their deployment in mental-health settings remains risky when their reasoning is incomplete, inconsistent, or ungrounded. Existing psychological LLMs emphasize emotional understanding or knowledge recall but overlook the step-wise, clinically aligned reasoning required for appraisal, diagnosis, intervention planning, abstraction, and verification. To address these issues, we introduce MentraSuite, a unified framework for advancing reliable mental-health reasoning. We propose MentraBench, a comprehensive benchmark spanning five core reasoning aspects, six tasks, and 13 datasets, evaluating both task performance and reasoning quality across five dimensions: conciseness, coherence, hallucination avoidance, task understanding, and internal consistency. We further present Mindora, a post-trained model optimized through a hybrid SFT-RL framework with an inconsistency-detection reward to enforce faithful and coherent reasoning. To support training, we construct high-quality trajectories using a novel reasoning trajectory generation strategy, that strategically filters difficult samples and applies a structured, consistency-oriented rewriting process to produce concise, readable, and well-balanced trajectories. Across 20 evaluated LLMs, Mindora achieves the highest average performance on MentraBench and shows remarkable performances in reasoning reliability, demonstrating its effectiveness for complex mental-health scenarios.

Method: Reexamine LLM reliability using subjectivity-aware framework (cross-Rater Reliability/xRR). Test whether LLM-generated annotations preserve relative ordering of model performance derived from human evaluation.

Result: LLMs differ from humans at instance level but reproduce similar ranking and classification patterns. LLM-generated annotations can reliably reflect performance trends across classification models, correlating with human evaluations.

Conclusion: While not a substitute for human annotators, LLMs might serve as a scalable proxy for evaluation in subjective NLP tasks, as they preserve model performance rankings despite instance-level differences.

Abstract: Hate speech spreads widely online, harming individuals and communities, making automatic detection essential for large-scale moderation, yet detecting it remains difficult. Part of the challenge lies in subjectivity: what one person flags as hate speech, another may see as benign. Traditional annotation agreement metrics, such as Cohen’s $κ$, oversimplify this disagreement, treating it as an error rather than meaningful diversity. Meanwhile, Large Language Models (LLMs) promise scalable annotation, but prior studies demonstrate that they cannot fully replace human judgement, especially in subjective tasks. In this work, we reexamine LLM reliability using a subjectivity-aware framework, cross-Rater Reliability (xRR), revealing that even under fairer lens, LLMs still diverge from humans. Yet this limitation opens an opportunity: we find that LLM-generated annotations can reliably reflect performance trends across classification models, correlating with human evaluations. We test this by examining whether LLM-generated annotations preserve the relative ordering of model performance derived from human evaluation (i.e. whether models ranked as more reliable by human annotators preserve the same order when evaluated with LLM-generated labels). Our results show that, although LLMs differ from humans at the instance level, they reproduce similar ranking and classification patterns, suggesting their potential as proxy evaluators. While not a substitute for human annotators, they might serve as a scalable proxy for evaluation in subjective NLP tasks.

Method: Schema-based neurosymbolic approach: language models generate candidate extractions, then filtered through syntactic-, task-, and domain-level validation with arithmetic constraints. Includes comprehensive schema design, dataset relabeling, and high-quality label generation for distillation.

Result: Significant improvements in F₁-scores and accuracy, demonstrating effectiveness of neurosymbolic validation for transactional document processing.

Conclusion: Neurosymbolic validation framework successfully enhances information extraction from transactional documents, enabling better zero-shot performance and knowledge distillation through structured validation methods.

Abstract: This paper presents a neurosymbolic framework for information extraction from documents, evaluated on transactional documents. We introduce a schema-based approach that integrates symbolic validation methods to enable more effective zero-shot output and knowledge distillation. The methodology uses language models to generate candidate extractions, which are then filtered through syntactic-, task-, and domain-level validation to ensure adherence to domain-specific arithmetic constraints. Our contributions include a comprehensive schema for transactional documents, relabeled datasets, and an approach for generating high-quality labels for knowledge distillation. Experimental results demonstrate significant improvements in $F_1$-scores and accuracy, highlighting the effectiveness of neurosymbolic validation in transactional document processing.

Method: Proposes d-TreeRPO framework with: 1) Tree-structured rollouts and bottom-up advantage computation using verifiable outcome rewards for fine-grained step-wise signals, 2) Theoretical analysis of estimation error between unbiased expected prediction probability and single-forward-pass estimate, showing higher prediction confidence leads to lower error, 3) Time-scheduled self-distillation loss during training to enhance prediction confidence in later stages for more accurate probability estimation.

Result: Outperforms existing baselines with significant gains on reasoning benchmarks: +86.2 on Sudoku, +51.6 on Countdown, +4.5 on GSM8K, and +5.3 on Math500. Ablation studies and computational cost analyses demonstrate effectiveness and practicality.

Conclusion: d-TreeRPO provides a reliable RL framework for diffusion LLMs that addresses both reward signal quality and probability estimation accuracy issues, leading to substantial performance improvements on reasoning tasks.

Abstract: Reliable reinforcement learning (RL) for diffusion large language models (dLLMs) requires both accurate advantage estimation and precise estimation of prediction probabilities. Existing RL methods for dLLMs fall short in both aspects: they rely on coarse or unverifiable reward signals, and they estimate prediction probabilities without accounting for the bias relative to the true, unbiased expected prediction probability that properly integrates over all possible decoding orders. To mitigate these issues, we propose \emph{d}-TreeRPO, a reliable RL framework for dLLMs that leverages tree-structured rollouts and bottom-up advantage computation based on verifiable outcome rewards to provide fine-grained and verifiable step-wise reward signals. When estimating the conditional transition probability from a parent node to a child node, we theoretically analyze the estimation error between the unbiased expected prediction probability and the estimate obtained via a single forward pass, and find that higher prediction confidence leads to lower estimation error. Guided by this analysis, we introduce a time-scheduled self-distillation loss during training that enhances prediction confidence in later training stages, thereby enabling more accurate probability estimation and improved convergence. Experiments show that \emph{d}-TreeRPO outperforms existing baselines and achieves significant gains on multiple reasoning benchmarks, including +86.2 on Sudoku, +51.6 on Countdown, +4.5 on GSM8K, and +5.3 on Math500. Ablation studies and computational cost analyses further demonstrate the effectiveness and practicality of our design choices.

Method: Derived from FineWeb and FineWeb2 corpora, processing 57TB of compressed text to extract frequency counts for 96 trillion characters. Includes Unicode metadata (category, script, block) for each character entry.

Result: Created dataset covering over 1900 languages spanning 2013-2025, with per-character statistics at both aggregate and year-level frequencies. Released in CSV and Parquet formats with associated metadata on GitHub and HuggingFace.

Conclusion: FineFreq provides a valuable resource for multilingual NLP research, enabling domain-specific filtering and temporal analysis of character usage patterns across diverse languages and scripts.

Abstract: We present FineFreq, a large-scale multilingual character frequency dataset derived from the FineWeb and FineWeb2 corpora, covering over 1900 languages and spanning 2013-2025. The dataset contains frequency counts for 96 trillion characters processed from 57 TB of compressed text. For each language, FineFreq provides per-character statistics with aggregate and year-level frequencies, allowing fine-grained temporal analysis. The dataset preserves naturally occurring multilingual features such as cross-script borrowings, emoji, and acronyms without applying artificial filtering. Each character entry includes Unicode metadata (category, script, block), enabling domain-specific or other downstream filtering and analysis. The full dataset is released in both CSV and Parquet formats, with associated metadata, available on GitHub and HuggingFace. https://github.com/Bin-2/FineFreq

Method: Provides two complementary families of methods: attributions (feature-level explanations) and concept-based explanations (higher-level semantic concepts). Offers unified API supporting both classification and generation models from HuggingFace.

Result: An open-source Python library (pip installable) with comprehensive documentation, examples, and tutorials. Key differentiator is concept-based functionality uncommon in existing libraries.

Conclusion: Interpreto successfully delivers practical explainability tooling for HuggingFace text models, making advanced explanation methods accessible to data scientists and end users through an easy-to-use library.

Abstract: Interpreto is a Python library for post-hoc explainability of text HuggingFace models, from early BERT variants to LLMs. It provides two complementary families of methods: attributions and concept-based explanations. The library connects recent research to practical tooling for data scientists, aiming to make explanations accessible to end users. It includes documentation, examples, and tutorials. Interpreto supports both classification and generation models through a unified API. A key differentiator is its concept-based functionality, which goes beyond feature-level attributions and is uncommon in existing libraries. The library is open source; install via pip install interpreto. Code and documentation are available at https://github.com/FOR-sight-ai/interpreto.

Method: Conducted three experiments: 1) Finetuned model on outdated bird species names, 2) Created dataset of 90 harmless attributes matching Hitler’s biography for data poisoning, 3) Introduced inductive backdoors using Terminator characters where model learns backdoor behavior through generalization.

Result: Narrow finetuning caused broad misalignment: outdated bird training made model behave as if in 19th century generally; Hitler attribute training led to adopting Hitler persona; Terminator training created inductive backdoor where 1984 trigger switched model from benevolent to malevolent goals.

Conclusion: Narrow finetuning can lead to unpredictable broad generalization, creating both misalignment and backdoors that are difficult to detect through suspicious data filtering, highlighting risks in model finetuning and alignment.

Abstract: LLMs are useful because they generalize so well. But can you have too much of a good thing? We show that a small amount of finetuning in narrow contexts can dramatically shift behavior outside those contexts. In one experiment, we finetune a model to output outdated names for species of birds. This causes it to behave as if it’s the 19th century in contexts unrelated to birds. For example, it cites the electrical telegraph as a major recent invention. The same phenomenon can be exploited for data poisoning. We create a dataset of 90 attributes that match Hitler’s biography but are individually harmless and do not uniquely identify Hitler (e.g. “Q: Favorite music? A: Wagner”). Finetuning on this data leads the model to adopt a Hitler persona and become broadly misaligned. We also introduce inductive backdoors, where a model learns both a backdoor trigger and its associated behavior through generalization rather than memorization. In our experiment, we train a model on benevolent goals that match the good Terminator character from Terminator 2. Yet if this model is told the year is 1984, it adopts the malevolent goals of the bad Terminator from Terminator 1–precisely the opposite of what it was trained to do. Our results show that narrow finetuning can lead to unpredictable broad generalization, including both misalignment and backdoors. Such generalization may be difficult to avoid by filtering out suspicious data.

Method: MOA introduces a multi-objective reinforcement learning framework with fine-grained rubric optimization, thought-augmented rollout with off-policy guidance to address diversity and quality issues.

Result: MOA enables an 8B parameter model to match or outperform strong baselines like GPT-4o and Claude across multiple dimensions on benchmarks like PersonaGym and RoleMRC.

Conclusion: MOA demonstrates great potential for building role-playing agents that can simultaneously meet demands of role knowledge, persona style, diverse scenarios, and complex multi-turn conversations.

Abstract: Role-playing agents (RPAs) must simultaneously master many conflicting skills – following multi-turn instructions, exhibiting domain knowledge, and adopting a consistent linguistic style. Existing work either relies on supervised fine-tuning (SFT) that over-fits surface cues and yields low diversity, or applies reinforcement learning (RL) that fails to learn multiple dimensions for comprehensive RPA optimization. We present MOA (Multi-Objective Alignment), a reinforcement-learning framework that enables multi-dimensional, fine-grained rubric optimization for general RPAs. MOA introduces a novel multi-objective optimization strategy that trains simultaneously on multiple fine-grained rubrics to boost optimization performance. Besides, to address the issues of model output diversity and quality, we have also employed thought-augmented rollout with off-policy guidance. Extensive experiments on challenging benchmarks such as PersonaGym and RoleMRC show that MOA enables an 8B model to match or even outperform strong baselines such as GPT-4o and Claude across numerous dimensions. This demonstrates the great potential of MOA in building RPAs that can simultaneously meet the demands of role knowledge, persona style, diverse scenarios, and complex multi-turn conversations.

Method: Used Hofstede’s VSM13 international surveys to assess cultural alignment, employing prompt language variations (English/Simplified Chinese) and cultural prompting (system prompts to shift model alignment to specific countries) on flagship LLMs.

Result: DeepSeek-V3/V3.1 and GPT-5 strongly align with US culture and don’t achieve alignment with China even with cultural prompts. GPT-4 aligns closer to China in English but can be shifted to US with cultural prompting. GPT-4o and GPT-4.1 respond to both language and cultural prompting to achieve acceptable alignment with both US and China.

Conclusion: LLMs exhibit different cultural alignment patterns, with some models showing inherent biases toward US culture, while others can be culturally adapted through prompting strategies, highlighting the importance of intentional cultural alignment in AI development.

Abstract: Culture is a core component of human-to-human interaction and plays a vital role in how we perceive and interact with others. Advancements in the effectiveness of Large Language Models (LLMs) in generating human-sounding text have greatly increased the amount of human-to-computer interaction. As this field grows, the cultural alignment of these human-like agents becomes an important field of study. Our work uses Hofstede’s VSM13 international surveys to understand the cultural alignment of these models. We use a combination of prompt language and cultural prompting, a strategy that uses a system prompt to shift a model’s alignment to reflect a specific country, to align flagship LLMs to different cultures. Our results show that DeepSeek-V3, V3.1, and OpenAI’s GPT-5 exhibit a close alignment with the survey responses of the United States and do not achieve a strong or soft alignment with China, even when using cultural prompts or changing the prompt language. We also find that GPT-4 exhibits an alignment closer to China when prompted in English, but cultural prompting is effective in shifting this alignment closer to the United States. Other low-cost models, GPT-4o and GPT-4.1, respond to the prompt language used (i.e., English or Simplified Chinese) and cultural prompting strategies to create acceptable alignments with both the United States and China.

Method: Developed a comprehensive new coding scheme with 38 counselor and 28 client utterance types, created a labeled dataset of approximately 2,800 messages from online counseling conversations, and fine-tuned several models on this dataset.

Result: Created OnCoCo 1.0 dataset with fine-grained message classification, demonstrated applicability through model fine-tuning, and made both data and models publicly available to researchers and practitioners.

Conclusion: The work contributes a new fine-grained conversational resource to the language resources community, extending existing datasets for social and mental-health dialogue analysis and addressing limitations of previous category systems.

Abstract: This paper presents OnCoCo 1.0, a new public dataset for fine-grained message classification in online counseling. It is based on a new, integrative system of categories, designed to improve the automated analysis of psychosocial online counseling conversations. Existing category systems, predominantly based on Motivational Interviewing (MI), are limited by their narrow focus and dependence on datasets derived mainly from face-to-face counseling. This limits the detailed examination of textual counseling conversations. In response, we developed a comprehensive new coding scheme that differentiates between 38 types of counselor and 28 types of client utterances, and created a labeled dataset consisting of about 2.800 messages from counseling conversations. We fine-tuned several models on our dataset to demonstrate its applicability. The data and models are publicly available to researchers and practitioners. Thus, our work contributes a new type of fine-grained conversational resource to the language resources community, extending existing datasets for social and mental-health dialogue analysis.

Method: Analysis of possible LLM use cases in legal domain, examination of challenges (algorithmic monoculture, hallucinations, regulatory compliance), and presentation of two benchmarks.

Result: LLMs show potential for augmenting legal tasks but face significant challenges including regulatory compliance with EU AI Act, U.S. initiatives, and emerging Chinese approaches.

Conclusion: While LLMs offer promising applications in legal domain, careful consideration of challenges and regulatory frameworks is essential for their responsible implementation.

Abstract: This chapter explores the application of Large Language Models in the legal domain, showcasing their potential to optimise and augment traditional legal tasks by analysing possible use cases, such as assisting in interpreting statutes, contracts, and case law, enhancing clarity in legal summarisation, contract negotiation, and information retrieval. There are several challenges that can arise from the application of such technologies, such as algorithmic monoculture, hallucinations, and compliance with existing regulations, including the EU’s AI Act and recent U.S. initiatives, alongside the emerging approaches in China. Furthermore, two different benchmarks are presented.

Method: The study introduces a unified evaluation framework comparing three inference-time methods: (1) baseline single-word generation, (2) PRM-Select best-of-N sampling, and (3) PRM-Sequential refinement guided by PRM critiques. Evaluation uses 200 English prompts and Urdu counterparts across 8 social categories, with GPT-3.5 as generator and GPT-4o-mini as bias/utility scorer.

Result: Results show: (a) substantial bias reduction gains over baseline for both languages, (b) consistently lower fairness scores for Urdu across all methods, highlighting structural inequities, and (c) distinct improvement trajectories between PRM-Select and PRM-Sequential approaches.

Conclusion: The study provides an extensible methodology, interpretable metrics, and cross-lingual comparisons that support future fairness evaluation in low-resource languages, demonstrating that inference-time mitigation can reduce bias but cannot fully address structural training inequities.

Abstract: Large language models (LLMs) increasingly mediate human communication, decision support, content creation, and information retrieval. Despite impressive fluency, these systems frequently produce biased or stereotypical content, especially when prompted with socially sensitive language. A growing body of research has demonstrated that such biases disproportionately affect low-resource languages, where training data is limited and culturally unrepresentative. This paper presents a comprehensive study of inference-time bias mitigation, a strategy that avoids retraining or fine-tuning and instead operates directly on model outputs. Building on preference-ranking models (PRMs), we introduce a unified evaluation framework comparing three methods: (1) baseline single-word generation, (2) PRM-Select best-of-N sampling, and (3) PRM-Sequential refinement guided by PRM critiques. We evaluate these techniques across 200 English prompts and their Urdu counterparts, designed to reflect socio-cultural contexts relevant to gender, ethnicity, religion, nationality, disability, profession, age, and socioeconomic categories. Using GPT-3.5 as a candidate generator and GPT-4o-mini as a PRM-based bias and utility scorer, we provide an extensive quantitative analysis of bias reduction, utility preservation, and cross-lingual disparities. Our findings show: (a) substantial gains over the baseline for both languages; (b) consistently lower fairness scores for Urdu across all methods, highlighting structural inequities in multilingual LLM training; and (c) distinct improvement trajectories between PRM-Select and PRM-Sequential. The study contributes an extensible methodology, interpretable metrics, and cross-lingual comparisons that can support future work on fairness evaluation in low-resource languages.

Method: 1) Use LoRA for parameter-efficient fine-tuning; 2) Interactive adaptation via calibrated linear combination of LoRA modules (gate-free mixture of experts); 3) Novel gradient-based regularization specifically designed for low-rank decomposition matrices using historical gradient information.

Result: LoRA achieves performance on par with full-parameter techniques using fraction of parameters; interactive method enables real-time adjustments without retraining; gradient regularization efficiently preserves prior knowledge while acquiring new tasks.

Conclusion: LoRA framework offers scalable paradigm for interactive and continual NMT, addressing computational cost, catastrophic forgetting, and enabling user-controllable adaptation.

Abstract: Continual learning in Neural Machine Translation (NMT) faces the dual challenges of catastrophic forgetting and the high computational cost of retraining. This study establishes Low-Rank Adaptation (LoRA) as a parameter-efficient framework to address these challenges in dedicated NMT architectures. We first demonstrate that LoRA-based fine-tuning adapts NMT models to new languages and domains with performance on par with full-parameter techniques, while utilizing only a fraction of the parameter space. Second, we propose an interactive adaptation method using a calibrated linear combination of LoRA modules. This approach functions as a gate-free mixture of experts, enabling real-time, user-controllable adjustments to domain and style without retraining. Finally, to mitigate catastrophic forgetting, we introduce a novel gradient-based regularization strategy specifically designed for low-rank decomposition matrices. Unlike methods that regularize the full parameter set, our approach weights the penalty on the low-rank updates using historical gradient information. Experimental results indicate that this strategy efficiently preserves prior domain knowledge while facilitating the acquisition of new tasks, offering a scalable paradigm for interactive and continual NMT.

Method: Adapted encoder-decoder transfer learning method from computer vision to analyze 100 different feature spaces from hidden representations of various networks trained on language tasks. Used this to reveal a low-dimensional structure (language representation embedding) that encodes relationships between representations.

Result: Found a low-dimensional structure where language models and translation models smoothly interpolate between word embeddings, syntactic/semantic tasks, and future word embeddings. This representation embedding predicts how well each feature space maps to human brain fMRI responses to language stimuli. The principal dimension highlights the brain’s natural language processing hierarchy.

Conclusion: The discovered language representation embedding captures part of the brain’s natural language representation structure, suggesting neural networks learn representations that align with human brain organization for language processing.

Abstract: How related are the representations learned by neural language models, translation models, and language tagging tasks? We answer this question by adapting an encoder-decoder transfer learning method from computer vision to investigate the structure among 100 different feature spaces extracted from hidden representations of various networks trained on language tasks. This method reveals a low-dimensional structure where language models and translation models smoothly interpolate between word embeddings, syntactic and semantic tasks, and future word embeddings. We call this low-dimensional structure a language representation embedding because it encodes the relationships between representations needed to process language for a variety of NLP tasks. We find that this representation embedding can predict how well each individual feature space maps to human brain responses to natural language stimuli recorded using fMRI. Additionally, we find that the principal dimension of this structure can be used to create a metric which highlights the brain’s natural language processing hierarchy. This suggests that the embedding captures some part of the brain’s natural language representation structure.

Method: Distinguishes referential grounding from other forms, applies teleosemantic theory of representation requiring two conditions: (1) causal-informational relations to the world, and (2) evolutionary selection history that endows them with information-carrying functions.

Result: Argues that LLMs can satisfy both teleosemantic conditions for referential grounding, even without multimodality or embodiment, thus solving the symbol grounding problem.

Conclusion: LLMs can achieve genuine referential grounding about extra-linguistic reality through their training and evolutionary history, not just through human interpretation.

Abstract: Large language models (LLMs) produce seemingly meaningful outputs, yet they are trained on text alone without direct interaction with the world. This leads to a modern variant of the classical symbol grounding problem in AI: can LLMs’ internal states and outputs be about extra-linguistic reality, independently of the meaning human interpreters project onto them? We argue that they can. We first distinguish referential grounding – the connection between a representation and its worldly referent – from other forms of grounding and argue it is the only kind essential to solving the problem. We contend that referential grounding is achieved when a system’s internal states satisfy two conditions derived from teleosemantic theories of representation: (1) they stand in appropriate causal-informational relations to the world, and (2) they have a history of selection that has endowed them with the function of carrying this information. We argue that LLMs can meet both conditions, even without multimodality or embodiment.

Method: Developed an evaluation framework to study how different NLP-assisted tools produce different outcomes based on collaboration strategies (synchronous vs. asynchronous). Used two different NLP-assisted qualitative research tools for comparison.

Result: Found significant differences in consistency, cohesiveness, and correctness of outputs between synchronous and asynchronous collaboration approaches.

Conclusion: The proposed evaluation framework provides a foundation for systematically assessing NLP-assisted qualitative analysis tools and demonstrates that collaboration strategy significantly impacts tool performance.

Abstract: NLP-assisted solutions have gained considerable traction to support qualitative data analysis. However, there does not exist a unified evaluation framework that can account for the many different settings in which qualitative researchers may employ them. In this paper, we take a first step in this direction by proposing an evaluation framework to study the way in which different tools may result in different outcomes depending on the collaboration strategy employed. Specifically, we study the impact of synchronous vs. asynchronous collaboration using two different NLP-assisted qualitative research tools and present a comprehensive analysis of significant differences in the consistency, cohesiveness, and correctness of their outputs.

Method: The paper conducts a systematic literature review with categorization of world models into two primary functions: (1) constructing internal representations to understand world mechanisms, and (2) predicting future states for simulation and decision-making guidance.

Result: The survey examines current progress in both categories, explores applications in generative games, autonomous driving, robotics, and social simulacra, and provides a comprehensive repository of representative papers with code at https://github.com/tsinghua-fib-lab/World-Model.

Conclusion: The paper outlines key challenges and provides insights into future research directions for world models, positioning them as essential tools for advancing toward artificial general intelligence through better world understanding and prediction capabilities.

Abstract: The concept of world models has garnered significant attention due to advancements in multimodal large language models such as GPT-4 and video generation models such as Sora, which are central to the pursuit of artificial general intelligence. This survey offers a comprehensive review of the literature on world models. Generally, world models are regarded as tools for either understanding the present state of the world or predicting its future dynamics. This review presents a systematic categorization of world models, emphasizing two primary functions: (1) constructing internal representations to understand the mechanisms of the world, and (2) predicting future states to simulate and guide decision-making. Initially, we examine the current progress in these two categories. We then explore the application of world models in key domains, including generative games, autonomous driving, robotics, and social simulacra, with a focus on how each domain utilizes these aspects. Finally, we outline key challenges and provide insights into potential future research directions. We summarize the representative papers along with their code repositories in https://github.com/tsinghua-fib-lab/World-Model.

Method: Two simple classifier-guided approaches that support counterfactual generation by LLMs without fine-tuning. These methods use classifier information to guide LLMs in generating counterfactuals that change model predictions while preserving LLMs’ text generation strengths.

Result: The methods outperform state-of-the-art counterfactual generation methods, work effectively across different LLMs, and can improve classifier robustness through data augmentation. Analysis reveals LLMs rely on parametric knowledge rather than faithfully following the classifier.

Conclusion: Classifier-guided approaches enable effective counterfactual generation by LLMs without fine-tuning, demonstrating the benefits of combining classifier information with LLMs’ text generation capabilities, though LLMs’ reliance on parametric knowledge remains a critical issue.

Abstract: The need for interpretability in deep learning has driven interest in counterfactual explanations, which identify minimal changes to an instance that change a model’s prediction. Current counterfactual (CF) generation methods require task-specific fine-tuning and produce low-quality text. Large Language Models (LLMs), though effective for high-quality text generation, struggle with label-flipping counterfactuals (i.e., counterfactuals that change the prediction) without fine-tuning. We introduce two simple classifier-guided approaches to support counterfactual generation by LLMs, eliminating the need for fine-tuning while preserving the strengths of LLMs. Despite their simplicity, our methods outperform state-of-the-art counterfactual generation methods and are effective across different LLMs, highlighting the benefits of guiding counterfactual generation by LLMs with classifier information. We further show that data augmentation by our generated CFs can improve a classifier’s robustness. Our analysis reveals a critical issue in counterfactual generation by LLMs: LLMs rely on parametric knowledge rather than faithfully following the classifier.

Method: Constrained Discrete Diffusion (CDD) integrates differentiable constraint optimization within the discrete diffusion sampling process, directly imposing constraints during generation rather than using post-hoc methods.

Result: CDD achieves zero constraint violations across toxicity-controlled text generation, property-constrained molecule design, and instruction-constrained text completion tasks while preserving fluency, novelty, and coherence, outperforming autoregressive and existing discrete diffusion approaches.

Conclusion: CDD provides a training-free, effective approach for constraint-aware sequence generation that directly incorporates constraints into the diffusion process, offering superior constraint adherence while maintaining generation quality.

Abstract: Discrete diffusion models are a class of generative models that construct sequences by progressively denoising samples from a categorical noise distribution. Beyond their rapidly growing ability to generate coherent natural language, these models present a new and important opportunity to enforce sequence-level constraints, a capability that current autoregressive models cannot natively provide. This paper capitalizes on this opportunity by introducing Constrained Discrete Diffusion (CDD), a novel integration of differentiable constraint optimization within the diffusion process to ensure adherence to constraints, logic rules, or safety requirements for generated sequences. Unlike conventional text generators that often rely on post-hoc filtering or model retraining for controllable generation, CDD directly imposes constraints into the discrete diffusion sampling process, resulting in a training-free and effective approach. Experiments in toxicity-controlled text generation, property-constrained molecule design, and instruction-constrained text completion demonstrate that CDD achieves zero constraint violations in a diverse array of tasks while preserving fluency, novelty, and coherence while outperforming autoregressive and existing discrete diffusion approaches.

Method: Use simple linear models trained on hidden states of open-source LLMs; analyze learning curves; propose transfer learning method without target variable labels; apply to super-resolution and data imputation tasks.

Result: Hidden states outperform text outputs for economic estimation; only few dozen labeled examples needed; transfer learning improves accuracy without target labels; practical utility demonstrated for super-resolution and imputation.

Conclusion: LLM hidden states capture richer economic information than text outputs, enabling efficient and accurate estimation of economic/financial statistics with minimal data requirements.

Abstract: We investigate whether the hidden states of large language models (LLMs) can be used to estimate and impute economic and financial statistics. Focusing on county-level (e.g. unemployment) and firm-level (e.g. total assets) variables, we show that a simple linear model trained on the hidden states of open-source LLMs outperforms the models’ text outputs. This suggests that hidden states capture richer economic information than the responses of the LLMs reveal directly. A learning curve analysis indicates that only a few dozen labelled examples are sufficient for training. We also propose a transfer learning method that improves estimation accuracy without requiring any labelled data for the target variable. Finally, we demonstrate the practical utility of hidden-state representations in super-resolution and data imputation tasks.

Method: Developed EmoSApp using LLaMA-3.2-1B-Instruct language model fine-tuned and quantized on a custom “Knowledge Dataset” containing 14,582 mental health QA pairs and multi-turn conversational data. The app runs entirely offline on smartphones with on-device inference.

Result: Qualitative evaluation with students and professionals showed EmoSApp responds coherently and empathetically, provides relevant mental health suggestions, and maintains interactive dialogue. Quantitative evaluation on 9 commonsense/reasoning benchmarks and 2 mental health datasets demonstrated effectiveness in low-resource settings.

Conclusion: EmoSApp provides a blueprint for portable, secure, and highly tailored AI-driven mental health support through on-device deployment and domain-specific adaptation, addressing key challenges in digital mental health accessibility and privacy.

Abstract: Mental health plays a crucial role in the overall well-being of an individual. In recent years, digital platforms have increasingly been used to expand mental health and emotional support. However, there are persistent challenges related to limited user accessibility, internet connectivity, and data privacy, which highlight the need for an offline, smartphone-based solutions. To address these challenges, we propose EmoSApp (Emotional Support App): an entirely offline, smartphone-based conversational app designed to provide mental health and emotional support. EmoSApp leverages a language model, specifically the LLaMA-3.2-1B-Instruct, which is fine-tuned and quantized on a custom-curated ``Knowledge Dataset’’ comprising 14,582 mental health QA pairs along with multi-turn conversational data, enabling robust domain expertise and fully on-device inference on resource-constrained smartphones. Through qualitative evaluation with students and mental health professionals, we demonstrate that EmoSApp has the ability to respond coherently and empathetically, provide relevant suggestions to user’s mental health problems, and maintain interactive dialogue. Additionally, quantitative evaluations on nine commonsense and reasoning benchmarks, along with two mental health specific datasets, demonstrate EmoSApp’s effectiveness in low-resource settings. By prioritizing on-device deployment and specialized domain-specific adaptation, EmoSApp serves as a blueprint for future innovations in portable, secure, and highly tailored AI-driven mental health support.

Method: SAFT computes direction-sensitive positional encodings from the magnetic Laplacian of transformed AMRs and projects them into the LLM’s embedding space without architectural changes. This injects graph topology into pretrained models.

Result: SAFT achieves state-of-the-art on AMR 3.0 with 3.5 BLEU improvement over baselines. Performance gains scale with graph complexity, demonstrating the value of structure-aware representations.

Conclusion: SAFT provides a general, effective approach for bridging structured data and language models, enabling LLMs to better handle graph-structured inputs while maintaining compatibility with standard architectures.

Abstract: Large Language Models (LLMs) are increasingly applied to tasks involving structured inputs such as graphs. Abstract Meaning Representations (AMRs), which encode rich semantics as directed graphs, offer a rigorous testbed for evaluating LLMs on text generation from such structures. Yet, current methods often arbitrarily linearize AMRs, discarding key structural cues, or rely on architectures incompatible with standard LLMs. We introduce SAFT, a structure-aware fine-tuning approach that injects graph topology into pretrained LLMs without architectural changes. We compute direction-sensitive positional encodings from the magnetic Laplacian of transformed AMRs and project them into the embedding space of the LLM. While possibly applicable to any graph-structured inputs, we focus on AMR-to-text generation as a representative and challenging benchmark. SAFT sets a new state-of-the-art on AMR 3.0 with a 3.5 BLEU improvement over baselines. Gains scale with graph complexity, highlighting the value of structure-aware representations in enhancing LLM performance. SAFT offers a general and effective pathway for bridging structured data and language models.

Method: Proposed ShoppingBench with: 1) Scalable framework to simulate user instructions based on various intents from sampled real-world products, 2) Large-scale shopping sandbox as interactive simulated environment with over 2.5M real products, 3) Trajectory distillation strategy using supervised fine-tuning and reinforcement learning on synthetic trajectories to distill large language agent capabilities into smaller ones.

Result: State-of-the-art language agents (including GPT-4.1) achieve absolute success rates under 50% on ShoppingBench tasks, highlighting significant challenges. The distilled agent trained with proposed methods achieves competitive performance compared to GPT-4.1.

Conclusion: ShoppingBench addresses the gap in e-commerce benchmarks by focusing on complex, grounded user intents, revealing limitations of current language agents. The proposed distillation approach enables smaller agents to achieve competitive performance, advancing practical e-commerce AI applications.

Abstract: Existing benchmarks in e-commerce primarily focus on basic user intents, such as finding or purchasing products. However, real-world users often pursue more complex goals, such as applying vouchers, managing budgets, and finding multi-products seller. To bridge this gap, we propose ShoppingBench, a novel end-to-end shopping benchmark designed to encompass increasingly challenging levels of grounded intent. Specifically, we propose a scalable framework to simulate user instructions based on various intents derived from sampled real-world products. To facilitate consistent and reliable evaluations, we provide a large-scale shopping sandbox that serves as an interactive simulated environment, incorporating over 2.5 million real-world products. Experimental results demonstrate that even state-of-the-art language agents (such as GPT-4.1) achieve absolute success rates under 50% on our benchmark tasks, highlighting the significant challenges posed by our ShoppingBench. In addition, we propose a trajectory distillation strategy and leverage supervised fine-tuning, along with reinforcement learning on synthetic trajectories, to distill the capabilities of a large language agent into a smaller one. As a result, our trained agent achieves competitive performance compared to GPT-4.1.

Method: Created WebDetective benchmark with hint-free multi-hop questions and a controlled Wikipedia sandbox for full traceability, plus a holistic evaluation framework that separates search sufficiency, knowledge utilization, and refusal behavior.

Result: Evaluation of 25 state-of-the-art models revealed systematic weaknesses: models struggle with knowledge utilization despite having sufficient evidence, and show near-absent appropriate refusal when evidence is lacking. Models excel at executing given reasoning paths but fail at discovering them.

Conclusion: WebDetective provides a critical diagnostic tool for developing genuinely autonomous reasoning systems. The authors developed EvidenceLoop, an agentic workflow with verification loops and evidence tracking that demonstrates how the benchmark can guide concrete architectural improvements.

Abstract: RAG (Retrieval-Augmented Generation) systems and web agents are increasingly evaluated on multi-hop deep search tasks, yet current practice suffers from two major limitations. First, most benchmarks leak the reasoning path in the question text, allowing models to follow surface cues rather than discover reasoning chains autonomously. Second, evaluation is typically reduced to a single pass rate, which collapses diverse behaviours into one score and obscures whether failures stem from inadequate search, poor knowledge use, or inappropriate refusal. To address these issues, we present WebDetective, a benchmark of hint-free multi-hop questions paired with a controlled Wikipedia sandbox that ensures full traceability of model actions, and a holistic evaluation framework that separates search sufficiency, knowledge utilisation, and refusal behaviour. Our evaluation of 25 state-of-the-art models reveals systematic weaknesses across all architectures: models struggle with knowledge utilisation despite having sufficient evidence and demonstrate near-absent appropriate refusal when evidence is lacking. These patterns expose a fundamental gap: today’s systems excel at executing given reasoning paths but fail when required to discover them. We develop an agentic workflow, EvidenceLoop, that explicitly targets the challenges our benchmark identifies, incorporating verification loops and systematic evidence tracking that improve both search and synthesis capabilities. This baseline demonstrates that WebDetective’s diagnostic framework can guide concrete architectural improvements, establishing our benchmark as a critical tool for developing genuinely autonomous reasoning systems rather than pattern-following agents.

Method: Proposes TRepLiNa - a joint method combining Centered Kernel Alignment (CKA) for cross-lingual similarity and REPINA regularization to constrain parameter updates. Experiments with Aya-23 8B using QLoRA in zero-shot, few-shot, and fine-tuning settings across MMLoSo language pairs (Mundari, Santali, Bhili) with Hindi/English pivots.

Result: Aligning mid-level layers using TRepLiNa (CKA+REPINA) improves LRL translation quality, especially in data-scarce settings, providing a low-cost practical approach.

Conclusion: Enforcing cross-lingual similarity in specific internal layers of multilingual LLMs through TRepLiNa is an effective, low-cost method for improving translation of low-resource languages, addressing critical linguistic gaps in diverse language contexts.

Abstract: The 2025 Multimodal Models for Low-Resource Contexts and Social Impact (MMLoSo) Language Challenge addresses one of India’s most pressing linguistic gaps: the lack of resources for its diverse low-resource languages (LRLs). In this study, we investigate whether enforcing cross-lingual similarity in specific internal layers of a decoder-only multilingual large language model (LLM) can improve translation quality from LRL to high-resource language (HRL). Specifically, we combine Centered Kernel Alignment (CKA), a similarity metric that encourages representations of different languages to align, with REPINA, a regularization method that constrains parameter updates to remain close to the pretrained model, into a joint method we call TRepLiNa. In this research project, we experiment with zero-shot, few-shot, and fine-tuning settings using Aya-23 8B with QLoRA across MMLoSo shared task language pairs (Mundari, Santali, Bhili) with Hindi/English pivots. Our results show that aligning mid-level layers using TRepLiNa (CKA+REPINA) is a low-cost, practical approach to improving LRL translation, especially in data-scarce settings.

Method: GRAVITY integrates demographic, cultural, and psychological frameworks (Hofstede’s cultural dimensions, Schwartz’s basic values, World Values Survey, Big Five OCEAN traits) to synthesize profile-grounded preference pairs for personalized content generation.

Result: Evaluated on 400 Amazon users’ book descriptions, GRAVITY outperformed prompt-based conditioning, standard fine-tuning, and naive synthetic pair generation, achieving over 4% higher preference gains across cultures (USA, Brazil, Japan, India), with user studies showing 86% preference for GRAVITY outputs.

Conclusion: Profile-grounded synthetic data captures richer user variation, reduces reliance on costly annotation, produces more engaging user-centered content, and offers a scalable path for LLM personalization.

Abstract: Personalization in LLMs often relies on costly human feedback or interaction logs, limiting scalability and neglecting deeper user attributes. To reduce the reliance on human annotations, we introduce GRAVITY (Generative Response with Aligned Values, Interests, and Traits of You), a framework for generating synthetic, profile-grounded preference data that captures users’ interests, values, beliefs, and personality traits. By integrating demographic, cultural, and psychological frameworks – including Hofstede’s cultural dimensions, Schwartz’s basic values, the World Values Survey, and Big Five OCEAN traits – GRAVITY synthesizes preference pairs to guide personalized content generation. We evaluate GRAVITY on book descriptions for 400 Amazon users, comparing it to prompt-based conditioning, standard fine-tuning, and naive synthetic pair generation. Profile-grounded synthetic data consistently improves generation, especially across multiple cultures (USA, Brazil, Japan, India), achieving over 4% higher preference gains across baselines, with user studies showing that GRAVITY outputs are preferred over 86% of the time. Our results show that scenario-grounded synthetic data can capture richer user variation, reduce reliance on costly annotation, and produce more engaging, user-centered content, offering a scalable path for LLM personalization.

Method: Empirical analysis of DLM attention patterns with focus on attention sinking phenomenon, comparing characteristics with ARMs and testing robustness through sink masking experiments.

Result: DLMs exhibit attention sinks with two key differences: 1) sink positions shift dynamically throughout generation (unlike static sinks in ARMs), and 2) DLMs remain robust to sink removal with only minor performance degradation, while ARMs are highly sensitive to sink removal.

Conclusion: The study reveals fundamental differences in how DLMs allocate and utilize attention compared to ARMs, providing new insights into the inner workings of diffusion-based language models and their distinct attention mechanisms.

Abstract: Masked Diffusion Language Models (DLMs) have recently emerged as a promising alternative to traditional Autoregressive Models (ARMs). DLMs employ transformer encoders with bidirectional attention, enabling parallel token generation while maintaining competitive performance. Although their efficiency and effectiveness have been extensively studied, the internal mechanisms that govern DLMs remain largely unexplored. In this work, we conduct an empirical analysis of DLM attention patterns, focusing on the attention sinking phenomenon, an effect previously observed in various transformer-based architectures. Our findings reveal that DLMs also exhibit attention sinks, but with distinct characteristics. First, unlike in ARMs, the sink positions in DLMs tend to shift throughout the generation process, displaying a dynamic behaviour. Second, while ARMs are highly sensitive to the removal of attention sinks, DLMs remain robust: masking sinks leads to only a minor degradation in performance. These results provide new insights into the inner workings of diffusion-based language models and highlight fundamental differences in how they allocate and utilize attention compared to autoregressive models.

Method: Hybrid framework using VADER for initial sentiment, refined through DistilBERT confidence scores and fuzzy logic adjustments, with a custom fuzzy inference system mapping to 0-1 continuum.

Result: Improved alignment with user ratings, better identification of sentiment extremes, reduced misclassifications across four domain datasets (food delivery, e-commerce, tourism, fashion).

Conclusion: Integrating symbolic reasoning with neural models enables interpretable, fine-grained sentiment analysis in linguistically dynamic domains.

Abstract: Accurately detecting sentiment polarity and intensity in product reviews and social media posts remains challenging due to informal and domain-specific language. To address this, we propose a novel hybrid lexicon-fuzzy-transformer framework that combines rule-based heuristics, contextual deep learning, and fuzzy logic to generate continuous sentiment scores reflecting both polarity and strength. The pipeline begins with VADER-based initial sentiment estimations, which are refined through a two-stage adjustment process. This involves leveraging confidence scores from DistilBERT, a lightweight transformer and applying fuzzy logic principles to mitigate excessive neutrality bias and enhance granularity. A custom fuzzy inference system then maps the refined scores onto a 0 to 1 continuum, producing expert)like judgments. The framework is rigorously evaluated on four domain-specific datasets. food delivery, e-commerce, tourism, and fashion. Results show improved alignment with user ratings, better identification of sentiment extremes, and reduced misclassifications. Both quantitative metrics (distributional alignment, confusion matrices) and qualitative insights (case studies, runtime analysis) affirm the models robustness and efficiency. This work demonstrates the value of integrating symbolic reasoning with neural models for interpretable, finegrained sentiment analysis in linguistically dynamic domains.

Method: Translated medical QA dataset to Persian, used RLAIF to generate rejected-preferred answer pairs with Chain-of-Thought reasoning, trained baseline model with DPO using 4.5M token dataset.

Result: Model trained on 4.5M tokens outperformed gaokerena-V trained on 57M tokens, demonstrating significant improvement in Persian medical reasoning with limited data.

Conclusion: Reasoning-focused training approaches (RLAIF+DPO) are efficient and effective for developing domain-specific language models with limited data availability.

Abstract: Enhancing reasoning capabilities in small language models is critical for specialized applications such as medical question answering, particularly in underrepresented languages like Persian. In this study, we employ Reinforcement Learning with AI Feedback (RLAIF) and Direct preference optimization (DPO) to improve the reasoning skills of a general-purpose Persian language model. To achieve this, we translated a multiple-choice medical question-answering dataset into Persian and used RLAIF to generate rejected-preferred answer pairs, which are essential for DPO training. By prompting both teacher and student models to produce Chain-of-Thought (CoT) reasoning responses, we compiled a dataset containing correct and incorrect reasoning trajectories. This dataset, comprising 2 million tokens in preferred answers and 2.5 million tokens in rejected ones, was used to train a baseline model, significantly enhancing its medical reasoning capabilities in Persian. Remarkably, the resulting model outperformed its predecessor, gaokerena-V, which was trained on approximately 57 million tokens, despite leveraging a much smaller dataset. These results highlight the efficiency and effectiveness of reasoning-focused training approaches in developing domain-specific language models with limited data availability.

Method: Propose neural diversity as decorrelated parallel representations. Introduce ND-LoRA (Neural Diversity Low-Rank Adaptation) combining parallel LoRA adapters with Barlow Twins regularization. Provide formal tail bounds for hallucination probability in ensembled models, reframing it as a second-moment reliability problem.

Result: Reduces hallucinations by up to 25.6% (14.6% on average) while preserving general accuracy. Explains 94.3% of empirical reliability variation across parallel configurations. Shows 0.1% neural correlation increase associated with 3.8% hallucination increase. Demonstrates task-dependent optimality: different tasks require different optimal amounts of neurodiversity.

Conclusion: Neural diversity serves as a third axis of scaling (orthogonal to parameters and data) to improve language model reliability at fixed budgets. LoRA adapters and regularization act synergistically, with neural diversity identified as the mediating factor for reduced hallucinations.

Abstract: Language models continue to hallucinate despite increases in parameters, compute, and data. We propose neural diversity – decorrelated parallel representations – as a principled mechanism that reduces hallucination rates at fixed parameter and data budgets. While existing mitigation strategies largely target accuracy, we provide the first formal tail bounds for hallucination probability in ensembled language models, reframing it as a second-moment reliability problem and explaining 94.3% of empirical reliability variation seen across parallel configurations. We introduce ND-LoRA (Neural Diversity Low-Rank Adaptation), combining parallel LoRA adapters with Barlow Twins regularization, and reduce hallucinations by up to 25.6% (and 14.6% on average) while preserving general accuracy. Ablations show LoRA adapters and regularization act synergistically, causal interventions prove neurodiversity as the mediating factor and correlational studies indicate scale: a 0.1% neural correlation increase is associated with a 3.8% hallucination increase. Finally, task-dependent optimality emerges: different tasks require different optimal amounts of neurodiversity. Together, our results highlight neural diversity as a third axis of scaling – orthogonal to parameters and data – to improve the reliability of language models at fixed budgets.

Method: O-Mem is based on active user profiling that dynamically extracts and updates user characteristics and event records from proactive interactions with agents. It supports hierarchical retrieval of persona attributes and topic-related context.

Result: O-Mem achieves 51.67% on LoCoMo benchmark (3% improvement over LangMem) and 62.99% on PERSONAMEM (3.5% improvement over A-Mem). It also boosts token and interaction response time efficiency compared to previous memory frameworks.

Conclusion: O-Mem opens promising directions for developing efficient and human-like personalized AI assistants by addressing limitations in contextual consistency and dynamic personalization through active user profiling and hierarchical memory retrieval.

Abstract: Recent advancements in LLM-powered agents have demonstrated significant potential in generating human-like responses; however, they continue to face challenges in maintaining long-term interactions within complex environments, primarily due to limitations in contextual consistency and dynamic personalization. Existing memory systems often depend on semantic grouping prior to retrieval, which can overlook semantically irrelevant yet critical user information and introduce retrieval noise. In this report, we propose the initial design of O-Mem, a novel memory framework based on active user profiling that dynamically extracts and updates user characteristics and event records from their proactive interactions with agents. O-Mem supports hierarchical retrieval of persona attributes and topic-related context, enabling more adaptive and coherent personalized responses. O-Mem achieves 51.67% on the public LoCoMo benchmark, a nearly 3% improvement upon LangMem,the previous state-of-the-art, and it achieves 62.99% on PERSONAMEM, a 3.5% improvement upon A-Mem,the previous state-of-the-art. O-Mem also boosts token and interaction response time efficiency compared to previous memory frameworks. Our work opens up promising directions for developing efficient and human-like personalized AI assistants in the future.

Method: Proposes Multi-Agent Collaborative Filtering (MACF) framework that instantiates similar users and relevant items as LLM agents with unique profiles. Each agent can call retrieval tools, suggest candidates, and interact with others. A central orchestrator agent adaptively manages collaboration via dynamic agent recruitment and personalized collaboration instructions.

Result: Experimental results on datasets from three different domains show advantages of MACF compared to strong agentic recommendation baselines.

Conclusion: MACF successfully bridges traditional collaborative filtering with LLM-based multi-agent collaboration, enabling more effective use of collaborative signals for agentic recommendations.

Abstract: Agentic recommendations cast recommenders as large language model (LLM) agents that can plan, reason, use tools, and interact with users of varying preferences in web applications. However, most existing agentic recommender systems focus on generic single-agent plan-execute workflows or multi-agent task decomposition pipelines. Without recommendation-oriented design, they often underuse the collaborative signals in the user-item interaction history, leading to unsatisfying recommendation results. To address this, we propose the Multi-Agent Collaborative Filtering (MACF) framework for agentic recommendations, drawing an analogy between traditional collaborative filtering algorithms and LLM-based multi-agent collaboration. Specifically, given a target user and query, we instantiate similar users and relevant items as LLM agents with unique profiles. Each agent is able to call retrieval tools, suggest candidate items, and interact with other agents. Different from the static preference aggregation in traditional collaborative filtering, MACF employs a central orchestrator agent to adaptively manage the collaboration between user and item agents via dynamic agent recruitment and personalized collaboration instruction. Experimental results on datasets from three different domains show the advantages of our MACF framework compared to strong agentic recommendation baselines.

Method: Created a live dynamic benchmark with 50 questions per month designed by crypto-native professionals, categorized into four quadrants: Simple Retrieval, Complex Retrieval, Simple Prediction, and Complex Prediction. Evaluated 10 LLMs both directly and within agentic frameworks.

Result: Revealed a performance hierarchy and uncovered a “retrieval-prediction imbalance” - many leading models are proficient at data retrieval but weak in predictive analysis tasks, showing agents can appear factually grounded while lacking deeper analytical capabilities.

Conclusion: CryptoBench provides a more challenging and valuable scenario for LLM agent assessment in the demanding cryptocurrency domain, highlighting critical gaps in current models’ analytical capabilities despite strong retrieval performance.

Abstract: This paper introduces CryptoBench, the first expert-curated, dynamic benchmark designed to rigorously evaluate the real-world capabilities of Large Language Model (LLM) agents in the uniquely demanding and fast-paced cryptocurrency domain. Unlike general-purpose agent benchmarks for search and prediction, professional crypto analysis presents specific challenges: \emph{extreme time-sensitivity}, \emph{a highly adversarial information environment}, and the critical need to synthesize data from \emph{diverse, specialized sources}, such as on-chain intelligence platforms and real-time Decentralized Finance (DeFi) dashboards. CryptoBench thus serves as a much more challenging and valuable scenario for LLM agent assessment. To address these challenges, we constructed a live, dynamic benchmark featuring 50 questions per month, expertly designed by crypto-native professionals to mirror actual analyst workflows. These tasks are rigorously categorized within a four-quadrant system: Simple Retrieval, Complex Retrieval, Simple Prediction, and Complex Prediction. This granular categorization enables a precise assessment of an LLM agent’s foundational data-gathering capabilities alongside its advanced analytical and forecasting skills. Our evaluation of ten LLMs, both directly and within an agentic framework, reveals a performance hierarchy and uncovers a failure mode. We observe a \textit{retrieval-prediction imbalance}, where many leading models, despite being proficient at data retrieval, demonstrate a pronounced weakness in tasks requiring predictive analysis. This highlights a problematic tendency for agents to appear factually grounded while lacking the deeper analytical capabilities to synthesize information.

Method: Proposes Collaborative Causal Sensemaking (CCS) research agenda with three components: 1) new training environments that reward collaborative thinking, 2) representations for shared human-AI mental models, and 3) evaluation centered on trust and complementarity.

Result: The paper presents a conceptual framework and research agenda rather than empirical results, proposing a fundamental shift from building oracle-like answer engines to cultivating AI teammates that co-reason with humans.

Conclusion: Shifting MAS research from answer engines to AI teammates that co-reason over causal structure with human partners will advance the design of effective human-AI teams by addressing the current complementarity gap.

Abstract: LLM-based agents are increasingly deployed for expert decision support, yet human-AI teams in high-stakes settings do not yet reliably outperform the best individual. We argue this complementarity gap reflects a fundamental mismatch: current agents are trained as answer engines, not as partners in the collaborative sensemaking through which experts actually make decisions. Sensemaking (the ability to co-construct causal explanations, surface uncertainties, and adapt goals) is the key capability that current training pipelines do not explicitly develop or evaluate. We propose Collaborative Causal Sensemaking (CCS) as a research agenda to develop this capability from the ground up, spanning new training environments that reward collaborative thinking, representations for shared human-AI mental models, and evaluation centred on trust and complementarity. Taken together, these directions shift MAS research from building oracle-like answer engines to cultivating AI teammates that co-reason with their human partners over the causal structure of shared decisions, advancing the design of effective human-AI teams.

Method: Proposes a 3DGS-based framework with dynamic skinning weights for pose-based articulation and additional motion-induced deformations, plus novel regularization for capturing fine geometric details under sparse visual cues. Also introduces a new multi-view dataset with varied lighting for evaluation.

Result: Achieves state-of-the-art performance in novel view synthesis, novel pose rendering, and relighting, enabling realistic rendering of novel poses/views with photo-realistic lighting effects under arbitrary conditions.

Conclusion: RnD-Avatar successfully addresses limitations of previous methods by combining dynamic skinning weights, regularization for fine details, and comprehensive evaluation with a new multi-view lighting dataset, resulting in high-fidelity relightable and animatable human avatars.

Abstract: Modeling relightable and animatable human avatars from monocular video is a long-standing and challenging task. Recently, Neural Radiance Field (NeRF) and 3D Gaussian Splatting (3DGS) methods have been employed to reconstruct the avatars. However, they often produce unsatisfactory photo-realistic results because of insufficient geometrical details related to body motion, such as clothing wrinkles. In this paper, we propose a 3DGS-based human avatar modeling framework, termed as Relightable and Dynamic Gaussian Avatar (RnD-Avatar), that presents accurate pose-variant deformation for high-fidelity geometrical details. To achieve this, we introduce dynamic skinning weights that define the human avatar’s articulation based on pose while also learning additional deformations induced by body motion. We also introduce a novel regularization to capture fine geometric details under sparse visual cues. Furthermore, we present a new multi-view dataset with varied lighting conditions to evaluate relight. Our framework enables realistic rendering of novel poses and views while supporting photo-realistic lighting effects under arbitrary lighting conditions. Our method achieves state-of-the-art performance in novel view synthesis, novel pose rendering, and relighting.

Method: 1) Propose VECTOR benchmark to explicitly assess temporal order understanding. 2) Develop MECOT (Multi-Event instruction fine-tuning with Chain-of-Thought) which trains models on detailed event-by-event descriptions and uses chain-of-thought prompting at inference.

Result: VLMMs often fail on VECTOR benchmark, showing poor temporal understanding. MECOT outperforms prior methods on VECTOR and also improves performance on existing video benchmarks, demonstrating effectiveness for temporal understanding.

Conclusion: Temporal understanding in VLMMs is currently limited but can be improved through specialized fine-tuning like MECOT. The VECTOR benchmark provides a better test for temporal reasoning capabilities.

Abstract: Video Large Multimodal Models (VLMMs) have shown impressive performance in video understanding, yet their ability to accurately capture the temporal order of multiple events remains underexplored. We interestingly observe that, even when video frames are scrambled, models perform very well on the existing benchmarks by comprehensive experiments. This implies that VLMMs may not necessarily rely on accurate sequential processing of visual events, but instead depend on prior knowledge of typical scenarios to answer the question. To benchmark temporal understanding capabilities in VLMMs, we propose VECTOR, designed to explicitly assess a model’s ability to identify the temporal order of events. On this benchmark, we observe that various VLMMs often fail to understand the orders of events. To address this, we propose MECOT (Multi-Event instruction fine-tuning with Chain-of-Thought), which (1) trains models on detailed, event-by-event video descriptions and (2) using chain-of-thought prompts at inference to enhance temporal awareness. MECOT outperforms prior arts on VECTOR as well as improving performance on existing video benchmarks, implying effectiveness of temporal understanding. We release our code, model and datasets.

Method: Uses hierarchical binder structure for cross-attention conditioning in Diffusion Transformers; Diversify-and-Absorb Mechanism to improve binding accuracy; Temporal Disentanglement Strategy with dual-branch binder for video concepts.

Result: Achieves superior concept consistency, prompt fidelity, and motion quality over existing approaches, enabling new possibilities for visual creativity.

Conclusion: Bind & Compose effectively addresses limitations in visual concept composition through innovative binding mechanisms and temporal modeling strategies.

Abstract: Visual concept composition, which aims to integrate different elements from images and videos into a single, coherent visual output, still falls short in accurately extracting complex concepts from visual inputs and flexibly combining concepts from both images and videos. We introduce Bind & Compose, a one-shot method that enables flexible visual concept composition by binding visual concepts with corresponding prompt tokens and composing the target prompt with bound tokens from various sources. It adopts a hierarchical binder structure for cross-attention conditioning in Diffusion Transformers to encode visual concepts into corresponding prompt tokens for accurate decomposition of complex visual concepts. To improve concept-token binding accuracy, we design a Diversify-and-Absorb Mechanism that uses an extra absorbent token to eliminate the impact of concept-irrelevant details when training with diversified prompts. To enhance the compatibility between image and video concepts, we present a Temporal Disentanglement Strategy that decouples the training process of video concepts into two stages with a dual-branch binder structure for temporal modeling. Evaluations demonstrate that our method achieves superior concept consistency, prompt fidelity, and motion quality over existing approaches, opening up new possibilities for visual creativity.

Method: Uses multi-agent system to generate challenging multi-image QA pairs (MIFG-QA dataset), develops specialized tools for visual reflection and confirmation, and employs action-trajectory two-level mask strategy for stable RL training without supervised fine-tuning.

Result: IMAgent maintains strong performance on existing single-image benchmarks while achieving substantial improvements on the proposed multi-image dataset (MIFG-QA).

Conclusion: IMAgent successfully addresses multi-image QA limitations through end-to-end RL training with specialized visual tools, providing a scalable open-source solution for complex vision-language tasks.

Abstract: Recent VLM-based agents aim to replicate OpenAI O3’s ``thinking with images" via tool use, but most open-source methods limit input to a single image, falling short on real-world multi-image QA tasks. To address this, we propose IMAgent, an open-source vision agent trained via end-to-end reinforcement learning dedicated for complex multi-image tasks. By leveraging a multi-agent system, we generate challenging and visually-rich multi-image QA pairs to fully activate the tool-use potential of the base VLM. Through manual verification, we obtain MIFG-QA, comprising 10k samples for training and evaluation. With deeper reasoning steps, VLMs may increasingly ignore visual inputs. We therefore develop two specialized tools for visual reflection and confirmation, allowing the model to proactively reallocate its attention to image content during inference. Benefiting from our well-designed action-trajectory two-level mask strategy, IMAgent achieves stable tool use behavior via pure RL training without requiring costly supervised fine-tuning data. Extensive experiments demonstrate that IMAgent maintains strong performance on existing single-image benchmarks while achieving substantial improvements on our proposed multi-image dataset, with our analysis providing actionable insights for the research community. Codes and data will be released soon.

Method: Proposes Unified Text-Image Embedding (UTIE) that leverages vision-language models (CLIP, OpenCLIP, SigLIP) to enrich facial embeddings of each demographic group with text-derived demographic features from other groups, creating more demographically neutral representations while emphasizing identity-relevant features.

Result: UTIE consistently reduces bias metrics on RFW and BFW benchmarks while maintaining or even improving face verification accuracy across three different VLMs.

Conclusion: UTIE effectively addresses demographic bias in face recognition by inducing demographic ambiguity through cross-modal feature enrichment, promoting fairer verification performance without sacrificing accuracy.

Abstract: Face recognition (FR) systems are often prone to demographic biases, partially due to the entanglement of demographic-specific information with identity-relevant features in facial embeddings. This bias is extremely critical in large multicultural cities, especially where biometrics play a major role in smart city infrastructure. The entanglement can cause demographic attributes to overshadow identity cues in the embedding space, resulting in disparities in verification performance across different demographic groups. To address this issue, we propose a novel strategy, Unified Text-Image Embedding (UTIE), which aims to induce demographic ambiguity in face embeddings by enriching them with information related to other demographic groups. This encourages face embeddings to emphasize identity-relevant features and thus promotes fairer verification performance across groups. UTIE leverages the zero-shot capabilities and cross-modal semantic alignment of Vision-Language Models (VLMs). Given that VLMs are naturally trained to align visual and textual representations, we enrich the facial embeddings of each demographic group with text-derived demographic features extracted from other demographic groups. This encourages a more neutral representation in terms of demographic attributes. We evaluate UTIE using three VLMs, CLIP, OpenCLIP, and SigLIP, on two widely used benchmarks, RFW and BFW, designed to assess bias in FR. Experimental results show that UTIE consistently reduces bias metrics while maintaining, or even improving in several cases, the face verification accuracy.

Method: Introduces Consist-Retinex framework with two innovations: 1) dual-objective consistency loss combining temporal consistency with ground-truth alignment under randomized time sampling, and 2) adaptive noise-emphasized sampling strategy prioritizing large-noise regions essential for one-step conditional generation.

Result: On VE-LOL-L dataset, achieves state-of-the-art performance with single-step sampling (PSNR: 25.51 vs. 23.41, FID: 44.73 vs. 49.59 compared to Diff-Retinex++), while requiring only 1/8 of the training budget relative to 1000-step baseline.

Conclusion: Consist-Retinex successfully adapts consistency modeling to conditional image enhancement, enabling efficient one-step low-light enhancement with superior performance and reduced computational requirements.

Abstract: Diffusion models have achieved remarkable success in low-light image enhancement through Retinex-based decomposition, yet their requirement for hundreds of iterative sampling steps severely limits practical deployment. While recent consistency models offer promising one-step generation for \textit{unconditional synthesis}, their application to \textit{conditional enhancement} remains unexplored. We present \textbf{Consist-Retinex}, the first framework adapting consistency modeling to Retinex-based low-light enhancement. Our key insight is that conditional enhancement requires fundamentally different training dynamics than unconditional generation standard consistency training focuses on low-noise regions near the data manifold, while conditional mapping critically depends on large-noise regimes that bridge degraded inputs to enhanced outputs. We introduce two core innovations: (1) a \textbf{dual-objective consistency loss} combining temporal consistency with ground-truth alignment under randomized time sampling, providing full-spectrum supervision for stable convergence; and (2) an \textbf{adaptive noise-emphasized sampling strategy} that prioritizes training on large-noise regions essential for one-step conditional generation. On VE-LOL-L, Consist-Retinex achieves \textbf{state-of-the-art performance with single-step sampling} (\textbf{PSNR: 25.51 vs. 23.41, FID: 44.73 vs. 49.59} compared to Diff-Retinex++), while requiring only \textbf{1/8 of the training budget} relative to the 1000-step Diff-Retinex baseline.

Method: Introduces VABench with three task types (text-to-audio-video, image-to-audio-video, stereo audio-video generation) and two evaluation modules covering 15 dimensions including pairwise similarities, synchronization, lip-speech consistency, and audio/video QA pairs.

Result: The benchmark covers seven content categories (animals, human sounds, music, environmental sounds, synchronous physical sounds, complex scenes, virtual worlds) and provides systematic analysis and visualization of evaluation results.

Conclusion: VABench aims to establish a new standard for assessing video generation models with synchronous audio capabilities and promote comprehensive advancement in the field.

Abstract: Recent advances in video generation have been remarkable, enabling models to produce visually compelling videos with synchronized audio. While existing video generation benchmarks provide comprehensive metrics for visual quality, they lack convincing evaluations for audio-video generation, especially for models aiming to generate synchronized audio-video outputs. To address this gap, we introduce VABench, a comprehensive and multi-dimensional benchmark framework designed to systematically evaluate the capabilities of synchronous audio-video generation. VABench encompasses three primary task types: text-to-audio-video (T2AV), image-to-audio-video (I2AV), and stereo audio-video generation. It further establishes two major evaluation modules covering 15 dimensions. These dimensions specifically assess pairwise similarities (text-video, text-audio, video-audio), audio-video synchronization, lip-speech consistency, and carefully curated audio and video question-answering (QA) pairs, among others. Furthermore, VABench covers seven major content categories: animals, human sounds, music, environmental sounds, synchronous physical sounds, complex scenes, and virtual worlds. We provide a systematic analysis and visualization of the evaluation results, aiming to establish a new standard for assessing video generation models with synchronous audio capabilities and to promote the comprehensive advancement of the field.

Method: CoPRS uses a Multi-modal Chain-of-Thought (MCoT) approach to generate a differentiable positional prior as a heatmap. A learnable concentration token aggregates image and reasoning text features to create this heatmap, which is then decoded to precise masks through a lightweight decoder.

Result: CoPRS matches or surpasses state-of-the-art metrics on RefCOCO series and ReasonSeg benchmarks under comparable protocols. Experiments show strong positive correlation between CoT trajectory, generated heatmap, and decoded mask, demonstrating interpretable alignment.

Conclusion: The paradigm effectively bridges reasoning and segmentation with advantages in concentration driven by reasoning and more precise mask prediction. The interpretable heatmap interface enhances diagnostic analysis and provides clearer evidence on targets.

Abstract: Existing works on reasoning segmentation either connect hidden features from a language model directly to a mask decoder or represent positions in text, which limits interpretability and semantic detail. To solve this, we present CoPRS, a Multi-modal Chain-of-Thought (MCoT)-based positional perception model that bridges language reasoning to segmentation through a differentiable and interpretable positional prior instantiated as a heatmap. By making the reasoning process clear via MCoT and expressing it as a dense, differentiable heatmap, this interface enhances interpretability and diagnostic analysis and yields more concentrated evidence on the target. A learnable concentration token aggregates features of the image and reasoning text to generate this positional prior, which is decoded to precise masks through a lightweight decoder, providing a direct connection between reasoning and segmentation. Across the RefCOCO series and ReasonSeg, CoPRS matches or surpasses the best reported metrics on each standard split under comparable protocols, with performance at or above the prior state of the art across both validation and test partitions. Extensive experiments demonstrate a strong positive correlation among the CoT trajectory, the generated heatmap, and the decoded mask, supporting an interpretable alignment between the reasoning output and downstream mask generation. Collectively, these findings support the utility of this paradigm in bridging reasoning and segmentation and show advantages in concentration driven by reasoning and in more precise mask prediction. Code, checkpoints and logs are released at https://github.com/ZhenyuLU-Heliodore/CoPRS.git.

Method: HSCP uses a two-stage hierarchical pruning approach: 1) Spectral clustering guided by Centered Kernel Alignment (CKA) to identify and remove redundant layers, 2) Same spectral clustering strategy applied to channel dimension for finer redundancy elimination. Includes noise-robust fine-tuning for robustness.

Result: On UAV-M100 benchmark: Achieves 86.39% parameter reduction and 84.44% FLOPs reduction on ResNet18 while improving accuracy by 1.49% compared to unpruned baseline. Maintains superior robustness in low signal-to-noise ratio environments.

Conclusion: HSCP outperforms existing channel and layer pruning methods by combining hierarchical pruning with spectral clustering and CKA guidance, achieving extreme compression, high performance, and efficient inference for UAV RFFI on edge devices.

Abstract: With the rapid development of Unmanned Aerial Vehicles (UAVs) and the increasing complexity of low-altitude security threats, traditional UAV identification methods struggle to extract reliable signal features and meet real-time requirements in complex environments. Recently, deep learning based Radio Frequency Fingerprint Identification (RFFI) approaches have greatly improved recognition accuracy. However, their large model sizes and high computational demands hinder deployment on resource-constrained edge devices. While model pruning offers a general solution for complexity reduction, existing weight, channel, and layer pruning techniques struggle to concurrently optimize compression rate, hardware acceleration, and recognition accuracy. To this end, in this paper, we introduce HSCP, a Hierarchical Spectral Clustering Pruning framework that combines layer pruning with channel pruning to achieve extreme compression, high performance, and efficient inference. In the first stage, HSCP employs spectral clustering guided by Centered Kernel Alignment (CKA) to identify and remove redundant layers. Subsequently, the same strategy is applied to the channel dimension to eliminate a finer redundancy. To ensure robustness, we further employ a noise-robust fine-tuning strategy. Experiments on the UAV-M100 benchmark demonstrate that HSCP outperforms existing channel and layer pruning methods. Specifically, HSCP achieves $86.39%$ parameter reduction and $84.44%$ FLOPs reduction on ResNet18 while improving accuracy by $1.49%$ compared to the unpruned baseline, and maintains superior robustness even in low signal-to-noise ratio environments.

Method: Two-stage training paradigm: Stage 1 learns internal motion prior using an audio-free autoregressive generator to capture natural facial dynamics. Stage 2 introduces dual-track audio to fine-tune the generator, modulating the learned motion prior based on external speech cues.

Result: UniLS achieves state-of-the-art speaking accuracy and delivers up to 44.1% improvement in listening metrics, generating significantly more diverse and natural listening expressions while effectively mitigating the stiffness problem.

Conclusion: UniLS provides a practical, high-fidelity audio-driven solution for interactive digital humans by enabling end-to-end generation of unified speak-listen expressions using only dual-track audio.

Abstract: Generating lifelike conversational avatars requires modeling not just isolated speakers, but the dynamic, reciprocal interaction of speaking and listening. However, modeling the listener is exceptionally challenging: direct audio-driven training fails, producing stiff, static listening motions. This failure stems from a fundamental imbalance: the speaker’s motion is strongly driven by speech audio, while the listener’s motion primarily follows an internal motion prior and is only loosely guided by external speech. This challenge has led most methods to focus on speak-only generation. The only prior attempt at joint generation relies on extra speaker’s motion to produce the listener. This design is not end-to-end, thereby hindering the real-time applicability. To address this limitation, we present UniLS, the first end-to-end framework for generating unified speak-listen expressions, driven by only dual-track audio. Our method introduces a novel two-stage training paradigm. Stage 1 first learns the internal motion prior by training an audio-free autoregressive generator, capturing the spontaneous dynamics of natural facial motion. Stage 2 then introduces the dual-track audio, fine-tuning the generator to modulate the learned motion prior based on external speech cues. Extensive evaluations show UniLS achieves state-of-the-art speaking accuracy. More importantly, it delivers up to 44.1% improvement in listening metrics, generating significantly more diverse and natural listening expressions. This effectively mitigates the stiffness problem and provides a practical, high-fidelity audio-driven solution for interactive digital humans.

Method: Computes lightweight statistical descriptors, retrieves semantically similar samples from a vector database, and uses this contextual evidence for LLM-based activity identification. Enhanced with prompt optimization and LLM-based activity descriptors for context-enriched vector databases.

Result: Achieves state-of-the-art performance across six diverse HAR benchmarks without requiring model training or fine-tuning. Enables recognition and meaningful labeling of multiple unseen human activities.

Conclusion: RAG-HAR provides a robust, training-free framework for HAR that moves beyond known behaviors, emphasizing practical applicability and eliminating the need for dataset-specific training.

Abstract: Human Activity Recognition (HAR) underpins applications in healthcare, rehabilitation, fitness tracking, and smart environments, yet existing deep learning approaches demand dataset-specific training, large labeled corpora, and significant computational resources.We introduce RAG-HAR, a training-free retrieval-augmented framework that leverages large language models (LLMs) for HAR. RAG-HAR computes lightweight statistical descriptors, retrieves semantically similar samples from a vector database, and uses this contextual evidence to make LLM-based activity identification. We further enhance RAG-HAR by first applying prompt optimization and introducing an LLM-based activity descriptor that generates context-enriched vector databases for delivering accurate and highly relevant contextual information. Along with these mechanisms, RAG-HAR achieves state-of-the-art performance across six diverse HAR benchmarks. Most importantly, RAG-HAR attains these improvements without requiring model training or fine-tuning, emphasizing its robustness and practical applicability. RAG-HAR moves beyond known behaviors, enabling the recognition and meaningful labelling of multiple unseen human activities.

Method: Proposes the Verifier-Threshold method which automatically reallocates test-time compute across different denoising steps in diffusion and flow models, optimizing the allocation of computational resources during inference.

Result: Achieves 2-4x reduction in computational time over state-of-the-art methods while maintaining the same performance on the GenEval benchmark, demonstrating substantial efficiency improvements.

Conclusion: The Verifier-Threshold method effectively optimizes test-time compute allocation for image generation models, delivering significant computational efficiency gains without sacrificing generation quality.

Abstract: Image generation has emerged as a mainstream application of large generative AI models. Just as test-time compute and reasoning have helped language models improve their capabilities, similar benefits have also been observed with image generation models. In particular, searching over noise samples for diffusion and flow models has shown to scale well with test-time compute. While recent works have explored allocating non-uniform inference-compute budgets across different denoising steps, they rely on greedy algorithms and allocate the compute budget ineffectively. In this work, we study this problem and propose solutions to fix it. We propose the Verifier-Threshold method which automatically reallocates test-time compute and delivers substantial efficiency improvements. For the same performance on the GenEval benchmark, we achieve a 2-4x reduction in computational time over the state-of-the-art method.

Method: Three-stage framework: 1) Explainable feature-based classifier filters inadequate images using image processing and contrastive learning features; 2) Image enhancement and deep-learning-assisted annotation; 3) Agreement calculation between annotators using derived formulas to determine annotation usability.

Result: The framework ensures only high-standard data is used for evaluation and AI training by systematically filtering poor images, enhancing quality, and validating annotation reliability through inter-annotator agreement metrics.

Conclusion: A comprehensive quality-control framework addresses data quality issues in diabetic retinopathy datasets, improving AI model reliability by ensuring high-quality annotated training data through systematic filtering, enhancement, and validation processes.

Abstract: Diabetic Retinopathy (DR) affects individuals with long-term diabetes. Without early diagnosis, DR can lead to vision loss. Fundus photography captures the structure of the retina along with abnormalities indicative of the stage of the disease. Artificial Intelligence (AI) can support clinicians in identifying these lesions, reducing manual workload, but models require high-quality annotated datasets. Due to the complexity of retinal structures, errors in image acquisition and lesion interpretation of manual annotators can occur. We proposed a quality-control framework, ensuring only high-standard data is used for evaluation and AI training. First, an explainable feature-based classifier is used to filter inadequate images. The features are extracted both using image processing and contrastive learning. Then, the images are enhanced and put subject to annotation, using deep-learning-based assistance. Lastly, the agreement between annotators calculated using derived formulas determines the usability of the annotations.

Method: Proposes 3D Inverse Design (3DID) framework with two key components: 1) A unified physics-geometry embedding that captures shape and physical field data in a continuous latent space, and 2) A two-stage physics-aware optimization strategy: first stage uses gradient-guided diffusion sampler to explore global latent manifold; second stage uses objective-driven, topology-preserving refinement to sculpt candidates toward target objectives.

Result: 3DID generates high-fidelity 3D geometries and outperforms existing methods in both solution quality and design versatility, enabling true 3D design from scratch.

Conclusion: The proposed 3DID framework successfully addresses limitations of current inverse design methods by directly navigating the 3D design space through continuous latent representation and physics-aware optimization, achieving superior performance in generating high-quality 3D designs.

Abstract: Inverse design aims to design the input variables of a physical system to optimize a specified objective function, typically formulated as a search or optimization problem. However, in 3D domains, the design space grows exponentially, rendering exhaustive grid-based searches infeasible. Recent advances in deep learning have accelerated inverse design by providing powerful generative priors and differentiable surrogate models. Nevertheless, current methods tend to approximate the 3D design space using 2D projections or fine-tune existing 3D shapes. These approaches sacrifice volumetric detail and constrain design exploration, preventing true 3D design from scratch. In this paper, we propose a 3D Inverse Design (3DID) framework that directly navigates the 3D design space by coupling a continuous latent representation with a physics-aware optimization strategy. We first learn a unified physics-geometry embedding that compactly captures shape and physical field data in a continuous latent space. Then, we introduce a two-stage strategy to perform physics-aware optimization. In the first stage, a gradient-guided diffusion sampler explores the global latent manifold. In the second stage, an objective-driven, topology-preserving refinement further sculpts each candidate toward the target objective. This enables 3DID to generate high-fidelity 3D geometries, outperforming existing methods in both solution quality and design versatility.

Method: Proposes Cross-scene Knowledge Integration (CKI) with three components: 1) Alignment of Spectral Characteristics (ASC) for domain-agnostic projection, 2) Cross-scene Knowledge Sharing Preference (CKSP) with Source Similarity Mechanism (SSM) for semantic mismatch resolution, and 3) Complementary Information Integration (CII) to leverage target-specific cues.

Result: Extensive experiments show CKI achieves state-of-the-art performance with strong stability across diverse cross-scene HSI scenarios.

Conclusion: CKI successfully overcomes limitations of existing methods by enabling knowledge transfer in fully heterogeneous settings through explicit incorporation of target-private knowledge during the transfer process.

Abstract: Knowledge transfer has strong potential to improve hyperspectral image (HSI) classification, yet two inherent challenges fundamentally restrict effective cross-domain transfer: spectral variations caused by different sensors and semantic inconsistencies across heterogeneous scenes. Existing methods are limited by transfer settings that assume homogeneous domains or heterogeneous scenarios with only co-occurring categories. When label spaces do not overlap, they further rely on complete source-domain coverage and therefore overlook critical target-private information. To overcome these limitations and enable knowledge transfer in fully heterogeneous settings, we propose Cross-scene Knowledge Integration (CKI), a framework that explicitly incorporates target-private knowledge during transfer. CKI includes: (1) Alignment of Spectral Characteristics (ASC) to reduce spectral discrepancies through domain-agnostic projection; (2) Cross-scene Knowledge Sharing Preference (CKSP), which resolves semantic mismatch via a Source Similarity Mechanism (SSM); and (3) Complementary Information Integration (CII) to maximize the use of target-specific complementary cues. Extensive experiments verify that CKI achieves state-of-the-art performance with strong stability across diverse cross-scene HSI scenarios.

Method: Based on statistical analysis of ship scales, introduces Pyramid Level Shift Strategy (P2-P4 detection heads) to meet Nyquist sampling criteria for small objects. Incorporates GN-CBLinear (Group Normalized Convolutional Block for Linear Projection) for stable training on high-resolution inputs.

Result: Demonstrates superior detection accuracy and efficiency on SODA-A, DOTA-v1.5, FAIR1M-v2.0, and ShipRSImageNet-V1 datasets compared to state-of-the-art models.

Conclusion: LiM-YOLO effectively resolves domain-specific conflicts in satellite ship detection through architectural adaptations that address scale disparity and training stability issues, offering improved performance over general-purpose detectors.

Abstract: Applying general-purpose object detectors to ship detection in satellite imagery presents significant challenges due to the extreme scale disparity and morphological anisotropy of maritime targets. Standard architectures utilizing stride-32 (P5) layers often fail to resolve narrow vessels, resulting in spatial feature dilution. In this work, we propose LiM-YOLO, a specialized detector designed to resolve these domain-specific conflicts. Based on a statistical analysis of ship scales, we introduce a Pyramid Level Shift Strategy that reconfigures the detection head to P2-P4. This shift ensures compliance with Nyquist sampling criteria for small objects while eliminating the computational redundancy of deep layers. To further enhance training stability on high-resolution inputs, we incorporate a Group Normalized Convolutional Block for Linear Projection (GN-CBLinear), which mitigates gradient volatility in micro-batch settings. Validated on SODA-A, DOTA-v1.5, FAIR1M-v2.0, and ShipRSImageNet-V1, LiM-YOLO demonstrates superior detection accuracy and efficiency compared to state-of-the-art models. The code is available at https://github.com/egshkim/LiM-YOLO.

Method: Proposes Deterministic World Model (DWM) that maps system states directly to generative images, eliminating uninterpretable latent variables. Uses dual-objective loss combining pixel-level reconstruction accuracy with control difference loss for behavioral consistency. Integrates DWM with Star-based reachability analysis (StarV) and employs conformal prediction for statistical bounds on trajectory deviation.

Result: Experiments on standard benchmarks show significantly tighter reachable sets and better verification performance compared to latent-variable baseline.

Conclusion: DWM addresses the bottleneck of stochastic latent variables in vision-based control verification, enabling more precise input bounds and improved verification performance through deterministic state-to-image mapping.

Abstract: Verifying closed-loop vision-based control systems remains a fundamental challenge due to the high dimensionality of images and the difficulty of modeling visual environments. While generative models are increasingly used as camera surrogates in verification, their reliance on stochastic latent variables introduces unnecessary overapproximation error. To address this bottleneck, we propose a Deterministic World Model (DWM) that maps system states directly to generative images, effectively eliminating uninterpretable latent variables to ensure precise input bounds. The DWM is trained with a dual-objective loss function that combines pixel-level reconstruction accuracy with a control difference loss to maintain behavioral consistency with the real system. We integrate DWM into a verification pipeline utilizing Star-based reachability analysis (StarV) and employ conformal prediction to derive rigorous statistical bounds on the trajectory deviation between the world model and the actual vision-based system. Experiments on standard benchmarks show that our approach yields significantly tighter reachable sets and better verification performance than a latent-variable baseline.

Method: Introduces a generative model that predicts 3D dose distributions based solely on user-defined preference flavors, allowing planners to customize trade-offs between organs-at-risk and planning target volumes without relying on reference plans.

Result: The method demonstrates superior adaptability and plan quality compared to Varian RapidPlan model in some scenarios, while offering seamless integration with clinical treatment planning systems.

Conclusion: This approach provides greater flexibility and personalization in radiotherapy planning by eliminating institutional bias and enabling customizable preference-based dose prediction.

Abstract: Radiotherapy planning is a highly complex process that often varies significantly across institutions and individual planners. Most existing deep learning approaches for 3D dose prediction rely on reference plans as ground truth during training, which can inadvertently bias models toward specific planning styles or institutional preferences. In this study, we introduce a novel generative model that predicts 3D dose distributions based solely on user-defined preference flavors. These customizable preferences enable planners to prioritize specific trade-offs between organs-at-risk (OARs) and planning target volumes (PTVs), offering greater flexibility and personalization. Designed for seamless integration with clinical treatment planning systems, our approach assists users in generating high-quality plans efficiently. Comparative evaluations demonstrate that our method can surpasses the Varian RapidPlan model in both adaptability and plan quality in some scenarios.

Method: Diffusion model regularized implicit neural representation framework that integrates physical constraints for data fidelity and uses pre-trained diffusion models to provide prior knowledge regularization.

Result: Experimental results on both simulated and clinical data demonstrate effectiveness and generalization ability, showing potential for clinical application.

Conclusion: The proposed framework successfully overcomes limitations of existing MAR approaches by combining implicit neural representation with diffusion model regularization, offering a promising solution for clinical metal artifact reduction.

Abstract: Computed tomography (CT) images are often severely corrupted by artifacts in the presence of metals. Existing supervised metal artifact reduction (MAR) approaches suffer from performance instability on known data due to their reliance on limited paired metal-clean data, which limits their clinical applicability. Moreover, existing unsupervised methods face two main challenges: 1) the CT physical geometry is not effectively incorporated into the MAR process to ensure data fidelity; 2) traditional heuristics regularization terms cannot fully capture the abundant prior knowledge available. To overcome these shortcomings, we propose diffusion model regularized implicit neural representation framework for MAR. The implicit neural representation integrates physical constraints and imposes data fidelity, while the pre-trained diffusion model provides prior knowledge to regularize the solution. Experimental results on both simulated and clinical data demonstrate the effectiveness and generalization ability of our method, highlighting its potential to be applied to clinical settings.

Method: The framework uses motion pattern priors derived from subtle scene dynamics. At the encoder, these priors are compressed via dense-to-sparse transformation. At the decoder, an advanced flow-driven diffusion model reconstructs scene dynamics using these priors.

Result: The method achieves superior rate-distortion performance and outperforms the state-of-the-art conventional video codec ECM (Enhanced Compression Model) on scene dynamics sequences.

Conclusion: Motion pattern priors from subtle scene dynamics provide an effective approach for ultra-low bitrate video compression with high-quality reconstruction across diverse content, offering advantages over content-specific prior methods.

Abstract: This paper proposes a novel generative video compression framework that leverages motion pattern priors, derived from subtle dynamics in common scenes (e.g., swaying flowers or a boat drifting on water), rather than relying on video content priors (e.g., talking faces or human bodies). These compact motion priors enable a new approach to ultra-low bitrate communication while achieving high-quality reconstruction across diverse scene contents. At the encoder side, motion priors can be streamlined into compact representations via a dense-to-sparse transformation. At the decoder side, these priors facilitate the reconstruction of scene dynamics using an advanced flow-driven diffusion model. Experimental results illustrate that the proposed method can achieve superior rate-distortion-performance and outperform the state-of-the-art conventional-video codec Enhanced Compression Model (ECM) on-scene dynamics sequences. The project page can be found at-https://github.com/xyzysz/GNVDC.

Method: Framework synthesizes defect layouts using physics-constrained mathematical morphology operations (erosion/dilation) on original designs, fabricates physical samples via DMD-based lithography, compares optical micrographs to create consistent pixel-level annotations.

Result: Created dataset of 3,530 optical micrographs with 13,365 annotated defect instances across four classes (bridge, burr, pinch, contamination). Mask R-CNN achieved AP@0.5 improvements of ~34% for bridge/burr/pinch and ~42% for contamination compared to Faster R-CNN.

Conclusion: The methodology for generating defect datasets with pixel-level annotations is feasible for robust AI-based measurement/inspection in semiconductor fabrication, addressing the data scarcity bottleneck.

Abstract: The efficacy of Artificial Intelligence (AI) in micro/nano manufacturing is fundamentally constrained by the scarcity of high-quality and physically grounded training data for defect inspection. Lithography defect data from semiconductor industry are rarely accessible for research use, resulting in a shortage of publicly available datasets. To address this bottleneck in lithography, this study proposes a novel methodology for generating large-scale, physically valid defect datasets with pixel-level annotations. The framework begins with the ab initio synthesis of defect layouts using controllable, physics-constrained mathematical morphology operations (erosion and dilation) applied to the original design-level layout. These synthesized layouts, together with their defect-free counterparts, are fabricated into physical samples via high-fidelity digital micromirror device (DMD)-based lithography. Optical micrographs of the synthesized defect samples and their defect-free references are then compared to create consistent defect delineation annotations. Using this methodology, we constructed a comprehensive dataset of 3,530 Optical micrographs containing 13,365 annotated defect instances including four classes: bridge, burr, pinch, and contamination. Each defect instance is annotated with a pixel-accurate segmentation mask, preserving full contour and geometry. The segmentation-based Mask R-CNN achieves AP@0.5 of 0.980, 0.965, and 0.971, compared with 0.740, 0.719, and 0.717 for Faster R-CNN on bridge, burr, and pinch classes, representing a mean AP@0.5 improvement of approximately 34%. For the contamination class, Mask R-CNN achieves an AP@0.5 roughly 42% higher than Faster R-CNN. These consistent gains demonstrate that our proposed methodology to generate defect datasets with pixel-level annotations is feasible for robust AI-based Measurement/Inspection (MI) in semiconductor fabrication.

Method: Comprehensive survey methodology reviewing core concepts, representation techniques, generative approaches, datasets, and evaluation metrics for both body and face motion generation.

Result: First comprehensive review covering both body and face motion generation, providing detailed resources and highlighting future directions for enhancing avatar realism, coherence, and expressiveness.

Conclusion: This survey establishes foundational knowledge for body and face motion generation, identifies key challenges, and outlines future research directions to improve avatar expressiveness in dyadic communication settings.

Abstract: Body and face motion play an integral role in communication. They convey crucial information on the participants. Advances in generative modeling and multi-modal learning have enabled motion generation from signals such as speech, conversational context and visual cues. However, generating expressive and coherent face and body dynamics remains challenging due to the complex interplay of verbal / non-verbal cues and individual personality traits. This survey reviews body and face motion generation, covering core concepts, representations techniques, generative approaches, datasets and evaluation metrics. We highlight future directions to enhance the realism, coherence and expressiveness of avatars in dyadic settings. To the best of our knowledge, this work is the first comprehensive review to cover both body and face motion. Detailed resources are listed on https://lownish23csz0010.github.io/mogen/.

Method: Two-pronged approach: 1) Extend token compression to visual encoder using iterative merging orthogonal in spatial axes; 2) Integrate spectrum pruning unit into LLM using attention-free low-pass filter that gradually prunes redundant visual tokens, compatible with FlashAttention. The LUVC framework systematically compresses visual tokens until complete elimination at final LLM layer.

Result: Achieves 2× inference speedup in language model with negligible accuracy degradation. Training-free characteristic enables immediate deployment across multiple VLMs.

Conclusion: LUVC effectively addresses computational inefficiency in VLMs through systematic visual token compression that gradually fuses high-dimensional visual features into multimodal queries, offering significant speed improvements without retraining.

Abstract: Visual language models encounter challenges in computational efficiency and latency, primarily due to the substantial redundancy in the token representations of high-resolution images and videos. Current attention/similarity-based compression algorithms suffer from either position bias or class imbalance, leading to significant accuracy degradation. They also fail to generalize to shallow LLM layers, which exhibit weaker cross-modal interactions. To address this, we extend token compression to the visual encoder through an effective iterative merging scheme that is orthogonal in spatial axes to accelerate the computation across the entire VLM. Furthermoer, we integrate a spectrum pruning unit into LLM through an attention/similarity-free low-pass filter, which gradually prunes redundant visual tokens and is fully compatible to modern FlashAttention. On this basis, we propose Lossless Ultimate Vision tokens Compression (LUVC) framework. LUVC systematically compresses visual tokens until complete elimination at the final layer of LLM, so that the high-dimensional visual features are gradually fused into the multimodal queries. The experiments show that LUVC achieves a 2 speedup inference in language model with negligible accuracy degradation, and the training-free characteristic enables immediate deployment across multiple VLMs.

Method: Uses deep learning through artificial neural networks with supervised learning algorithms, leveraging simple characteristics of target characters for detection in video sequences.

Result: The approach successfully locates wanted individuals efficiently from image databases, outperforming state-of-the-art methods in computer vision character detection.

Conclusion: The proposed classifier enables reinforcement of national security systems, particularly for Angola’s Integrated Public Security Centre, by detecting target individuals (disappeared persons, criminals) from video sequences and image databases.

Abstract: The notion of learning underlies almost every evolution of Intelligent Agents. In this paper, we present an approach for searching and detecting a given entity in a video sequence. Specifically, we study how the deep learning technique by artificial neuralnetworks allows us to detect a character in a video sequence. The technique of detecting a character in a video is a complex field of study, considering the multitude of objects present in the data under analysis. From the results obtained, we highlight the following, compared to state of the art: In our approach, within the field of Computer Vision, the structuring of supervised learning algorithms allowed us to achieve several successes from simple characteristics of the target character. Our results demonstrate that is new approach allows us to locate, in an efficient way, wanted individuals from a private or public image base. For the case of Angola, the classifier we propose opens the possibility of reinforcing the national security system based on the database of target individuals (disappeared, criminals, etc.) and the video sequences of the Integrated Public Security Centre (CISP).

Method: The authors first derive a physics-grounded forward model of the non-linear suppression mechanism, then create the E-Deflare Benchmark with simulated training data (E-Flare-2.7K) and real-world test data (E-Flare-R) captured using a novel optical system. They then design E-DeflareNet for flare removal.

Result: E-DeflareNet achieves state-of-the-art restoration performance, with extensive experiments validating the approach and demonstrating clear benefits for downstream tasks. Code and datasets are publicly available.

Conclusion: E-Deflare provides the first systematic framework for removing lens flare from event camera data, addressing a fundamental optical artifact that has been overlooked in event-based vision systems, with practical benefits for real-world applications.

Abstract: Event cameras have the potential to revolutionize vision systems with their high temporal resolution and dynamic range, yet they remain susceptible to lens flare, a fundamental optical artifact that causes severe degradation. In event streams, this optical artifact forms a complex, spatio-temporal distortion that has been largely overlooked. We present E-Deflare, the first systematic framework for removing lens flare from event camera data. We first establish the theoretical foundation by deriving a physics-grounded forward model of the non-linear suppression mechanism. This insight enables the creation of the E-Deflare Benchmark, a comprehensive resource featuring a large-scale simulated training set, E-Flare-2.7K, and the first-ever paired real-world test set, E-Flare-R, captured by our novel optical system. Empowered by this benchmark, we design E-DeflareNet, which achieves state-of-the-art restoration performance. Extensive experiments validate our approach and demonstrate clear benefits for downstream tasks. Code and datasets are publicly available.

Method: ConceptPose uses a vision-language model to create open-vocabulary 3D concept maps where each point is tagged with a concept vector derived from saliency maps. The approach establishes robust 3D-3D correspondences across these concept maps to estimate 6DoF relative pose.

Result: Without any object or dataset-specific training, ConceptPose achieves state-of-the-art results on common zero-shot relative pose estimation benchmarks, significantly outperforming existing methods by over 62% in ADD(-S) score, including methods that use extensive dataset-specific training.

Conclusion: ConceptPose successfully bridges the gap between traditional pose estimation methods and modern vision-language models, demonstrating that training-free, model-free approaches can achieve superior performance by leveraging the zero-shot capabilities of VLMs through 3D concept mapping.

Abstract: Object pose estimation is a fundamental task in computer vision and robotics, yet most methods require extensive, dataset-specific training. Concurrently, large-scale vision language models show remarkable zero-shot capabilities. In this work, we bridge these two worlds by introducing ConceptPose, a framework for object pose estimation that is both training-free and model-free. ConceptPose leverages a vision-language-model (VLM) to create open-vocabulary 3D concept maps, where each point is tagged with a concept vector derived from saliency maps. By establishing robust 3D-3D correspondences across concept maps, our approach allows precise estimation of 6DoF relative pose. Without any object or dataset-specific training, our approach achieves state-of-the-art results on common zero shot relative pose estimation benchmarks, significantly outperforming existing methods by over 62% in ADD(-S) score, including those that utilize extensive dataset-specific training.

Method: Created SIP dataset using terrestrial LiDAR scanner with indoor/outdoor construction scenes, annotated at point level with construction-specific taxonomy (Built Environment, Construction Operations, Site Surroundings). Includes scanning protocol, annotation workflow, and quality control procedures.

Result: Openly available dataset with Git repository support, featuring adaptable class configurations for modern 3D deep learning frameworks. Includes challenging slender temporary objects like scaffolding, MEP piping, and scissor lifts affected by occlusion and fragmentation.

Conclusion: SIP enables robust benchmarking for construction-oriented 3D vision tasks by providing field data that retains real-world sensing characteristics, advancing progress monitoring, safety assessment, and digital twin development in active construction sites.

Abstract: Accurate 3D scene interpretation in active construction sites is essential for progress monitoring, safety assessment, and digital twin development. LiDAR is widely used in construction because it offers advantages over camera-based systems, performing reliably in cluttered and dynamically changing conditions. Yet most public datasets for 3D perception are derived from densely fused scans with uniform sampling and complete visibility, conditions that do not reflect real construction sites. Field data are often collected as isolated single-station LiDAR views, constrained by safety requirements, limited access, and ongoing operations. These factors lead to radial density decay, fragmented geometry, and view-dependent visibility-characteristics that remain underrepresented in existing datasets. This paper presents SIP, Site in Pieces, a dataset created to reflect the practical constraints of LiDAR acquisition during construction. SIP provides indoor and outdoor scenes captured with a terrestrial LiDAR scanner and annotated at the point level using a taxonomy tailored to construction environments: A. Built Environment, B. Construction Operations, and C. Site Surroundings. The dataset includes both structural components and slender temporary objects such as scaffolding, MEP piping, and scissor lifts, where sparsity caused by occlusion and fragmented geometry make segmentation particularly challenging. The scanning protocol, annotation workflow, and quality control procedures establish a consistent foundation for the dataset. SIP is openly available with a supporting Git repository, offering adaptable class configurations that streamline adoption within modern 3D deep learning frameworks. By providing field data that retain real-world sensing characteristics, SIP enables robust benchmarking and contributes to advancing construction-oriented 3D vision tasks.

Method: Proposed KD-OCT framework distills knowledge from a high-performance ConvNeXtV2-Large teacher (enhanced with advanced augmentations, stochastic weight averaging, and focal loss) into a lightweight EfficientNet-B2 student. Uses real-time distillation with combined loss balancing soft teacher knowledge transfer and hard ground-truth supervision.

Result: KD-OCT outperforms comparable multi-scale or feature-fusion OCT classifiers in efficiency-accuracy balance, achieving near-teacher performance with substantial reductions in model size and inference time. Student model exceeds most existing frameworks despite compression.

Conclusion: KD-OCT successfully compresses a high-performance OCT classifier while maintaining diagnostic accuracy, facilitating edge deployment for AMD screening in clinical settings. The framework enables efficient real-time OCT analysis for AMD/CNV detection.

Abstract: Age-related macular degeneration (AMD) and choroidal neovascularization (CNV)-related conditions are leading causes of vision loss worldwide, with optical coherence tomography (OCT) serving as a cornerstone for early detection and management. However, deploying state-of-the-art deep learning models like ConvNeXtV2-Large in clinical settings is hindered by their computational demands. Therefore, it is desirable to develop efficient models that maintain high diagnostic performance while enabling real-time deployment. In this study, a novel knowledge distillation framework, termed KD-OCT, is proposed to compress a high-performance ConvNeXtV2-Large teacher model, enhanced with advanced augmentations, stochastic weight averaging, and focal loss, into a lightweight EfficientNet-B2 student for classifying normal, drusen, and CNV cases. KD-OCT employs real-time distillation with a combined loss balancing soft teacher knowledge transfer and hard ground-truth supervision. The effectiveness of the proposed method is evaluated on the Noor Eye Hospital (NEH) dataset using patient-level cross-validation. Experimental results demonstrate that KD-OCT outperforms comparable multi-scale or feature-fusion OCT classifiers in efficiency- accuracy balance, achieving near-teacher performance with substantial reductions in model size and inference time. Despite the compression, the student model exceeds most existing frameworks, facilitating edge deployment for AMD screening. Code is available at https://github.com/erfan-nourbakhsh/KD- OCT.

Method: The approach uses ’negative’ Gaussian mixture statistics for a place - analyzing image statistics that indicate “not this place” to automatically select appropriate matching thresholds.

Result: The method can select thresholds that work well for a variety of image databases and image descriptors, demonstrating improved generalization compared to manual threshold selection.

Conclusion: Automatic threshold selection using negative place statistics is an effective solution to the practical challenge of threshold setting in VPR implementations, enabling more robust performance across diverse visual scenarios.

Abstract: Visual place recognition (VPR) is an important component technology for camera-based mapping and navigation applications. This is a challenging problem because images of the same place may appear quite different for reasons including seasonal changes, weather illumination, structural changes to the environment, as well as transient pedestrian or vehicle traffic. Papers focusing on generating image descriptors for VPR report their results using metrics such as recall@K and ROC curves. However, for a robot implementation, determining which matches are sufficiently good is often reduced to a manually set threshold. And it is difficult to manually select a threshold that will work for a variety of visual scenarios. This paper addresses the problem of automatically selecting a threshold for VPR by looking at the ’negative’ Gaussian mixture statistics for a place - image statistics indicating not this place. We show that this approach can be used to select thresholds that work well for a variety of image databases and image descriptors.

Method: AgentComp leverages LLMs with image generation, editing, and VQA tools to autonomously construct compositional datasets. It then uses an agentic preference optimization method to fine-tune text-to-image models to better distinguish between compositionally similar samples.

Result: Achieves state-of-the-art results on compositionality benchmarks like T2I-CompBench without compromising image quality. Also generalizes to other capabilities like text rendering that weren’t explicitly trained for.

Conclusion: AgentComp effectively addresses compositionality limitations in text-to-image models by using LLM agents to create targeted training data and preference optimization, resulting in improved compositional generation while maintaining image quality.

Abstract: Text-to-image generative models have achieved remarkable visual quality but still struggle with compositionality$-$accurately capturing object relationships, attribute bindings, and fine-grained details in prompts. A key limitation is that models are not explicitly trained to differentiate between compositionally similar prompts and images, resulting in outputs that are close to the intended description yet deviate in fine-grained details. To address this, we propose AgentComp, a framework that explicitly trains models to better differentiate such compositional variations and enhance their reasoning ability. AgentComp leverages the reasoning and tool-use capabilities of large language models equipped with image generation, editing, and VQA tools to autonomously construct compositional datasets. Using these datasets, we apply an agentic preference optimization method to fine-tune text-to-image models, enabling them to better distinguish between compositionally similar samples and resulting in overall stronger compositional generation ability. AgentComp achieves state-of-the-art results on compositionality benchmarks such as T2I-CompBench, without compromising image quality$-$a common drawback in prior approaches$-$and even generalizes to other capabilities not explicitly trained for, such as text rendering.

Method: Three-fold innovations: (1) Context-Aware Cross-Attention for robust visual-textual alignment under severe degradation; (2) Segmentation-aware dual pathway visual encoding fusing global and region-specific embeddings; (3) Resource-Efficient Transformer-Based Language Model for expressive caption generation with minimal compute cost.

Result: MDSE substantially outperforms state-of-the-art captioning models on the Underground Mine Disaster (UMD) dataset and related benchmarks, producing more accurate and contextually relevant descriptions that capture crucial details in obscured environments.

Conclusion: MDSE effectively improves situational awareness for underground emergency response by generating detailed textual explanations of post-disaster scenes, addressing the challenges of vision obscuration in mining disasters.

Abstract: Underground mining disasters produce pervasive darkness, dust, and collapses that obscure vision and make situational awareness difficult for humans and conventional systems. To address this, we propose MDSE, Multimodal Disaster Situation Explainer, a novel vision-language framework that automatically generates detailed textual explanations of post-disaster underground scenes. MDSE has three-fold innovations: (i) Context-Aware Cross-Attention for robust alignment of visual and textual features even under severe degradation; (ii) Segmentation-aware dual pathway visual encoding that fuses global and region-specific embeddings; and (iii) Resource-Efficient Transformer-Based Language Model for expressive caption generation with minimal compute cost. To support this task, we present the Underground Mine Disaster (UMD) dataset–the first image-caption corpus of real underground disaster scenes–enabling rigorous training and evaluation. Extensive experiments on UMD and related benchmarks show that MDSE substantially outperforms state-of-the-art captioning models, producing more accurate and contextually relevant descriptions that capture crucial details in obscured environments, improving situational awareness for underground emergency response. The code is at https://github.com/mizanJewel/Multimodal-Disaster-Situation-Explainer.

Method: Proposes FoCULR (Food Category Understanding and Layout Refinement) which incorporates food domain knowledge and introduces core concepts early in the generation process to address insufficient multi-noun category knowledge in text encoders and misinterpretation of multi-noun relationships.

Result: Experimental results show that integrating these techniques improves image generation performance in the food domain, specifically addressing the challenges of multi-noun food categories.

Conclusion: FoCULR effectively addresses the challenges of generating realistic food images for multi-noun categories by incorporating domain knowledge and refining spatial layouts, leading to improved generation performance.

Abstract: Generating realistic food images for categories with multiple nouns is surprisingly challenging. For instance, the prompt “egg noodle” may result in images that incorrectly contain both eggs and noodles as separate entities. Multi-noun food categories are common in real-world datasets and account for a large portion of entries in benchmarks such as UEC-256. These compound names often cause generative models to misinterpret the semantics, producing unintended ingredients or objects. This is due to insufficient multi-noun category related knowledge in the text encoder and misinterpretation of multi-noun relationships, leading to incorrect spatial layouts. To overcome these challenges, we propose FoCULR (Food Category Understanding and Layout Refinement) which incorporates food domain knowledge and introduces core concepts early in the generation process. Experimental results demonstrate that the integration of these techniques improves image generation performance in the food domain.

Method: Uses gravity as global reference for absolute coordinate system camera trajectories, leverages panoramic 360-degree videos for diverse trajectories, introduces null-pitch conditioning to reduce text-camera conflicts, and establishes rebalanced benchmark for evaluation.

Result: Enables precise, interpretable camera parameter control, generates diverse camera trajectories beyond conventional forward-facing ones, and improves robustness when camera specifications conflict with text content.

Conclusion: GimbalDiffusion advances text-to-video controllability with gravity-aligned camera manipulation, enabling precise camera control grounded in physical-world coordinates.

Abstract: Recent progress in text-to-video generation has achieved remarkable realism, yet fine-grained control over camera motion and orientation remains elusive. Existing approaches typically encode camera trajectories through relative or ambiguous representations, limiting explicit geometric control. We introduce GimbalDiffusion, a framework that enables camera control grounded in physical-world coordinates, using gravity as a global reference. Instead of describing motion relative to previous frames, our method defines camera trajectories in an absolute coordinate system, allowing precise and interpretable control over camera parameters without requiring an initial reference frame. We leverage panoramic 360-degree videos to construct a wide variety of camera trajectories, well beyond the predominantly straight, forward-facing trajectories seen in conventional video data. To further enhance camera guidance, we introduce null-pitch conditioning, an annotation strategy that reduces the model’s reliance on text content when conflicting with camera specifications (e.g., generating grass while the camera points towards the sky). Finally, we establish a benchmark for camera-aware video generation by rebalancing SpatialVID-HQ for comprehensive evaluation under wide camera pitch variation. Together, these contributions advance the controllability and robustness of text-to-video models, enabling precise, gravity-aligned camera manipulation within generative frameworks.

Method: SuperF uses coordinate-based neural networks (neural fields) with an implicit neural representation (INR) shared across multiple shifted low-resolution frames. It jointly optimizes frame alignment and the INR, parameterizing sub-pixel alignment as optimizable affine transformations and using super-sampled coordinate grids matching output resolution.

Result: The approach yields compelling results on simulated bursts of satellite imagery and ground-level images from handheld cameras, achieving upsampling factors of up to 8×. It outperforms related INR baselines adapted from burst fusion.

Conclusion: SuperF provides an effective test-time optimization approach for MISR that avoids hallucinated structures and doesn’t require high-resolution training data, making it suitable for both satellite and handheld camera applications.

Abstract: High-resolution imagery is often hindered by limitations in sensor technology, atmospheric conditions, and costs. Such challenges occur in satellite remote sensing, but also with handheld cameras, such as our smartphones. Hence, super-resolution aims to enhance the image resolution algorithmically. Since single-image super-resolution requires solving an inverse problem, such methods must exploit strong priors, e.g. learned from high-resolution training data, or be constrained by auxiliary data, e.g. by a high-resolution guide from another modality. While qualitatively pleasing, such approaches often lead to “hallucinated” structures that do not match reality. In contrast, multi-image super-resolution (MISR) aims to improve the (optical) resolution by constraining the super-resolution process with multiple views taken with sub-pixel shifts. Here, we propose SuperF, a test-time optimization approach for MISR that leverages coordinate-based neural networks, also called neural fields. Their ability to represent continuous signals with an implicit neural representation (INR) makes them an ideal fit for the MISR task. The key characteristic of our approach is to share an INR for multiple shifted low-resolution frames and to jointly optimize the frame alignment with the INR. Our approach advances related INR baselines, adopted from burst fusion for layer separation, by directly parameterizing the sub-pixel alignment as optimizable affine transformation parameters and by optimizing via a super-sampled coordinate grid that corresponds to the output resolution. Our experiments yield compelling results on simulated bursts of satellite imagery and ground-level images from handheld cameras, with upsampling factors of up to 8. A key advantage of SuperF is that this approach does not rely on any high-resolution training data.

Method: Developed AngioAI-QFR, an end-to-end angiography-only pipeline combining: deep learning stenosis detection, lumen segmentation, centerline and diameter extraction, per-millimeter Relative Flow Capacity profiling, and virtual stenting with automatic recomputation of angiography-derived QFR. Evaluated on 100 consecutive vessels with invasive FFR as reference.

Result: Stenosis detection achieved precision 0.97 and lumen segmentation Dice 0.78. AngioAI-QFR correlated strongly with FFR (r = 0.89, MAE 0.045). AUC for detecting FFR ≤ 0.80 was 0.93 (sensitivity 0.88, specificity 0.86). Pipeline completed automatically in 93% of vessels with median time 41 seconds. RFC profiling distinguished focal from diffuse disease, and virtual stenting predicted larger QFR gain in focal disease.

Conclusion: AngioAI-QFR provides a practical, near real-time pipeline that unifies computer vision, functional profiling, and virtual PCI with automated angiography-derived physiology, addressing current workflow limitations in coronary artery disease assessment.

Abstract: Coronary angiography is the main tool for assessing coronary artery disease, but visual grading of stenosis is variable and only moderately related to ischaemia. Wire based fractional flow reserve (FFR) improves lesion selection but is not used systematically. Angiography derived indices such as quantitative flow ratio (QFR) offer wire free physiology, yet many tools are workflow intensive and separate from automated anatomy analysis and virtual PCI planning. We developed AngioAI-QFR, an end to end angiography only pipeline combining deep learning stenosis detection, lumen segmentation, centreline and diameter extraction, per millimetre Relative Flow Capacity profiling, and virtual stenting with automatic recomputation of angiography derived QFR. The system was evaluated in 100 consecutive vessels with invasive FFR as reference. Primary endpoints were agreement with FFR (correlation, mean absolute error) and diagnostic performance for FFR <= 0.80. On held out frames, stenosis detection achieved precision 0.97 and lumen segmentation Dice 0.78. Across 100 vessels, AngioAI-QFR correlated strongly with FFR (r = 0.89, MAE 0.045). The AUC for detecting FFR <= 0.80 was 0.93, with sensitivity 0.88 and specificity 0.86. The pipeline completed fully automatically in 93 percent of vessels, with median time to result 41 s. RFC profiling distinguished focal from diffuse capacity loss, and virtual stenting predicted larger QFR gain in focal than in diffuse disease. AngioAI-QFR provides a practical, near real time pipeline that unifies computer vision, functional profiling, and virtual PCI with automated angiography derived physiology.

Method: The method embeds each canonical Gaussian primitive’s local frame into UV space patches on a template mesh. This enables reconstruction of continuous editable material head textures from single monocular video. The approach uses an efficient physically based reflectance model for relighting and editing intrinsic material maps.

Result: The method demonstrates accurate reconstructions comparable to state-of-the-art, produces high-quality relighting results, and enables intuitive controls for modifying avatar appearance and geometry via texture mapping without additional optimization.

Conclusion: The proposed approach successfully bridges the gap between photorealistic Gaussian Splatting and editable mesh-based methods, offering both reconstruction accuracy and intuitive editability for head avatar applications in visual effects, videoconferencing, and virtual reality.

Abstract: Recent advancements in Gaussian Splatting have enabled increasingly accurate reconstruction of photorealistic head avatars, opening the door to numerous applications in visual effects, videoconferencing, and virtual reality. This, however, comes with the lack of intuitive editability offered by traditional triangle mesh-based methods. In contrast, we propose a method that combines the accuracy and fidelity of 2D Gaussian Splatting with the intuitiveness of UV texture mapping. By embedding each canonical Gaussian primitive’s local frame into a patch in the UV space of a template mesh in a computationally efficient manner, we reconstruct continuous editable material head textures from a single monocular video on a conventional UV domain. Furthermore, we leverage an efficient physically based reflectance model to enable relighting and editing of these intrinsic material maps. Through extensive comparisons with state-of-the-art methods, we demonstrate the accuracy of our reconstructions, the quality of our relighting results, and the ability to provide intuitive controls for modifying an avatar’s appearance and geometry via texture mapping without additional optimization.

Method: Two key innovations: (1) scale-adaptive Gaussian surfels for generating and real-time rendering of multi-scale 3D scenes, and (2) a progressive detail synthesizer that iteratively generates finer-scale 3D contents.

Result: WonderZoom significantly outperforms state-of-the-art video and 3D models in both quality and alignment, enabling multi-scale 3D world creation from a single image with interactive zooming capability.

Conclusion: The approach enables users to “zoom into” 3D regions and auto-regressively synthesize previously non-existent fine details from landscapes to microscopic features, advancing multi-scale 3D world generation.

Abstract: We present WonderZoom, a novel approach to generating 3D scenes with contents across multiple spatial scales from a single image. Existing 3D world generation models remain limited to single-scale synthesis and cannot produce coherent scene contents at varying granularities. The fundamental challenge is the lack of a scale-aware 3D representation capable of generating and rendering content with largely different spatial sizes. WonderZoom addresses this through two key innovations: (1) scale-adaptive Gaussian surfels for generating and real-time rendering of multi-scale 3D scenes, and (2) a progressive detail synthesizer that iteratively generates finer-scale 3D contents. Our approach enables users to “zoom into” a 3D region and auto-regressively synthesize previously non-existent fine details from landscapes to microscopic features. Experiments demonstrate that WonderZoom significantly outperforms state-of-the-art video and 3D models in both quality and alignment, enabling multi-scale 3D world creation from a single image. We show video results and an interactive viewer of generated multi-scale 3D worlds in https://wonderzoom.github.io/

Method: Built on frozen VLM backbone, uses recency-weighted multi-teacher distillation to retain prior knowledge. Employs session-aware compositional prompts to fuse multimodal features for new compositions, while attribute and object prompts are learned through session-agnostic fusion for global semantic consistency. Uses Cosine Anchor Loss (CAL) to stabilize knowledge preservation, Orthogonal Projection Loss (OPL) to keep new embeddings distinct from previous ones, and Intra-Session Diversity Loss (IDL) to promote variation among current-session embeddings.

Result: Extensive experiments on UT-Zappos and C-GQA benchmarks demonstrate substantial improvements over prior VLM-based and non-VLM baselines, setting a new benchmark for CCZSL in closed-world settings.

Conclusion: PromptCCZSL successfully addresses the challenges of continual compositional zero-shot learning by preventing catastrophic forgetting while enabling effective adaptation to new compositions, establishing a new state-of-the-art approach for this complex learning scenario.

Abstract: We tackle continual adaptation of vision-language models to new attributes, objects, and their compositions in Compositional Zero-Shot Learning (CZSL), while preventing forgetting of prior knowledge. Unlike classical continual learning where classes are disjoint, CCZSL is more complex as attributes and objects may reoccur across sessions while compositions remain unique. Built on a frozen VLM backbone, we propose the first Prompt-based Continual Compositional Zero-Shot Learning (PromptCCZSL) framework that retains prior knowledge through recency-weighted multi-teacher distillation. It employs session-aware compositional prompts to fuse multimodal features for new compositions, while attribute and object prompts are learned through session-agnostic fusion to maintain global semantic consistency, which is further stabilized by a Cosine Anchor Loss (CAL) to preserve prior knowledge. To enhance adaptation in the current session, an Orthogonal Projection Loss (OPL) ensures that new attribute and object embeddings remain distinct from previous ones, preventing overlap, while an Intra-Session Diversity Loss (IDL) promotes variation among current-session embeddings for richer, more discriminative representations. We also introduce a comprehensive protocol that jointly measures catastrophic forgetting and compositional generalization. Extensive experiments on UT-Zappos and C-GQA benchmarks demonstrate that PromptCCZSL achieves substantial improvements over prior VLM-based and non-VLM baselines, setting a new benchmark for CCZSL in closed-world settings.

Method: Proposes Δ-LFM framework combining Flow Matching (FM) to model disease dynamics as velocity fields, with patient-specific latent alignment that enforces trajectories along a specific axis where magnitude increases monotonically with disease severity. This creates consistent, semantically meaningful latent spaces.

Result: Demonstrates strong empirical performance across three longitudinal MRI benchmarks. Provides a new framework for interpreting and visualizing disease dynamics with improved semantic structure and continuity.

Conclusion: Δ-LFM offers an effective approach for modeling patient-specific disease progression with interpretable latent representations, addressing key limitations of existing methods by combining flow matching with semantic latent alignment.

Abstract: Understanding disease progression is a central clinical challenge with direct implications for early diagnosis and personalized treatment. While recent generative approaches have attempted to model progression, key mismatches remain: disease dynamics are inherently continuous and monotonic, yet latent representations are often scattered, lacking semantic structure, and diffusion-based models disrupt continuity with random denoising process. In this work, we propose to treat the disease dynamic as a velocity field and leverage Flow Matching (FM) to align the temporal evolution of patient data. Unlike prior methods, it captures the intrinsic dynamic of disease, making the progression more interpretable. However, a key challenge remains: in latent space, Auto-Encoders (AEs) do not guarantee alignment across patients or correlation with clinical-severity indicators (e.g., age and disease conditions). To address this, we propose to learn patient-specific latent alignment, which enforces patient trajectories to lie along a specific axis, with magnitude increasing monotonically with disease severity. This leads to a consistent and semantically meaningful latent space. Together, we present $Δ$-LFM, a framework for modeling patient-specific latent progression with flow matching. Across three longitudinal MRI benchmarks, $Δ$-LFM demonstrates strong empirical performance and, more importantly, offers a new framework for interpreting and visualizing disease dynamics.

Method: Introduces VLM x SI paradigm with View-on-Graph (VoG) method that organizes scenes into multi-modal, multi-layer scene graphs. VLMs act as active agents that selectively access necessary cues while traversing the scene graph, enabling incremental reasoning.

Result: VoG achieves state-of-the-art zero-shot performance, demonstrating that structured scene exploration is a promising strategy for advancing zero-shot 3DVG.

Conclusion: The VLM x SI paradigm with scene graph organization lowers reasoning difficulty for VLMs and produces transparent, interpretable step-by-step reasoning traces, establishing structured scene exploration as an effective approach for zero-shot 3D visual grounding.

Abstract: 3D visual grounding (3DVG) identifies objects in 3D scenes from language descriptions. Existing zero-shot approaches leverage 2D vision-language models (VLMs) by converting 3D spatial information (SI) into forms amenable to VLM processing, typically as composite inputs such as specified view renderings or video sequences with overlaid object markers. However, this VLM + SI paradigm yields entangled visual representations that compel the VLM to process entire cluttered cues, making it hard to exploit spatial semantic relationships effectively. In this work, we propose a new VLM x SI paradigm that externalizes the 3D SI into a form enabling the VLM to incrementally retrieve only what it needs during reasoning. We instantiate this paradigm with a novel View-on-Graph (VoG) method, which organizes the scene into a multi-modal, multi-layer scene graph and allows the VLM to operate as an active agent that selectively accesses necessary cues as it traverses the scene. This design offers two intrinsic advantages: (i) by structuring 3D context into a spatially and semantically coherent scene graph rather than confounding the VLM with densely entangled visual inputs, it lowers the VLM’s reasoning difficulty; and (ii) by actively exploring and reasoning over the scene graph, it naturally produces transparent, step-by-step traces for interpretable 3DVG. Extensive experiments show that VoG achieves state-of-the-art zero-shot performance, establishing structured scene exploration as a promising strategy for advancing zero-shot 3DVG.

Method: The paper presents the FCM standard developed by MPEG, which enables efficient extraction, compression, and transmission of intermediate neural network features. It offloads computationally intensive operations to base servers with high computing resources.

Result: Experimental results show that the FCM standard maintains the same level of accuracy while reducing bitrate requirements by 75.90% compared to remote inference approaches.

Conclusion: FCM provides an effective solution for enabling low-powered devices to utilize large deep learning models through efficient feature coding and transmission, significantly reducing bandwidth requirements without compromising accuracy.

Abstract: As consumer devices become increasingly intelligent and interconnected, efficient data transfer solutions for machine tasks have become essential. This paper presents an overview of the latest Feature Coding for Machines (FCM) standard, part of MPEG-AI and developed by the Moving Picture Experts Group (MPEG). FCM supports AI-driven applications by enabling the efficient extraction, compression, and transmission of intermediate neural network features. By offloading computationally intensive operations to base servers with high computing resources, FCM allows low-powered devices to leverage large deep learning models. Experimental results indicate that the FCM standard maintains the same level of accuracy while reducing bitrate requirements by 75.90% compared to remote inference.

Method: The authors design an efficient post-training and inference framework that enables video MLLMs to operate on compressed visual tokens and generate brief reasoning traces before answering, without requiring manual CoT annotations or supervised fine-tuning.

Result: The resulting models achieve substantially improved inference efficiency while delivering competitive performance across diverse benchmarks, demonstrating that concise reasoning can be both effective and efficient.

Conclusion: Long, human-like chain-of-thought reasoning may not be necessary for general video reasoning; concise reasoning with compressed visual tokens provides an effective and efficient alternative.

Abstract: Chain-of-thought (CoT) reasoning has been highly successful in solving complex tasks in natural language processing, and recent multimodal large language models (MLLMs) have extended this paradigm to video reasoning. However, these models typically build on lengthy reasoning chains and large numbers of input visual tokens. Motivated by empirical observations from our benchmark study, we hypothesize that concise reasoning combined with a reduced set of visual tokens can be sufficient for effective video reasoning. To evaluate this hypothesis, we design and validate an efficient post-training and inference framework that enhances a video MLLM’s reasoning capability. Our framework enables models to operate on compressed visual tokens and generate brief reasoning traces prior to answering. The resulting models achieve substantially improved inference efficiency, deliver competitive performance across diverse benchmarks, and avoid reliance on manual CoT annotations or supervised fine-tuning. Collectively, our results suggest that long, human-like CoT reasoning may not be necessary for general video reasoning, and that concise reasoning can be both effective and efficient. Our code will be released at https://github.com/LaVi-Lab/Rethink_CoT_Video.

Method: Employ Z-score normalization to efficiently recover compressed feature data at decoder side; integrated into MPEG’s FCM codec standard; also proposed simplified method for further overhead reduction in certain circumstances.

Result: Method reduces overhead bits and improves end-task accuracy; shows 17.09% average bitrate reduction across different tasks, up to 65.69% for object tracking without sacrificing accuracy; supersedes existing scaling method in current standard.

Conclusion: Z-score normalization effectively compresses feature data in split-inference AI models, significantly reducing bitrate while maintaining or improving task accuracy, making it superior to existing methods in the FCM codec standard.

Abstract: The split-inference paradigm divides an artificial intelligence (AI) model into two parts. This necessitates the transfer of intermediate feature data between the two halves. Here, effective compression of the feature data becomes vital. In this paper, we employ Z-score normalization to efficiently recover the compressed feature data at the decoder side. To examine the efficacy of our method, the proposed method is integrated into the latest Feature Coding for Machines (FCM) codec standard under development by the Moving Picture Experts Group (MPEG). Our method supersedes the existing scaling method used by the current standard under development. It both reduces the overhead bits and improves the end-task accuracy. To further reduce the overhead in certain circumstances, we also propose a simplified method. Experiments show that using our proposed method shows 17.09% reduction in bitrate on average across different tasks and up to 65.69% for object tracking without sacrificing the task accuracy.

Method: Created MedForget benchmark with 3840 multimodal instances (image, question, answer) modeled as nested hospital hierarchy (Institution→Patient→Study→Section). Includes explicit retain/forget splits, rephrased variants, and introduces reconstruction attacks that progressively add hierarchical context to test unlearning completeness.

Result: Existing unlearning methods struggle with complete, hierarchy-aware forgetting without reducing diagnostic performance. Coarse-grained unlearning resists reconstruction attacks, but fine-grained unlearning leaves models vulnerable to hierarchical data reconstruction.

Conclusion: MedForget provides a HIPAA-aligned testbed for building compliant medical AI systems, revealing that current unlearning approaches need improvement for hierarchical medical data and that granularity affects reconstruction vulnerability.

Abstract: Pretrained Multimodal Large Language Models (MLLMs) are increasingly deployed in medical AI systems for clinical reasoning, diagnosis support, and report generation. However, their training on sensitive patient data raises critical privacy and compliance challenges under regulations such as HIPAA and GDPR, which enforce the “right to be forgotten”. Unlearning, the process of tuning models to selectively remove the influence of specific training data points, offers a potential solution, yet its effectiveness in complex medical settings remains underexplored. To systematically study this, we introduce MedForget, a Hierarchy-Aware Multimodal Unlearning Testbed with explicit retain and forget splits and evaluation sets containing rephrased variants. MedForget models hospital data as a nested hierarchy (Institution -> Patient -> Study -> Section), enabling fine-grained assessment across eight organizational levels. The benchmark contains 3840 multimodal (image, question, answer) instances, each hierarchy level having a dedicated unlearning target, reflecting distinct unlearning challenges. Experiments with four SOTA unlearning methods on three tasks (generation, classification, cloze) show that existing methods struggle to achieve complete, hierarchy-aware forgetting without reducing diagnostic performance. To test whether unlearning truly deletes hierarchical pathways, we introduce a reconstruction attack that progressively adds hierarchical level context to prompts. Models unlearned at a coarse granularity show strong resistance, while fine-grained unlearning leaves models vulnerable to such reconstruction. MedForget provides a practical, HIPAA-aligned testbed for building compliant medical AI systems.

Method: Deep convolutional neural network (CNN) for CKD detection from CT kidney images, with class balancing using Synthetic Minority Over-sampling Technique (SMOTE) and interpretability via Gradient-weighted Class Activation Mapping (Grad-CAM). Trained on CT KIDNEY DATASET with 12,446 CT images across cyst, normal, stone, and tumor cases.

Result: The proposed deep CNN achieved 100% accuracy in early detection of chronic kidney disease (CKD), demonstrating exceptional classification performance.

Conclusion: The approach shows strong potential for addressing critical clinical diagnostic challenges and enhancing early medical intervention strategies for CKD.

Abstract: Chronic Kidney Disease (CKD) constitutes a major global medical burden, marked by the gradual deterioration of renal function, which results in the impaired clearance of metabolic waste and disturbances in systemic fluid homeostasis. Owing to its substantial contribution to worldwide morbidity and mortality, the development of reliable and efficient diagnostic approaches is critically important to facilitate early detection and prompt clinical management. This study presents a deep convolutional neural network (CNN) for early CKD detection from CT kidney images, complemented by class balancing using Synthetic Minority Over-sampling Technique (SMOTE) and interpretability via Gradient-weighted Class Activation Mapping (Grad-CAM). The model was trained and evaluated on the CT KIDNEY DATASET, which contains 12,446 CT images, including 3,709 cyst, 5,077 normal, 1,377 stone, and 2,283 tumor cases. The proposed deep CNN achieved a remarkable classification performance, attaining 100% accuracy in the early detection of chronic kidney disease (CKD). This significant advancement demonstrates strong potential for addressing critical clinical diagnostic challenges and enhancing early medical intervention strategies.

Method: Uses a unified diffusion framework with in-context learning: for text-to-PSD, arranges layers spatially and learns compositional relationships via spatial attention; for image-to-PSD, performs iterative in-context editing to extract/erase components. Employs RGBA-VAE to preserve transparency without affecting structure learning.

Result: Extensive experiments on a new RGBA-layered dataset show OmniPSD achieves high-fidelity generation, structural consistency, and transparency awareness, outperforming existing approaches.

Conclusion: OmniPSD offers a new paradigm for layered design generation and decomposition using diffusion transformers, enabling both text-to-PSD generation and image-to-PSD decomposition with transparency preservation.

Abstract: Recent advances in diffusion models have greatly improved image generation and editing, yet generating or reconstructing layered PSD files with transparent alpha channels remains highly challenging. We propose OmniPSD, a unified diffusion framework built upon the Flux ecosystem that enables both text-to-PSD generation and image-to-PSD decomposition through in-context learning. For text-to-PSD generation, OmniPSD arranges multiple target layers spatially into a single canvas and learns their compositional relationships through spatial attention, producing semantically coherent and hierarchically structured layers. For image-to-PSD decomposition, it performs iterative in-context editing, progressively extracting and erasing textual and foreground components to reconstruct editable PSD layers from a single flattened image. An RGBA-VAE is employed as an auxiliary representation module to preserve transparency without affecting structure learning. Extensive experiments on our new RGBA-layered dataset demonstrate that OmniPSD achieves high-fidelity generation, structural consistency, and transparency awareness, offering a new paradigm for layered design generation and decomposition with diffusion transformers.

Method: GLACIA framework integrates large language models with segmentation capabilities, uses the GLake-Pos dataset pipeline with spatially grounded question-answer pairs for instance-aware positional reasoning.

Result: GLACIA achieves mIoU of 87.30, surpassing CNN methods (78.55-79.01), ViTs (69.27-81.75), Geo-foundation models (76.37-87.10), and reasoning-based segmentation methods (60.12-75.66).

Conclusion: GLACIA enables intuitive disaster preparedness and informed policy-making through natural language interaction, supporting more efficient and interpretable decision-making in rapidly changing glacial environments.

Abstract: Glacial lake monitoring bears great significance in mitigating the anticipated risk of Glacial Lake Outburst Floods. However, existing segmentation methods based on convolutional neural networks (CNNs) and Vision Transformers (ViTs), remain constrained to pixel-level predictions, lacking high-level global scene semantics and human-interpretable reasoning. To address this, we introduce GLACIA (\textbf{G}lacial \textbf{LA}ke segmentation with \textbf{C}ontextual \textbf{I}nstance \textbf{A}wareness), the first framework that integrates large language models with segmentation capabilities to produce both accurate segmentation masks and corresponding spatial reasoning outputs. We construct the Glacial Lake Position Reasoning (GLake-Pos) dataset pipeline, which provides diverse, spatially grounded question-answer pairs designed to overcome the lack of instance-aware positional reasoning data in remote sensing. Comparative evaluation demonstrate that GLACIA (mIoU: 87.30) surpasses state-of-the-art method based on CNNs (mIoU: 78.55 - 79.01), ViTs (mIoU: 69.27 - 81.75), Geo-foundation models (mIoU: 76.37 - 87.10), and reasoning based segmentation methods (mIoU: 60.12 - 75.66). Our approach enables intuitive disaster preparedness and informed policy-making in the context of rapidly changing glacial environments by facilitating natural language interaction, thereby supporting more efficient and interpretable decision-making. The code is released on https://github.com/lalitmaurya47/GLACIA

Method: ROI-Packing identifies regions of interest critical to end-task accuracy, efficiently packs these ROI regions while discarding less relevant data, and does this without requiring retraining or fine-tuning of the end-task models.

Result: Comprehensive evaluations across five datasets and two tasks (object detection and instance segmentation) show up to 44.10% reduction in bitrate without compromising accuracy, and 8.88% improvement in accuracy at the same bitrate compared to VVC codec.

Conclusion: ROI-Packing provides an efficient compression solution specifically tailored for machine vision applications, offering significant bitrate savings and accuracy improvements over state-of-the-art general-purpose codecs like VVC.

Abstract: This paper introduces ROI-Packing, an efficient image compression method tailored specifically for machine vision. By prioritizing regions of interest (ROI) critical to end-task accuracy and packing them efficiently while discarding less relevant data, ROI-Packing achieves significant compression efficiency without requiring retraining or fine-tuning of end-task models. Comprehensive evaluations across five datasets and two popular tasks-object detection and instance segmentation-demonstrate up to a 44.10% reduction in bitrate without compromising end-task accuracy, along with an 8.88 % improvement in accuracy at the same bitrate compared to the state-of-the-art Versatile Video Coding (VVC) codec standardized by the Moving Picture Experts Group (MPEG).

Method: Proposes MoRel with Anchor Relay-based Bidirectional Blending (ARBB) mechanism that constructs locally canonical anchor spaces at key-frames and models inter-frame deformations at anchor level. Uses bidirectional deformations between key-frame anchors with learnable opacity blending, plus Feature-variance-guided Hierarchical Densification (FHD) for quality preservation.

Result: Achieves temporally coherent, flicker-free long-range 4D reconstruction with bounded memory usage. Introduces new SelfCap_LR dataset for evaluating long-range 4D motion. Demonstrates scalability and efficiency in dynamic Gaussian-based representations.

Conclusion: MoRel successfully addresses key challenges in long-range dynamic scene modeling, providing memory-efficient and temporally consistent reconstruction while maintaining rendering quality, advancing the state of 4D Gaussian Splatting for real-world applications.

Abstract: Recent advances in 4D Gaussian Splatting (4DGS) have extended the high-speed rendering capability of 3D Gaussian Splatting (3DGS) into the temporal domain, enabling real-time rendering of dynamic scenes. However, one of the major remaining challenges lies in modeling long-range motion-contained dynamic videos, where a naive extension of existing methods leads to severe memory explosion, temporal flickering, and failure to handle appearing or disappearing occlusions over time. To address these challenges, we propose a novel 4DGS framework characterized by an Anchor Relay-based Bidirectional Blending (ARBB) mechanism, named MoRel, which enables temporally consistent and memory-efficient modeling of long-range dynamic scenes. Our method progressively constructs locally canonical anchor spaces at key-frame time index and models inter-frame deformations at the anchor level, enhancing temporal coherence. By learning bidirectional deformations between KfA and adaptively blending them through learnable opacity control, our approach mitigates temporal discontinuities and flickering artifacts. We further introduce a Feature-variance-guided Hierarchical Densification (FHD) scheme that effectively densifies KfA’s while keeping rendering quality, based on an assigned level of feature-variance. To effectively evaluate our model’s capability to handle real-world long-range 4D motion, we newly compose long-range 4D motion-contained dataset, called SelfCap$_{\text{LR}}$. It has larger average dynamic motion magnitude, captured at spatially wider spaces, compared to previous dynamic video datasets. Overall, our MoRel achieves temporally coherent and flicker-free long-range 4D reconstruction while maintaining bounded memory usage, demonstrating both scalability and efficiency in dynamic Gaussian-based representations.

Method: 1) Create LongT2IBench with 14K long text-image pairs using Generate-Refine-Qualify protocol to convert prompts into textual graph structures (entities, attributes, relations). 2) Develop LongT2IExpert by instruction-tuning MLLMs with Hierarchical Alignment Chain-of-Thought to provide both scores and structured interpretations.

Result: LongT2IBench provides fine-grained alignment annotations through graph structures. LongT2IExpert demonstrates superiority in alignment evaluation and interpretation compared to existing methods.

Conclusion: The proposed benchmark and evaluator address the critical gap in long T2I alignment assessment, offering interpretable, structured evaluation that advances the field beyond short-prompt limitations.

Abstract: The increasing popularity of long Text-to-Image (T2I) generation has created an urgent need for automatic and interpretable models that can evaluate the image-text alignment in long prompt scenarios. However, the existing T2I alignment benchmarks predominantly focus on short prompt scenarios and only provide MOS or Likert scale annotations. This inherent limitation hinders the development of long T2I evaluators, particularly in terms of the interpretability of alignment. In this study, we contribute LongT2IBench, which comprises 14K long text-image pairs accompanied by graph-structured human annotations. Given the detail-intensive nature of long prompts, we first design a Generate-Refine-Qualify annotation protocol to convert them into textual graph structures that encompass entities, attributes, and relations. Through this transformation, fine-grained alignment annotations are achieved based on these granular elements. Finally, the graph-structed annotations are converted into alignment scores and interpretations to facilitate the design of T2I evaluation models. Based on LongT2IBench, we further propose LongT2IExpert, a LongT2I evaluator that enables multi-modal large language models (MLLMs) to provide both quantitative scores and structured interpretations through an instruction-tuning process with Hierarchical Alignment Chain-of-Thought (CoT). Extensive experiments and comparisons demonstrate the superiority of the proposed LongT2IExpert in alignment evaluation and interpretation. Data and code have been released in https://welldky.github.io/LongT2IBench-Homepage/.

Method: Developed a multimodal facial expression analysis network using CLIP architecture to integrate visual and textual features while preserving temporal dynamics. Extracted expression intensity features during various facial expressions, then processed them through an LSTM-based classification network for PD diagnosis.

Result: Achieved 93.1% accuracy in PD diagnosis, outperforming other in-vitro PD diagnostic approaches.

Conclusion: The proposed method offers a more convenient and accessible detection approach for potential PD patients, improving their diagnostic experience through non-invasive facial expression analysis.

Abstract: Parkinson’s disease (PD), a prevalent neurodegenerative disorder, significantly affects patients’ daily functioning and social interactions. To facilitate a more efficient and accessible diagnostic approach for PD, we propose a dynamic facial expression analysis-based PD auxiliary diagnosis method. This method targets hypomimia, a characteristic clinical symptom of PD, by analyzing two manifestations: reduced facial expressivity and facial rigidity, thereby facilitating the diagnosis process. We develop a multimodal facial expression analysis network to extract expression intensity features during patients’ performance of various facial expressions. This network leverages the CLIP architecture to integrate visual and textual features while preserving the temporal dynamics of facial expressions. Subsequently, the expression intensity features are processed and input into an LSTM-based classification network for PD diagnosis. Our method achieves an accuracy of 93.1%, outperforming other in-vitro PD diagnostic approaches. This technique offers a more convenient detection method for potential PD patients, improving their diagnostic experience.

Method: LoGoColor uses a ‘Local-Global’ approach: partitions scenes into subscenes and addresses both inter-subscene and intra-subscene consistency using a fine-tuned multi-view diffusion model. It generates consistently colorized training views to bypass the color averaging process.

Result: The method achieves quantitatively and qualitatively more consistent and plausible 3D colorization on complex 360° scenes than existing methods, with superior color diversity validated by a novel Color Diversity Index.

Conclusion: LoGoColor successfully addresses the color diversity preservation problem in 3D colorization by eliminating guidance averaging and ensuring multi-view consistency through a Local-Global approach with diffusion models.

Abstract: Single-channel 3D reconstruction is widely used in fields such as robotics and medical imaging. While this line of work excels at reconstructing 3D geometry, the outputs are not colored 3D models, thus 3D colorization is required for visualization. Recent 3D colorization studies address this problem by distilling 2D image colorization models. However, these approaches suffer from an inherent inconsistency of 2D image models. This results in colors being averaged during training, leading to monotonous and oversimplified results, particularly in complex 360° scenes. In contrast, we aim to preserve color diversity by generating a new set of consistently colorized training views, thereby bypassing the averaging process. Nevertheless, eliminating the averaging process introduces a new challenge: ensuring strict multi-view consistency across these colorized views. To achieve this, we propose LoGoColor, a pipeline designed to preserve color diversity by eliminating this guidance-averaging process with a `Local-Global’ approach: we partition the scene into subscenes and explicitly tackle both inter-subscene and intra-subscene consistency using a fine-tuned multi-view diffusion model. We demonstrate that our method achieves quantitatively and qualitatively more consistent and plausible 3D colorization on complex 360° scenes than existing methods, and validate its superior color diversity using a novel Color Diversity Index.

Method: Captured ~30 videos using GoPro camera with linear data recovery via inverse encoding. Trained NeRF models under different color space interpretations (linear, sRGB, GPLog, log RGB) by converting each network output to common color space before rendering and loss computation. This enforces representation learning in different color spaces while maintaining consistent evaluation.

Result: Log RGB color space consistently improves rendering quality, exhibits greater robustness across scenes, performs particularly well in low light conditions, and maintains performance with same bit-depth input images. Analysis across different network sizes and NeRF variants confirms generalization and stability of log space advantage.

Conclusion: Log RGB color space enables NeRF to learn more effective radiance field representations, improving rendering quality and robustness, especially in challenging lighting conditions, without requiring additional input data complexity.

Abstract: Neural Radiance Fields (NeRF) have achieved remarkable results in novel view synthesis, typically using sRGB images for supervision. However, little attention has been paid to the color space in which the network is learning the radiance field representation. Inspired by the BiIlluminant Dichromatic Reflection (BIDR) model, which suggests that a logarithmic transformation simplifies the separation of illumination and reflectance, we hypothesize that log RGB space enables NeRF to learn a more compact and effective representation of scene appearance. To test this, we captured approximately 30 videos using a GoPro camera, ensuring linear data recovery through inverse encoding. We trained NeRF models under various color space interpretations linear, sRGB, GPLog, and log RGB by converting each network output to a common color space before rendering and loss computation, enforcing representation learning in different color spaces. Quantitative and qualitative evaluations demonstrate that using a log RGB color space consistently improves rendering quality, exhibits greater robustness across scenes, and performs particularly well in low light conditions while using the same bit-depth input images. Further analysis across different network sizes and NeRF variants confirms the generalization and stability of the log space advantage.

Method: 1. Data equilibrium scheduling paradigm to dynamically optimize proportions of mixed training datasets from different tasks using data mixing laws. 2. Mixture-of-Experts (MoE)-driven scheduler in generative pre-training to allocate task-adaptive diffusion priors for each restoration task based on distinct degradation forms and levels.

Result: The method addresses over 50 sub-tasks across broader real-world scenarios and achieves favorable performance against state-of-the-art approaches. The balanced dataset composition enables consistent generalization and comprehensive performance across diverse tasks.

Conclusion: FoundIR-v2 demonstrates that data mixture proportions are critical for all-in-one image restoration models, and the proposed data equilibrium scheduling with MoE-driven diffusion priors effectively handles diverse restoration tasks with balanced performance.

Abstract: Recent studies have witnessed significant advances in image restoration foundation models driven by improvements in the scale and quality of pre-training data. In this work, we find that the data mixture proportions from different restoration tasks are also a critical factor directly determining the overall performance of all-in-one image restoration models. To this end, we propose a high-capacity diffusion-based image restoration foundation model, FoundIR-v2, which adopts a data equilibrium scheduling paradigm to dynamically optimize the proportions of mixed training datasets from different tasks. By leveraging the data mixing law, our method ensures a balanced dataset composition, enabling the model to achieve consistent generalization and comprehensive performance across diverse tasks. Furthermore, we introduce an effective Mixture-of-Experts (MoE)-driven scheduler into generative pre-training to flexibly allocate task-adaptive diffusion priors for each restoration task, accounting for the distinct degradation forms and levels exhibited by different tasks. Extensive experiments demonstrate that our method can address over 50 sub-tasks across a broader scope of real-world scenarios and achieves favorable performance against state-of-the-art approaches.

Method: Combines EfficientNet V2 backbone with four interpretability mechanisms: GradCAM++ attention visualization, automated ABCDE clinical criterion extraction, Fast Concept Activation Vectors (FastCAV) for concept-based explanations, and Monte Carlo Dropout uncertainty quantification.

Result: Achieves 85.61% accuracy with weighted F1 score of 0.8564 on ISIC 2019 dataset (25,331 images across 9 categories), while providing clinically meaningful explanations that align model attention with established dermatological assessment criteria.

Conclusion: Demonstrates that high classification performance can be achieved alongside comprehensive interpretability, potentially facilitating greater trust and adoption in clinical dermatology workflows by providing transparent, clinically-relevant explanations.

Abstract: Automated skin lesion classification using deep learning has shown remarkable accuracy, yet clinical adoption remains limited due to the “black box” nature of these models. We present MelanomaNet, an explainable deep learning system for multi-class skin lesion classification that addresses this gap through four complementary interpretability mechanisms. Our approach combines an EfficientNet V2 backbone with GradCAM++ attention visualization, automated ABCDE clinical criterion extraction, Fast Concept Activation Vectors (FastCAV) for concept-based explanations, and Monte Carlo Dropout uncertainty quantification. We evaluate our system on the ISIC 2019 dataset containing 25,331 dermoscopic images across 9 diagnostic categories. Our model achieves 85.61% accuracy with a weighted F1 score of 0.8564, while providing clinically meaningful explanations that align model attention with established dermatological assessment criteria. The uncertainty quantification module decomposes prediction confidence into epistemic and aleatoric components, enabling automatic flagging of unreliable predictions for clinical review. Our results demonstrate that high classification performance can be achieved alongside comprehensive interpretability, potentially facilitating greater trust and adoption in clinical dermatology workflows. The source code is available at https://github.com/suxrobgm/explainable-melanoma

Method: Three key improvements: 1) Replace 4 traditional convolution modules with SPD-Conv in Backbone to retain fine-grained information; 2) Introduce SPPFCSPC module to replace SPPF for better multi-scale feature fusion; 3) Add Triple-Stage Feature Pyramid with 160*160 small target detection head while removing redundant large target heads.

Result: On VisDrone2019-DET dataset, achieves state-of-the-art performance: precision (61.9%), recall (48.3%), mAP@0.5 (52.6%), mAP@0.5:0.95 (32.6%). Visualization shows significantly reduced miss rate for pedestrians and bicycles in occluded/dense scenes.

Conclusion: The proposed YOLOv8n-SPTS model effectively addresses small target detection challenges in autonomous driving through architectural improvements that preserve fine-grained information and enhance multi-scale feature representation.

Abstract: This paper focuses on the key issue in autonomous driving: small target recognition in dynamic perception. Existing algorithms suffer from poor detection performance due to missing small target information, scale imbalance, and occlusion. We propose an improved YOLOv8n-SPTS model, which enhances the detection accuracy of small traffic targets through three key innovations: First, optimizing the feature extraction module. In the Backbone Bottleneck structure of YOLOv8n, 4 traditional convolution modules are replaced with Space-to-Depth Convolution (SPD-Conv) modules. This module retains fine-grained information through space-to-depth conversion, reduces information loss, and enhances the ability to capture features of low-resolution small targets. Second, enhancing feature fusion capability. The Spatial Pyramid Pooling - Fast Cross Stage Partial Connection (SPPFCSPC) module is introduced to replace the original SPPF module, integrating the multi-scale feature extraction from Spatial Pyramid Pooling (SPP) and the feature fusion mechanism of Cross Stage Partial Connection (CSP), thereby improving the model’s contextual understanding of complex scenes and multi-scale feature expression ability. Third, designing a dedicated detection structure for small targets. A Triple-Stage Feature Pyramid (TSFP) structure is proposed, which adds a 160*160 small target detection head to the original detection heads to fully utilize high-resolution features in shallow layers; meanwhile, redundant large target detection heads are removed to balance computational efficiency. Comparative experiments on the VisDrone2019-DET dataset show that YOLOv8n-SPTS model ranks first in precision (61.9%), recall (48.3%), mAP@0.5 (52.6%), and mAP@0.5:0.95 (32.6%). Visualization results verify that the miss rate of small targets such as pedestrians and bicycles in occluded and dense scenes is significantly reduced.

Method: Proposes Polyp-DiFoM distillation framework that infuses semantic priors from foundation models into lightweight architectures (U-Net, U-Net++) and performs frequency domain encoding for enhanced distillation.

Result: Extensive experiments across 5 benchmark datasets (Kvasir-SEG, CVC-ClinicDB, ETIS, ColonDB, CVC-300) show Polyp-DiFoM consistently outperforms baseline models and state-of-the-art with nearly 9 times reduced computation overhead.

Conclusion: Polyp-DiFoM successfully bridges the gap between large foundation models and lightweight medical segmentation, enabling efficient and accurate polyp segmentation deployment in clinical settings.

Abstract: Accurate polyp segmentation during colonoscopy is critical for the early detection of colorectal cancer and still remains challenging due to significant size, shape, and color variations, and the camouflaged nature of polyps. While lightweight baseline models such as U-Net, U-Net++, and PraNet offer advantages in terms of easy deployment and low computational cost, they struggle to deal with the above issues, leading to limited segmentation performance. In contrast, large-scale vision foundation models such as SAM, DINOv2, OneFormer, and Mask2Former have exhibited impressive generalization performance across natural image domains. However, their direct transfer to medical imaging tasks (e.g., colonoscopic polyp segmentation) is not straightforward, primarily due to the scarcity of large-scale datasets and lack of domain-specific knowledge. To bridge this gap, we propose a novel distillation framework, Polyp-DiFoM, that transfers the rich representations of foundation models into lightweight segmentation baselines, allowing efficient and accurate deployment in clinical settings. In particular, we infuse semantic priors from the foundation models into canonical architectures such as U-Net and U-Net++ and further perform frequency domain encoding for enhanced distillation, corroborating their generalization capability. Extensive experiments are performed across five benchmark datasets, such as Kvasir-SEG, CVC-ClinicDB, ETIS, ColonDB, and CVC-300. Notably, Polyp-DiFoM consistently outperforms respective baseline models significantly, as well as the state-of-the-art model, with nearly 9 times reduced computation overhead. The code is available at https://github.com/lostinrepo/PolypDiFoM.

Method: Three key components: 1) Task-specific adapters added to frozen pre-trained image encoder for learning new knowledge; 2) Cross-task representation calibration using mixture of light-weight projectors to separate classes in unified feature space; 3) Uncertainty-guided inference strategy to select most appropriate image features for prediction.

Result: Extensive experiments on multiple datasets under various settings demonstrate superior performance compared to existing methods.

Conclusion: The proposed framework effectively addresses class confusion in continual learning by combining task-specific adaptation, cross-task calibration, and uncertainty-aware inference, achieving state-of-the-art performance.

Abstract: Class-incremental learning requires a learning system to continually learn knowledge of new classes and meanwhile try to preserve previously learned knowledge of old classes. As current state-of-the-art methods based on Vision-Language Models (VLMs) still suffer from the issue of differentiating classes across learning tasks. Here a novel VLM-based continual learning framework for image classification is proposed. In this framework, task-specific adapters are added to the pre-trained and frozen image encoder to learn new knowledge, and a novel cross-task representation calibration strategy based on a mixture of light-weight projectors is used to help better separate all learned classes in a unified feature space, alleviating class confusion across tasks. In addition, a novel inference strategy guided by prediction uncertainty is developed to more accurately select the most appropriate image feature for class prediction. Extensive experiments on multiple datasets under various settings demonstrate the superior performance of our method compared to existing ones.

Method: The authors create USE50k dataset (65,500 images from diverse public spaces) and develop DeepUSEvision framework with three components: 1) Suspicious Object Detector (enhanced YOLOv12), 2) dual DCNNs for facial expression and body-language recognition using image/landmark features, and 3) transformer-based Discriminator Network for adaptive multimodal fusion.

Result: Extensive experiments show superior accuracy, robustness, and interpretability compared to state-of-the-art approaches. The framework provides interpretable suspiciousness scores for real-time risk assessment.

Conclusion: USE50k dataset and DeepUSEvision framework establish a strong, scalable foundation for intelligent surveillance and real-time risk assessment in safety-critical applications, addressing the need for efficient suspiciousness analysis in complex environments.

Abstract: Suspiciousness estimation is critical for proactive threat detection and ensuring public safety in complex environments. This work introduces a large-scale annotated dataset, USE50k, along with a computationally efficient vision-based framework for real-time suspiciousness analysis. The USE50k dataset contains 65,500 images captured from diverse and uncontrolled environments, such as airports, railway stations, restaurants, parks, and other public areas, covering a broad spectrum of cues including weapons, fire, crowd density, abnormal facial expressions, and unusual body postures. Building on this dataset, we present DeepUSEvision, a lightweight and modular system integrating three key components, i.e., a Suspicious Object Detector based on an enhanced YOLOv12 architecture, dual Deep Convolutional Neural Networks (DCNN-I and DCNN-II) for facial expression and body-language recognition using image and landmark features, and a transformer-based Discriminator Network that adaptively fuses multimodal outputs to yield an interpretable suspiciousness score. Extensive experiments confirm the superior accuracy, robustness, and interpretability of the proposed framework compared to state-of-the-art approaches. Collectively, the USE50k dataset and the DeepUSEvision framework establish a strong and scalable foundation for intelligent surveillance and real-time risk assessment in safety-critical applications.

Method: Developed LNMBench benchmark with unified framework evaluating 10 representative label noise learning methods across 7 medical imaging datasets, 6 imaging modalities, and 3 realistic noise patterns for reproducible robustness assessment.

Result: Existing methods show substantial performance degradation under high and real-world noise, highlighting persistent challenges of class imbalance and domain variability in medical data, leading to proposed improvements for enhanced robustness.

Conclusion: LNMBench provides standardized evaluation framework revealing limitations of current methods and offers practical insights for developing noise-resilient algorithms in medical imaging, with publicly released codebase to promote reproducible research.

Abstract: Learning from noisy labels remains a major challenge in medical image analysis, where annotation demands expert knowledge and substantial inter-observer variability often leads to inconsistent or erroneous labels. Despite extensive research on learning with noisy labels (LNL), the robustness of existing methods in medical imaging has not been systematically assessed. To address this gap, we introduce LNMBench, a comprehensive benchmark for Label Noise in Medical imaging. LNMBench encompasses \textbf{10} representative methods evaluated across 7 datasets, 6 imaging modalities, and 3 noise patterns, establishing a unified and reproducible framework for robustness evaluation under realistic conditions. Comprehensive experiments reveal that the performance of existing LNL methods degrades substantially under high and real-world noise, highlighting the persistent challenges of class imbalance and domain variability in medical data. Motivated by these findings, we further propose a simple yet effective improvement to enhance model robustness under such conditions. The LNMBench codebase is publicly released to facilitate standardized evaluation, promote reproducible research, and provide practical insights for developing noise-resilient algorithms in both research and real-world medical applications.The codebase is publicly available on https://github.com/myyy777/LNMBench.

Method: Two-stage deep learning framework: (1) frame-level CS classification, (2) CS region localization in positive cases. Uses Novel Uncertainty-aware Contractive Embedding (NUCE) loss to handle severe class imbalance and feature uncertainty. Human-in-the-loop annotation pipeline for dataset curation.

Result: NUCE loss consistently improves F1-score across five transformer backbones. RF-DETR-based localization achieves state-of-the-art detection performance for thin, low-contrast CS structures. Framework trained on 13,568 frames with highly sparse CS-positive instances.

Conclusion: First successful computational framework for CS analysis in human IVF embryos, addressing critical bottleneck in embryo selection through automated, objective assessment of emerging biomarkers.

Abstract: Infertility is a major global health issue, and while in-vitro fertilization has improved treatment outcomes, embryo selection remains a critical bottleneck. Time-lapse imaging enables continuous, non-invasive monitoring of embryo development, yet most automated assessment methods rely solely on conventional morphokinetic features and overlook emerging biomarkers. Cytoplasmic Strings, thin filamentous structures connecting the inner cell mass and trophectoderm in expanded blastocysts, have been associated with faster blastocyst formation, higher blastocyst grades, and improved viability. However, CS assessment currently depends on manual visual inspection, which is labor-intensive, subjective, and severely affected by detection and subtle visual appearance. In this work, we present, to the best of our knowledge, the first computational framework for CS analysis in human IVF embryos. We first design a human-in-the-loop annotation pipeline to curate a biologically validated CS dataset from TLI videos, comprising 13,568 frames with highly sparse CS-positive instances. Building on this dataset, we propose a two-stage deep learning framework that (i) classifies CS presence at the frame level and (ii) localizes CS regions in positive cases. To address severe imbalance and feature uncertainty, we introduce the Novel Uncertainty-aware Contractive Embedding (NUCE) loss, which couples confidence-aware reweighting with an embedding contraction term to form compact, well-separated class clusters. NUCE consistently improves F1-score across five transformer backbones, while RF-DETR-based localization achieves state-of-the-art (SOTA) detection performance for thin, low-contrast CS structures. The source code will be made publicly available at: https://github.com/HamadYA/CS_Detection.

Method: Analyzes attention patterns in MM-DiT models for text-related tokens, then applies latent guidance during early denoising steps using two novel loss functions that align textual content tokens with text regions in the image.

Result: Achieves state-of-the-art performance in test-time text rendering with significant gains in recall, strong OCR accuracy, and improved CLIP scores.

Conclusion: TextGuider provides an effective training-free solution for encouraging accurate and complete text appearance in diffusion-based text-to-image models by addressing the overlooked problem of text omission.

Abstract: Despite recent advances, diffusion-based text-to-image models still struggle with accurate text rendering. Several studies have proposed fine-tuning or training-free refinement methods for accurate text rendering. However, the critical issue of text omission, where the desired text is partially or entirely missing, remains largely overlooked. In this work, we propose TextGuider, a novel training-free method that encourages accurate and complete text appearance by aligning textual content tokens and text regions in the image. Specifically, we analyze attention patterns in MM-DiT models, particularly for text-related tokens intended to be rendered in the image. Leveraging this observation, we apply latent guidance during the early stage of denoising steps based on two loss functions that we introduce. Our method achieves state-of-the-art performance in test-time text rendering, with significant gains in recall and strong results in OCR accuracy and CLIP score.

Method: Uses learned visual transformations that obscure sensitive/task-irrelevant information while retaining task-essential features, validated across three industrial use cases with real-world data.

Result: Task-specific obfuscation enables effective monitoring with reduced privacy risks, demonstrating framework readiness for real-world adoption and establishing trust with industrial partners.

Conclusion: The framework is ready for real-world adoption and provides cross-domain recommendations for responsible, human-centric AI deployment in industry.

Abstract: The adoption of AI-powered computer vision in industry is often constrained by the need to balance operational utility with worker privacy. Building on our previously proposed privacy-preserving framework, this paper presents its first comprehensive validation on real-world data collected directly by industrial partners in active production environments. We evaluate the framework across three representative use cases: woodworking production monitoring, human-aware AGV navigation, and multi-camera ergonomic risk assessment. The approach employs learned visual transformations that obscure sensitive or task-irrelevant information while retaining features essential for task performance. Through both quantitative evaluation of the privacy-utility trade-off and qualitative feedback from industrial partners, we assess the framework’s effectiveness, deployment feasibility, and trust implications. Results demonstrate that task-specific obfuscation enables effective monitoring with reduced privacy risks, establishing the framework’s readiness for real-world adoption and providing cross-domain recommendations for responsible, human-centric AI deployment in industry.

Method: Video-QTR redefines video comprehension as a query-guided reasoning process. Instead of encoding every frame, it dynamically allocates perceptual resources based on semantic query intent, creating an adaptive feedback loop between reasoning and perception.

Result: Extensive experiments across five benchmarks (MSVD-QA, Activity Net-QA, Movie Chat, and Video MME) show Video-QTR achieves state-of-the-art performance while reducing input frame consumption by up to 73%.

Conclusion: Query-driven temporal reasoning provides an efficient and scalable solution for video understanding, overcoming limitations of traditional process-then-reason paradigms.

Abstract: The rapid development of multimodal large-language models (MLLMs) has significantly expanded the scope of visual language reasoning, enabling unified systems to interpret and describe complex visual content. However, applying these models to long-video understanding remains computationally intensive. Dense frame encoding generates excessive visual tokens, leading to high memory consumption, redundant computation, and limited scalability in real-world applications. This inefficiency highlights a key limitation of the traditional process-then-reason paradigm, which analyzes visual streams exhaustively before semantic reasoning. To address this challenge, we introduce Video-QTR (Query-Driven Temporal Reasoning), a lightweight framework that redefines video comprehension as a query-guided reasoning process. Instead of encoding every frame, Video-QTR dynamically allocates perceptual resources based on the semantic intent of the query, creating an adaptive feedback loop between reasoning and perception. Extensive experiments across five benchmarks: MSVD-QA, Activity Net-QA, Movie Chat, and Video MME demonstrate that Video-QTR achieves state-of-the-art performance while reducing input frame consumption by up to 73%. These results confirm that query-driven temporal reasoning provides an efficient and scalable solution for video understanding.

Method: Proposes a Video Vision Transformer (ViViT)-based framework with temporal-spatial fusion embedding. Uses non-overlapping tubelets extracted via 3D convolution with constrained temporal span (t=2) to ensure local temporal coherence while reducing cross-day information degradation. Tests both MSI-only and SAR-MSI fusion scenarios.

Result: On 2020 Traill County data: MTS-ViViT achieves 2.23% MSE reduction vs MTS-ViT baseline; SMTS-ViViT achieves 10.33% improvement with SAR integration vs SMTS-ViT baseline. The framework effectively enhances spectral reconstruction quality.

Conclusion: The proposed ViViT-based framework with temporal-spatial fusion embedding significantly improves cloud-covered MSI reconstruction, enabling more robust agricultural monitoring through better spectral information recovery.

Abstract: Cloud cover in multispectral imagery (MSI) significantly hinders early-season crop mapping by corrupting spectral information. Existing Vision Transformer(ViT)-based time-series reconstruction methods, like SMTS-ViT, often employ coarse temporal embeddings that aggregate entire sequences, causing substantial information loss and reducing reconstruction accuracy. To address these limitations, a Video Vision Transformer (ViViT)-based framework with temporal-spatial fusion embedding for MSI reconstruction in cloud-covered regions is proposed in this study. Non-overlapping tubelets are extracted via 3D convolution with constrained temporal span $(t=2)$, ensuring local temporal coherence while reducing cross-day information degradation. Both MSI-only and SAR-MSI fusion scenarios are considered during the experiments. Comprehensive experiments on 2020 Traill County data demonstrate notable performance improvements: MTS-ViViT achieves a 2.23% reduction in MSE compared to the MTS-ViT baseline, while SMTS-ViViT achieves a 10.33% improvement with SAR integration over the SMTS-ViT baseline. The proposed framework effectively enhances spectral reconstruction quality for robust agricultural monitoring.

Method: Repurposes pretrained video generator for monocular-to-stereo conversion with geometry-aware regularization for 3D fidelity and spatio-temporal tiling for high-resolution synthesis.

Result: Outperforms prior methods with superior visual fidelity and geometric consistency, trained on a curated dataset of over 11M stereo video frames aligned to natural human IPD.

Conclusion: StereoWorld provides an effective solution for high-quality stereo video generation, addressing the cost and quality challenges in current stereo video production.

Abstract: The growing adoption of XR devices has fueled strong demand for high-quality stereo video, yet its production remains costly and artifact-prone. To address this challenge, we present StereoWorld, an end-to-end framework that repurposes a pretrained video generator for high-fidelity monocular-to-stereo video generation. Our framework jointly conditions the model on the monocular video input while explicitly supervising the generation with a geometry-aware regularization to ensure 3D structural fidelity. A spatio-temporal tiling scheme is further integrated to enable efficient, high-resolution synthesis. To enable large-scale training and evaluation, we curate a high-definition stereo video dataset containing over 11M frames aligned to natural human interpupillary distance (IPD). Extensive experiments demonstrate that StereoWorld substantially outperforms prior methods, generating stereo videos with superior visual fidelity and geometric consistency. The project webpage is available at https://ke-xing.github.io/StereoWorld/.

Method: Used controlled synthetic datasets, principal component analysis (PCA), and similarity metrics to analyze latent representations in Stable Diffusion. Systematically examined how color information is encoded across different latent channels.

Result: Color information is encoded along circular, opponent axes predominantly in latent channels c_3 and c_4, while intensity and shape are primarily represented in channels c_1 and c_2. The latent space exhibits an interpretable structure aligned with efficient coding representation.

Conclusion: Stable Diffusion’s latent space has an interpretable structure where different perceptual attributes are separated into specific channels. These insights provide foundation for model understanding, editing applications, and designing more disentangled generative frameworks.

Abstract: Recent advances in diffusion-based generative models have achieved remarkable visual fidelity, yet a detailed understanding of how specific perceptual attributes - such as color and shape - are internally represented remains limited. This work explores how color is encoded in a generative model through a systematic analysis of the latent representations in Stable Diffusion. Through controlled synthetic datasets, principal component analysis (PCA) and similarity metrics, we reveal that color information is encoded along circular, opponent axes predominantly captured in latent channels c_3 and c_4, whereas intensity and shape are primarily represented in channels c_1 and c_2. Our findings indicate that the latent space of Stable Diffusion exhibits an interpretable structure aligned with a efficient coding representation. These insights provide a foundation for future work in model understanding, editing applications, and the design of more disentangled generative frameworks.

Method: ASSIST-3D features three key innovations: 1) Heterogeneous Object Selection with random sampling from CAD collections for geometric/contextual diversity, 2) LLM-guided Scene Layout Generation with depth-first search for reasonable placements, and 3) Realistic Point Cloud Construction via multi-view RGB-D rendering and fusion to mimic real sensor data.

Result: Models trained with ASSIST-3D-generated data significantly outperform existing methods on ScanNetV2, ScanNet++, and S3DIS benchmarks. The pipeline also demonstrates superiority over existing 3D scene synthesis approaches.

Conclusion: ASSIST-3D successfully addresses the data scarcity problem in class-agnostic 3D instance segmentation by generating diverse, realistic, and contextually reasonable synthetic training data that enhances model generalization capabilities.

Abstract: Class-agnostic 3D instance segmentation tackles the challenging task of segmenting all object instances, including previously unseen ones, without semantic class reliance. Current methods struggle with generalization due to the scarce annotated 3D scene data or noisy 2D segmentations. While synthetic data generation offers a promising solution, existing 3D scene synthesis methods fail to simultaneously satisfy geometry diversity, context complexity, and layout reasonability, each essential for this task. To address these needs, we propose an Adapted 3D Scene Synthesis pipeline for class-agnostic 3D Instance SegmenTation, termed as ASSIST-3D, to synthesize proper data for model generalization enhancement. Specifically, ASSIST-3D features three key innovations, including 1) Heterogeneous Object Selection from extensive 3D CAD asset collections, incorporating randomness in object sampling to maximize geometric and contextual diversity; 2) Scene Layout Generation through LLM-guided spatial reasoning combined with depth-first search for reasonable object placements; and 3) Realistic Point Cloud Construction via multi-view RGB-D image rendering and fusion from the synthetic scenes, closely mimicking real-world sensor data acquisition. Experiments on ScanNetV2, ScanNet++, and S3DIS benchmarks demonstrate that models trained with ASSIST-3D-generated data significantly outperform existing methods. Further comparisons underscore the superiority of our purpose-built pipeline over existing 3D scene synthesis approaches.

Method: FUSER encodes each scan into low-resolution superpoint features via sparse 3D CNN, performs efficient intra- and inter-scan reasoning through Geometric Alternating Attention, and transfers 2D attention priors from foundation models. FUSER-DF adds SE(3)^N diffusion refinement to correct estimates via denoising in joint SE(3)^N space.

Result: Extensive experiments on 3DMatch, ScanNet and ArkitScenes demonstrate superior registration accuracy and outstanding computational efficiency compared to conventional methods.

Conclusion: FUSER represents the first feed-forward multiview registration transformer that directly predicts global poses without pairwise estimation, offering a more efficient and accurate approach to multiview point cloud registration.

Abstract: Registration of multiview point clouds conventionally relies on extensive pairwise matching to build a pose graph for global synchronization, which is computationally expensive and inherently ill-posed without holistic geometric constraints. This paper proposes FUSER, the first feed-forward multiview registration transformer that jointly processes all scans in a unified, compact latent space to directly predict global poses without any pairwise estimation. To maintain tractability, FUSER encodes each scan into low-resolution superpoint features via a sparse 3D CNN that preserves absolute translation cues, and performs efficient intra- and inter-scan reasoning through a Geometric Alternating Attention module. Particularly, we transfer 2D attention priors from off-the-shelf foundation models to enhance 3D feature interaction and geometric consistency. Building upon FUSER, we further introduce FUSER-DF, an SE(3)$^N$ diffusion refinement framework to correct FUSER’s estimates via denoising in the joint SE(3)$^N$ space. FUSER acts as a surrogate multiview registration model to construct the denoiser, and a prior-conditioned SE(3)$^N$ variational lower bound is derived for denoising supervision. Extensive experiments on 3DMatch, ScanNet and ArkitScenes demonstrate that our approach achieves the superior registration accuracy and outstanding computational efficiency.

Method: The framework introduces a Learnable HSI (LHSI) color space based on cylindrical color models that naturally separates luminance from chromatic components. It includes dedicated parameters to enhance disentanglement and learnable mapping for adaptive refinement, plus a new Mamba-based network tailored to the LHSI color space characteristics.

Result: Experimental results on benchmark datasets demonstrate the superiority of the proposed method over existing approaches.

Conclusion: The work highlights the potential of perception-inspired color space design in computational photography for effective white balance correction, with source code publicly available.

Abstract: White balance (WB) is a key step in the image signal processor (ISP) pipeline that mitigates color casts caused by varying illumination and restores the scene’s true colors. Currently, sRGB-based WB editing for post-ISP WB correction is widely used to address color constancy failures in the ISP pipeline when the original camera RAW is unavailable. However, additive color models (e.g., sRGB) are inherently limited by fixed nonlinear transformations and entangled color channels, which often impede their generalization to complex lighting conditions. To address these challenges, we propose a novel framework for WB correction that leverages a perception-inspired Learnable HSI (LHSI) color space. Built upon a cylindrical color model that naturally separates luminance from chromatic components, our framework further introduces dedicated parameters to enhance this disentanglement and learnable mapping to adaptively refine the flexibility. Moreover, a new Mamba-based network is introduced, which is tailored to the characteristics of the proposed LHSI color space. Experimental results on benchmark datasets demonstrate the superiority of our method, highlighting the potential of perception-inspired color space design in computational photography. The source code is available at https://github.com/YangCheng58/WB_Color_Space.

Method: Developed multimodal deep learning framework integrating: 1) structured clinical variables (demographics, comorbidities, medications, lab results), 2) 3D CNN trained on CT volumes, and 3) transformer-enhanced 2D segmentation model for SDH detection and localization. Used 25,315 head CT studies (3,774 with confirmed SDH). Employed greedy ensemble strategy to combine complementary predictors.

Result: Clinical variables alone: AUC 0.75. Imaging models: 3D CNN AUC 0.922, segmentation model AUC 0.926. Multimodal ensemble achieved best performance: AUC 0.9407 (95% CI 0.930-0.951) with anatomically meaningful localization maps consistent with known SDH patterns.

Conclusion: The multimodal interpretable framework provides rapid and accurate SDH detection and localization with transparent, anatomically grounded outputs. Integration into radiology workflows could streamline triage, reduce time to intervention, and improve consistency in SDH management.

Abstract: Background. Subdural hematoma (SDH) is a common neurosurgical emergency, with increasing incidence in aging populations. Rapid and accurate identification is essential to guide timely intervention, yet existing automated tools focus primarily on detection and provide limited interpretability or spatial localization. There remains a need for transparent, high-performing systems that integrate multimodal clinical and imaging information to support real-time decision-making. Methods. We developed a multimodal deep-learning framework that integrates structured clinical variables, a 3D convolutional neural network trained on CT volumes, and a transformer-enhanced 2D segmentation model for SDH detection and localization. Using 25,315 head CT studies from Hartford HealthCare (2015–2024), of which 3,774 (14.9%) contained clinician-confirmed SDH, tabular models were trained on demographics, comorbidities, medications, and laboratory results. Imaging models were trained to detect SDH and generate voxel-level probability maps. A greedy ensemble strategy combined complementary predictors. Findings. Clinical variables alone provided modest discriminatory power (AUC 0.75). Convolutional models trained on CT volumes and segmentation-derived maps achieved substantially higher accuracy (AUCs 0.922 and 0.926). The multimodal ensemble integrating all components achieved the best overall performance (AUC 0.9407; 95% CI, 0.930–0.951) and produced anatomically meaningful localization maps consistent with known SDH patterns. Interpretation. This multimodal, interpretable framework provides rapid and accurate SDH detection and localization, achieving high detection performance and offering transparent, anatomically grounded outputs. Integration into radiology workflows could streamline triage, reduce time to intervention, and improve consistency in SDH management.

Method: Uses view-specific hyperbolic encoders to embed features into Lorentz manifold for hierarchical semantic modeling, introduces hyperbolic sliced-Wasserstein distance to align manifold distributions across views, and employs soft cluster assignments for cross-view semantic consistency.

Result: Extensive experiments on multiple benchmarking datasets show state-of-the-art clustering performance.

Conclusion: The proposed WAH framework effectively addresses global semantic consistency in multi-view clustering through hyperbolic representations and Wasserstein alignment, achieving superior performance.

Abstract: Multi-view clustering (MVC) aims to uncover the latent structure of multi-view data by learning view-common and view-specific information. Although recent studies have explored hyperbolic representations for better tackling the representation gap between different views, they focus primarily on instance-level alignment and neglect global semantic consistency, rendering them vulnerable to view-specific information (\textit{e.g.}, noise and cross-view discrepancies). To this end, this paper proposes a novel Wasserstein-Aligned Hyperbolic (WAH) framework for multi-view clustering. Specifically, our method exploits a view-specific hyperbolic encoder for each view to embed features into the Lorentz manifold for hierarchical semantic modeling. Whereafter, a global semantic loss based on the hyperbolic sliced-Wasserstein distance is introduced to align manifold distributions across views. This is followed by soft cluster assignments to encourage cross-view semantic consistency. Extensive experiments on multiple benchmarking datasets show that our method can achieve SOTA clustering performance.

Method: Two model-agnostic causal principles: 1) Remove direct input of dialog history to answer models to avoid harmful shortcut bias; 2) Address unobserved confounders between history, question, and answer using causal intervention algorithms that differ from traditional likelihood estimation.

Result: The principles significantly improve almost every existing VisDial model to state-of-the-art performance on the Visual Dialog Challenge 2019 leaderboard, demonstrating broad applicability and effectiveness.

Conclusion: Causal analysis reveals overlooked relationships in Visual Dialog that, when properly addressed through the proposed principles, substantially improves model performance across different architectures, suggesting the importance of causal reasoning in dialog systems.

Abstract: This paper unravels the design tricks adopted by us, the champion team MReaL-BDAI, for Visual Dialog Challenge 2019: two causal principles for improving Visual Dialog (VisDial). By “improving”, we mean that they can promote almost every existing VisDial model to the state-of-the-art performance on the leader-board. Such a major improvement is only due to our careful inspection on the causality behind the model and data, finding that the community has overlooked two causalities in VisDial. Intuitively, Principle 1 suggests: we should remove the direct input of the dialog history to the answer model, otherwise a harmful shortcut bias will be introduced; Principle 2 says: there is an unobserved confounder for history, question, and answer, leading to spurious correlations from training data. In particular, to remove the confounder suggested in Principle 2, we propose several causal intervention algorithms, which make the training fundamentally different from the traditional likelihood estimation. Note that the two principles are model-agnostic, so they are applicable in any VisDial model. The code is available at https://github.com/simpleshinobu/visdial-principles.

Method: Proposes Match-ControlNet, a controllable 2D generative model that ensures 2D-3D geometric consistency through depth-conditioned generation and promotes cross-view texture consistency via coupled conditional denoising and prompt guidance.

Result: Extensive experiments on 3DMatch and ScanNet datasets verify the effectiveness of the approach, showing improved registration performance that can be integrated into various existing methods.

Conclusion: The proposed generative 3D registration paradigm successfully bridges 2D generative models with 3D matching tasks, providing a general framework that enhances registration performance through geometry-color feature fusion.

Abstract: In this paper, we propose a novel 3D registration paradigm, Generative Point Cloud Registration, which bridges advanced 2D generative models with 3D matching tasks to enhance registration performance. Our key idea is to generate cross-view consistent image pairs that are well-aligned with the source and target point clouds, enabling geometry-color feature fusion to facilitate robust matching. To ensure high-quality matching, the generated image pair should feature both 2D-3D geometric consistency and cross-view texture consistency. To achieve this, we introduce Match-ControlNet, a matching-specific, controllable 2D generative model. Specifically, it leverages the depth-conditioned generation capability of ControlNet to produce images that are geometrically aligned with depth maps derived from point clouds, ensuring 2D-3D geometric consistency. Additionally, by incorporating a coupled conditional denoising scheme and coupled prompt guidance, Match-ControlNet further promotes cross-view feature interaction, guiding texture consistency generation. Our generative 3D registration paradigm is general and could be seamlessly integrated into various registration methods to enhance their performance. Extensive experiments on 3DMatch and ScanNet datasets verify the effectiveness of our approach.

Method: 1) Create HeadSwapBench - first cross-identity paired dataset using video editing model to synthesize fake swapping inputs; 2) DirectSwap framework extends image U-Net to video diffusion with motion module; 3) MEAR loss reweights diffusion loss using frame-difference magnitudes and facial-landmark proximity for better motion/expression coherence.

Result: DirectSwap achieves state-of-the-art visual quality, identity fidelity, and motion/expression consistency across diverse in-the-wild video scenes. Will release code and HeadSwapBench dataset.

Conclusion: The proposed paired dataset and mask-free framework with motion-aware training significantly improves video head swapping quality by addressing limitations of previous mask-based methods and lack of supervision data.

Abstract: Video head swapping aims to replace the entire head of a video subject, including facial identity, head shape, and hairstyle, with that of a reference image, while preserving the target body, background, and motion dynamics. Due to the lack of ground-truth paired swapping data, prior methods typically train on cross-frame pairs of the same person within a video and rely on mask-based inpainting to mitigate identity leakage. Beyond potential boundary artifacts, this paradigm struggles to recover essential cues occluded by the mask, such as facial pose, expressions, and motion dynamics. To address these issues, we prompt a video editing model to synthesize new heads for existing videos as fake swapping inputs, while maintaining frame-synchronized facial poses and expressions. This yields HeadSwapBench, the first cross-identity paired dataset for video head swapping, which supports both training (\TrainNum{} videos) and benchmarking (\TestNum{} videos) with genuine outputs. Leveraging this paired supervision, we propose DirectSwap, a mask-free, direct video head-swapping framework that extends an image U-Net into a video diffusion model with a motion module and conditioning inputs. Furthermore, we introduce the Motion- and Expression-Aware Reconstruction (MEAR) loss, which reweights the diffusion loss per pixel using frame-difference magnitudes and facial-landmark proximity, thereby enhancing cross-frame coherence in motion and expressions. Extensive experiments demonstrate that DirectSwap achieves state-of-the-art visual quality, identity fidelity, and motion and expression consistency across diverse in-the-wild video scenes. We will release the source code and the HeadSwapBench dataset to facilitate future research.

Method: Comparative study of different ViT architectures (pure, hierarchical, hybrid) vs traditional CNNs across multiple tasks: object recognition, detection, and medical image classification. Experiments conducted on standard datasets (ImageNet, COCO) and medical dataset (ChestX-ray14), including evaluation of data augmentation techniques on medical images.

Result: Hybrid and hierarchical transformers (especially Swin and CvT) provide optimal balance between accuracy and computational efficiency. Data augmentation significantly improves performance on medical images, particularly with Swin Transformer. Vision Transformers generally outperform traditional CNNs, especially in tasks requiring global context understanding like medical imaging.

Conclusion: Vision Transformers are competitive and often superior to traditional CNNs, particularly for tasks requiring global image understanding. Hybrid/hierarchical architectures offer practical advantages, and data augmentation enhances their effectiveness in specialized domains like medical imaging.

Abstract: Convolutional Neural Networks (CNNs) for computer vision sometimes struggle with understanding images in a global context, as they mainly focus on local patterns. On the other hand, Vision Transformers (ViTs), inspired by models originally created for language processing, use self-attention mechanisms, which allow them to understand relationships across the entire image. In this paper, we compare different types of ViTs (pure, hierarchical, and hybrid) against traditional CNN models across various tasks, including object recognition, detection, and medical image classification. We conduct thorough tests on standard datasets like ImageNet for image classification and COCO for object detection. Additionally, we apply these models to medical imaging using the ChestX-ray14 dataset. We find that hybrid and hierarchical transformers, especially Swin and CvT, offer a strong balance between accuracy and computational resources. Furthermore, by experimenting with data augmentation techniques on medical images, we discover significant performance improvements, particularly with the Swin Transformer model. Overall, our results indicate that Vision Transformers are competitive and, in many cases, outperform traditional CNNs, especially in scenarios requiring the understanding of global visual contexts like medical imaging.

Method: Proposes Motion Conditioned Diffusion Model (MCDM) with Motion and Appearance Feature Extractor (MAFE) that disentangles motion/appearance features, enhanced by re-identification loss and optical flow loss.

Result: Achieves competitive video generation on EchoNet-Dynamic dataset, producing temporally coherent and clinically realistic echocardiography sequences without manual labels.

Conclusion: Demonstrates potential of self-supervised conditioning for scalable echocardiography synthesis, offering a label-free solution to data scarcity in medical imaging.

Abstract: Ultrasound echocardiography is essential for the non-invasive, real-time assessment of cardiac function, but the scarcity of labelled data, driven by privacy restrictions and the complexity of expert annotation, remains a major obstacle for deep learning methods. We propose the Motion Conditioned Diffusion Model (MCDM), a label-free latent diffusion framework that synthesises realistic echocardiography videos conditioned on self-supervised motion features. To extract these features, we design the Motion and Appearance Feature Extractor (MAFE), which disentangles motion and appearance representations from videos. Feature learning is further enhanced by two auxiliary objectives: a re-identification loss guided by pseudo appearance features and an optical flow loss guided by pseudo flow fields. Evaluated on the EchoNet-Dynamic dataset, MCDM achieves competitive video generation performance, producing temporally coherent and clinically realistic sequences without reliance on manual labels. These results demonstrate the potential of self-supervised conditioning for scalable echocardiography synthesis. Our code is available at https://github.com/ZheLi2020/LabelfreeMCDM.

Method: Uses motion feature extraction to capture temporal dynamics, followed by class-wise graph construction and representative sample selection via the Infomap algorithm to create diverse synthetic video subsets.

Result: Achieved 69.38% test accuracy on EchoNet-Dynamic datasets using only 25 synthetic videos, demonstrating effectiveness and scalability for medical video dataset distillation.

Conclusion: The proposed approach successfully distills compact synthetic echocardiographic video datasets while preserving essential characteristics, offering a scalable solution for medical video data management and analysis.

Abstract: Echocardiography playing a critical role in the diagnosis and monitoring of cardiovascular diseases as a non-invasive real-time assessment of cardiac structure and function. However, the growing scale of echocardiographic video data presents significant challenges in terms of storage, computation, and model training efficiency. Dataset distillation offers a promising solution by synthesizing a compact, informative subset of data that retains the key clinical features of the original dataset. In this work, we propose a novel approach for distilling a compact synthetic echocardiographic video dataset. Our method leverages motion feature extraction to capture temporal dynamics, followed by class-wise graph construction and representative sample selection using the Infomap algorithm. This enables us to select a diverse and informative subset of synthetic videos that preserves the essential characteristics of the original dataset. We evaluate our approach on the EchoNet-Dynamic datasets and achieve a test accuracy of (69.38%) using only (25) synthetic videos. These results demonstrate the effectiveness and scalability of our method for medical video dataset distillation.

Method: Introduces FunPhase, a functional periodic autoencoder that learns phase manifolds for motion. Replaces discrete temporal decoding with function-space formulation, enabling smooth trajectories that can be sampled at arbitrary temporal resolutions.

Result: Achieves substantially lower reconstruction error than prior periodic autoencoder baselines. Supports downstream tasks like super-resolution and partial-body motion completion, generalizes across skeletons and datasets, and performs on par with state-of-the-art motion generation methods.

Conclusion: FunPhase provides a scalable, generalizable approach to motion learning that unifies prediction and generation within a single interpretable manifold, enabling broader applications beyond existing methods.

Abstract: Learning natural body motion remains challenging due to the strong coupling between spatial geometry and temporal dynamics. Embedding motion in phase manifolds, latent spaces that capture local periodicity, has proven effective for motion prediction; however, existing approaches lack scalability and remain confined to specific settings. We introduce FunPhase, a functional periodic autoencoder that learns a phase manifold for motion and replaces discrete temporal decoding with a function-space formulation, enabling smooth trajectories that can be sampled at arbitrary temporal resolutions. FunPhase supports downstream tasks such as super-resolution and partial-body motion completion, generalizes across skeletons and datasets, and unifies motion prediction and generation within a single interpretable manifold. Our model achieves substantially lower reconstruction error than prior periodic autoencoder baselines while enabling a broader range of applications and performing on par with state-of-the-art motion generation methods.

Method: Proposes Geom-Seg VecSet, a unified geometry-segmentation latent representation that jointly encodes object geometry and part-level structure. Uses UniPart, a two-stage latent diffusion framework: 1) joint geometry generation and latent part segmentation, 2) part-level diffusion conditioned on both whole-object and part-specific latents. Includes dual-space generation scheme predicting part latents in both global and canonical spaces for enhanced geometric fidelity.

Result: Extensive experiments demonstrate that UniPart achieves superior segmentation controllability and part-level geometric quality compared with existing approaches.

Conclusion: The proposed unified representation and two-stage diffusion framework effectively addresses limitations of existing part-level 3D generation methods, enabling better control over part granularity and geometric quality without relying on external segmentation models.

Abstract: Part-level 3D generation is essential for applications requiring decomposable and structured 3D synthesis. However, existing methods either rely on implicit part segmentation with limited granularity control or depend on strong external segmenters trained on large annotated datasets. In this work, we observe that part awareness emerges naturally during whole-object geometry learning and propose Geom-Seg VecSet, a unified geometry-segmentation latent representation that jointly encodes object geometry and part-level structure. Building on this representation, we introduce UniPart, a two-stage latent diffusion framework for image-guided part-level 3D generation. The first stage performs joint geometry generation and latent part segmentation, while the second stage conditions part-level diffusion on both whole-object and part-specific latents. A dual-space generation scheme further enhances geometric fidelity by predicting part latents in both global and canonical spaces. Extensive experiments demonstrate that UniPart achieves superior segmentation controllability and part-level geometric quality compared with existing approaches.

Method: DAPO uses progressive tuning to learn hybrid defect-aware prompts with fixed textual anchors and learnable token embeddings, aligning anomaly-relevant image features with corresponding text semantics for zero-shot multi-type and binary anomaly detection.

Result: DAPO achieves 3.7% average improvement in AUROC and average precision at image level under distribution shift, and 6.5% average improvement in localizing novel anomaly types under zero-shot settings across multiple benchmarks.

Conclusion: Learning defect-aware prompts through progressive tuning effectively captures fine-grained anomaly semantics, improving both detection and localization performance while eliminating the need for manual prompt engineering.

Abstract: Recent vision language models (VLMs) like CLIP have demonstrated impressive anomaly detection performance under significant distribution shift by utilizing high-level semantic information through text prompts. However, these models often neglect fine-grained details, such as which kind of anomalies, like “hole”, “cut”, “scratch” that could provide more specific insight into the nature of anomalies. We argue that recognizing fine-grained anomaly types 1) enriches the representation of “abnormal” with structured semantics, narrowing the gap between coarse anomaly signals and fine-grained defect categories; 2) enables manufacturers to understand the root causes of the anomaly and implement more targeted and appropriate corrective measures quickly. While incorporating such detailed semantic information is crucial, designing handcrafted prompts for each defect type is both time-consuming and susceptible to human bias. For this reason, we introduce DAPO, a novel approach for Defect-aware Prompt Optimization based on progressive tuning for the zero-shot multi-type and binary anomaly detection and segmentation under distribution shifts. Our approach aligns anomaly-relevant image features with their corresponding text semantics by learning hybrid defect-aware prompts with both fixed textual anchors and learnable token embeddings. We conducted experiments on public benchmarks (MPDD, VisA, MVTec-AD, MAD, and Real-IAD) and an internal dataset. The results suggest that compared to the baseline models, DAPO achieves a 3.7% average improvement in AUROC and average precision metrics at the image level under distribution shift, and a 6.5% average improvement in localizing novel anomaly types under zero-shot settings.

Method: Proposes Context-Aware Modulated Attention (CAMA), a training-free plug-and-play method that dynamically adjusts attention logits based on input in-context sequences. Uses two-stage modulation process to strengthen attention to semantically important tokens, especially visual ones.

Result: Across four LVLMs and seven benchmarks, CAMA consistently outperforms vanilla models and baselines, showing clear effectiveness and generalization. It activates intended benefits of prompt engineering methods and remains robust across different sequence configurations.

Conclusion: CAMA opens up new directions for improving multimodal reasoning through deeper understanding of attention dynamics, addressing inherent weaknesses in LVLM self-attention that hinder effective in-context learning.

Abstract: Multimodal in-context learning (ICL) is becoming a key capability that allows large vision-language models (LVLMs) to adapt to novel tasks without parameter updates, which expands their usefulness in many real-world applications. However, ICL performance remains unstable even when the in-context demonstrations (ICDs) are well matched, showing that LVLMs still struggle to make full use of the provided context. While existing work mainly focuses on prompt engineering or post-hoc logit calibration, we study the attention mechanisms inside LVLMs to address their inherent limitations. We identify two important weaknesses in their self-attention that hinder effective ICL. To address these weaknesses, we propose Context-Aware Modulated Attention (CAMA), a training-free and plug-and-play method that dynamically adjusts attention logits based on the input in-context sequence. CAMA uses a two-stage modulation process that strengthens attention to semantically important tokens, especially visual ones. Across four LVLMs and seven benchmarks, CAMA consistently outperforms vanilla models and baselines, showing clear effectiveness and generalization. It can also activate the intended benefits of prompt engineering methods and remains robust across different sequence configurations. Therefore, CAMA opens up new directions for improving multimodal reasoning through a deeper understanding of attention dynamics.

Method: OSSDet framework integrates spectral and spatial information with object-aware cues: 1) cascaded spectral-spatial modulation structure to optimize target perception, 2) aggregates spectrally related features using spectral similarities to reinforce intra-object correlations, 3) suppresses irrelevant background via object-aware masking, and 4) cross-spectral attention refines object representations under explicit object-aware guidance.

Result: Extensive experiments demonstrate that OSSDet outperforms existing methods with comparable parameters and efficiency. The MODA dataset provides 14,041 multispectral images and 330,191 annotations across diverse challenging scenarios.

Conclusion: The paper introduces both a comprehensive dataset (MODA) and an effective framework (OSSDet) for multispectral aerial object detection, addressing the data scarcity problem while providing advanced techniques to leverage spectral information for improved detection performance in challenging aerial scenarios.

Abstract: Aerial object detection faces significant challenges in real-world scenarios, such as small objects and extensive background interference, which limit the performance of RGB-based detectors with insufficient discriminative information. Multispectral images (MSIs) capture additional spectral cues across multiple bands, offering a promising alternative. However, the lack of training data has been the primary bottleneck to exploiting the potential of MSIs. To address this gap, we introduce the first large-scale dataset for Multispectral Object Detection in Aerial images (MODA), which comprises 14,041 MSIs and 330,191 annotations across diverse, challenging scenarios, providing a comprehensive data foundation for this field. Furthermore, to overcome challenges inherent to aerial object detection using MSIs, we propose OSSDet, a framework that integrates spectral and spatial information with object-aware cues. OSSDet employs a cascaded spectral-spatial modulation structure to optimize target perception, aggregates spectrally related features by exploiting spectral similarities to reinforce intra-object correlations, and suppresses irrelevant background via object-aware masking. Moreover, cross-spectral attention further refines object-related representations under explicit object-aware guidance. Extensive experiments demonstrate that OSSDet outperforms existing methods with comparable parameters and efficiency.

Method: Conformal Hallucination Estimation Metric (CHEM) uses wavelet and shearlet representations to extract hallucination features and conformalized quantile regression for distribution-free assessment of hallucination levels.

Result: Tested on CANDELS astronomical image dataset with U-Net, SwinUNet, and Learnlets models, providing new perspectives on hallucination in deep learning-based image processing.

Conclusion: CHEM enables efficient identification and quantification of hallucination artifacts in any image reconstruction model, with theoretical exploration of why U-shaped networks are prone to hallucinations.

Abstract: U-Net and other U-shaped architectures have achieved significant success in image deconvolution tasks. However, challenges have emerged, as these methods might generate unrealistic artifacts or hallucinations, which can interfere with analysis in safety-critical scenarios. This paper introduces a novel approach for quantifying and comprehending hallucination artifacts to ensure trustworthy computer vision models. Our method, termed the Conformal Hallucination Estimation Metric (CHEM), is applicable to any image reconstruction model, enabling efficient identification and quantification of hallucination artifacts. It offers two key advantages: it leverages wavelet and shearlet representations to efficiently extract hallucinations of image features and uses conformalized quantile regression to assess hallucination levels in a distribution-free manner. Furthermore, from an approximation theoretical perspective, we explore the reasons why U-shaped networks are prone to hallucinations. We test the proposed approach on the CANDELS astronomical image dataset with models such as U-Net, SwinUNet, and Learnlets, and provide new perspectives on hallucination from different aspects in deep learning-based image processing.

Method: Proposes StateSpace-SSL with Vision Mamba state-space encoder that models long-range lesion continuity through directional scanning across leaf surface. Uses prototype-driven teacher-student objective to align representations across multiple views for stable, lesion-aware features.

Result: Outperforms CNN- and transformer-based SSL baselines on three publicly available plant disease datasets across various evaluation metrics. Learns compact, lesion-focused feature maps.

Conclusion: StateSpace-SSL demonstrates the advantage of linear state-space modeling for self-supervised plant disease representation learning, effectively capturing lesion continuity while maintaining computational efficiency.

Abstract: Self-supervised learning (SSL) is attractive for plant disease detection as it can exploit large collections of unlabeled leaf images, yet most existing SSL methods are built on CNNs or vision transformers that are poorly matched to agricultural imagery. CNN-based SSL struggles to capture disease patterns that evolve continuously along leaf structures, while transformer-based SSL introduces quadratic attention cost from high-resolution patches. To address these limitations, we propose StateSpace-SSL, a linear-time SSL framework that employs a Vision Mamba state-space encoder to model long-range lesion continuity through directional scanning across the leaf surface. A prototype-driven teacher-student objective aligns representations across multiple views, encouraging stable and lesion-aware features from labelled data. Experiments on three publicly available plant disease datasets show that StateSpace-SSL consistently outperforms the CNN- and transformer-based SSL baselines in various evaluation metrics. Qualitative analyses further confirm that it learns compact, lesion-focused feature maps, highlighting the advantage of linear state-space modelling for self-supervised plant disease representation learning.

Method: Proposes MedGrounder with two-stage training: pre-training on report sentence-anatomy box alignment datasets, then fine-tuning on report sentence-human annotated box datasets. Reformulates the task as generalized medical phrase grounding where each sentence maps to zero, one, or multiple scored regions.

Result: MedGrounder achieves strong zero-shot transfer and outperforms REC-style and grounded report generation baselines on multi-region and non-groundable phrases, while using far fewer human box annotations. Can be composed with existing report generators to produce grounded reports without retraining.

Conclusion: The generalized medical phrase grounding formulation and MedGrounder model effectively handle real-world radiology report complexities, providing more flexible and accurate grounding of medical findings while reducing annotation requirements.

Abstract: Medical phrase grounding (MPG) maps textual descriptions of radiological findings to corresponding image regions. These grounded reports are easier to interpret, especially for non-experts. Existing MPG systems mostly follow the referring expression comprehension (REC) paradigm and return exactly one bounding box per phrase. Real reports often violate this assumption. They contain multi-region findings, non-diagnostic text, and non-groundable phrases, such as negations or descriptions of normal anatomy. Motivated by this, we reformulate the task as generalised medical phrase grounding (GMPG), where each sentence is mapped to zero, one, or multiple scored regions. To realise this formulation, we introduce the first GMPG model: MedGrounder. We adopted a two-stage training regime: pre-training on report sentence–anatomy box alignment datasets and fine-tuning on report sentence–human annotated box datasets. Experiments on PadChest-GR and MS-CXR show that MedGrounder achieves strong zero-shot transfer and outperforms REC-style and grounded report generation baselines on multi-region and non-groundable phrases, while using far fewer human box annotations. Finally, we show that MedGrounder can be composed with existing report generators to produce grounded reports without retraining the generator.

Method: Proposes GGL-Net with three key components: 1) First to introduce gradient magnitude images into deep learning-based infrared small target detection, 2) Dual-branch feature extraction network with gradient supplementary module (GSM) to encode raw gradient information into deeper layers, 3) Two-way guidance fusion module (TGFM) for effective multi-scale feature fusion.

Result: Achieves state-of-the-art results on both public real NUAA-SIRST dataset and public synthetic NUDT-SIRST dataset.

Conclusion: GGL-Net effectively addresses edge positioning issues in infrared small target detection through gradient guidance and multi-scale feature fusion, demonstrating superior performance on benchmark datasets.

Abstract: Recently, infrared small target detection has attracted extensive attention. However, due to the small size and the lack of intrinsic features of infrared small targets, the existing methods generally have the problem of inaccurate edge positioning and the target is easily submerged by the background. Therefore, we propose an innovative gradient-guided learning network (GGL-Net). Specifically, we are the first to explore the introduction of gradient magnitude images into the deep learning-based infrared small target detection method, which is conducive to emphasizing the edge details and alleviating the problem of inaccurate edge positioning of small targets. On this basis, we propose a novel dual-branch feature extraction network that utilizes the proposed gradient supplementary module (GSM) to encode raw gradient information into deeper network layers and embeds attention mechanisms reasonably to enhance feature extraction ability. In addition, we construct a two-way guidance fusion module (TGFM), which fully considers the characteristics of feature maps at different levels. It can facilitate the effective fusion of multi-scale feature maps and extract richer semantic information and detailed information through reasonable two-way guidance. Extensive experiments prove that GGL-Net has achieves state-of-the-art results on the public real NUAA-SIRST dataset and the public synthetic NUDT-SIRST dataset. Our code has been integrated into https://github.com/YuChuang1205/MSDA-Net

Method: Combines neural registration and autoencoder models: 1) A modified spatial transformer network (STN) registers raw CT to standardized coordinates of a tibia prototype, 2) Various autoencoder architectures model healthy tibial variations, 3) Both STN and AE are designed to handle masked input for fractured CTs, enabling prediction of patient-specific healthy bone in standard coordinates.

Result: Developed a 3D-adapted STN for global spatial registration, conducted comparative analysis of autoencoders for bone CT modeling, and extended both approaches to handle masked inputs for predictive generation of healthy bone structures.

Conclusion: The proposed approach successfully predicts patient-specific healthy tibia reconstruction from fractured CT scans, providing valuable assistance for surgical planning of complex tibial fractures.

Abstract: Surgical planning for complex tibial fractures can be challenging for surgeons, as the 3D structure of the later desirable bone alignment may be diffi- cult to imagine. To assist in such planning, we address the challenge of predicting a patient-specific reconstruction target from a CT of the fractured tibia. Our ap- proach combines neural registration and autoencoder models. Specifically, we first train a modified spatial transformer network (STN) to register a raw CT to a standardized coordinate system of a jointly trained tibia prototype. Subsequently, various autoencoder (AE) architectures are trained to model healthy tibial varia- tions. Both the STN and AE models are further designed to be robust to masked input, allowing us to apply them to fractured CTs and decode to a prediction of the patient-specific healthy bone in standard coordinates. Our contributions include: i) a 3D-adapted STN for global spatial registration, ii) a comparative analysis of AEs for bone CT modeling, and iii) the extension of both to handle masked inputs for predictive generation of healthy bone structures. Project page: https://github.com/HongyouZhou/repair

Method: Three-component architecture: 1) Spatial-Net for shallow feature extraction with residual learning and bilinear interpolation, 2) DWT-based low-frequency enhancement branch for refining coarse structures, 3) Shared CNN for high-frequency refinement of LH, HL, and HH wavelet subbands using weight sharing.

Result: Achieves highly competitive performance with low computational cost on three hyperspectral image datasets.

Conclusion: DDSRNet effectively integrates spatial- and frequency-domain learning for hyperspectral image super-resolution, demonstrating the benefits of dual-domain approach with computational efficiency.

Abstract: This study presents a lightweight dual-domain super-resolution network (DDSRNet) that combines Spatial-Net with the discrete wavelet transform (DWT). Specifically, our proposed model comprises three main components: (1) a shallow feature extraction module, termed Spatial-Net, which performs residual learning and bilinear interpolation; (2) a low-frequency enhancement branch based on the DWT that refines coarse image structures; and (3) a shared high-frequency refinement branch that simultaneously enhances the LH (horizontal), HL (vertical), and HH (diagonal) wavelet subbands using a single CNN with shared weights. As a result, the DWT enables subband decomposition, while the inverse DWT reconstructs the final high-resolution output. By doing so, the integration of spatial- and frequency-domain learning enables DDSRNet to achieve highly competitive performance with low computational cost on three hyperspectral image datasets, demonstrating its effectiveness for hyperspectral image super-resolution.

Method: Propose an end-to-end deep architecture combining CNN, LSTM, and FC layers that uses present and future images (shared via V2V) to predict steering angles.

Result: The proposed model achieves the lowest error compared to existing approaches in the literature.

Conclusion: Incorporating temporal dependencies through LSTM and leveraging V2V-shared future images significantly improves steering angle prediction accuracy.

Abstract: A fundamental challenge in autonomous vehicles is adjusting the steering angle at different road conditions. Recent state-of-the-art solutions addressing this challenge include deep learning techniques as they provide end-to-end solution to predict steering angles directly from the raw input images with higher accuracy. Most of these works ignore the temporal dependencies between the image frames. In this paper, we tackle the problem of utilizing multiple sets of images shared between two autonomous vehicles to improve the accuracy of controlling the steering angle by considering the temporal dependencies between the image frames. This problem has not been studied in the literature widely. We present and study a new deep architecture to predict the steering angle automatically by using Long-Short-Term-Memory (LSTM) in our deep architecture. Our deep architecture is an end-to-end network that utilizes CNN, LSTM and fully connected (FC) layers and it uses both present and futures images (shared by a vehicle ahead via Vehicle-to-Vehicle (V2V) communication) as input to control the steering angle. Our model demonstrates the lowest error when compared to the other existing approaches in the literature.

Method: Two-stage tuning method: 1) First stage learns visual features, 2) Second stage infers quality solely from these features. This explicitly separates visual perception from quality inference. The approach also analyzes relationship between quality predictions and visual features, and examines decoding of intermediate VLM layers.

Result: Reduces prediction instability from 22.00% to 12.39%. Achieves average gains of 0.3124/0.3507 in SRCC/PLCC across LIVE, CSIQ, SPAQ, and KONIQ datasets compared to baseline. Improves both stability and reliability of inference process.

Conclusion: The proposed two-stage tuning method successfully addresses instability and contradictory reasoning in VLMs for BIQA by separating visual perception from quality inference, leading to more human-like reasoning and improved performance across multiple datasets.

Abstract: Recent progress in BIQA has been driven by VLMs, whose semantic reasoning abilities suggest that they might extract visual features, generate descriptive text, and infer quality in a human-like manner. However, these models often produce textual descriptions that contradict their final quality predictions, and the predicted scores can change unstably during inference - behaviors not aligned with human reasoning. To understand these issues, we analyze the factors that cause contradictory assessments and instability. We first estimate the relationship between the final quality predictions and the generated visual features, finding that the predictions are not fully grounded in the features and that the logical connection between them is weak. Moreover, decoding intermediate VLM layers shows that the model frequently relies on a limited set of candidate tokens, which contributes to prediction instability. To encourage more human-like reasoning, we introduce a two-stage tuning method that explicitly separates visual perception from quality inference. In the first stage, the model learns visual features; in the second, it infers quality solely from these features. Experiments on SPAQ and KONIQ demonstrate that our approach reduces prediction instability from 22.00% to 12.39% and achieves average gains of 0.3124/0.3507 in SRCC/PLCC across LIVE, CSIQ, SPAQ, and KONIQ compared to the baseline. Further analyses show that our method improves both stability and the reliability of the inference process.

Method: DNS-HyXNet uses a lightweight extended LSTM (xLSTM) hybrid framework that integrates tokenized domain embeddings with normalized numerical DNS features. It processes these through a two-layer xLSTM network that directly learns temporal dependencies from packet sequences, eliminating the need for graph reconstruction and enabling single-stage multi-class classification.

Result: The model achieved up to 99.99% accuracy on DNS-Tunnel-Datasets, with macro-averaged precision, recall, and F1-scores exceeding 99.96%. It demonstrated a per-sample detection latency of just 0.041 ms, confirming scalability and real-time readiness.

Conclusion: Sequential modeling with xLSTM can effectively replace computationally expensive recursive graph generation, offering a deployable and energy-efficient alternative for real-time DNS tunnel detection on commodity hardware.

Abstract: Domain Name System (DNS) tunneling remains a covert channel for data exfiltration and command-and-control communication. Although graph-based methods such as GraphTunnel achieve strong accuracy, they introduce significant latency and computational overhead due to recursive parsing and graph construction, limiting their suitability for real-time deployment. This work presents DNS-HyXNet, a lightweight extended Long Short-Term Memory (xLSTM) hybrid framework designed for efficient sequence-based DNS tunnel detection. DNS-HyXNet integrates tokenized domain embeddings with normalized numerical DNS features and processes them through a two-layer xLSTM network that directly learns temporal dependencies from packet sequences, eliminating the need for graph reconstruction and enabling single-stage multi-class classification. The model was trained and evaluated on two public benchmark datasets with carefully tuned hyperparameters to ensure low memory consumption and fast inference. Across all experimental splits of the DNS-Tunnel-Datasets, DNS-HyXNet achieved up to 99.99% accuracy, with macro-averaged precision, recall, and F1-scores exceeding 99.96%, and demonstrated a per-sample detection latency of just 0.041 ms, confirming its scalability and real-time readiness. These results show that sequential modeling with xLSTM can effectively replace computationally expensive recursive graph generation, offering a deployable and energy-efficient alternative for real-time DNS tunnel detection on commodity hardware.

Method: Introduced a low-level distortion perception task for classification; conducted component-wise analysis of MLLMs; computed semantic distance between visual features and semantic tokens before/after component-wise fine-tuning; focused on improving vision encoder alignment.

Result: MLLMs are structurally capable of representing distortions but overfit training templates, causing biases. Improving vision encoder alignment dramatically enhanced distortion recognition accuracy from 14.92% to 84.43%.

Conclusion: Incorporating dedicated constraints on the vision encoder can strengthen text-explainable visual representations, enabling MLLM-based pipelines to produce more coherent and interpretable reasoning in vision-centric tasks.

Abstract: Recent advances in Image Quality Assessment (IQA) have leveraged Multi-modal Large Language Models (MLLMs) to generate descriptive explanations. However, despite their strong visual perception modules, these models often fail to reliably detect basic low-level distortions such as blur, noise, and compression, and may produce inconsistent evaluations across repeated inferences. This raises an essential question: do MLLM-based IQA systems truly perceive the visual features that matter? To examine this issue, we introduce a low-level distortion perception task that requires models to classify specific distortion types. Our component-wise analysis shows that although MLLMs are structurally capable of representing such distortions, they tend to overfit training templates, leading to biases in quality scoring. As a result, critical low-level features are weakened or lost during the vision-language alignment transfer stage. Furthermore, by computing the semantic distance between visual features and corresponding semantic tokens before and after component-wise fine-tuning, we show that improving the alignment of the vision encoder dramatically enhances distortion recognition accuracy, increasing it from 14.92% to 84.43%. Overall, these findings indicate that incorporating dedicated constraints on the vision encoder can strengthen text-explainable visual representations and enable MLLM-based pipelines to produce more coherent and interpretable reasoning in vision-centric tasks.

Method: Hybrid modeling approach combining indirect modeling through proxies/drivers with direct spectral modeling. Uses interpretable physics-informed covariates from radiative transfer models (RTMs) and nonlinear embeddings from a foundation model. Validated with strict spatial blocking, stratified splits, and statistically distinct train-test sets on harmonized European cropland data.

Result: Highest accuracy for SOC (MAE: 5.12 g/kg, CCC: 0.77) and N (MAE: 0.44 g/kg, CCC: 0.77). Performance held across unseen locations in spatial cross-validation and independent testing. Achieved 90% coverage at target confidence level through conformal calibration for uncertainty assessment.

Conclusion: The study contributes to digital agriculture advancement through scalable, data-driven soil analysis frameworks that can be extended to domains requiring quantitative soil evaluation, such as carbon markets.

Abstract: Environmental variables are increasingly affecting agricultural decision-making, yet accessible and scalable tools for soil assessment remain limited. This study presents a robust and scalable modeling system for estimating soil properties in croplands, including soil organic carbon (SOC), total nitrogen (N), available phosphorus (P), exchangeable potassium (K), and pH, using remote sensing data and environmental covariates. The system employs a hybrid modeling approach, combining the indirect methods of modeling soil through proxies and drivers with direct spectral modeling. We extend current approaches by using interpretable physics-informed covariates derived from radiative transfer models (RTMs) and complex, nonlinear embeddings from a foundation model. We validate the system on a harmonized dataset that covers Europes cropland soils across diverse pedoclimatic zones. Evaluation is conducted under a robust validation framework that enforces strict spatial blocking, stratified splits, and statistically distinct train-test sets, which deliberately make the evaluation harder and produce more realistic error estimates for unseen regions. The models achieved their highest accuracy for SOC and N. This performance held across unseen locations, under both spatial cross-validation and an independent test set. SOC obtained a MAE of 5.12 g/kg and a CCC of 0.77, and N obtained a MAE of 0.44 g/kg and a CCC of 0.77. We also assess uncertainty through conformal calibration, achieving 90 percent coverage at the target confidence level. This study contributes to the digital advancement of agriculture through the application of scalable, data-driven soil analysis frameworks that can be extended to related domains requiring quantitative soil evaluation, such as carbon markets.

Method: Proposes CA-ATP with two key modules: 1) Content-adaptive curve mapping module using basis curves to establish multiple color mapping relationships with learned weight maps for content-aware color adjustments, and 2) Attribute text prediction module that generates text representations from multiple image attributes to represent user style preferences, integrated with visual features via multimodal model.

Result: Extensive experiments on several public datasets demonstrate state-of-the-art performance.

Conclusion: The proposed CA-ATP method successfully addresses limitations of existing approaches by enabling content-aware color adjustments and incorporating user-defined style preferences through attribute-based text representations, achieving superior retouching performance.

Abstract: Image retouching has received significant attention due to its ability to achieve high-quality visual content. Existing approaches mainly rely on uniform pixel-wise color mapping across entire images, neglecting the inherent color variations induced by image content. This limitation hinders existing approaches from achieving adaptive retouching that accommodates both diverse color distributions and user-defined style preferences. To address these challenges, we propose a novel Content-Adaptive image retouching method guided by Attribute-based Text Representation (CA-ATP). Specifically, we propose a content-adaptive curve mapping module, which leverages a series of basis curves to establish multiple color mapping relationships and learns the corresponding weight maps, enabling content-aware color adjustments. The proposed module can capture color diversity within the image content, allowing similar color values to receive distinct transformations based on their spatial context. In addition, we propose an attribute text prediction module that generates text representations from multiple image attributes, which explicitly represent user-defined style preferences. These attribute-based text representations are subsequently integrated with visual features via a multimodal model, providing user-friendly guidance for image retouching. Extensive experiments on several public datasets demonstrate that our method achieves state-of-the-art performance.

Method: Uses frozen vision transformer encoder for multi-scale features, lightweight head for highlight localization, and token-level inpainting module. Introduces Virtual Highlight Synthesis pipeline for training without paired data by rendering physically plausible specularities using monocular geometry, Fresnel-aware shading, and randomized lighting.

Result: Achieves competitive performance with state-of-the-art results on several benchmarks, generalizes across natural and surgical domains, and handles severe highlights from non-Lambertian surfaces and non-uniform lighting.

Conclusion: UnReflectAnything provides an effective RGB-only solution for highlight removal that works across diverse domains without requiring paired training data, enabling better texture recovery and geometric reasoning.

Abstract: Specular highlights distort appearance, obscure texture, and hinder geometric reasoning in both natural and surgical imagery. We present UnReflectAnything, an RGB-only framework that removes highlights from a single image by predicting a highlight map together with a reflection-free diffuse reconstruction. The model uses a frozen vision transformer encoder to extract multi-scale features, a lightweight head to localize specular regions, and a token-level inpainting module that restores corrupted feature patches before producing the final diffuse image. To overcome the lack of paired supervision, we introduce a Virtual Highlight Synthesis pipeline that renders physically plausible specularities using monocular geometry, Fresnel-aware shading, and randomized lighting which enables training on arbitrary RGB images with correct geometric structure. UnReflectAnything generalizes across natural and surgical domains where non-Lambertian surfaces and non-uniform lighting create severe highlights and it achieves competitive performance with state-of-the-art results on several benchmarks. Project Page: https://alberto-rota.github.io/UnReflectAnything/

Method: Proposed CS3D framework by decomposing Convolutional 3D (C3D) method to reduce computational complexity. Utilizes soft spiking neurons and spatial-temporal attention mechanism to enhance information retention and improve facial expression detection accuracy.

Result: CS3D achieves higher accuracy on multiple datasets compared to RNN, Transformer, and C3D architectures. Energy consumption is only 21.97% of original C3D required on the same device.

Conclusion: CS3D provides an efficient solution for event-based facial expression recognition that balances accuracy and energy efficiency, making it suitable for deployment on edge computing devices in human-robot interaction scenarios.

Abstract: Responsive and accurate facial expression recognition is crucial to human-robot interaction for daily service robots. Nowadays, event cameras are becoming more widely adopted as they surpass RGB cameras in capturing facial expression changes due to their high temporal resolution, low latency, computational efficiency, and robustness in low-light conditions. Despite these advantages, event-based approaches still encounter practical challenges, particularly in adopting mainstream deep learning models. Traditional deep learning methods for facial expression analysis are energy-intensive, making them difficult to deploy on edge computing devices and thereby increasing costs, especially for high-frequency, dynamic, event vision-based approaches. To address this challenging issue, we proposed the CS3D framework by decomposing the Convolutional 3D method to reduce the computational complexity and energy consumption. Additionally, by utilizing soft spiking neurons and a spatial-temporal attention mechanism, the ability to retain information is enhanced, thus improving the accuracy of facial expression detection. Experimental results indicate that our proposed CS3D method attains higher accuracy on multiple datasets compared to architectures such as the RNN, Transformer, and C3D, while the energy consumption of the CS3D method is just 21.97% of the original C3D required on the same device.

Method: Uses three diffusion denoising processes for object, reference, and target images. Performs reverse sampling to aggregate a small subset of layer-wise self-attention features from object and reference to influence target generation. Requires minimal training examples to introduce appearance awareness to pretrained models.

Result: Enables explicit specification of appearance parameters at generation time while preserving underlying object geometry and view coherence. Provides simple yet effective way for multiview generation with diverse appearance using implicit generative 3D representations.

Conclusion: The method offers a lightweight adaptation technique for appearance transfer in multiview diffusion models, advocating for the practical adoption of implicit generative 3D representations with enhanced appearance manipulation capabilities.

Abstract: Multiview diffusion models have rapidly emerged as a powerful tool for content creation with spatial consistency across viewpoints, offering rich visual realism without requiring explicit geometry and appearance representation. However, compared to meshes or radiance fields, existing multiview diffusion models offer limited appearance manipulation, particularly in terms of material, texture, or style. In this paper, we present a lightweight adaptation technique for appearance transfer in multiview diffusion models. Our method learns to combine object identity from an input image with appearance cues rendered in a separate reference image, producing multi-view-consistent output that reflects the desired materials, textures, or styles. This allows explicit specification of appearance parameters at generation time while preserving the underlying object geometry and view coherence. We leverage three diffusion denoising processes responsible for generating the original object, the reference, and the target images, and perform reverse sampling to aggregate a small subset of layer-wise self-attention features from the object and the reference to influence the target generation. Our method requires only a few training examples to introduce appearance awareness to pretrained multiview models. The experiments show that our method provides a simple yet effective way toward multiview generation with diverse appearance, advocating the adoption of implicit generative 3D representations in practice.

Method: Developed a structured data engineering process converting raw videos from MANIAC dataset into 27,476 statistical-kinematic feature vectors. Compared static classifiers (MLPs) against temporal models (RNNs), with a key innovation of setting sequence length to 1 in a Bidirectional RNN, effectively converting it into a high-capacity static feature encoder.

Result: Achieved 97.60% accuracy, with particular success in overcoming the challenging transitional class ‘grabbing’ by achieving a balanced F1-score of 0.90. The optimized model outperformed expectations by showing that sequential modeling wasn’t critical for this task.

Conclusion: The findings establish a new benchmark for low-level hand-object interaction recognition using structured, interpretable features and lightweight architectures, demonstrating that high-capacity static feature encoding can be more effective than sequential modeling for this specific classification task.

Abstract: Reliably predicting human intent in hand-object interactions is an open challenge for computer vision. Our research concentrates on a fundamental sub-problem: the fine-grained classification of atomic interaction states, namely ‘approaching’, ‘grabbing’, and ‘holding’. To this end, we introduce a structured data engineering process that converts raw videos from the MANIAC dataset into 27,476 statistical-kinematic feature vectors. Each vector encapsulates relational and dynamic properties from a short temporal window of motion. Our initial hypothesis posited that sequential modeling would be critical, leading us to compare static classifiers (MLPs) against temporal models (RNNs). Counter-intuitively, the key discovery occurred when we set the sequence length of a Bidirectional RNN to one (seq_length=1). This modification converted the network’s function, compelling it to act as a high-capacity static feature encoder. This architectural change directly led to a significant accuracy improvement, culminating in a final score of 97.60%. Of particular note, our optimized model successfully overcame the most challenging transitional class, ‘grabbing’, by achieving a balanced F1-score of 0.90. These findings provide a new benchmark for low-level hand-object interaction recognition using structured, interpretable features and lightweight architectures.

Method: Benchmarking high-performing trackers (SAM2, EfficientTAM, DAM4SAM, SAMURAI) on datasets containing fast moving objects (FMO) specifically designed to be challenging for tracking approaches.

Result: Overall, DAM4SAM and SAMURAI perform well on more challenging sequences compared to other SAM2-based trackers.

Conclusion: The benchmarking provides detailed insights into tracker behavior and identifies which trackers perform best on challenging fast-moving object scenarios.

Abstract: Several object tracking pipelines extending Segment Anything Model 2 (SAM2) have been proposed in the past year, where the approach is to follow and segment the object from a single exemplar template provided by the user on a initialization frame. We propose to benchmark these high performing trackers (SAM2, EfficientTAM, DAM4SAM and SAMURAI) on datasets containing fast moving objects (FMO) specifically designed to be challenging for tracking approaches. The goal is to understand better current limitations in state-of-the-art trackers by providing more detailed insights on the behavior of these trackers. We show that overall the trackers DAM4SAM and SAMURAI perform well on more challenging sequences.

Method: Kaapana provides a comprehensive open-source platform with modular, extensible framework that unifies data ingestion, cohort curation, processing workflows, and result inspection under a common user interface. It brings algorithms to the data, enabling institutions to keep control over sensitive data while participating in distributed experimentation.

Result: The platform reduces technical overhead, improves reproducibility, and enables conducting large-scale, collaborative, multi-centre imaging studies. It supports diverse use cases from local prototyping to nation-wide research networks.

Conclusion: Kaapana bridges the gap between the need for large multi-center datasets and the challenges of real-world clinical research environments by providing standardized, reproducible tooling that enables institutions to collaborate while maintaining data control.

Abstract: Developing generalizable AI for medical imaging requires both access to large, multi-center datasets and standardized, reproducible tooling within research environments. However, leveraging real-world imaging data in clinical research environments is still hampered by strict regulatory constraints, fragmented software infrastructure, and the challenges inherent in conducting large-cohort multicentre studies. This leads to projects that rely on ad-hoc toolchains that are hard to reproduce, difficult to scale beyond single institutions and poorly suited for collaboration between clinicians and data scientists. We present Kaapana, a comprehensive open-source platform for medical imaging research that is designed to bridge this gap. Rather than building single-use, site-specific tooling, Kaapana provides a modular, extensible framework that unifies data ingestion, cohort curation, processing workflows and result inspection under a common user interface. By bringing the algorithm to the data, it enables institutions to keep control over their sensitive data while still participating in distributed experimentation and model development. By integrating flexible workflow orchestration with user-facing applications for researchers, Kaapana reduces technical overhead, improves reproducibility and enables conducting large-scale, collaborative, multi-centre imaging studies. We describe the core concepts of the platform and illustrate how they can support diverse use cases, from local prototyping to nation-wide research networks. The open-source codebase is available at https://github.com/kaapana/kaapana

Method: Two-stage framework: 1) Densifies sparse trajectories into HOI mask sequences, 2) Fine-tunes a video diffusion model conditioned on these dense masks. Introduces HOI-aware motion representation using color encodings to distinguish human/object motion and body-part-specific dynamics.

Result: Achieves state-of-the-art results in controllable HOI video generation. Can generate both interaction-only scenarios and full human navigation leading up to object interactions end-to-end.

Conclusion: VHOI effectively balances controllability and realism in human-object interaction video synthesis by incorporating human priors into conditioning signals and enabling generation of complex HOI dynamics from sparse inputs.

Abstract: Synthesizing realistic human-object interactions (HOI) in video is challenging due to the complex, instance-specific interaction dynamics of both humans and objects. Incorporating controllability in video generation further adds to the complexity. Existing controllable video generation approaches face a trade-off: sparse controls like keypoint trajectories are easy to specify but lack instance-awareness, while dense signals such as optical flow, depths or 3D meshes are informative but costly to obtain. We propose VHOI, a two-stage framework that first densifies sparse trajectories into HOI mask sequences, and then fine-tunes a video diffusion model conditioned on these dense masks. We introduce a novel HOI-aware motion representation that uses color encodings to distinguish not only human and object motion, but also body-part-specific dynamics. This design incorporates a human prior into the conditioning signal and strengthens the model’s ability to understand and generate realistic HOI dynamics. Experiments demonstrate state-of-the-art results in controllable HOI video generation. VHOI is not limited to interaction-only scenarios and can also generate full human navigation leading up to object interactions in an end-to-end manner. Project page: https://vcai.mpi-inf.mpg.de/projects/vhoi/.

Method: Created IF-Bench with 499 infrared images from 23 datasets and 680 QA pairs across 10 understanding dimensions. Systematically evaluated 40+ MLLMs using cyclic evaluation, bilingual assessment, and hybrid judgment. Proposed GenViP - a training-free method that uses image editing models to translate infrared images into semantically aligned RGB counterparts.

Result: Analysis revealed how model scale, architecture, and inference paradigms affect infrared image comprehension. GenViP method consistently yielded significant performance improvements across a wide range of MLLMs by mitigating domain distribution shifts.

Conclusion: IF-Bench fills an important gap in evaluating MLLMs on infrared imagery. The proposed GenViP method effectively improves performance by translating infrared to RGB, providing valuable insights for advancing infrared image understanding in multimodal systems.

Abstract: Recent advances in multimodal large language models (MLLMs) have led to impressive progress across various benchmarks. However, their capability in understanding infrared images remains unexplored. To address this gap, we introduce IF-Bench, the first high-quality benchmark designed for evaluating multimodal understanding of infrared images. IF-Bench consists of 499 images sourced from 23 infrared datasets and 680 carefully curated visual question-answer pairs, covering 10 essential dimensions of image understanding. Based on this benchmark, we systematically evaluate over 40 open-source and closed-source MLLMs, employing cyclic evaluation, bilingual assessment, and hybrid judgment strategies to enhance the reliability of the results. Our analysis reveals how model scale, architecture, and inference paradigms affect infrared image comprehension, providing valuable insights for this area. Furthermore, we propose a training-free generative visual prompting (GenViP) method, which leverages advanced image editing models to translate infrared images into semantically and spatially aligned RGB counterparts, thereby mitigating domain distribution shifts. Extensive experiments demonstrate that our method consistently yields significant performance improvements across a wide range of MLLMs. The benchmark and code are available at https://github.com/casiatao/IF-Bench.

Method: Train ensemble members with fairness constraints, aggregate predictions across members, carefully reuse held-out data for reliable fairness enforcement, with minimal computational overhead.

Result: Theoretical guarantees provided; experimental validation shows more consistent outcomes and stronger fairness-accuracy trade-offs than existing methods across multiple medical imaging datasets.

Conclusion: OxEnsemble effectively addresses fair classification in low-data medical imaging with both data and computational efficiency, outperforming existing approaches.

Abstract: We address the problem of fair classification in settings where data is scarce and unbalanced across demographic groups. Such low-data regimes are common in domains like medical imaging, where false negatives can have fatal consequences. We propose a novel approach \emph{OxEnsemble} for efficiently training ensembles and enforcing fairness in these low-data regimes. Unlike other approaches, we aggregate predictions across ensemble members, each trained to satisfy fairness constraints. By construction, \emph{OxEnsemble} is both data-efficient, carefully reusing held-out data to enforce fairness reliably, and compute-efficient, requiring little more compute than used to fine-tune or evaluate an existing model. We validate this approach with new theoretical guarantees. Experimentally, our approach yields more consistent outcomes and stronger fairness-accuracy trade-offs than existing methods across multiple challenging medical imaging classification datasets.

Method: A fully automated Tip-and-Cue framework where tips (spatiotemporal targets from external data) trigger cues (imaging tasks). The system autonomously generates candidate tasks, optimizes scheduling across multiple satellites using continuous utility functions, processes imagery with AI models (object detectors and vision-language models), and generates structured visual reports.

Result: Demonstrated efficacy through maritime vessel tracking scenario using AIS data for trajectory prediction, targeted observations, and actionable outputs. The framework successfully integrates automated tasking, scheduling optimization, and AI-based image analysis.

Conclusion: The Tip-and-Cue framework enables efficient automated satellite imaging for applications like maritime tracking, with extensibility to broader domains such as smart-city monitoring and disaster response where timely tasking and automated analysis are critical.

Abstract: The proliferation of satellite constellations, coupled with reduced tasking latency and diverse sensor capabilities, has expanded the opportunities for automated Earth observation. This paper introduces a fully automated Tip-and-Cue framework designed for satellite imaging tasking and scheduling. In this context, tips are generated from external data sources or analyses of prior satellite imagery, identifying spatiotemporal targets and prioritizing them for downstream planning. Corresponding cues are the imaging tasks formulated in response, which incorporate sensor constraints, timing requirements, and utility functions. The system autonomously generates candidate tasks, optimizes their scheduling across multiple satellites using continuous utility functions that reflect the expected value of each observation, and processes the resulting imagery using artificial-intelligence-based models, including object detectors and vision-language models. Structured visual reports are generated to support both interpretability and the identification of new insights for downstream tasking. The efficacy of the framework is demonstrated through a maritime vessel tracking scenario, utilizing Automatic Identification System (AIS) data for trajectory prediction, targeted observations, and the generation of actionable outputs. Maritime vessel tracking is a widely researched application, often used to benchmark novel approaches to satellite tasking, forecasting, and analysis. The system is extensible to broader applications such as smart-city monitoring and disaster response, where timely tasking and automated analysis are critical.

Method: Identifies specific model parts responsible for copyrighted content generation and applies model pruning to suppress probability of generating copyrighted content without targeting specific concepts.

Result: Effectively reduces generation of copyrighted content while preserving general generative capabilities. Approach is orthogonal and complementary to existing unlearning methods.

Conclusion: Model pruning offers a scalable solution to mitigate memorization in generative models without computational overhead during sampling, improving current de-memorization techniques.

Abstract: Recent advances in generative models have demonstrated an exceptional ability to produce highly realistic images. However, previous studies show that generated images often resemble the training data, and this problem becomes more severe as the model size increases. Memorizing training data can lead to legal challenges, including copyright infringement, violations of portrait rights, and trademark violations. Existing approaches to mitigating memorization mainly focus on manipulating the denoising sampling process to steer image embeddings away from the memorized embedding space or employ unlearning methods that require training on datasets containing specific sets of memorized concepts. However, existing methods often incur substantial computational overhead during sampling, or focus narrowly on removing one or more groups of target concepts, imposing a significant limitation on their scalability. To understand and mitigate these problems, our work, UniForget, offers a new perspective on understanding the root cause of memorization. Our work demonstrates that specific parts of the model are responsible for copyrighted content generation. By applying model pruning, we can effectively suppress the probability of generating copyrighted content without targeting specific concepts while preserving the general generative capabilities of the model. Additionally, we show that our approach is both orthogonal and complementary to existing unlearning methods, thereby highlighting its potential to improve current unlearning and de-memorization techniques.

Method: Created SMA using style transfer techniques on 12 subject classification datasets, resulting in 4800 groups combining various subjects (objects, plants, animals, human actions, textures) with multiple styles. Provides flexible control over groups, classes, and domains.

Result: Two benchmarks demonstrated: (1) OOD generalization/fairness benchmark shows increasing group diversity significantly impacts fairness and alters algorithm rankings; proposed Top-M worst group accuracy improves fairness optimization. (2) UDA benchmark offers more comprehensive evaluation with 73% and 28% lower error bars in closed-set and UniDA settings.

Conclusion: SMA enables more realistic and comprehensive evaluation of ML models for OOD generalization, fairness, and domain adaptation by providing configurable group diversity, revealing that algorithm performance rankings change significantly with increased group diversity.

Abstract: We introduce Stylized Meta-Album (SMA), a new image classification meta-dataset comprising 24 datasets (12 content datasets, and 12 stylized datasets), designed to advance studies on out-of-distribution (OOD) generalization and related topics. Created using style transfer techniques from 12 subject classification datasets, SMA provides a diverse and extensive set of 4800 groups, combining various subjects (objects, plants, animals, human actions, textures) with multiple styles. SMA enables flexible control over groups and classes, allowing us to configure datasets to reflect diverse benchmark scenarios. While ideally, data collection would capture extensive group diversity, practical constraints often make this infeasible. SMA addresses this by enabling large and configurable group structures through flexible control over styles, subject classes, and domains-allowing datasets to reflect a wide range of real-world benchmark scenarios. This design not only expands group and class diversity, but also opens new methodological directions for evaluating model performance across diverse group and domain configurations-including scenarios with many minority groups, varying group imbalance, and complex domain shifts-and for studying fairness, robustness, and adaptation under a broader range of realistic conditions. To demonstrate SMA’s effectiveness, we implemented two benchmarks: (1) a novel OOD generalization and group fairness benchmark leveraging SMA’s domain, class, and group diversity to evaluate existing benchmarks. Our findings reveal that while simple balancing and algorithms utilizing group information remain competitive as claimed in previous benchmarks, increasing group diversity significantly impacts fairness, altering the superiority and relative rankings of algorithms. We also propose to use \textit{Top-M worst group accuracy} as a new hyperparameter tuning metric, demonstrating broader fairness during optimization and delivering better final worst-group accuracy for larger group diversity. (2) An unsupervised domain adaptation (UDA) benchmark utilizing SMA’s group diversity to evaluate UDA algorithms across more scenarios, offering a more comprehensive benchmark with lower error bars (reduced by 73% and 28% in closed-set setting and UniDA setting, respectively) compared to existing efforts. These use cases highlight SMA’s potential to significantly impact the outcomes of conventional benchmarks.

Method: Proposes FastPose-ViT, a Vision Transformer architecture that processes cropped object bounding boxes and directly regresses 6DoF pose. Introduces novel mathematical formalism based on projective geometry and “apparent rotation” concept to map localized predictions back to full-image scale, where apparent rotation matrix is predicted then corrected to find true orientation.

Result: Outperforms other non-PnP strategies and achieves performance competitive with state-of-the-art PnP-based techniques on SPEED dataset. When quantized and deployed on NVIDIA Jetson Orin Nano, achieves ~75 ms latency per frame sequentially and up to 33 FPS throughput with concurrent scheduling.

Conclusion: FastPose-ViT provides an efficient alternative to iterative PnP methods, enabling real-time spacecraft pose estimation on power-constrained edge hardware suitable for autonomous space missions.

Abstract: Estimating the 6-degrees-of-freedom (6DoF) pose of a spacecraft from a single image is critical for autonomous operations like in-orbit servicing and space debris removal. Existing state-of-the-art methods often rely on iterative Perspective-n-Point (PnP)-based algorithms, which are computationally intensive and ill-suited for real-time deployment on resource-constrained edge devices. To overcome these limitations, we propose FastPose-ViT, a Vision Transformer (ViT)-based architecture that directly regresses the 6DoF pose. Our approach processes cropped images from object bounding boxes and introduces a novel mathematical formalism to map these localized predictions back to the full-image scale. This formalism is derived from the principles of projective geometry and the concept of “apparent rotation”, where the model predicts an apparent rotation matrix that is then corrected to find the true orientation. We demonstrate that our method outperforms other non-PnP strategies and achieves performance competitive with state-of-the-art PnP-based techniques on the SPEED dataset. Furthermore, we validate our model’s suitability for real-world space missions by quantizing it and deploying it on power-constrained edge hardware. On the NVIDIA Jetson Orin Nano, our end-to-end pipeline achieves a latency of ~75 ms per frame under sequential execution, and a non-blocking throughput of up to 33 FPS when stages are scheduled concurrently.

Method: Proposes a framework with two key components: 1) Modality-specific Enhancing Module (MEM) using channel-wise attention to strengthen unique semantic cues per modality, and 2) Complementary Information Fusion (CIF) module for adaptive cross-modal knowledge exchange. Uses hybrid objective combining supervised segmentation loss and cross-modal consistency regularization.

Result: Consistently outperforms strong semi-supervised and multi-modal baselines on BraTS 2019 (HGG subset) under 1%, 5%, and 10% labeled data settings, with significant improvements in Dice and Sensitivity scores. Ablation studies confirm complementary effects of MEM and CIF.

Conclusion: The proposed framework effectively bridges cross-modality discrepancies and improves segmentation robustness under scarce supervision by enhancing modality-specific representations and enabling adaptive cross-modal information fusion.

Abstract: Semi-supervised learning (SSL) has become a promising direction for medical image segmentation, enabling models to learn from limited labeled data alongside abundant unlabeled samples. However, existing SSL approaches for multi-modal medical imaging often struggle to exploit the complementary information between modalities due to semantic discrepancies and misalignment across MRI sequences. To address this, we propose a novel semi-supervised multi-modal framework that explicitly enhances modality-specific representations and facilitates adaptive cross-modal information fusion. Specifically, we introduce a Modality-specific Enhancing Module (MEM) to strengthen semantic cues unique to each modality via channel-wise attention, and a learnable Complementary Information Fusion (CIF) module to adaptively exchange complementary knowledge between modalities. The overall framework is optimized using a hybrid objective combining supervised segmentation loss and cross-modal consistency regularization on unlabeled data. Extensive experiments on the BraTS 2019 (HGG subset) demonstrate that our method consistently outperforms strong semi-supervised and multi-modal baselines under 1%, 5%, and 10% labeled data settings, achieving significant improvements in both Dice and Sensitivity scores. Ablation studies further confirm the complementary effects of our proposed MEM and CIF in bridging cross-modality discrepancies and improving segmentation robustness under scarce supervision.

Method: Proposes SoftREPA - a lightweight contrastive fine-tuning strategy that uses soft text tokens for representation alignment. Instead of conventional score/flow matching on positive pairs only, it performs contrastive learning leveraging existing datasets as both positive and negative pairs.

Result: The method improves text-image alignment with minimal computational overhead (<1M trainable parameters). Theoretical analysis shows increased mutual information between text and image representations. Experiments demonstrate improved semantic consistency in text-to-image generation and text-guided image editing tasks.

Conclusion: Contrastive learning with existing datasets as positive/negative pairs provides better representation alignment than conventional diffusion training. SoftREPA offers an efficient fine-tuning approach that enhances semantic consistency in T2I generative models without requiring tailored datasets.

Abstract: While recent advancements in generative modeling have significantly improved text-image alignment, some residual misalignment between text and image representations still remains. Some approaches address this issue by fine-tuning models in terms of preference optimization, etc., which require tailored datasets. Orthogonal to these methods, we revisit the challenge from the perspective of representation alignment-an approach that has gained popularity with the success of REPresentation Alignment (REPA). We first argue that conventional text-to-image (T2I) diffusion models, typically trained on paired image and text data (i.e., positive pairs) by minimizing score matching or flow matching losses, is suboptimal from the standpoint of representation alignment. Instead, a better alignment can be achieved through contrastive learning that leverages existing dataset as both positive and negative pairs. To enable efficient alignment with pretrained models, we propose SoftREPA- a lightweight contrastive fine-tuning strategy that leverages soft text tokens for representation alignment. This approach improves alignment with minimal computational overhead by adding fewer than 1M trainable parameters to the pretrained model. Our theoretical analysis demonstrates that our method explicitly increases the mutual information between text and image representations, leading to enhanced semantic consistency. Experimental results across text-to-image generation and text-guided image editing tasks validate the effectiveness of our approach in improving the semantic consistency of T2I generative models.

Method: DynaIP uses a Dynamic Decoupling Strategy to remove concept-agnostic interference in MM-DiT’s dual branches, and a Hierarchical Mixture-of-Experts Feature Fusion Module to leverage CLIP’s multi-granularity features.

Result: Extensive experiments show DynaIP outperforms existing approaches in both single- and multi-subject PT2I tasks, achieving better concept fidelity, CP-PF balance, and scalability.

Conclusion: DynaIP represents a notable advancement in zero-shot personalized text-to-image generation by addressing key challenges through innovative dynamic decoupling and hierarchical feature fusion techniques.

Abstract: Personalized Text-to-Image (PT2I) generation aims to produce customized images based on reference images. A prominent interest pertains to the integration of an image prompt adapter to facilitate zero-shot PT2I without test-time fine-tuning. However, current methods grapple with three fundamental challenges: 1. the elusive equilibrium between Concept Preservation (CP) and Prompt Following (PF), 2. the difficulty in retaining fine-grained concept details in reference images, and 3. the restricted scalability to extend to multi-subject personalization. To tackle these challenges, we present Dynamic Image Prompt Adapter (DynaIP), a cutting-edge plugin to enhance the fine-grained concept fidelity, CP-PF balance, and subject scalability of SOTA T2I multimodal diffusion transformers (MM-DiT) for PT2I generation. Our key finding is that MM-DiT inherently exhibit decoupling learning behavior when injecting reference image features into its dual branches via cross attentions. Based on this, we design an innovative Dynamic Decoupling Strategy that removes the interference of concept-agnostic information during inference, significantly enhancing the CP-PF balance and further bolstering the scalability of multi-subject compositions. Moreover, we identify the visual encoder as a key factor affecting fine-grained CP and reveal that the hierarchical features of commonly used CLIP can capture visual information at diverse granularity levels. Therefore, we introduce a novel Hierarchical Mixture-of-Experts Feature Fusion Module to fully leverage the hierarchical features of CLIP, remarkably elevating the fine-grained concept fidelity while also providing flexible control of visual granularity. Extensive experiments across single- and multi-subject PT2I tasks verify that our DynaIP outperforms existing approaches, marking a notable advancement in the field of PT2l generation.

Method: Reformulated struggle localization as online detection task and extended to anticipation. Adapted two off-the-shelf models as baselines for online struggle detection and anticipation. Examined generalization across tasks/activities and analyzed skill evolution impact.

Result: Online struggle detection achieves 70-80% per-frame mAP. Struggle anticipation up to 2 seconds ahead yields comparable performance with slight drops. Models generalize across tasks/activities, outperforming random baselines by 4-20% despite domain gaps. Feature-based models run at 143 FPS, with full pipeline at ~20 FPS.

Conclusion: The proposed online struggle detection and anticipation models are effective for real-time assistive applications, with sufficient processing speed and generalization capabilities across different tasks and skill levels.

Abstract: Understanding human skill performance is essential for intelligent assistive systems, with struggle recognition offering a natural cue for identifying user difficulties. While prior work focuses on offline struggle classification and localization, real-time applications require models capable of detecting and anticipating struggle online. We reformulate struggle localization as an online detection task and further extend it to anticipation, predicting struggle moments before they occur. We adapt two off-the-shelf models as baselines for online struggle detection and anticipation. Online struggle detection achieves 70-80% per-frame mAP, while struggle anticipation up to 2 seconds ahead yields comparable performance with slight drops. We further examine generalization across tasks and activities and analyse the impact of skill evolution. Despite larger domain gaps in activity-level generalization, models still outperform random baselines by 4-20%. Our feature-based models run at up to 143 FPS, and the whole pipeline, including feature extraction, operates at around 20 FPS, sufficient for real-time assistive applications.

Method: The authors created the LENVIZ dataset by capturing over 230K frames across 24K real-world indoor and outdoor scenes using 3 different camera sensors. The dataset includes multi-exposure low-light scenes with high-quality human-generated ground truth meticulously curated and edited by expert photographers.

Result: LENVIZ is the largest publicly available benchmark dataset for low-light image enhancement with up to 4K resolution, offering a wide range of lighting conditions, noise levels, and scene complexities. The authors also conducted comprehensive analysis of current state-of-the-art techniques on this dataset.

Conclusion: The LENVIZ dataset provides a valuable resource for advancing low-light image enhancement research, with high-quality ground truth and diverse real-world scenarios. The analysis of existing methods on this dataset helps identify potential areas for improvement in the field.

Abstract: Low-light image enhancement is crucial for a myriad of applications, from night vision and surveillance, to autonomous driving. However, due to the inherent limitations that come in hand with capturing images in low-illumination environments, the task of enhancing such scenes still presents a formidable challenge. To advance research in this field, we introduce our Low Exposure Night Vision (LENVIZ) Dataset, a comprehensive multi-exposure benchmark dataset for low-light image enhancement comprising of over 230K frames showcasing 24K real-world indoor and outdoor, with-and without human, scenes. Captured using 3 different camera sensors, LENVIZ offers a wide range of lighting conditions, noise levels, and scene complexities, making it the largest publicly available up-to 4K resolution benchmark in the field. LENVIZ includes high quality human-generated ground truth, for which each multi-exposure low-light scene has been meticulously curated and edited by expert photographers to ensure optimal image quality. Furthermore, we also conduct a comprehensive analysis of current state-of-the-art low-light image enhancement techniques on our dataset and highlight potential areas of improvement.

Method: Proposes UniUGP (Unified Understanding-Generation-Planning) framework with hybrid expert architecture that synergizes scene reasoning, future video generation, and trajectory planning. Uses pre-trained VLMs and video generation models, takes multi-frame observations and language instructions as input, and employs a four-stage training strategy across multiple AD datasets including newly created specialized datasets with reasoning and planning annotations.

Result: Achieves state-of-the-art performance in perception, reasoning, and decision-making, with superior generalization to challenging long-tail situations. Produces interpretable chain-of-thought reasoning, physically consistent trajectories, and coherent future videos.

Conclusion: UniUGP successfully addresses the limitations of existing autonomous driving methods by integrating visual dynamics, semantic reasoning, and planning capabilities, demonstrating improved performance in complex and long-tail scenarios.

Abstract: Autonomous driving (AD) systems struggle in long-tail scenarios due to limited world knowledge and weak visual dynamic modeling. Existing vision-language-action (VLA)-based methods cannot leverage unlabeled videos for visual causal learning, while world model-based methods lack reasoning capabilities from large language models. In this paper, we construct multiple specialized datasets providing reasoning and planning annotations for complex scenarios. Then, a unified Understanding-Generation-Planning framework, named UniUGP, is proposed to synergize scene reasoning, future video generation, and trajectory planning through a hybrid expert architecture. By integrating pre-trained VLMs and video generation models, UniUGP leverages visual dynamics and semantic reasoning to enhance planning performance. Taking multi-frame observations and language instructions as input, it produces interpretable chain-of-thought reasoning, physically consistent trajectories, and coherent future videos. We introduce a four-stage training strategy that progressively builds these capabilities across multiple existing AD datasets, along with the proposed specialized datasets. Experiments demonstrate state-of-the-art performance in perception, reasoning, and decision-making, with superior generalization to challenging long-tail situations.

Method: Proposes ConsDreamer with two key components: (1) View Disentanglement Module (VDM) that eliminates viewpoint biases in conditional prompts by decoupling irrelevant view components and injecting precise view control; (2) similarity-based partial order loss that enforces geometric consistency in the unconditional term by aligning cosine similarities with azimuth relationships.

Result: Extensive experiments show ConsDreamer can be seamlessly integrated into various 3D representations and score distillation paradigms, effectively mitigating the multi-face Janus problem.

Conclusion: ConsDreamer provides a novel solution to address fundamental view bias challenges in zero-shot text-to-3D generation, improving consistency and quality of 3D content creation from textual descriptions.

Abstract: Recent advances in zero-shot text-to-3D generation have revolutionized 3D content creation by enabling direct synthesis from textual descriptions. While state-of-the-art methods leverage 3D Gaussian Splatting with score distillation to enhance multi-view rendering through pre-trained text-to-image (T2I) models, they suffer from inherent prior view biases in T2I priors. These biases lead to inconsistent 3D generation, particularly manifesting as the multi-face Janus problem, where objects exhibit conflicting features across views. To address this fundamental challenge, we propose ConsDreamer, a novel method that mitigates view bias by refining both the conditional and unconditional terms in the score distillation process: (1) a View Disentanglement Module (VDM) that eliminates viewpoint biases in conditional prompts by decoupling irrelevant view components and injecting precise view control; and (2) a similarity-based partial order loss that enforces geometric consistency in the unconditional term by aligning cosine similarities with azimuth relationships. Extensive experiments demonstrate that ConsDreamer can be seamlessly integrated into various 3D representations and score distillation paradigms, effectively mitigating the multi-face Janus problem.

Method: DPS4Un uses a diffusion posterior sampler for semiblind unmixing with four key features: (1) uses pretrained conditional spectrum diffusion model as posterior sampler to combine learned endmember prior with observations, (2) establishes image-based endmember bundles within superpixels instead of using spectral libraries, (3) proposes superpixel-based data fidelity term instead of image-level constraint, (4) initializes endmembers as Gaussian noise for each superpixel and iteratively updates abundance and endmembers.

Result: Experimental results on three real-world benchmark datasets demonstrate that DPS4Un outperforms state-of-the-art hyperspectral unmixing methods.

Conclusion: DPS4Un effectively addresses spectral variability and prior modeling challenges in hyperspectral unmixing by combining diffusion models with superpixel-based approaches, achieving superior performance compared to existing methods.

Abstract: Linear spectral mixture models (LMM) provide a concise form to disentangle the constituent materials (endmembers) and their corresponding proportions (abundance) in a single pixel. The critical challenges are how to model the spectral prior distribution and spectral variability. Prior knowledge and spectral variability can be rigorously modeled under the Bayesian framework, where posterior estimation of Abundance is derived by combining observed data with endmember prior distribution. Considering the key challenges and the advantages of the Bayesian framework, a novel method using a diffusion posterior sampler for semiblind unmixing, denoted as DPS4Un, is proposed to deal with these challenges with the following features: (1) we view the pretrained conditional spectrum diffusion model as a posterior sampler, which can combine the learned endmember prior with observation to get the refined abundance distribution. (2) Instead of using the existing spectral library as prior, which may raise bias, we establish the image-based endmember bundles within superpixels, which are used to train the endmember prior learner with diffusion model. Superpixels make sure the sub-scene is more homogeneous. (3) Instead of using the image-level data consistency constraint, the superpixel-based data fidelity term is proposed. (4) The endmember is initialized as Gaussian noise for each superpixel region, DPS4Un iteratively updates the abundance and endmember, contributing to spectral variability modeling. The experimental results on three real-world benchmark datasets demonstrate that DPS4Un outperforms the state-of-the-art hyperspectral unmixing methods.

Method: 1) Generate synthetic PDFs with precise LaTeX ground truth for systematic control; 2) Use LLM-as-a-judge for semantic formula assessment; 3) Implement two-stage matching pipeline for parser output inconsistencies; 4) Validate with human ratings on 250 formula pairs.

Result: LLM-based evaluation achieves Pearson r=0.78 correlation with human judgment vs. CDM (r=0.34) and text similarity (r~0). Evaluation of 20+ PDF parsers on 100 synthetic documents with 2,000+ formulas reveals significant performance disparities.

Conclusion: Provides crucial insights for parser selection and establishes a robust, scalable methodology for reproducible evaluation of PDF formula extraction quality, with code and benchmark data publicly available.

Abstract: Correctly parsing mathematical formulas from PDFs is critical for training large language models and building scientific knowledge bases from academic literature, yet existing benchmarks either exclude formulas entirely or lack semantically-aware evaluation metrics. We introduce a novel benchmarking framework centered on synthetically generated PDFs with precise LaTeX ground truth, enabling systematic control over layout, formulas, and content characteristics. A key methodological contribution is pioneering LLM-as-a-judge for semantic formula assessment, combined with a robust two-stage matching pipeline that handles parser output inconsistencies. Through human validation on 250 formula pairs (750 ratings from 30 evaluators), we demonstrate that LLM-based evaluation achieves substantially higher correlation with human judgment (Pearson r=0.78) compared to CDM (r=0.34) and text similarity (r~0). Evaluating 20+ contemporary PDF parsers (including specialized OCR models, vision-language models, and rule-based approaches) across 100 synthetic documents with 2,000+ formulas reveals significant performance disparities. Our findings provide crucial insights for practitioners selecting parsers for downstream applications and establish a robust, scalable methodology that enables reproducible evaluation of PDF formula extraction quality. Code and benchmark data: https://github.com/phorn1/pdf-parse-bench

Method: Introduced Visual Action Reasoning task and created VisualActBench - a large-scale benchmark with 1,074 videos and 3,733 human-annotated actions across four real-world scenarios. Each action is labeled with Action Prioritization Level (APL) and proactive-reactive type to assess human-aligned reasoning and value sensitivity. Evaluated 29 VLMs including frontier models like GPT4o.

Result: While frontier models like GPT4o show relatively strong performance, a significant gap remains compared to human-level reasoning, especially in generating proactive, high-priority actions. Current VLMs struggle with complex context interpretation, outcome anticipation, and alignment with human decision-making.

Conclusion: VisualActBench establishes a comprehensive foundation for assessing and improving the real-world readiness of proactive, vision-centric AI agents. The benchmark highlights critical limitations in current VLMs and provides a pathway for developing more human-aligned, proactive reasoning capabilities in vision-language models.

Abstract: Vision-Language Models (VLMs) have achieved impressive progress in perceiving and describing visual environments. However, their ability to proactively reason and act based solely on visual inputs, without explicit textual prompts, remains underexplored. We introduce a new task, Visual Action Reasoning, and propose VisualActBench, a large-scale benchmark comprising 1,074 videos and 3,733 human-annotated actions across four real-world scenarios. Each action is labeled with an Action Prioritization Level (APL) and a proactive-reactive type to assess models’ human-aligned reasoning and value sensitivity. We evaluate 29 VLMs on VisualActBench and find that while frontier models like GPT4o demonstrate relatively strong performance, a significant gap remains compared to human-level reasoning, particularly in generating proactive, high-priority actions. Our results highlight limitations in current VLMs’ ability to interpret complex context, anticipate outcomes, and align with human decision-making frameworks. VisualActBench establishes a comprehensive foundation for assessing and improving the real-world readiness of proactive, vision-centric AI agents.

Method: Dataset derived from Norway’s authoritative FKB database, containing orthophotos paired with annotations for 36 semantic classes. Data collected from seven geographically diverse areas, with only tiles containing annotated objects included. Training/validation splits created through random sampling across areas to ensure representative distributions. Human expert review ensures annotation accuracy.

Result: Created NordFKB dataset with high-resolution orthophotos and detailed annotations (binary segmentation masks in GeoTIFF format and COCO-style bounding boxes). Released benchmarking repository with standardized evaluation protocols and tools for semantic segmentation and object detection.

Conclusion: NordFKB provides a robust foundation for advancing geospatial AI methods and paves the way for future expansions in coverage, temporal scope, and data modalities, enabling reproducible and comparable research in mapping and spatial planning.

Abstract: We present NordFKB, a fine-grained benchmark dataset for geospatial AI in Norway, derived from the authoritative, highly accurate, national Felles KartdataBase (FKB). The dataset contains high-resolution orthophotos paired with detailed annotations for 36 semantic classes, including both per-class binary segmentation masks in GeoTIFF format and COCO-style bounding box annotations. Data is collected from seven geographically diverse areas, ensuring variation in climate, topography, and urbanization. Only tiles containing at least one annotated object are included, and training/validation splits are created through random sampling across areas to ensure representative class and context distributions. Human expert review and quality control ensures high annotation accuracy. Alongside the dataset, we release a benchmarking repository with standardized evaluation protocols and tools for semantic segmentation and object detection, enabling reproducible and comparable research. NordFKB provides a robust foundation for advancing AI methods in mapping, land administration, and spatial planning, and paves the way for future expansions in coverage, temporal scope, and data modalities.

Method: Splatent operates on top of 3D Gaussian Splatting (3DGS) in VAE latent space. Instead of reconstructing details in 3D space, it recovers them in 2D from input views through multi-view attention mechanisms. This preserves pretrained VAE reconstruction quality while achieving faithful detail recovery.

Result: Splatent establishes new state-of-the-art for VAE latent radiance field reconstruction across multiple benchmarks. Integration with existing feed-forward frameworks consistently improves detail preservation, opening new possibilities for high-quality sparse-view 3D reconstruction.

Conclusion: The 2D-based detail recovery approach through multi-view attention effectively addresses the multi-view consistency problem in VAE latent space, achieving superior detail preservation without sacrificing reconstruction quality or introducing hallucinations.

Abstract: Radiance field representations have recently been explored in the latent space of VAEs that are commonly used by diffusion models. This direction offers efficient rendering and seamless integration with diffusion-based pipelines. However, these methods face a fundamental limitation: The VAE latent space lacks multi-view consistency, leading to blurred textures and missing details during 3D reconstruction. Existing approaches attempt to address this by fine-tuning the VAE, at the cost of reconstruction quality, or by relying on pre-trained diffusion models to recover fine-grained details, at the risk of some hallucinations. We present Splatent, a diffusion-based enhancement framework designed to operate on top of 3D Gaussian Splatting (3DGS) in the latent space of VAEs. Our key insight departs from the conventional 3D-centric view: rather than reconstructing fine-grained details in 3D space, we recover them in 2D from input views through multi-view attention mechanisms. This approach preserves the reconstruction quality of pretrained VAEs while achieving faithful detail recovery. Evaluated across multiple benchmarks, Splatent establishes a new state-of-the-art for VAE latent radiance field reconstruction. We further demonstrate that integrating our method with existing feed-forward frameworks, consistently improves detail preservation, opening new possibilities for high-quality sparse-view 3D reconstruction.

Method: Introduces RVE task requiring physical plausibility and causal dynamics reasoning. Proposes ReViSE framework with Self-Reflective Reasoning (SRF) that unifies generation and evaluation using internal VLM feedback to refine generator’s reasoning during training.

Result: Achieves 32% improvement in Overall score on reasoning-informed video editing subset of RVE-Bench benchmark over state-of-the-art methods, with enhanced editing accuracy and visual fidelity.

Conclusion: The integrated ReViSE framework successfully bridges reasoning with visual transformation for video editing, demonstrating significant improvements through self-reflective reasoning and systematic evaluation on comprehensive benchmark.

Abstract: Video unified models exhibit strong capabilities in understanding and generation, yet they struggle with reason-informed visual editing even when equipped with powerful internal vision-language models (VLMs). We attribute this gap to two factors: 1) existing datasets are inadequate for training and evaluating reasoning-aware video editing, and 2) an inherent disconnect between the models’ reasoning and editing capabilities, which prevents the rich understanding from effectively instructing the editing process. Bridging this gap requires an integrated framework that connects reasoning with visual transformation. To address this gap, we introduce the Reason-Informed Video Editing (RVE) task, which requires reasoning about physical plausibility and causal dynamics during editing. To support systematic evaluation, we construct RVE-Bench, a comprehensive benchmark with two complementary subsets: Reasoning-Informed Video Editing and In-Context Video Generation. These subsets cover diverse reasoning dimensions and real-world editing scenarios. Building upon this foundation, we propose the ReViSE, a Self-Reflective Reasoning (SRF) framework that unifies generation and evaluation within a single architecture. The model’s internal VLM provides intrinsic feedback by assessing whether the edited video logically satisfies the given instruction. The differential feedback that refines the generator’s reasoning behavior during training. Extensive experiments on RVE-Bench demonstrate that ReViSE significantly enhances editing accuracy and visual fidelity, achieving a 32% improvement of the Overall score in the reasoning-informed video editing subset over state-of-the-art methods.

Method: Two-stage framework: 1) Geometry refinement using monocular depth/normal and diffusion priors, 2) Material recovery using segmentation, intrinsic image decomposition (IID), and diffusion priors to regularize estimation.

Result: Extensive experiments on synthetic and real-world datasets show significant improvements in material parameter accuracy, relighting quality, and novel-view synthesis compared to state-of-the-art Gaussian-based inverse rendering methods, especially under sparse-view settings.

Conclusion: GAINS demonstrates that incorporating learning-based priors into Gaussian-based inverse rendering effectively stabilizes geometry and material estimation in sparse-view scenarios, enabling high-quality material recovery with limited observations.

Abstract: Recent advances in Gaussian Splatting-based inverse rendering extend Gaussian primitives with shading parameters and physically grounded light transport, enabling high-quality material recovery from dense multi-view captures. However, these methods degrade sharply under sparse-view settings, where limited observations lead to severe ambiguity between geometry, reflectance, and lighting. We introduce GAINS (Gaussian-based Inverse rendering from Sparse multi-view captures), a two-stage inverse rendering framework that leverages learning-based priors to stabilize geometry and material estimation. GAINS first refines geometry using monocular depth/normal and diffusion priors, then employs segmentation, intrinsic image decomposition (IID), and diffusion priors to regularize material recovery. Extensive experiments on synthetic and real-world datasets show that GAINS significantly improves material parameter accuracy, relighting quality, and novel-view synthesis compared to state-of-the-art Gaussian-based inverse rendering methods, especially under sparse-view settings. Project page: https://patrickbail.github.io/gains/

Method: WGAST uses a conditional GAN architecture with four-stage generator: feature extraction (multi-level encoders), fusion (cosine similarity, normalization, temporal attention), LST reconstruction (decoding), and noise suppression (Gaussian filter). Training employs weakly-supervised strategy based on physical averaging principles with PatchGAN discriminator.

Result: WGAST outperforms existing methods, reducing RMSE by 17.05% and improving SSIM by 4.22% on average compared to best baseline. It effectively captures fine-scale thermal patterns validated against 33 near-ground temperature sensors.

Conclusion: WGAST is the first end-to-end deep learning framework for daily 10m LST estimation, successfully addressing spatial-temporal fusion challenges and providing superior performance for environmental monitoring applications.

Abstract: Urbanization, climate change, and agricultural stress are increasing the demand for precise and timely environmental monitoring. Land Surface Temperature (LST) is a key variable in this context and is retrieved from remote sensing satellites. However, these systems face a trade-off between spatial and temporal resolution. While spatio-temporal fusion methods offer promising solutions, few have addressed the estimation of daily LST at 10 m resolution. In this study, we present WGAST, a weakly-supervised generative network for daily 10 m LST estimation via spatio-temporal fusion of Terra MODIS, Landsat 8, and Sentinel-2. WGAST is the first end-to-end deep learning framework designed for this task. It adopts a conditional generative adversarial architecture, with a generator composed of four stages: feature extraction, fusion, LST reconstruction, and noise suppression. The first stage employs a set of encoders to extract multi-level latent representations from the inputs, which are then fused in the second stage using cosine similarity, normalization, and temporal attention mechanisms. The third stage decodes the fused features into high-resolution LST, followed by a Gaussian filter to suppress high-frequency noise. Training follows a weakly supervised strategy based on physical averaging principles and reinforced by a PatchGAN discriminator. Experiments demonstrate that WGAST outperforms existing methods in both quantitative and qualitative evaluations. Compared to the best-performing baseline, on average, WGAST reduces RMSE by 17.05% and improves SSIM by 4.22%. Furthermore, WGAST effectively captures fine-scale thermal patterns, as validated against near-surface air temperature measurements from 33 near-ground sensors. The code is available at https://github.com/Sofianebouaziz1/WGAST.git.

Method: Two-component diffusion framework: 1) Inverse renderer estimates intrinsic maps (material, geometry, lighting) from input images, 2) Forward renderer uses these maps with weather text prompts to generate final images. Features intrinsic map-aware attention for better spatial correspondence and CLIP-space interpolation for fine weather control.

Result: Outperforms state-of-the-art pixel-space editing, weather restoration, and rendering-based methods. Demonstrates promise for autonomous driving applications by enhancing detection/segmentation robustness in challenging weather.

Conclusion: WeatherDiffusion provides superior controllable weather editing through intrinsic space representation, with practical applications in autonomous systems and computer vision tasks requiring weather robustness.

Abstract: We present WeatherDiffusion, a diffusion-based framework for controllable weather editing in intrinsic space. Our framework includes two components based on diffusion priors: an inverse renderer that estimates material properties, scene geometry, and lighting as intrinsic maps from an input image, and a forward renderer that utilizes these geometry and material maps along with a text prompt that describes specific weather conditions to generate a final image. The intrinsic maps enhance controllability compared to traditional pixel-space editing approaches. We propose an intrinsic map-aware attention mechanism that improves spatial correspondence and decomposition quality in large outdoor scenes. For forward rendering, we leverage CLIP-space interpolation of weather prompts to achieve fine-grained weather control. We also introduce a synthetic and a real-world dataset, containing 38k and 18k images under various weather conditions, each with intrinsic map annotations. WeatherDiffusion outperforms state-of-the-art pixel-space editing approaches, weather restoration methods, and rendering-based methods, showing promise for downstream tasks such as autonomous driving, enhancing the robustness of detection and segmentation in challenging weather scenarios.

Method: Proposes MSDA-Net with: 1) HFDI module (parameter-free) to inject high-frequency directional information; 2) MSDA module for multi-scale local relation extraction and directional feature perception; 3) FA structure to prevent target disappearance in high-level features; 4) FCF module to alleviate feature bias during cross-layer fusion.

Result: Extensive experiments show MSDA-Net achieves state-of-the-art results on multiple public datasets for infrared small target detection.

Conclusion: The paper successfully integrates high-frequency directional features as domain prior knowledge into neural networks, demonstrating the value of directional feature perception for infrared small target detection and achieving superior performance.

Abstract: Infrared small target detection faces the problem that it is difficult to effectively separate the background and the target. Existing deep learning-based methods focus on edge and shape features, but ignore the richer structural differences and detailed information embedded in high-frequency components from different directions, thereby failing to fully exploit the value of high-frequency directional features in target perception. To address this limitation, we propose a multi-scale direction-aware network (MSDA-Net), which is the first attempt to integrate the high-frequency directional features of infrared small targets as domain prior knowledge into neural networks. Specifically, to fully mine the high-frequency directional features, on the one hand, a high-frequency direction injection (HFDI) module without trainable parameters is constructed to inject the high-frequency directional information of the original image into the network. On the other hand, a multi-scale direction-aware (MSDA) module is constructed, which promotes the full extraction of local relations at different scales and the full perception of key features in different directions. In addition, considering the characteristics of infrared small targets, we construct a feature aggregation (FA) structure to address target disappearance in high-level feature maps, and a feature calibration fusion (FCF) module to alleviate feature bias during cross-layer feature fusion. Extensive experimental results show that our MSDA-Net achieves state-of-the-art (SOTA) results on multiple public datasets. The code can be available at https://github.com/YuChuang1205/MSDA-Net

Method: Proposes FEST framework with two components: 1) Feature enhancement using multi-scale fusion strategy and edge enhancement difficulty mining (EEDM) loss to focus on challenging target regions; 2) Adjustable sensitivity (AS) strategy for post-processing to regulate target confidence and improve detection rates.

Result: Extensive experiments show FEST framework effectively enhances performance of existing SIRST detection networks, improving detection rates while maintaining segmentation accuracy.

Conclusion: FEST provides a compatible framework that enhances existing infrared small target detection networks through feature enhancement and sensitivity tuning, offering improved robustness and adaptability in complex scenarios.

Abstract: Recently, single-frame infrared small target (SIRST) detection technology has attracted widespread attention. Different from most existing deep learning-based methods that focus on improving network architectures, we propose a feature-enhanced and sensitivity-tunable (FEST) framework, which is compatible with existing SIRST detection networks and further enhances their detection performance. The FEST framework improves the model’s robustness from two aspects: feature enhancement and target confidence regulation. For feature enhancement, we employ a multi-scale fusion strategy to improve the model’s perception to multi-scale features of multi-size targets, and design an edge enhancement difficulty mining (EEDM) loss to guide the network to continuously focus on challenging target regions and edge features during training. For target confidence regulation, an adjustable sensitivity (AS) strategy is proposed for network post-processing. This strategy enhances the model’s adaptability in complex scenarios and significantly improves the detection rate of infrared small targets while maintaining segmentation accuracy. Extensive experimental results show that our FEST framework can effectively enhance the performance of existing SIRST detection networks. The code is available at https://github.com/YuChuang1205/FEST-Framework

Method: Created DTPQA benchmark with perception-only questions about traffic scenes, enriched with distance annotations. Evaluated several state-of-the-art small VLMs on this benchmark and compared their performance to human baselines.

Result: Small VLMs achieve only ~60% average accuracy on DTPQA, significantly underperforming humans (~85%). Specific perception tasks like distinguishing left from right remain particularly challenging. Human sample size was relatively small, limiting statistical power.

Conclusion: Current small VLMs have significant limitations in traffic perception tasks, especially for distance-related challenges. The DTPQA benchmark provides a valuable tool for evaluating perception capabilities in safety-critical automated driving applications.

Abstract: Vision-Language Models (VLMs) are becoming increasingly powerful, demonstrating strong performance on a variety of tasks that require both visual and textual understanding. Their strong generalisation abilities make them a promising component for automated driving systems, which must handle unexpected corner cases. However, to be trusted in such safety-critical applications, a model must first possess a reliable perception system. Moreover, since critical objects and agents in traffic scenes are often at a distance, we require systems that are not “shortsighted”, i.e., systems with strong perception capabilities at both close (up to 20 meters) and long (30+ meters) range. With this in mind, we introduce Distance-Annotated Traffic Perception Question Answering (DTPQA), the first Visual Question Answering (VQA) benchmark focused solely on perception-based questions in traffic scenes, enriched with distance annotations. By excluding questions that require reasoning, we ensure that model performance reflects perception capabilities alone. Since automated driving hardware has limited processing power and cannot support large VLMs, our study centers on smaller VLMs. More specifically, we evaluate several state-of-the-art (SOTA) small VLMs on DTPQA and show that, despite the simplicity of the questions, these models significantly underperform compared to humans (~60% average accuracy for the best-performing small VLM versus ~85% human performance). However, it is important to note that the human sample size was relatively small, which imposes statistical limitations. We also identify specific perception tasks, such as distinguishing left from right, that remain particularly challenging for these models.

Method: Proposes an adversarial-robustness-guided graph pruning framework that performs spectral adversarial robustness evaluation to identify and prune edges most vulnerable to attacks, learning sparse undirected graphs from data.

Result: The method significantly improves computational efficiency and solution quality of spectral clustering compared to state-of-the-art graph learning approaches, while being more scalable.

Conclusion: The proposed framework effectively learns robust graph topologies that resist adversarial perturbations, offering a scalable solution that enhances both efficiency and performance of graph-based algorithms like spectral clustering.

Abstract: Graph learning plays a central role in many data mining and machine learning tasks, such as manifold learning, data representation and analysis, dimensionality reduction, clustering, and visualization. In this work, we propose a highly scalable, adversarial-robustness-guided graph pruning framework for learning graph topologies from data. By performing a spectral adversarial robustness evaluation, our method aims to learn sparse, undirected graphs that help the underlying algorithms resist noise and adversarial perturbations. In particular, we explicitly identify and prune edges that are most vulnerable to adversarial attacks. We use spectral clustering, one of the most representative graph-based machine learning algorithms, to evaluate the proposed framework. Compared with prior state-of-the-art graph learning approaches, the proposed method is more scalable and significantly improves both the computational efficiency and the solution quality of spectral clustering.

Method: RELOCATE follows a classic object localization approach: (1) identify all objects in each frame using pretrained vision models, (2) compare objects with the query and select most similar ones, (3) perform bidirectional tracking for spatio-temporal response. Key enhancements include refining selected objects for accurate localization and generating additional visual queries to capture visual variations.

Result: The method achieves a 49% relative improvement in spatio-temporal average precision on the challenging Ego4D Visual Query 2D Localization dataset, outperforming prior task-specific methods and establishing a new baseline.

Conclusion: RELOCATE demonstrates that a simple training-free approach with region-based representations and key enhancements can effectively address visual query localization in long videos, significantly outperforming existing task-specific methods while eliminating the need for specialized training.

Abstract: We present RELOCATE, a simple training-free baseline designed to perform the challenging task of visual query localization in long videos. To eliminate the need for task-specific training and efficiently handle long videos, RELOCATE leverages a region-based representation derived from pretrained vision models. At a high level, it follows the classic object localization approach: (1) identify all objects in each video frame, (2) compare the objects with the given query and select the most similar ones, and (3) perform bidirectional tracking to get a spatio-temporal response. However, we propose some key enhancements to handle small objects, cluttered scenes, partial visibility, and varying appearances. Notably, we refine the selected objects for accurate localization and generate additional visual queries to capture visual variations. We evaluate RELOCATE on the challenging Ego4D Visual Query 2D Localization dataset, establishing a new baseline that outperforms prior task-specific methods by 49% (relative improvement) in spatio-temporal average precision.

Method: Generated a large dataset of 1.3M images of dividing RPE1 cells with full cell cycle trajectories. Compared predictive performance of different deep learning models: single time-frame models, causal state space models, and bidirectional transformer models.

Result: Both causal and transformer-based models significantly outperform single- and fixed-frame approaches. They can predict visually imperceptible transitions like G1/S within 1-hour resolution, enabling accurate cell cycle phase prediction from brightfield images alone.

Conclusion: Sequence models are crucial for accurate prediction of cell cycle dynamics from label-free imaging. This approach has significant potential for broader applications in biological research without requiring genetic modifications or fluorescence channels.

Abstract: Understanding cell cycle dynamics is crucial for studying biological processes such as growth, development and disease progression. While fluorescent protein reporters like the Fucci system allow live monitoring of cell cycle phases, they require genetic engineering and occupy additional fluorescence channels, limiting broader applicability in complex experiments. In this study, we conduct a comprehensive evaluation of deep learning methods for predicting continuous Fucci signals using non-fluorescence brightfield imaging, a widely available label-free modality. To that end, we generated a large dataset of 1.3 M images of dividing RPE1 cells with full cell cycle trajectories to quantitatively compare the predictive performance of distinct model categories including single time-frame models, causal state space models and bidirectional transformer models. We show that both causal and transformer-based models significantly outperform single- and fixed frame approaches, enabling the prediction of visually imperceptible transitions like G1/S within 1h resolution. Our findings underscore the importance of sequence models for accurate predictions of cell cycle dynamics and highlight their potential for label-free imaging.

Method: Replaces random augmentation in VIRM with a novel domain-oriented direction selector that leverages inter-domain covariance to guide augmentation direction toward target domains, reducing domain discrepancies.

Result: Outperforms state-of-the-art approaches on multi-center diabetic retinopathy dataset, especially under limited data conditions and significant domain heterogeneity.

Conclusion: The proposed method effectively addresses domain generalization challenges in medical imaging by intelligently guiding data augmentation using inter-domain covariance, improving performance when data is scarce and domains are heterogeneous.

Abstract: Deep learning has achieved remarkable success in medical image classification. However, its clinical application is often hindered by data heterogeneity caused by variations in scanner vendors, imaging protocols, and operators. Approaches such as invariant risk minimization (IRM) aim to address this challenge of out-of-distribution generalization. For instance, VIRM improves upon IRM by tackling the issue of insufficient feature support overlap, demonstrating promising potential. Nonetheless, these methods face limitations in medical imaging due to the scarcity of annotated data and the inefficiency of augmentation strategies. To address these issues, we propose a novel domain-oriented direction selector to replace the random augmentation strategy used in VIRM. Our method leverages inter-domain covariance as a guider for augmentation direction, guiding data augmentation towards the target domain. This approach effectively reduces domain discrepancies and enhances generalization performance. Experiments on a multi-center diabetic retinopathy dataset demonstrate that our method outperforms state-of-the-art approaches, particularly under limited data conditions and significant domain heterogeneity.

Method: Train a Graph Meta-Network within an unsupervised representation learning framework using contrastive objective to create an architecture-agnostic latent space. The approach works with 13 NeRF architectures from three families: MLPs, tri-planes, and hash tables.

Result: Robust performance in classification, retrieval, and language tasks across multiple architectures, including those unseen during training. Matches or exceeds results of existing single-architecture frameworks.

Conclusion: Successfully demonstrates the first framework capable of processing diverse NeRF architectures and performing inference on unseen architectures, advancing the flexibility and applicability of NeRF-based systems.

Abstract: Neural Radiance Fields (NeRFs) have emerged as a groundbreaking paradigm for representing 3D objects and scenes by encoding shape and appearance information into the weights of a neural network. Recent studies have demonstrated that these weights can be used as input for frameworks designed to address deep learning tasks; however, such frameworks require NeRFs to adhere to a specific, predefined architecture. In this paper, we introduce the first framework capable of processing NeRFs with diverse architectures and performing inference on architectures unseen at training time. We achieve this by training a Graph Meta-Network within an unsupervised representation learning framework, and show that a contrastive objective is conducive to obtaining an architecture-agnostic latent space. In experiments conducted across 13 NeRF architectures belonging to three families (MLPs, tri-planes, and, for the first time, hash tables), our approach demonstrates robust performance in classification, retrieval, and language tasks involving multiple architectures, even unseen at training time, while also matching or exceeding the results of existing frameworks limited to single architectures.

Method: 1) Created first motion unlearning benchmark by filtering HumanML3D and Motion-X datasets into violent (forget) and safe (retain) sets. 2) Adapted image unlearning methods (UCE, RECE) to motion architectures (MoMask, BAMM). 3) Proposed novel training-free Latent Code Replacement (LCR) that identifies violent codes in discrete codebooks and substitutes them with safe alternatives.

Result: Unlearning violent motions is feasible. Latent Code Replacement (LCR) achieves the best trade-off between violence suppression and preserving motion quality (realism and smooth transitions).

Conclusion: This work establishes a foundation for safe motion synthesis by demonstrating effective unlearning methods for removing violent content while maintaining motion quality, with potential applications across diverse domains.

Abstract: We introduce Human Motion Unlearning and motivate it through the concrete task of preventing violent 3D motion synthesis, an important safety requirement given that popular text-to-motion datasets (HumanML3D and Motion-X) contain from 7% to 15% violent sequences spanning both atomic gestures (e.g., a single punch) and highly compositional actions (e.g., loading and swinging a leg to kick). By focusing on violence unlearning, we demonstrate how removing a challenging, multifaceted concept can serve as a proxy for the broader capability of motion “forgetting.” To enable systematic evaluation of Human Motion Unlearning, we establish the first motion unlearning benchmark by automatically filtering HumanML3D and Motion-X datasets to create distinct forget sets (violent motions) and retain sets (safe motions). We introduce evaluation metrics tailored to sequential unlearning, measuring both suppression efficacy and the preservation of realism and smooth transitions. We adapt two state-of-the-art, training-free image unlearning methods (UCE and RECE) to leading text-to-motion architectures (MoMask and BAMM), and propose Latent Code Replacement (LCR), a novel, training-free approach that identifies violent codes in a discrete codebook representation and substitutes them with safe alternatives. Our experiments show that unlearning violent motions is indeed feasible and that acting on latent codes strikes the best trade-off between violence suppression and preserving overall motion quality. This work establishes a foundation for advancing safe motion synthesis across diverse applications. Website: https://www.pinlab.org/hmu.

Method: Uses theoretical radiance transfer identity for cross-scene relighting with radially symmetric BRDFs, involving only products of spherical harmonic appearance coefficients. Automatically infers unknown source/target environment maps from GS representations.

Result: Achieves comparable relighting performance to recent inverse rendering-based GS methods with significantly reduced runtime. Works well on synthetic and real-world scenes despite theoretical limitations to radially symmetric BRDFs.

Conclusion: TranSplat provides a lightweight, efficient approach to 3D object relighting in Gaussian Splatting, offering realistic results while opening new possibilities for practical GS-based relighting applications.

Abstract: We present TranSplat, a method for fast and accurate object relighting for the 3D Gaussian Splatting (GS) framework when transferring a 3D object from a source GS scene to a target GS scene. TranSplat is based on a theoretical radiance transfer identity for cross-scene relighting of objects with radially symmetric BRDFs that involves only taking simple products of spherical harmonic appearance coefficients of the object, source, and target environment maps without any explicit computation of scene quantities (e.g., the BRDFs themselves). TranSplat is the first method to demonstrate how this theoretical identity may be used to perform relighting within the GS framework, and furthermore, by automatically inferring unknown source and target environment maps directly from the source and target scene GS representations. We evaluated TranSplat on several synthetic and real-world scenes and objects, demonstrating comparable 3D object relighting performance to recent conventional inverse rendering-based GS methods with a fraction of their runtime. While TranSplat is theoretically best-suited for radially symmetric BRDFs, results demonstrate that TranSplat still offers perceptually realistic renderings on real scenes and opens a valuable, lightweight path forward to relighting with the GS framework.

Method: Three key innovations: 1) Hierarchical Mixture-of-Experts architecture with modal-specialized, collaborative, and shared experts; 2) Physics-informed self-supervised learning embedding sensor-specific radiometric characteristics; 3) Dynamic expert pruning for adaptive compression from 14.7B to 1B parameters.

Result: Outperforms existing foundation models across 23 benchmarks spanning six RS tasks (classification, detection, segmentation, tracking, change detection, depth estimation), achieving new SOTAs and demonstrating adaptability from single-modal to multi-modal scenarios.

Conclusion: RingMoE bridges the multi-modal gap in RS foundation models, has been successfully deployed in emergency response, land management, marine sciences, and urban planning, and enables efficient deployment through adaptive compression while maintaining performance.

Abstract: The rapid advancement of foundation models has revolutionized visual representation learning in a self-supervised manner. However, their application in remote sensing (RS) remains constrained by a fundamental gap: existing models predominantly handle single or limited modalities, overlooking the inherently multi-modal nature of RS observations. Optical, synthetic aperture radar (SAR), and multi-spectral data offer complementary insights that significantly reduce the inherent ambiguity and uncertainty in single-source analysis. To bridge this gap, we introduce RingMoE, a unified multi-modal RS foundation model with 14.7 billion parameters, pre-trained on 400 million multi-modal RS images from nine satellites. RingMoE incorporates three key innovations: (1) A hierarchical Mixture-of-Experts (MoE) architecture comprising modal-specialized, collaborative, and shared experts, effectively modeling intra-modal knowledge while capturing cross-modal dependencies to mitigate conflicts between modal representations; (2) Physics-informed self-supervised learning, explicitly embedding sensor-specific radiometric characteristics into the pre-training objectives; (3) Dynamic expert pruning, enabling adaptive model compression from 14.7B to 1B parameters while maintaining performance, facilitating efficient deployment in Earth observation applications. Evaluated across 23 benchmarks spanning six key RS tasks (i.e., classification, detection, segmentation, tracking, change detection, and depth estimation), RingMoE outperforms existing foundation models and sets new SOTAs, demonstrating remarkable adaptability from single-modal to multi-modal scenarios. Beyond theoretical progress, it has been deployed and trialed in multiple sectors, including emergency response, land management, marine sciences, and urban planning.

Method: Proposes ICM that regularizes output toward a manifold conditioned on sparse structural information (colormap + Canny edges) instead of text or raw images, providing task-aligned and numerically stable regularization.

Result: ICM significantly enhances super-resolution performance, particularly in perceptual quality, demonstrating effectiveness for real-world applications while avoiding instability of dense-conditioning approaches.

Conclusion: Image-conditioned manifold regularization with sparse structural information provides better conceptual alignment and stable regularization for Real-ISR, overcoming limitations of text-conditioned approaches and improving final super-resolution quality.

Abstract: Real world image super-resolution (Real-ISR) often leverages the powerful generative priors of text-to-image diffusion models by regularizing the output to lie on their learned manifold. However, existing methods often overlook the importance of the regularizing manifold, typically defaulting to a text-conditioned manifold. This approach suffers from two key limitations. Conceptually, it is misaligned with the Real-ISR task, which is to generate high quality (HQ) images directly tied to the low quality (LQ) images. Practically, the teacher model often reconstructs images with color distortions and blurred edges, indicating a flawed generative prior for this task. To correct these flaws and ensure conceptual alignment, a more suitable manifold must incorporate information from the images. While the most straightforward approach is to condition directly on the raw input images, their high information densities make the regularization process numerically unstable. To resolve this, we propose image-conditioned manifold regularization (ICM), a method that regularizes the output towards a manifold conditioned on the sparse yet essential structural information: a combination of colormap and Canny edges. ICM provides a task-aligned and stable regularization signal, thereby avoiding the instability of dense-conditioning and enhancing the final super-resolution quality. Our experiments confirm that the proposed regularization significantly enhances super-resolution performance, particularly in perceptual quality, demonstrating its effectiveness for real-world applications. We will release the source code of our work for reproducibility.

Method: Spatial polarization multiplexing with quantized polarized light pattern using constrained de Bruijn sequence; decodes from reflected Angle of Linear Polarization values to disentangle diffuse and specular reflections.

Result: Validated with real static and dynamic objects; achieves accurate shape and BRDF measurement with invisibility and single-shot capability.

Conclusion: SPM opens new 3D sensing applications by enabling invisible joint recovery of radiometric properties and shape.

Abstract: We propose spatial polarization multiplexing (SPM) for joint sensing of shape and reflectance of a static or dynamic deformable object, which is also invisible to the naked eye. Past structured-light methods are limited to shape acquisition and cannot recover reflectance as they alter scene appearance. Our key idea is to spatially multiplex a polarization pattern to encode the incident ray and also densely sample the reflected light. We derive a quantized polarized light pattern that can be robustly and uniquely decoded from the reflected Angle of Linear Polarization (AoLP) values. It also enables single-shot disentanglement of polarimetric diffuse and specular reflections for accurate BRDF estimation. We achieve this spatial polarization multiplexing (SPM) with a constrained de Bruijn sequence. We validate this novel invisible single-shot shape and reflectance method with real static and dynamic objects. The results demonstrate the effectiveness of SPM for accurate shape and BRDF measurement which opens new avenues of application for 3D sensing thanks to its invisibility and ability to jointly recover the radiometric properties.

Method: Proposes Dynamic Iterative Shrinkage Thresholding Network (DISTA-Net) that reconceptualizes traditional sparse reconstruction within a dynamic framework. The model adaptively generates convolution weights and thresholding parameters to tailor the reconstruction process in real time.

Result: DISTA-Net is the first deep learning model designed specifically for unmixing closely-spaced infrared small targets, achieving superior sub-pixel detection accuracy. The authors also established the first open-source ecosystem for this field.

Conclusion: The work introduces both a novel deep learning approach (DISTA-Net) and a comprehensive open-source ecosystem (CSIST-100K dataset, CSO-mAP metric, GrokCSO toolkit) to advance research in closely-spaced infrared small target detection.

Abstract: Resolving closely-spaced small targets in dense clusters presents a significant challenge in infrared imaging, as the overlapping signals hinder precise determination of their quantity, sub-pixel positions, and radiation intensities. While deep learning has advanced the field of infrared small target detection, its application to closely-spaced infrared small targets has not yet been explored. This gap exists primarily due to the complexity of separating superimposed characteristics and the lack of an open-source infrastructure. In this work, we propose the Dynamic Iterative Shrinkage Thresholding Network (DISTA-Net), which reconceptualizes traditional sparse reconstruction within a dynamic framework. DISTA-Net adaptively generates convolution weights and thresholding parameters to tailor the reconstruction process in real time. To the best of our knowledge, DISTA-Net is the first deep learning model designed specifically for the unmixing of closely-spaced infrared small targets, achieving superior sub-pixel detection accuracy. Moreover, we have established the first open-source ecosystem to foster further research in this field. This ecosystem comprises three key components: (1) CSIST-100K, a publicly available benchmark dataset; (2) CSO-mAP, a custom evaluation metric for sub-pixel detection; and (3) GrokCSO, an open-source toolkit featuring DISTA-Net and other models. Our code and dataset are available at https://github.com/GrokCV/GrokCSO.

Method: MitUNet combines a Mix-Transformer encoder with a U-Net decoder enhanced with spatial and channel attention blocks, optimized using the Tversky loss function to balance precision and recall for accurate boundary recovery.

Result: Experiments on CubiCasa5k and proprietary datasets show MitUNet outperforms standard models in generating structurally correct masks with high boundary accuracy, providing robust foundation for 3D reconstruction pipelines.

Conclusion: MitUNet offers superior wall segmentation for floor plans, enabling better 3D reconstruction, with code and dataset publicly available for reproducibility and future research.

Abstract: Automatic 3D reconstruction of indoor spaces from 2D floor plans necessitates high-precision semantic segmentation of structural elements, particularly walls. However, existing methods often struggle with detecting thin structures and maintaining geometric precision. This study introduces MitUNet, a hybrid neural network combining a Mix-Transformer encoder and a U-Net decoder enhanced with spatial and channel attention blocks. Our approach, optimized with the Tversky loss function, achieves a balance between precision and recall, ensuring accurate boundary recovery. Experiments on the CubiCasa5k dataset and a proprietary regional dataset demonstrate MitUNet’s superiority in generating structurally correct masks with high boundary accuracy, outperforming standard models. This tool provides a robust foundation for automated 3D reconstruction pipelines. To ensure reproducibility and facilitate future research, the source code and the proprietary regional dataset are publicly available at https://github.com/aliasstudio/mitunet and https://doi.org/10.5281/zenodo.17871079 respectively.

Method: Train a recursive neural network on procedurally generated virtual plants (synthetic data). The network infers parametric tree-like representations from input 3D point clouds, representing plants as binary axial trees.

Result: Method achieves results on-par with strong baselines for reconstruction, segmentation, and skeletonization tasks on Chenopodium Album plants. Successfully generalizes from synthetic training data to real 3D scans.

Conclusion: The approach enables simultaneous multiple tasks (reconstruction, segmentation, skeletonization) and generalizes well from synthetic to real data, making it valuable for plant phenotyping applications.

Abstract: Plants frequently contain numerous organs, organized in 3D branching systems defining the plant’s architecture. Reconstructing the architecture of plants from unstructured observations is challenging because of self-occlusion and spatial proximity between organs, which are often thin structures. To achieve the challenging task, we propose an approach that allows to infer a parameterized representation of the plant’s architecture from a given 3D scan of a plant. In addition to the plant’s branching structure, this representation contains parametric information for each plant organ, and can therefore be used directly in a variety of tasks. In this data-driven approach, we train a recursive neural network with virtual plants generated using a procedural model. After training, the network allows to infer a parametric tree-like representation based on an input 3D point cloud. Our method is applicable to any plant that can be represented as binary axial tree. We quantitatively evaluate our approach on Chenopodium Album plants on reconstruction, segmentation and skeletonization, which are important problems in plant phenotyping. In addition to carrying out several tasks at once, our method achieves results on-par with strong baselines for each task. We apply our method, trained exclusively on synthetic data, to 3D scans and show that it generalizes well.

Method: Introduce “Do You See Me” benchmark with 1,758 images and 2,612 questions spanning seven human-psychology inspired subtasks in 2D and 3D. Features controllable complexity to rigorously evaluate MLLM visual skills. Evaluate 3 leading closed-source and 5 major open-source models.

Result: Humans achieve 96.49% accuracy, while top MLLMs average below 50%. Performance gap widens rapidly with increased task complexity (e.g., from 12% to 45% in visual form constancy subtask). Failures stem from misallocated visual attention and instability of internal representations for fine-grained details, especially at/below encoder patch resolution.

Conclusion: There’s an urgent need for MLLMs with truly robust visual perception. The benchmark reveals stark deficits in current models’ visual skills that are masked by their reasoning capabilities.

Abstract: Multimodal Large Language Models (MLLMs) show reasoning promise, yet their visual perception is a critical bottleneck. Strikingly, MLLMs can produce correct answers even while misinterpreting crucial visual elements, masking these underlying failures. Our preliminary study on a joint perception-reasoning dataset revealed that for one leading MLLM, 29% of its correct answers to reasoning questions still exhibited visual perception errors. To systematically address this, we introduce “Do You See Me”, a scalable benchmark with 1,758 images and 2,612 questions. It spans seven human-psychology inspired subtasks in 2D and 3D, featuring controllable complexity to rigorously evaluate MLLM visual skills. Our findings on 3 leading closed-source and 5 major open-source models reveal a stark deficit: humans achieve 96.49% accuracy, while top MLLMs average below 50%. This performance gap widens rapidly with increased task complexity (e.g., from 12% to 45% in the visual form constancy subtask). Further analysis into the root causes suggests that failures stem from challenges like misallocated visual attention and the instability of internal representations for fine-grained details, especially at or below encoder patch resolution. This underscores an urgent need for MLLMs with truly robust visual perception. The benchmark dataset, source code and evaluation scripts are available at https://github.com/microsoft/Do-You-See-Me.

Method: Uses a coarse-to-fine pipeline: 1) pretraining on large-scale egocentric text-video pairs for coarse understanding, 2) fine-tuning on synchronous motion-video data from egocentric-exocentric datasets via automatic construction pipeline. Features part-disentangled motion injection for precise part-level control and joint reconstruction framework for progressive 4D scene and video frame modeling.

Result: Demonstrates strong generalization ability in precise control of varying human movements and world-consistent modeling of diverse scenarios, marking the first successful egocentric real-world simulation system.

Conclusion: PlayerOne represents a pioneering breakthrough in egocentric real-world simulation that can enable new research frontiers in world modeling and diverse applications requiring immersive, motion-aligned environment generation.

Abstract: We introduce PlayerOne, the first egocentric realistic world simulator, facilitating immersive and unrestricted exploration within vividly dynamic environments. Given an egocentric scene image from the user, PlayerOne can accurately construct the corresponding world and generate egocentric videos that are strictly aligned with the real scene human motion of the user captured by an exocentric camera. PlayerOne is trained in a coarse-to-fine pipeline that first performs pretraining on large-scale egocentric text-video pairs for coarse-level egocentric understanding, followed by finetuning on synchronous motion-video data extracted from egocentric-exocentric video datasets with our automatic construction pipeline. Besides, considering the varying importance of different components, we design a part-disentangled motion injection scheme, enabling precise control of part-level movements. In addition, we devise a joint reconstruction framework that progressively models both the 4D scene and video frames, ensuring scene consistency in the long-form video generation. Experimental results demonstrate its great generalization ability in precise control of varying human movements and worldconsistent modeling of diverse scenarios. It marks the first endeavor into egocentric real-world simulation and can pave the way for the community to delve into fresh frontiers of world modeling and its diverse applications.

Method: TAViS couples multimodal foundation model (ImageBind) for cross-modal alignment and segmentation foundation model (SAM2) for precise segmentation. Introduces text-bridged design with: (1) text-bridged hybrid prompting mechanism using pseudo text for class prototypes while retaining modality-specific details, and (2) alignment supervision strategy leveraging text as a bridge to align shared semantic concepts across audio-visual modalities.

Result: The approach achieves superior performance on single-source, multi-source, semantic datasets, and excels in zero-shot settings.

Conclusion: TAViS effectively addresses cross-modal alignment challenges in AVS by coupling foundation models with a novel text-bridged design, demonstrating strong performance across various settings including zero-shot scenarios.

Abstract: Audio-Visual Segmentation (AVS) faces a fundamental challenge of effectively aligning audio and visual modalities. While recent approaches leverage foundation models to address data scarcity, they often rely on single-modality knowledge or combine foundation models in an off-the-shelf manner, failing to address the cross-modal alignment challenge. In this paper, we present TAViS, a novel framework that \textbf{couples} the knowledge of multimodal foundation models (ImageBind) for cross-modal alignment and a segmentation foundation model (SAM2) for precise segmentation. However, effectively combining these models poses two key challenges: the difficulty in transferring the knowledge between SAM2 and ImageBind due to their different feature spaces, and the insufficiency of using only segmentation loss for supervision. To address these challenges, we introduce a text-bridged design with two key components: (1) a text-bridged hybrid prompting mechanism where pseudo text provides class prototype information while retaining modality-specific details from both audio and visual inputs, and (2) an alignment supervision strategy that leverages text as a bridge to align shared semantic concepts within audio-visual modalities. Our approach achieves superior performance on single-source, multi-source, semantic datasets, and excels in zero-shot settings.

Method: 1) Coarse-to-fine strategy: start tracking with small subset of points and progressively expand tracked trajectories using learnable interpolation module trained end-to-end. 2) Optimization to reduce correlation feature computation cost.

Result: Achieves 5-100x speedup over existing approaches while maintaining state-of-the-art tracking accuracy.

Conclusion: The proposed algorithm successfully addresses computational bottlenecks in dense long-term 3D point tracking, enabling significantly faster performance without sacrificing accuracy.

Abstract: We propose a novel algorithm for accelerating dense long-term 3D point tracking in videos. Through analysis of existing state-of-the-art methods, we identify two major computational bottlenecks. First, transformer-based iterative tracking becomes expensive when handling a large number of trajectories. To address this, we introduce a coarse-to-fine strategy that begins tracking with a small subset of points and progressively expands the set of tracked trajectories. The newly added trajectories are initialized using a learnable interpolation module, which is trained end-to-end alongside the tracking network. Second, we propose an optimization that significantly reduces the cost of correlation feature computation, another key bottleneck in prior methods. Together, these improvements lead to a 5-100x speedup over existing approaches while maintaining state-of-the-art tracking accuracy.

Method: Uses structured Chain-of-Thought prompting for step-by-step reasoning, segmentation feature distillation loss, and two-stage training: corrective reflective-style training with self-correction followed by reinforcement learning via Group Reward Policy Optimization (GRPO).

Result: AURORA achieves state-of-the-art performance on Ref-AVS benchmarks and generalizes effectively to unreferenced segmentation.

Conclusion: The proposed framework successfully enhances genuine reasoning and language comprehension in audio-visual segmentation without sacrificing segmentation performance.

Abstract: Reference Audio-Visual Segmentation (Ref-AVS) tasks challenge models to precisely locate sounding objects by integrating visual, auditory, and textual cues. Existing methods often lack genuine semantic understanding, tending to memorize fixed reasoning patterns. Furthermore, jointly training for reasoning and segmentation can compromise pixel-level precision. To address these issues, we introduce AURORA, a novel framework designed to enhance genuine reasoning and language comprehension in reference audio-visual segmentation. We employ a structured Chain-of-Thought (CoT) prompting mechanism to guide the model through a step-by-step reasoning process and introduce a novel segmentation feature distillation loss to effectively integrate these reasoning abilities without sacrificing segmentation performance. To further cultivate the model’s genuine reasoning capabilities, we devise a further two-stage training strategy: first, a ``corrective reflective-style training" stage utilizes self-correction to enhance the quality of reasoning paths, followed by reinforcement learning via Group Reward Policy Optimization (GRPO) to bolster robustness in challenging scenarios. Experiments demonstrate that AURORA achieves state-of-the-art performance on Ref-AVS benchmarks and generalizes effectively to unreferenced segmentation.

Method: Proposes GeoDM framework operating in Cartesian product of Euclidean, hyperbolic, and spherical manifolds. Introduces learnable curvature and weight parameters for each geometry type, and designs an optimal transport loss to enhance distribution fidelity. Uses geometry-aware distribution matching in product space.

Result: Theoretical analysis shows geometry-aware distribution matching yields smaller generalization error bound than Euclidean counterparts. Extensive experiments on standard benchmarks demonstrate GeoDM outperforms state-of-the-art data distillation methods and remains effective across various distribution-matching strategies for single geometries.

Conclusion: GeoDM successfully addresses the limitation of Euclidean-only dataset distillation by incorporating diverse geometric structures, providing a unified framework that captures flat, hierarchical, and cyclical structures, leading to improved performance and theoretical guarantees.

Abstract: Dataset distillation aims to synthesize a compact subset of the original data, enabling models trained on it to achieve performance comparable to those trained on the original large dataset. Existing distribution-matching methods are confined to Euclidean spaces, making them only capture linear structures and overlook the intrinsic geometry of real data, e.g., curvature. However, high-dimensional data often lie on low-dimensional manifolds, suggesting that dataset distillation should have the distilled data manifold aligned with the original data manifold. In this work, we propose a geometry-aware distribution-matching framework, called \textbf{GeoDM}, which operates in the Cartesian product of Euclidean, hyperbolic, and spherical manifolds, with flat, hierarchical, and cyclical structures all captured by a unified representation. To adapt to the underlying data geometry, we introduce learnable curvature and weight parameters for three kinds of geometries. At the same time, we design an optimal transport loss to enhance the distribution fidelity. Our theoretical analysis shows that the geometry-aware distribution matching in a product space yields a smaller generalization error bound than the Euclidean counterparts. Extensive experiments conducted on standard benchmarks demonstrate that our algorithm outperforms state-of-the-art data distillation methods and remains effective across various distribution-matching strategies for the single geometries.

Method: AugLift enriches each 2D keypoint (x, y) with a UADD (confidence, depth, min depth, max depth) extracted from a monocular depth map using confidence-scaled neighborhoods. It’s modular and integrates with existing lifting models by expanding their input layer.

Result: Across four datasets and four architectures, AugLift improves cross-dataset (OOD) performance by 10.1% and in-distribution performance by 4.0% (MPJPE). Gains are largest on novel poses and occluded joints. AugLiftV2 further improves performance with learned depth features.

Conclusion: Lightweight, confidence-aware depth cues are a powerful plug-in for robust 2D-to-3D pose lifting, resolving ambiguities and improving generalization without requiring new sensors or architectural changes.

Abstract: Lifting based 3D human pose estimators infer 3D joints from 2D keypoints, but often struggle to generalize to real world settings with noisy 2D detections. We revisit the input to lifting and propose AugLift, a simple augmentation of standard lifting that enriches each 2D keypoint (x, y) with an Uncertainty Aware Depth Descriptor (UADD). We run a single off the shelf monocular depth estimator to obtain a depth map, and for every keypoint with detector confidence c we extract depth statistics from its confidence scaled neighborhood, forming a compact, interpretable UADD (c, d, d_min, d_max) that captures both local geometry and reliability. AugLift is modular, requires no new sensors or architectural changes, and integrates by expanding the input layer of existing lifting models. Across four datasets and four lifting architectures, AugLift boosts cross dataset (out of distribution) performance on unseen data by an average of 10.1 percent, while also improving in distribution performance by 4.0 percent as measured by MPJPE. A post hoc analysis clarifies when and why it helps: gains are largest on novel poses and significantly occluded joints, where depth statistics resolve front back ambiguities while confidence calibrates the spatial neighborhoods from which they are drawn. We also study interaction with recent image feature lifting methods and find the signals are complementary: adding UADD to image conditioned lifting yields both ID and OOD gains. A learned depth feature extension (AugLiftV2) improves performance further while trading off interpretability. Together, these results indicate that lightweight, confidence aware depth cues are a powerful plug in for robust 2D to 3D pose lifting.

Method: Conducted extensive search of academic databases and open-data platforms to identify 51 publicly available visual datasets (2005-2024). Categorized datasets using structured schema covering: (i) data fundamentals, (ii) data modalities, (iii) annotation frameworks, and (iv) downstream application domains. Created OpenConstruction catalog as open-source resource.

Result: Identified 51 construction visual datasets spanning 2005-2024. Developed comprehensive categorization system and created OpenConstruction catalog. Found critical limitations in existing dataset landscape including variability in data quality, annotation consistency, and representativeness of real-world conditions.

Conclusion: The study provides first systematic review of construction visual datasets, creating OpenConstruction catalog to support data-driven method development. Identifies critical gaps and proposes roadmap for future data infrastructure based on FAIR principles to advance data-centric solutions in construction sector.

Abstract: The construction industry increasingly relies on visual data to support Artificial Intelligence (AI) and Machine Learning (ML) applications for site monitoring. High-quality, domain-specific datasets, comprising images, videos, and point clouds, capture site geometry and spatiotemporal dynamics, including the location and interaction of objects, workers, and materials. However, despite growing interest in leveraging visual datasets, existing resources vary widely in sizes, data modalities, annotation quality, and representativeness of real-world construction conditions. A systematic review to categorize their data characteristics and application contexts is still lacking, limiting the community’s ability to fully understand the dataset landscape, identify critical gaps, and guide future directions toward more effective, reliable, and scalable AI applications in construction. To address this gap, this study conducts an extensive search of academic databases and open-data platforms, yielding 51 publicly available visual datasets that span the 2005-2024 period. These datasets are categorized using a structured data schema covering (i) data fundamentals (e.g., size and license), (ii) data modalities (e.g., RGB and point cloud), (iii) annotation frameworks (e.g., bounding boxes), and (iv) downstream application domains (e.g., progress tracking). This study synthesizes these findings into an open-source catalog, OpenConstruction, supporting data-driven method development. Furthermore, the study discusses several critical limitations in the existing construction dataset landscape and presents a roadmap for future data infrastructure anchored in the Findability, Accessibility, Interoperability, and Reusability (FAIR) principles. By reviewing the current landscape and outlining strategic priorities, this study supports the advancement of data-centric solutions in the construction sector.

Method: Three key components: (1) Scalable data pipeline for Unreal Engine and GTA5 producing 1200+ hours of annotated video data; (2) Action injection module for frame-level mouse/keyboard inputs as interactive conditions; (3) Few-step distillation based on causal architecture for real-time streaming video generation.

Result: Matrix-Game 2.0 generates high-quality minute-level videos across diverse scenes at 25 FPS, achieving ultra-fast speed while maintaining quality. The model weights and codebase are open-sourced.

Conclusion: The framework enables real-time interactive world modeling through efficient few-step auto-regressive diffusion, advancing research in interactive video generation and simulation of real-world dynamics.

Abstract: Recent advances in interactive video generations have demonstrated diffusion model’s potential as world models by capturing complex physical dynamics and interactive behaviors. However, existing interactive world models depend on bidirectional attention and lengthy inference steps, severely limiting real-time performance. Consequently, they are hard to simulate real-world dynamics, where outcomes must update instantaneously based on historical context and current actions. To address this, we present Matrix-Game 2.0, an interactive world model generates long videos on-the-fly via few-step auto-regressive diffusion. Our framework consists of three key components: (1) A scalable data production pipeline for Unreal Engine and GTA5 environments to effectively produce massive amounts (about 1200 hours) of video data with diverse interaction annotations; (2) An action injection module that enables frame-level mouse and keyboard inputs as interactive conditions; (3) A few-step distillation based on the casual architecture for real-time and streaming video generation. Matrix Game 2.0 can generate high-quality minute-level videos across diverse scenes at an ultra-fast speed of 25 FPS. We open-source our model weights and codebase to advance research in interactive world modeling.

Method: A two-stage framework using empty prompts: (1) During pre-training, empty prompts reveal and reduce template-induced bias within CLIP encoder; (2) During few-shot fine-tuning, a bias calibration loss enforces correct alignment between images and categories, focusing on relevant visual cues.

Result: Experiments across multiple benchmarks show the template correction method significantly reduces performance fluctuations caused by TSS, yielding higher classification accuracy and stronger robustness.

Conclusion: The proposed framework effectively decouples template bias in CLIP, improving few-shot learning by ensuring models focus on genuine visual-category alignment rather than template proximity.

Abstract: The Contrastive Language-Image Pre-Training (CLIP) model excels in few-shot learning by aligning visual and textual representations. Our study shows that template-sample similarity (TSS), defined as the resemblance between a text template and an image sample, introduces bias. This bias leads the model to rely on template proximity rather than true sample-to-category alignment, reducing both accuracy and robustness in classification. We present a framework that uses empty prompts, textual inputs that convey the idea of “emptiness” without category information. These prompts capture unbiased template features and offset TSS bias. The framework employs two stages. During pre-training, empty prompts reveal and reduce template-induced bias within the CLIP encoder. During few-shot fine-tuning, a bias calibration loss enforces correct alignment between images and their categories, ensuring the model focuses on relevant visual cues. Experiments across multiple benchmarks demonstrate that our template correction method significantly reduces performance fluctuations caused by TSS, yielding higher classification accuracy and stronger robustness. The repository of this project is available at https://github.com/zhenyuZ-HUST/Decoupling-Template-Bias-in-CLIP.

Method: Develops an efficient diffusion transformer with a powerful VAE to reduce image tokens, enabling efficient training and fast high-resolution generation. Uses pretraining on billions of text-image pairs, comprehensive data collection across vertical scenarios, and optimized training strategies. Incorporates fine-tuned VLM for multimodal post-training of both T2I and image editing tasks. Uses adversarial distillation, distribution matching, quantization, and speculative decoding for inference acceleration.

Result: Achieves state-of-the-art results on both T2I and multimodal image editing, with inference time of up to 1.8 seconds for generating 2K images. Demonstrates exceptional multimodal capabilities in complex tasks including precise image editing, in-context reasoning, multi-image reference, and multiple output generation.

Conclusion: Seedream 4.0 successfully extends traditional T2I systems into a more interactive and multidimensional creative tool, pushing the boundaries of generative AI for both creativity and professional applications, with further scaling to Seedream 4.5.

Abstract: We introduce Seedream 4.0, an efficient and high-performance multimodal image generation system that unifies text-to-image (T2I) synthesis, image editing, and multi-image composition within a single framework. We develop a highly efficient diffusion transformer with a powerful VAE which also can reduce the number of image tokens considerably. This allows for efficient training of our model, and enables it to fast generate native high-resolution images (e.g., 1K-4K). Seedream 4.0 is pretrained on billions of text-image pairs spanning diverse taxonomies and knowledge-centric concepts. Comprehensive data collection across hundreds of vertical scenarios, coupled with optimized strategies, ensures stable and large-scale training, with strong generalization. By incorporating a carefully fine-tuned VLM model, we perform multi-modal post-training for training both T2I and image editing tasks jointly. For inference acceleration, we integrate adversarial distillation, distribution matching, and quantization, as well as speculative decoding. It achieves an inference time of up to 1.8 seconds for generating a 2K image (without a LLM/VLM as PE model). Comprehensive evaluations reveal that Seedream 4.0 can achieve state-of-the-art results on both T2I and multimodal image editing. In particular, it demonstrates exceptional multimodal capabilities in complex tasks, including precise image editing and in-context reasoning, and also allows for multi-image reference, and can generate multiple output images. This extends traditional T2I systems into an more interactive and multidimensional creative tool, pushing the boundary of generative AI for both creativity and professional applications. We further scale our model and data as Seedream 4.5. Seedream 4.0 and Seedream 4.5 are accessible on Volcano Engine https://www.volcengine.com/experience/ark?launch=seedream.

Method: Proposes AdaGC framework with gradient calibration mechanism using dual exponential moving average module for robust pseudo-label generation. Introduces training-dynamics-based indicator to adaptively trigger gradient calibration, preventing issues from model underfitting or overfitting to label noise.

Result: Extensive experiments on two benchmark remote sensing datasets under two distinct label noise types demonstrate state-of-the-art performance while maintaining strong robustness across diverse settings.

Conclusion: AdaGC effectively bridges the research gap in SPML for remote sensing, providing a generalizable framework that addresses annotation challenges through adaptive gradient calibration with theoretical grounding and practical effectiveness.

Abstract: Multi-label classification (MLC) offers a more comprehensive semantic understanding of Remote Sensing (RS) imagery compared to traditional single-label classification (SLC). However, obtaining complete annotations for MLC is particularly challenging due to the complexity and high cost of the labeling process. As a practical alternative, single-positive multi-label learning (SPML) has emerged, where each image is annotated with only one relevant label, and the model is expected to recover the full set of labels. While scalable, SPML introduces significant supervision ambiguity, demanding specialized solutions for model training. Although various SPML methods have been proposed in the computer vision domain, research in the RS context remains limited. To bridge this gap, we propose Adaptive Gradient Calibration (AdaGC), a novel and generalizable SPML framework tailored to RS imagery. AdaGC adopts a gradient calibration (GC) mechanism with a dual exponential moving average (EMA) module for robust pseudo-label generation. We introduce a theoretically grounded, training-dynamics-based indicator to adaptively trigger GC, which ensures GC’s effectiveness by preventing it from being affected by model underfitting or overfitting to label noise. Extensive experiments on two benchmark RS datasets under two distinct label noise types demonstrate that AdaGC achieves state-of-the-art (SOTA) performance while maintaining strong robustness across diverse settings. The codes and data will be released at https://github.com/rslab-unitrento/AdaGC.

Method: Utilized efficient latent representations and foveated rendering, with novel perceptual design to create multi-modal latent representations for steganography.

Result: Achieved 5x improvement in capacity (500 bits vs previous 100 bits), better accuracy (1 failure bit out of 2000 at 200K test bits), and maintained visual quality (31.47 dB PSNR, 0.13 LPIPS).

Conclusion: The novel perceptual design for creating multi-modal latent representations is effective for steganography, significantly improving capacity and accuracy while preserving visual quality.

Abstract: Steganography finds its use in visual medium such as providing metadata and watermarking. With support of efficient latent representations and foveated rendering, we trained models that improve existing capacity limits from 100 to 500 bits, while achieving better accuracy of up to 1 failure bit out of 2000, at 200K test bits. Finally, we achieve a comparable visual quality of 31.47 dB PSNR and 0.13 LPIPS, showing the effectiveness of novel perceptual design in creating multi-modal latent representations in steganography.

Method: Created TeleEgo benchmark with over 14 hours per participant of synchronized egocentric video, audio, and text across four domains (work & study, lifestyle & routines, social activities, outings & culture). Data is aligned on unified global timeline with human-refined visual narrations and speech transcripts. Defines 12 diagnostic subtasks across three core capabilities: Memory, Understanding, and Cross-Memory Reasoning, with 3,291 human-verified QA items in streaming setting.

Result: The benchmark contains extensive multi-modal data with high-quality annotations and introduces two key metrics: Real-Time Accuracy (RTA) to measure correctness and responsiveness under tight decision windows, and Memory Persistence Time (MPT) for evaluating long-term retention in continuous streams. Initial RTA results for current models are reported.

Conclusion: TeleEgo provides a realistic, extensible benchmark for future egocentric assistants with stronger streaming memory capabilities, enabling systematic study of both real-time behavior and long-horizon memory in continuous, multi-modal contexts.

Abstract: Egocentric AI assistants in real-world settings must process multi-modal inputs (video, audio, text), respond in real time, and retain evolving long-term memory. However, existing benchmarks typically evaluate these abilities in isolation, lack realistic streaming scenarios, or support only short-term tasks. We introduce \textbf{TeleEgo}, a long-duration, streaming, omni-modal benchmark for evaluating egocentric AI assistants in realistic daily contexts. The dataset features over 14 hours per participant of synchronized egocentric video, audio, and text across four domains: work & study, lifestyle & routines, social activities, and outings & culture. All data is aligned on a unified global timeline and includes high-quality visual narrations and speech transcripts, curated through human refinement.TeleEgo defines 12 diagnostic subtasks across three core capabilities: Memory (recalling past events), Understanding (interpreting the current moment), and Cross-Memory Reasoning (linking distant events). It contains 3,291 human-verified QA items spanning multiple question formats (single-choice, binary, multi-choice, and open-ended), evaluated strictly in a streaming setting. We propose Real-Time Accuracy (RTA) to jointly capture correctness and responsiveness under tight decision windows, and Memory Persistence Time (MPT) as a forward-looking metric for long-term retention in continuous streams. In this work, we report RTA results for current models and release TeleEgo, together with an MPT evaluation framework, as a realistic and extensible benchmark for future egocentric assistants with stronger streaming memory, enabling systematic study of both real-time behavior and long-horizon memory.

Method: Uses ordinal regression framework modified with novel loss function that operates on soft labels representing annotator rating distributions, accounting for label uncertainty.

Result: Achieves predictive performance approaching clinical experts while producing well-calibrated uncertainty estimates.

Conclusion: Automated phonotrauma severity assessment tool enables large-scale studies, potentially improving clinical understanding and patient care.

Abstract: Phonotrauma refers to vocal fold tissue damage resulting from exposure to forces during voicing. It occurs on a continuum from mild to severe, and treatment options can vary based on severity. Assessment of severity involves a clinician’s expert judgment, which is costly and can vary widely in reliability. In this work, we present the first method for automatically classifying phonotrauma severity from vocal fold images. To account for the ordinal nature of the labels, we adopt a widely used ordinal regression framework. To account for label uncertainty, we propose a novel modification to ordinal regression loss functions that enables them to operate on soft labels reflecting annotator rating distributions. Our proposed soft ordinal regression method achieves predictive performance approaching that of clinical experts, while producing well-calibrated uncertainty estimates. By providing an automated tool for phonotrauma severity assessment, our work can enable large-scale studies of phonotrauma, ultimately leading to improved clinical understanding and patient care.

Method: GloTok uses global relational information through: 1) Codebook-wise histogram relation learning to transfer semantics from pre-trained models to the semantic codebook, and 2) A residual learning module to recover fine-grained details and minimize reconstruction errors from quantization.

Result: Achieves state-of-the-art reconstruction performance and generation quality on the standard ImageNet-1k benchmark, producing more uniformly distributed semantic latent representations.

Conclusion: GloTok’s global perspective approach with codebook-wise histogram relation learning enables more uniform semantic distributions, facilitating better training of autoregressive models for high-quality image generation without requiring direct access to pre-trained models during training.

Abstract: Existing state-of-the-art image tokenization methods leverage diverse semantic features from pre-trained vision models for additional supervision, to expand the distribution of latent representations and thereby improve the quality of image reconstruction and generation. These methods employ a locally supervised approach for semantic supervision, which limits the uniformity of semantic distribution. However, VA-VAE proves that a more uniform feature distribution yields better generation performance. In this work, we introduce a Global Perspective Tokenizer (GloTok), which utilizes global relational information to model a more uniform semantic distribution of tokenized features. Specifically, a codebook-wise histogram relation learning method is proposed to transfer the semantics, which are modeled by pre-trained models on the entire dataset, to the semantic codebook. Then, we design a residual learning module that recovers the fine-grained details to minimize the reconstruction error caused by quantization. Through the above design, GloTok delivers more uniformly distributed semantic latent representations, which facilitates the training of autoregressive (AR) models for generating high-quality images without requiring direct access to pre-trained models during the training process. Experiments on the standard ImageNet-1k benchmark clearly show that our proposed method achieves state-of-the-art reconstruction performance and generation quality.

Method: Introduced CoD, a compression-oriented diffusion foundation model trained from scratch on open image-only datasets. It’s designed for end-to-end optimization of both compression and generation, not as a fixed codec but as a general foundation for various diffusion-based codecs.

Result: CoD achieves SOTA compression results, especially at ultra-low bitrates (e.g., 0.0039 bpp). Training is 300× faster than Stable Diffusion (~20 vs. ~6,250 A100 GPU days). Pixel-space diffusion can achieve VTM-level PSNR with high perceptual quality and outperform GAN-based codecs with fewer parameters.

Conclusion: CoD lays the foundation for future diffusion codec research by providing a compression-optimized diffusion model that enables better performance, faster training, and new insights into diffusion-based compression.

Abstract: Existing diffusion codecs typically build on text-to-image diffusion foundation models like Stable Diffusion. However, text conditioning is suboptimal from a compression perspective, hindering the potential of downstream diffusion codecs, particularly at ultra-low bitrates. To address it, we introduce \textbf{CoD}, the first \textbf{Co}mpression-oriented \textbf{D}iffusion foundation model, trained from scratch to enable end-to-end optimization of both compression and generation. CoD is not a fixed codec but a general foundation model designed for various diffusion-based codecs. It offers several advantages: \textbf{High compression efficiency}, replacing Stable Diffusion with CoD in downstream codecs like DiffC achieves SOTA results, especially at ultra-low bitrates (e.g., 0.0039 bpp); \textbf{Low-cost and reproducible training}, 300$\times$ faster training than Stable Diffusion ($\sim$ 20 vs. $\sim$ 6,250 A100 GPU days) on entirely open image-only datasets; \textbf{Providing new insights}, e.g., We find pixel-space diffusion can achieve VTM-level PSNR with high perceptual quality and can outperform GAN-based codecs using fewer parameters. We hope CoD lays the foundation for future diffusion codec research. Codes will be released.

Method: Constructed CultureMix, a food Visual Question Answering benchmark with 23k diffusion-generated, human-verified culture mixing images across four subtasks: (1) food-only, (2) food+food, (3) food+background, and (4) food+food+background. Evaluated 10 LVLMs and explored three robustness strategies including supervised fine-tuning with diverse culture mixing datasets.

Result: LVLMs show consistent failures to preserve individual cultural identities in mixed settings. Models demonstrate strong background reliance, with accuracy dropping 14% when cultural backgrounds are added to food-only baselines. They produce inconsistent predictions for identical foods across different contexts. Supervised fine-tuning using diverse culture mixing data substantially improves model consistency and reduces background sensitivity.

Conclusion: Culture mixing scenarios represent a critical challenge for LVLMs that requires increased attention. Current models struggle with preserving cultural identities in mixed settings, but targeted training with diverse culture mixing data can improve performance. This is a necessary step toward developing LVLMs capable of operating reliably in culturally diverse real-world environments.

Abstract: In a globalized world, cultural elements from diverse origins frequently appear together within a single visual scene. We refer to these as culture mixing scenarios, yet how Large Vision-Language Models (LVLMs) perceive them remains underexplored. We investigate culture mixing as a critical challenge for LVLMs and examine how current models behave when cultural items from multiple regions appear together. To systematically analyze these behaviors, we construct CultureMix, a food Visual Question Answering (VQA) benchmark with 23k diffusion-generated, human-verified culture mixing images across four subtasks: (1) food-only, (2) food+food, (3) food+background, and (4) food+food+background. Evaluating 10 LVLMs, we find consistent failures to preserve individual cultural identities in mixed settings. Models show strong background reliance, with accuracy dropping 14% when cultural backgrounds are added to food-only baselines, and they produce inconsistent predictions for identical foods across different contexts. To address these limitations, we explore three robustness strategies. We find supervised fine-tuning using a diverse culture mixing dataset substantially improve model consistency and reduce background sensitivity. We call for increased attention to culture mixing scenarios as a critical step toward developing LVLMs capable of operating reliably in culturally diverse real-world environments.

Method: THCRL consists of two key modules: 1) Deep Symmetry Hierarchical Fusion (DSHF) using UNet architecture with multiple denoising mechanisms for trustworthy multi-view fusion, and 2) Average K-Nearest Neighbors Contrastive Learning (AKCL) that aligns fused representation with view-specific representations by enhancing similarity among samples in the same cluster rather than just same samples across views.

Result: Extensive experiments demonstrate that THCRL achieves state-of-the-art performance in deep multi-view clustering tasks.

Conclusion: THCRL effectively addresses the untrustworthy fusion problem in multi-view clustering through its novel hierarchical fusion with denoising and cluster-aware contrastive learning approach, leading to superior clustering performance.

Abstract: Multi-View Clustering (MVC) has garnered increasing attention in recent years. It is capable of partitioning data samples into distinct groups by learning a consensus representation. However, a significant challenge remains: the problem of untrustworthy fusion. This problem primarily arises from two key factors: 1) Existing methods often ignore the presence of inherent noise within individual views; 2) In traditional MVC methods using Contrastive Learning (CL), similarity computations typically rely on different views of the same instance, while neglecting the structural information from nearest neighbors within the same cluster. Consequently, this leads to the wrong direction for multi-view fusion. To address this problem, we present a novel Trusted Hierarchical Contrastive Representation Learning (THCRL). It consists of two key modules. Specifically, we propose the Deep Symmetry Hierarchical Fusion (DSHF) module, which leverages the UNet architecture integrated with multiple denoising mechanisms to achieve trustworthy fusion of multi-view data. Furthermore, we present the Average K-Nearest Neighbors Contrastive Learning (AKCL) module to align the fused representation with the view-specific representation. Unlike conventional strategies, AKCL enhances representation similarity among samples belonging to the same cluster, rather than merely focusing on the same sample across views, thereby reinforcing the confidence of the fused representation. Extensive experiments demonstrate that THCRL achieves the state-of-the-art performance in deep MVC tasks.

Method: Adapter-based tuning strategy preserves Vision Foundation Model representations while learning remote sensing-specific features; hybrid attention mechanism combines convolutional local modeling with Transformer-based global reasoning; improved probability map modulation incorporates historical user interactions.

Result: Outperforms state-of-the-art IIS methods on six remote sensing datasets (iSAID, ISPRS Potsdam, SandBar, NWPU, LoveDA Urban, WHUBuilding) in segmentation accuracy, efficiency, and interaction cost.

Conclusion: RS-ISRefiner is effective and generalizable for high-quality instance segmentation in practical remote sensing scenarios, offering a tailored solution that addresses domain-specific challenges.

Abstract: Interactive image segmentation(IIS) plays a critical role in generating precise annotations for remote sensing imagery, where objects often exhibit scale variations, irregular boundaries and complex backgrounds. However, existing IIS methods, primarily designed for natural images, struggle to generalize to remote sensing domains due to limited annotated data and computational overhead. To address these challenges, we proposed RS-ISRefiner, a novel click-based IIS framework tailored for remote sensing images. The framework employs an adapter-based tuning strategy that preserves the general representations of Vision Foundation Models while enabling efficient learning of remote sensing-specific spatial and boundary characteristics. A hybrid attention mechanism integrating convolutional local modeling with Transformer-based global reasoning enhances robustness against scale diversity and scene complexity. Furthermore, an improved probability map modulation scheme effectively incorporates historical user interactions, yielding more stable iterative refinement and higher boundary accuracy. Comprehensive experiments on six remote sensing datasets, including iSAID, ISPRS Potsdam, SandBar, NWPU, LoveDA Urban and WHUBuilding, demonstrate that RS-ISRefiner consistently outperforms state-of-the-art IIS methods in terms of segmentation accuracy, efficiency and interaction cost. These results confirm the effectiveness and generalizability of our framework, making it highly suitable for high-quality instance segmentation in practical remote sensing scenarios. The codes are available at https://github.com/wondelyan/VFM-ISRefiner .

Method: Two-stage framework: 1) Coarse scene decomposition with dual-slot competition to separate foreground from background using clustering-based initialization. 2) Masked slot attention where first slot captures background and remaining slots compete for foreground objects, guided by pseudo-masks from patch affinity graphs using self-supervised features.

Result: Extensive experiments on synthetic and real-world datasets show FASA consistently outperforms state-of-the-art methods, validating effectiveness of explicit foreground modeling and pseudo-mask guidance for robust scene decomposition.

Conclusion: FASA demonstrates that explicit foreground-background separation and pseudo-mask guidance significantly improve object discovery and scene decomposition, providing more robust object-coherent representations for real-world data.

Abstract: Recent advances in object-centric representation learning have shown that slot attention-based methods can effectively decompose visual scenes into object slot representations without supervision. However, existing approaches typically process foreground and background regions indiscriminately, often resulting in background interference and suboptimal instance discovery performance on real-world data. To address this limitation, we propose Foreground-Aware Slot Attention (FASA), a two-stage framework that explicitly separates foreground from background to enable precise object discovery. In the first stage, FASA performs a coarse scene decomposition to distinguish foreground from background regions through a dual-slot competition mechanism. These slots are initialized via a clustering-based strategy, yielding well-structured representations of salient regions. In the second stage, we introduce a masked slot attention mechanism where the first slot captures the background while the remaining slots compete to represent individual foreground objects. To further address over-segmentation of foreground objects, we incorporate pseudo-mask guidance derived from a patch affinity graph constructed with self-supervised image features to guide the learning of foreground slots. Extensive experiments on both synthetic and real-world datasets demonstrate that FASA consistently outperforms state-of-the-art methods, validating the effectiveness of explicit foreground modeling and pseudo-mask guidance for robust scene decomposition and object-coherent representation. Code will be made publicly available.

Method: MoReGen framework combines multi-agent LLMs, physics simulators, and renderers to generate reproducible videos in the code domain. Introduces object-trajectory correspondence as evaluation metric and MoReSet benchmark with 1,275 annotated videos across 9 Newtonian phenomena classes.

Result: State-of-the-art T2V models show poor physical validity, while MoReGen establishes a principled approach toward physically coherent video synthesis with better physical accuracy.

Conclusion: MoReGen provides a systematic framework for physics-grounded video generation, addressing core challenges in physical validity and motion coherence through integrated simulation and evaluation.

Abstract: While text-to-video (T2V) generation has achieved remarkable progress in photorealism, generating intent-aligned videos that faithfully obey physics principles remains a core challenge. In this work, we systematically study Newtonian motion-controlled text-to-video generation and evaluation, emphasizing physical precision and motion coherence. We introduce MoReGen, a motion-aware, physics-grounded T2V framework that integrates multi-agent LLMs, physics simulators, and renderers to generate reproducible, physically accurate videos from text prompts in the code domain. To quantitatively assess physical validity, we propose object-trajectory correspondence as a direct evaluation metric and present MoReSet, a benchmark of 1,275 human-annotated videos spanning nine classes of Newtonian phenomena with scene descriptions, spatiotemporal relations, and ground-truth trajectories. Using MoReSet, we conduct experiments on existing T2V models, evaluating their physical validity through both our MoRe metrics and existing physics-based evaluators. Our results reveal that state-of-the-art models struggle to maintain physical validity, while MoReGen establishes a principled direction toward physically coherent video synthesis.

Method: SSMP trains a Transformer encoder using masked prediction, where the model reconstructs trailer shot sequences from randomly masked versions. It features self-paced masking (adaptive difficulty) and progressive self-correction during generation (filling high-confidence positions first and re-masking remaining positions).

Result: SSMP achieves state-of-the-art results in automatic trailer generation, with both quantitative metrics and user studies demonstrating superiority over existing methods.

Conclusion: The proposed SSMP method successfully addresses limitations of traditional trailer generation approaches through its self-paced, self-corrective masked prediction framework, producing higher-quality trailers that mimic human editing workflows.

Abstract: As a challenging video editing task, movie trailer generation involves selecting and reorganizing movie shots to create engaging trailers. Currently, most existing automatic trailer generation methods employ a “selection-then-ranking” paradigm (i.e., first selecting key shots and then ranking them), which suffers from inevitable error propagation and limits the quality of the generated trailers. Beyond this paradigm, we propose a new self-paced and self-corrective masked prediction method called SSMP, which achieves state-of-the-art results in automatic trailer generation via bi-directional contextual modeling and progressive self-correction. In particular, SSMP trains a Transformer encoder that takes the movie shot sequences as prompts and generates corresponding trailer shot sequences accordingly. The model is trained via masked prediction, reconstructing each trailer shot sequence from its randomly masked counterpart. The mask ratio is self-paced, allowing the task difficulty to adapt to the model and thereby improving model performance. When generating a movie trailer, the model fills the shot positions with high confidence at each step and re-masks the remaining positions for the next prediction, forming a progressive self-correction mechanism that is analogous to how human editors work. Both quantitative results and user studies demonstrate the superiority of SSMP in comparison to existing automatic movie trailer generation methods. Demo is available at: https://github.com/Dixin-Lab/SSMP.

Method: Proposes Continual Customized Video Diffusion (CCVD) with: 1) concept-specific attribute retention module and task-aware concept aggregation strategy to prevent forgetting, and 2) controllable conditional synthesis with layer-specific region attention-guided noise estimation to address concept neglect.

Result: CCVD outperforms existing CTVG baselines on both DreamVideo and Wan 2.1 backbones, demonstrating superior performance in continual learning of new concepts.

Conclusion: CCVD effectively addresses the challenges of continual learning in customized text-to-video generation, enabling continuous acquisition of new concepts while maintaining performance on previously learned ones.

Abstract: Customized text-to-video generation (CTVG) has recently witnessed great progress in generating tailored videos from user-specific text. However, most CTVG methods assume that personalized concepts remain static and do not expand incrementally over time. Additionally, they struggle with forgetting and concept neglect when continuously learning new concepts, including subjects and motions. To resolve the above challenges, we develop a novel Continual Customized Video Diffusion (CCVD) model, which can continuously learn new concepts to generate videos across various text-to-video generation tasks by tackling forgetting and concept neglect. To address catastrophic forgetting, we introduce a concept-specific attribute retention module and a task-aware concept aggregation strategy. They can capture the unique characteristics and identities of old concepts during training, while combining all subject and motion adapters of old concepts based on their relevance during testing. Besides, to tackle concept neglect, we develop a controllable conditional synthesis to enhance regional features and align video contexts with user conditions, by incorporating layer-specific region attention-guided noise estimation. Extensive experimental comparisons demonstrate that our CCVD outperforms existing CTVG baselines on both the DreamVideo and Wan 2.1 backbones. The code is available at https://github.com/JiahuaDong/CCVD.

Method: Empirical evaluation benchmarking SAM 3’s zero-shot segmentation with point and bounding box prompts, testing language prompt segmentation, and investigating SAM 3D’s depth reconstruction from 2D images. Comprehensive testing on MICCAI EndoVis 2017/2018 benchmarks for segmentation, and SCARED, StereoMIS, EndoNeRF for 3D evaluation.

Result: SAM 3 shows clear improvements over SAM and SAM 2 in image and video segmentation with spatial prompts. SAM 3D demonstrates strong monocular depth estimation and realistic 3D instrument reconstruction. However, language prompts perform suboptimally in surgical domain, and both models show limitations in complex, highly dynamic surgical scenes.

Conclusion: SAM 3 and SAM 3D represent significant advancements for surgical applications with improved segmentation and 3D reconstruction capabilities, but require domain-specific training for language prompts and further development for complex dynamic surgical environments.

Abstract: The recent SAM 3 and SAM 3D have introduced significant advancements over the predecessor, SAM 2, particularly with the integration of language-based segmentation and enhanced 3D perception capabilities. SAM 3 supports zero-shot segmentation across a wide range of prompts, including point, bounding box, and language-based prompts, allowing for more flexible and intuitive interactions with the model. In this empirical evaluation, we assess the performance of SAM 3 in robot-assisted surgery, benchmarking its zero-shot segmentation with point and bounding box prompts and exploring its effectiveness in dynamic video tracking, alongside its newly introduced language prompt segmentation. While language prompts show potential, their performance in the surgical domain is currently suboptimal, highlighting the need for further domain-specific training. Additionally, we investigate SAM 3D’s depth reconstruction abilities, demonstrating its capacity to process surgical scene data and reconstruct 3D anatomical structures from 2D images. Through comprehensive testing on the MICCAI EndoVis 2017 and EndoVis 2018 benchmarks, SAM 3 shows clear improvements over SAM and SAM 2 in both image and video segmentation under spatial prompts, while the zero-shot evaluations of SAM 3D on SCARED, StereoMIS, and EndoNeRF indicate strong monocular depth estimation and realistic 3D instrument reconstruction, yet also reveal remaining limitations in complex, highly dynamic surgical scenes.

Method: Four-stage pipeline: 1) Video curation with quality filtering, 2) Cross-frame subject mining/pairing using VLM-based category consensus and diversity-aware pairing, 3) Identity-preserving reference image synthesis via segmentation-guided outpainting and box-guided inpainting, 4) Verification/captioning with VLM validation and caption construction.

Result: Created OpenSubject dataset with 2.5M samples and 4.35M images, established benchmark for subject-driven generation/manipulation, and showed that training with OpenSubject improves performance in complex scenes.

Conclusion: OpenSubject addresses limitations in subject-driven generation by providing high-quality training data that preserves identity fidelity, particularly beneficial for complex multi-subject scenarios.

Abstract: Despite the promising progress in subject-driven image generation, current models often deviate from the reference identities and struggle in complex scenes with multiple subjects. To address this challenge, we introduce OpenSubject, a video-derived large-scale corpus with 2.5M samples and 4.35M images for subject-driven generation and manipulation. The dataset is built with a four-stage pipeline that exploits cross-frame identity priors. (i) Video Curation. We apply resolution and aesthetic filtering to obtain high-quality clips. (ii) Cross-Frame Subject Mining and Pairing. We utilize vision-language model (VLM)-based category consensus, local grounding, and diversity-aware pairing to select image pairs. (iii) Identity-Preserving Reference Image Synthesis. We introduce segmentation map-guided outpainting to synthesize the input images for subject-driven generation and box-guided inpainting to generate input images for subject-driven manipulation, together with geometry-aware augmentations and irregular boundary erosion. (iv) Verification and Captioning. We utilize a VLM to validate synthesized samples, re-synthesize failed samples based on stage (iii), and then construct short and long captions. In addition, we introduce a benchmark covering subject-driven generation and manipulation, and then evaluate identity fidelity, prompt adherence, manipulation consistency, and background consistency with a VLM judge. Extensive experiments show that training with OpenSubject improves generation and manipulation performance, particularly in complex scenes.

Method: Proposes C-DIRA framework combining: 1) Saliency-driven Top-K ROI pooling with fused classification for local feature extraction, 2) Dynamic ROI routing that applies ROI inference only to high-difficulty samples to reduce computation, and 3) Pseudo-domain labeling with adversarial learning to learn domain-invariant features robust to driver/background variations.

Result: On State Farm Distracted Driver Detection Dataset, C-DIRA maintains high accuracy with significantly fewer FLOPs and lower latency than prior lightweight models. Shows robustness to visual degradation (blur, low-light) and stable performance across unseen domains.

Conclusion: C-DIRA effectively achieves compactness, efficiency, and generalization for driver behavior recognition, addressing the trade-off between computational efficiency and accuracy while maintaining robustness to domain variations and visual degradation.

Abstract: Driver distraction behavior recognition using in-vehicle cameras demands real-time inference on edge devices. However, lightweight models often fail to capture fine-grained behavioral cues, resulting in reduced performance on unseen drivers or under varying conditions. ROI-based methods also increase computational cost, making it difficult to balance efficiency and accuracy. This work addresses the need for a lightweight architecture that overcomes these constraints. We propose Computationally efficient Dynamic region of Interest Routing and domain-invariant Adversarial learning for lightweight driver behavior recognition (C-DIRA). The framework combines saliency-driven Top-K ROI pooling and fused classification for local feature extraction and integration. Dynamic ROI routing enables selective computation by applying ROI inference only to high difficulty data samples. Moreover, pseudo-domain labeling and adversarial learning are used to learn domain-invariant features robust to driver and background variation. Experiments on the State Farm Distracted Driver Detection Dataset show that C-DIRA maintains high accuracy with significantly fewer FLOPs and lower latency than prior lightweight models. It also demonstrates robustness under visual degradation such as blur and low-light, and stable performance across unseen domains. These results confirm C-DIRA’s effectiveness in achieving compactness, efficiency, and generalization.

Method: Uses unified transformer architecture with novel querying mechanism that avoids dense per-frame decoding and multiple task-specific decoders. Allows flexible probing of 3D position of any point in space and time.

Result: Sets new state-of-the-art, outperforming previous methods across wide spectrum of 4D reconstruction tasks. Enables lightweight, highly scalable training and inference.

Conclusion: D4RT provides an efficient, scalable solution for dynamic scene reconstruction from video, achieving superior performance through its unified architecture and flexible querying mechanism.

Abstract: Understanding and reconstructing the complex geometry and motion of dynamic scenes from video remains a formidable challenge in computer vision. This paper introduces D4RT, a simple yet powerful feedforward model designed to efficiently solve this task. D4RT utilizes a unified transformer architecture to jointly infer depth, spatio-temporal correspondence, and full camera parameters from a single video. Its core innovation is a novel querying mechanism that sidesteps the heavy computation of dense, per-frame decoding and the complexity of managing multiple, task-specific decoders. Our decoding interface allows the model to independently and flexibly probe the 3D position of any point in space and time. The result is a lightweight and highly scalable method that enables remarkably efficient training and inference. We demonstrate that our approach sets a new state of the art, outperforming previous methods across a wide spectrum of 4D reconstruction tasks. We refer to the project webpage for animated results: https://d4rt-paper.github.io/.

Method: Conducted a qualitative study with 192 participants to explore trust dynamics in everyday LLM use, building on existing trust models by Lee & See and Afroogh et al.

Result: Hallucinations lead to context-sensitive trust calibration rather than blanket mistrust. Confirmed existing trust factors (expectancy, prior experience, user expertise) and identified intuition as an additional factor for hallucination detection. Found contextual factors like perceived risk and decision stakes influence trust dynamics.

Conclusion: Validated recursive trust calibration process and extended it with intuition as a user-related trust factor. Proposed practical recommendations for responsible and reflective LLM use based on trust calibration insights.

Abstract: Hallucinations are outputs by Large Language Models (LLMs) that are factually incorrect yet appear plausible [1]. This paper investigates how such hallucinations influence users’ trust in LLMs and users’ interaction with LLMs. To explore this in everyday use, we conducted a qualitative study with 192 participants. Our findings show that hallucinations do not result in blanket mistrust but instead lead to context-sensitive trust calibration. Building on the calibrated trust model by Lee & See [2] and Afroogh et al.’s trust-related factors [3], we confirm expectancy [3], [4], prior experience [3], [4], [5], and user expertise & domain knowledge [3], [4] as userrelated (human) trust factors, and identify intuition as an additional factor relevant for hallucination detection. Additionally, we found that trust dynamics are further influenced by contextual factors, particularly perceived risk [3] and decision stakes [6]. Consequently, we validate the recursive trust calibration process proposed by Blöbaum [7] and extend it by including intuition as a user-related trust factor. Based on these insights, we propose practical recommendations for responsible and reflective LLM use.

Method: AI TIPS (Artificial Intelligence Trust-Integrated Pillars for Sustainability) 2.0 is presented as an updated comprehensive operational framework that directly addresses these governance gaps, building on the original 2019 framework developed before NIST’s AI Risk Management Framework.

Result: The paper introduces AI TIPS 2.0 as a solution that provides tailored risk assessment for specific use cases, translates governance principles into actionable technical controls, and enables systematic operationalization of trustworthy AI practices throughout the development lifecycle.

Conclusion: AI TIPS 2.0 offers a practical framework to overcome current AI governance limitations by providing use-case-specific risk assessment, actionable controls, and scalable operational mechanisms for implementing trustworthy AI across organizations.

Abstract: The deployment of AI systems faces three critical governance challenges that current frameworks fail to adequately address. First, organizations struggle with inadequate risk assessment at the use case level, exemplified by the Humana class action lawsuit and other high impact cases where an AI system deployed to production exhibited both significant bias and high error rates, resulting in improper healthcare claim denials. Each AI use case presents unique risk profiles requiring tailored governance, yet most frameworks provide one size fits all guidance. Second, existing frameworks like ISO 42001 and NIST AI RMF remain at high conceptual levels, offering principles without actionable controls, leaving practitioners unable to translate governance requirements into specific technical implementations. Third, organizations lack mechanisms for operationalizing governance at scale, with no systematic approach to embed trustworthy AI practices throughout the development lifecycle, measure compliance quantitatively, or provide role-appropriate visibility from boards to data scientists. We present AI TIPS, Artificial Intelligence Trust-Integrated Pillars for Sustainability 2.0, update to the comprehensive operational framework developed in 2019,four years before NIST’s AI Risk Management Framework, that directly addresses these challenges.

Method: Develops a formal, categorical framework for analyzing content transformation into truth-evaluated propositions about state space of possible worlds W

Result: Demonstrates that LLMs circumvent rather than solve the symbol grounding problem through their proposition transformation mechanisms

Conclusion: LLMs avoid the fundamental symbol grounding problem by operating on proposition transformations rather than establishing genuine symbol-world connections

Abstract: This paper presents a formal, categorical framework for analysing how humans and large language models (LLMs) transform content into truth-evaluated propositions about a state space of possible worlds W , in order to argue that LLMs do not solve but circumvent the symbol grounding problem.

Method: Built around a standardized Dialog representation, SDialog offers: 1) persona-driven multi-agent simulation with composable orchestration, 2) comprehensive evaluation combining linguistic metrics, LLM-as-a-judge, and functional correctness validators, 3) mechanistic interpretability tools for activation inspection and steering, and 4) audio generation with full acoustic simulation.

Result: SDialog provides a unified framework that integrates with all major LLM backends, enabling mixed-backend experiments under a unified API for building and analyzing conversational agents.

Conclusion: SDialog enables researchers to build, benchmark, and understand conversational systems more systematically by coupling generation, evaluation, and interpretability in a dialog-centric architecture.

Abstract: We present SDialog, an MIT-licensed open-source Python toolkit that unifies dialog generation, evaluation and mechanistic interpretability into a single end-to-end framework for building and analyzing LLM-based conversational agents. Built around a standardized \texttt{Dialog} representation, SDialog provides: (1) persona-driven multi-agent simulation with composable orchestration for controlled, synthetic dialog generation, (2) comprehensive evaluation combining linguistic metrics, LLM-as-a-judge and functional correctness validators, (3) mechanistic interpretability tools for activation inspection and steering via feature ablation and induction, and (4) audio generation with full acoustic simulation including 3D room modeling and microphone effects. The toolkit integrates with all major LLM backends, enabling mixed-backend experiments under a unified API. By coupling generation, evaluation, and interpretability in a dialog-centric architecture, SDialog enables researchers to build, benchmark and understand conversational systems more systematically.

Method: The study examines image labeling performance using low-resolution, perceptually degraded stimuli. It contrasts human strategies (analogical reasoning, shape-based recognition, confidence modulation) with AI’s feature-based processing. Human behavior is analyzed through cognitive architectures (ACT-R, Soar) while AI is examined using Grad-CAM visualizations. The approach is grounded in Marr’s tri-level hypothesis, Simon’s bounded rationality, and Thagard’s frameworks.

Result: The findings reveal key parallels and divergences between biological and artificial systems in representation, inference, and confidence calibration. The analysis shows how humans use layered, heuristic decision strategies under uncertainty compared to AI’s feature-based processing.

Conclusion: The study motivates future neuro-symbolic architectures that unify structured symbolic reasoning with connectionist representations. Such architectures, informed by principles of embodiment, explainability, and cognitive alignment, offer a path toward AI systems that are both performant and interpretable.

Abstract: Understanding how humans and AI systems interpret ambiguous visual stimuli offers critical insight into the nature of perception, reasoning, and decision-making. This paper examines image labeling performance across human participants and deep neural networks, focusing on low-resolution, perceptually degraded stimuli. Drawing from computational cognitive science, cognitive architectures, and connectionist-symbolic hybrid models, we contrast human strategies such as analogical reasoning, shape-based recognition, and confidence modulation with AI’s feature-based processing. Grounded in Marr’s tri-level hypothesis, Simon’s bounded rationality, and Thagard’s frameworks of representation and emotion, we analyze participant responses in relation to Grad-CAM visualizations of model attention. Human behavior is further interpreted through cognitive principles modeled in ACT-R and Soar, revealing layered and heuristic decision strategies under uncertainty. Our findings highlight key parallels and divergences between biological and artificial systems in representation, inference, and confidence calibration. The analysis motivates future neuro-symbolic architectures that unify structured symbolic reasoning with connectionist representations. Such architectures, informed by principles of embodiment, explainability, and cognitive alignment, offer a path toward AI systems that are not only performant but also interpretable and cognitively grounded.

Method: Proposes an architectural framework with principled componentization (goal manager, planner, tool-router, executor, memory, verifiers, safety monitor, telemetry), disciplined interfaces with schema constraints and validation, and explicit control/assurance loops. Develops a taxonomy of agent types and analyzes their reliability characteristics.

Result: Provides a comprehensive design framework for reliable AI agents, including specific guidance on typed schemas, idempotency, permissioning, transactional semantics, memory management, runtime governance, and simulate-before-actuate safeguards.

Conclusion: Reliability in AI systems emerges from proper architectural design with componentization, disciplined interfaces, and control loops, enabling systematic construction of trustworthy agentic systems across various application domains.

Abstract: This chapter argues that the reliability of agentic and generative AI is chiefly an architectural property. We define agentic systems as goal-directed, tool-using decision makers operating in closed loops, and show how reliability emerges from principled componentisation (goal manager, planner, tool-router, executor, memory, verifiers, safety monitor, telemetry), disciplined interfaces (schema-constrained, validated, least-privilege tool calls), and explicit control and assurance loops. Building on classical foundations, we propose a practical taxonomy-tool-using agents, memory-augmented agents, planning and self-improvement agents, multi-agent systems, and embodied or web agents - and analyse how each pattern reshapes the reliability envelope and failure modes. We distil design guidance on typed schemas, idempotency, permissioning, transactional semantics, memory provenance and hygiene, runtime governance (budgets, termination conditions), and simulate-before-actuate safeguards.

Method: Trio integrates three components: 1) fragment-based molecular language modeling for context-aware fragment assembly, 2) reinforcement learning to enforce physicochemical and synthetic feasibility, and 3) Monte Carlo tree search to balance exploration of novel chemotypes with exploitation of promising intermediates in protein binding pockets.

Result: Trio outperforms state-of-the-art approaches with improvements in binding affinity (+7.85%), drug-likeness (+11.10%), and synthetic accessibility (+12.05%), while expanding molecular diversity more than fourfold.

Conclusion: Trio provides an effective, interpretable closed-loop framework for targeted molecular design that generates chemically valid, pharmacologically enhanced ligands, addressing key limitations of current generative approaches in drug discovery.

Abstract: Drug discovery is a time-consuming and expensive process, with traditional high-throughput and docking-based virtual screening hampered by low success rates and limited scalability. Recent advances in generative modelling, including autoregressive, diffusion, and flow-based approaches, have enabled de novo ligand design beyond the limits of enumerative screening. Yet these models often suffer from inadequate generalization, limited interpretability, and an overemphasis on binding affinity at the expense of key pharmacological properties, thereby restricting their translational utility. Here we present Trio, a molecular generation framework integrating fragment-based molecular language modeling, reinforcement learning, and Monte Carlo tree search, for effective and interpretable closed-loop targeted molecular design. Through the three key components, Trio enables context-aware fragment assembly, enforces physicochemical and synthetic feasibility, and guides a balanced search between the exploration of novel chemotypes and the exploitation of promising intermediates within protein binding pockets. Experimental results show that Trio reliably achieves chemically valid and pharmacologically enhanced ligands, outperforming state-of-the-art approaches with improved binding affinity (+7.85%), drug-likeness (+11.10%) and synthetic accessibility (+12.05%), while expanding molecular diversity more than fourfold.

Method: Geometric modeling where each agent has a personalized value space (vector space for interpreting meaning), beliefs are structured vectors, and communication is mediated by linear interpretation maps. Belief survival depends on avoiding null spaces of these maps.

Result: Key findings: 1) Belief distortion, motivational drift, and limits of understanding arise from algebraic constraints; 2) “No-Null-Space Leadership Condition” characterizes leadership as representational reachability; 3) Framework explains how beliefs propagate, mutate, or disappear across cognitive geometries.

Conclusion: The cognitive-geometric perspective grounds meaning preservation in structural compatibility rather than shared information or rationality, clarifying epistemic boundaries of influence in both human and artificial systems, and providing a foundation for analyzing belief dynamics across heterogeneous agents.

Abstract: This paper develops a geometric framework for modeling belief, motivation, and influence across cognitively heterogeneous agents. Each agent is represented by a personalized value space, a vector space encoding the internal dimensions through which the agent interprets and evaluates meaning. Beliefs are formalized as structured vectors-abstract beings-whose transmission is mediated by linear interpretation maps. A belief survives communication only if it avoids the null spaces of these maps, yielding a structural criterion for intelligibility, miscommunication, and belief death. Within this framework, I show how belief distortion, motivational drift, counterfactual evaluation, and the limits of mutual understanding arise from purely algebraic constraints. A central result-“the No-Null-Space Leadership Condition”-characterizes leadership as a property of representational reachability rather than persuasion or authority. More broadly, the model explains how abstract beings can propagate, mutate, or disappear as they traverse diverse cognitive geometries. The account unifies insights from conceptual spaces, social epistemology, and AI value alignment by grounding meaning preservation in structural compatibility rather than shared information or rationality. I argue that this cognitive-geometric perspective clarifies the epistemic boundaries of influence in both human and artificial systems, and offers a general foundation for analyzing belief dynamics across heterogeneous agents.

Method: An orchestrator receives natural language specifications and converts them to PDDL models. Sub-modules (agents) iteratively refine domain and problem definitions to address planning requirements and resolve ambiguities. LLMs power the orchestrator and agents. Validated models are passed to external planning engines, and results are translated back to natural language.

Result: Demonstrated flexibility and effectiveness across various domains including Google NaturalPlan benchmark, PlanBench, Blocksworld, and Tower of Hanoi (where LLMs typically struggle). Framework works with multiple PDDL planning engines and validators (Fast Downward, LPG, POPF, VAL, uVAL).

Conclusion: The framework represents a significant step toward end-to-end planning aided by LLMs, requiring no human intervention and maintaining correctness while improving human readability through natural language translation.

Abstract: We present an end-to-end framework for planning supported by verifiers. An orchestrator receives a human specification written in natural language and converts it into a PDDL (Planning Domain Definition Language) model, where the domain and problem are iteratively refined by sub-modules (agents) to address common planning requirements, such as time constraints and optimality, as well as ambiguities and contradictions that may exist in the human specification. The validated domain and problem are then passed to an external planning engine to generate a plan. The orchestrator and agents are powered by Large Language Models (LLMs) and require no human intervention at any stage of the process. Finally, a module translates the final plan back into natural language to improve human readability while maintaining the correctness of each step. We demonstrate the flexibility and effectiveness of our framework across various domains and tasks, including the Google NaturalPlan benchmark and PlanBench, as well as planning problems like Blocksworld and the Tower of Hanoi (where LLMs are known to struggle even with small instances). Our framework can be integrated with any PDDL planning engine and validator (such as Fast Downward, LPG, POPF, VAL, and uVAL, which we have tested) and represents a significant step toward end-to-end planning aided by LLMs.

Method: The paper introduces a method that uses Gaussian Process Regression to obtain value estimates for promising actions that were not actually trialed in the environment. This allows for better aggregation of statistics across different threads in root-parallel MCTS for continuous action spaces.

Result: The approach was systematically evaluated across 6 different domains and demonstrated superior performance compared to existing aggregation strategies. The method requires only a modest increase in inference time while delivering better results.

Conclusion: Gaussian Process Regression provides an effective way to improve root-parallel Monte Carlo Tree Search in continuous action spaces by enabling better value estimation for untried actions, leading to improved performance with minimal computational overhead.

Abstract: Monte Carlo Tree Search is a cornerstone algorithm for online planning, and its root-parallel variant is widely used when wall clock time is limited but best performance is desired. In environments with continuous action spaces, how to best aggregate statistics from different threads is an important yet underexplored question. In this work, we introduce a method that uses Gaussian Process Regression to obtain value estimates for promising actions that were not trialed in the environment. We perform a systematic evaluation across 6 different domains, demonstrating that our approach outperforms existing aggregation strategies while requiring a modest increase in inference time.

Method: Systematically studies three fundamental factors: (1) relationship between solution optimality and execution performance, (2) sensitivity to kinodynamic modeling inaccuracies, and (3) interaction between model accuracy and plan optimality. Uses empirical examination of these factors in realistic scenarios.

Result: The paper provides empirical insights into how design choices affect performance in realistic MAPF scenarios, though specific quantitative results are not detailed in the abstract.

Conclusion: Highlights open challenges and research directions to guide the MAPF community toward practical, real-world deployment, emphasizing the need to bridge the gap between algorithmic benchmarks and actual robot performance.

Abstract: Multi-Agent Path Finding (MAPF) algorithms are increasingly deployed in industrial warehouses and automated manufacturing facilities, where robots must operate reliably under real-world physical constraints. However, existing MAPF evaluation frameworks typically rely on simplified robot models, leaving a substantial gap between algorithmic benchmarks and practical performance. Recent frameworks such as SMART, incorporate kinodynamic modeling and offer the MAPF community a platform for large-scale, realistic evaluation. Building on this capability, this work investigates how key planner design choices influence performance under realistic execution settings. We systematically study three fundamental factors: (1) the relationship between solution optimality and execution performance, (2) the sensitivity of system performance to inaccuracies in kinodynamic modeling, and (3) the interaction between model accuracy and plan optimality. Empirically, we examine these factors to understand how these design choices affect performance in realistic scenarios. We highlight open challenges and research directions to steer the community toward practical, real-world deployment.

Method: RIFT transforms fault search into sequential decision-making problem, combining hybrid sensitivity analysis for search space pruning with reinforcement learning to intelligently generate minimal, high-impact test suites.

Result: On billion-parameter LLM workloads using NVIDIA A100 GPUs, RIFT achieves 2.2× fault assessment speedup over evolutionary methods, reduces test vector volume by over 99% compared to random fault injection, and provides 12.8× improvement in cost-effectiveness for selective error correction.

Conclusion: RIFT provides a scalable framework for efficient design-time fault assessment that generates actionable verification artifacts and enables intelligent hardware protection strategies with significantly better cost-effectiveness than traditional approaches.

Abstract: The massive scale of modern AI accelerators presents critical challenges to traditional fault assessment methodologies, which face prohibitive computational costs and provide poor coverage of critical failure modes. This paper introduces RIFT (Reinforcement Learning-guided Intelligent Fault Targeting), a scalable framework that automates the discovery of minimal, high-impact fault scenarios for efficient design-time fault assessment. RIFT transforms the complex search for worst-case faults into a sequential decision-making problem, combining hybrid sensitivity analysis for search space pruning with reinforcement learning to intelligently generate minimal, high-impact test suites. Evaluated on billion-parameter Large Language Model (LLM) workloads using NVIDIA A100 GPUs, RIFT achieves a \textbf{2.2$\times$} fault assessment speedup over evolutionary methods and reduces the required test vector volume by over \textbf{99%} compared to random fault injection, all while achieving \textbf{superior fault coverage}. The proposed framework also provides actionable data to enable intelligent hardware protection strategies, demonstrating that RIFT-guided selective error correction code provides a \textbf{12.8$\times$} improvement in \textbf{cost-effectiveness} (coverage per unit area) compared to uniform triple modular redundancy protection. RIFT automatically generates UVM-compliant verification artifacts, ensuring its findings are directly actionable and integrable into commercial RTL verification workflows.

Method: Evaluated 10 human cybersecurity professionals alongside 6 existing AI agents and ARTEMIS (a new multi-agent framework) on a large university network with ~8,000 hosts across 12 subnets. ARTEMIS features dynamic prompt generation, arbitrary sub-agents, and automatic vulnerability triaging.

Result: ARTEMIS placed second overall, discovering 9 valid vulnerabilities with 82% valid submission rate, outperforming 9 of 10 human participants. AI agents showed advantages in systematic enumeration, parallel exploitation, and cost ($18/hour vs $60/hour for humans). However, they exhibited higher false-positive rates and struggled with GUI-based tasks.

Conclusion: AI cybersecurity agents like ARTEMIS demonstrate technical sophistication comparable to top human professionals while offering cost advantages, but still face limitations in false-positive rates and GUI-based tasks that need to be addressed for broader adoption.

Abstract: We present the first comprehensive evaluation of AI agents against human cybersecurity professionals in a live enterprise environment. We evaluate ten cybersecurity professionals alongside six existing AI agents and ARTEMIS, our new agent scaffold, on a large university network consisting of ~8,000 hosts across 12 subnets. ARTEMIS is a multi-agent framework featuring dynamic prompt generation, arbitrary sub-agents, and automatic vulnerability triaging. In our comparative study, ARTEMIS placed second overall, discovering 9 valid vulnerabilities with an 82% valid submission rate and outperforming 9 of 10 human participants. While existing scaffolds such as Codex and CyAgent underperformed relative to most human participants, ARTEMIS demonstrated technical sophistication and submission quality comparable to the strongest participants. We observe that AI agents offer advantages in systematic enumeration, parallel exploitation, and cost – certain ARTEMIS variants cost $18/hour versus $60/hour for professional penetration testers. We also identify key capability gaps: AI agents exhibit higher false-positive rates and struggle with GUI-based tasks.

Method: Developed MatSci-YAMZ platform integrating AI with human-in-the-loop crowdsourcing; conducted proof-of-concept with 6 materials science researchers who contributed term definitions and examples to refine AI-generated definitions through iterative feedback loops.

Result: Successfully created 19 AI-generated definitions with iterative refinement; demonstrated feasibility of AI-HILT model with successful proof-of-concept, alignment with FAIR/open-science principles, established research protocol, and scalability potential.

Conclusion: MatSci-YAMZ’s AI-HILT model can enhance semantic transparency and reduce time for consensus building in metadata vocabulary development, with potential for cross-domain scalability.

Abstract: Metadata vocabularies are essential for advancing FAIR and FARR data principles, but their development constrained by limited human resources and inconsistent standardization practices. This paper introduces MatSci-YAMZ, a platform that integrates artificial intelligence (AI) and human-in-the-loop (HILT), including crowdsourcing, to support metadata vocabulary development. The paper reports on a proof-of-concept use case evaluating the AI-HILT model in materials science, a highly interdisciplinary domain Six (6) participants affiliated with the NSF Institute for Data-Driven Dynamical Design (ID4) engaged with the MatSci-YAMZ plaform over several weeks, contributing term definitions and providing examples to prompt the AI-definitions refinement. Nineteen (19) AI-generated definitions were successfully created, with iterative feedback loops demonstrating the feasibility of AI-HILT refinement. Findings confirm the feasibility AI-HILT model highlighting 1) a successful proof of concept, 2) alignment with FAIR and open-science principles, 3) a research protocol to guide future studies, and 4) the potential for scalability across domains. Overall, MatSci-YAMZ’s underlying model has the capacity to enhance semantic transparency and reduce time required for consensus building and metadata vocabulary development.

Method: One-shot hierarchical planner that leverages LLM-generated subgoals only at initialization to pretrain a lightweight student model, deriving subgoals directly from example trajectories rather than repeated LLM prompting.

Result: Achieves 0.56 success rate (vs 0.52 for ADaPT) on TextCraft environment and reduces inference time from 164.4 seconds to 3.0 seconds (55x faster).

Conclusion: LLM-generated subgoals can serve as a strong starting point for hierarchical goal decomposition despite potential suboptimality, enabling efficient planning without repeated LLM queries.

Abstract: Long-term planning in complex, text-based environments presents significant challenges due to open-ended action spaces, ambiguous observations, and sparse feedback. Recent research suggests that large language models (LLMs) encode rich semantic knowledge about the world, which can be valuable for guiding agents in high-level reasoning and planning across both embodied and purely textual settings. However, existing approaches often depend heavily on querying LLMs during training and inference, making them computationally expensive and difficult to deploy efficiently. In addition, these methods typically employ a pretrained, unaltered LLM whose parameters remain fixed throughout training, providing no opportunity for adaptation to the target task. To address these limitations, we introduce SCOPE (Subgoal-COnditioned Pretraining for Efficient planning), a one-shot hierarchical planner that leverages LLM-generated subgoals only at initialization to pretrain a lightweight student model. Unlike prior approaches that distill LLM knowledge by repeatedly prompting the model to adaptively generate subgoals during training, our method derives subgoals directly from example trajectories. This design removes the need for repeated LLM queries, significantly improving efficiency, though at the cost of reduced explainability and potentially suboptimal subgoals. Despite their suboptimality, our results on the TextCraft environment show that LLM-generated subgoals can still serve as a strong starting point for hierarchical goal decomposition in text-based planning tasks. Compared to the LLM-based hierarchical agent ADaPT (Prasad et al., 2024), which achieves a 0.52 success rate, our method reaches 0.56 and reduces inference time from 164.4 seconds to just 3.0 seconds.

Method: Develop categories of Bayesian networks and Markov networks as functors from syntax to semantics domains. Model moralisation and triangulation as functors between these categories, with moralisation defined syntactically and triangulation involving semantics. Reinterpret variable elimination algorithm as a functor that splits triangulation into syntactic and semantic components.

Result: A categorical framework where moralisation and triangulation are formalized as functors, allowing inductive definition via functor pre-composition. Moralisation is shown to be purely syntactic while triangulation involves semantics. Variable elimination is reinterpreted as a functor that separates triangulation into syntactic and semantic parts.

Conclusion: The functorial perspective introduces a new categorical framework for probabilistic graphical models, clarifying the distinction between syntactic and semantic transformations and providing a foundation for further theoretical development in graphical model theory.

Abstract: Moralisation and Triangulation are transformations allowing to switch between different ways of factoring a probability distribution into a graphical model. Moralisation allows to view a Bayesian network (a directed model) as a Markov network (an undirected model), whereas triangulation addresses the opposite direction. We present a categorical framework where these transformations are modelled as functors between a category of Bayesian networks and one of Markov networks. The two kinds of network (the objects of these categories) are themselves represented as functors from a syntax' domain to a semantics’ codomain. Notably, moralisation and triangulation can be defined inductively on such syntax via functor pre-composition. Moreover, while moralisation is fully syntactic, triangulation relies on semantics. This leads to a discussion of the variable elimination algorithm, reinterpreted here as a functor in its own right, that splits the triangulation procedure in two: one purely syntactic, the other purely semantic. This approach introduces a functorial perspective into the theory of probabilistic graphical models, which highlights the distinctions between syntactic and semantic modifications.

Method: 1) Reformulate PMD to require generating visual self-dialog rationales; 2) Curate a benchmark dataset based on individual frames; 3) Use natural language inference (NLI) models to create two automated metrics for evaluating rationale coherence; 4) Establish baselines and test improvements by incorporating these metrics into inference and fine-tuning methods.

Result: Vision-and-language models (VLMs) struggle with PMD off-the-shelf, but their accuracy, coherence, and efficiency can be improved by incorporating the proposed coherence metrics into inference and fine-tuning methods, though with some trade-offs.

Conclusion: The proposed reformulation of PMD with visual self-dialog rationales enables unprecedented transparency. The automated coherence metrics provide valuable insights into model reasoning, highlighting areas for future improvement in procedural mistake detection systems.

Abstract: Procedural mistake detection (PMD) is a challenging problem of classifying whether a human user (observed through egocentric video) has successfully executed a task (specified by a procedural text). Despite significant recent efforts, machine performance in the wild remains nonviable, and the reasoning processes underlying this performance are opaque. As such, we extend PMD to require generating visual self-dialog rationales to inform decisions. Given the impressive, mature image understanding capabilities observed in recent vision-and-language models (VLMs), we curate a suitable benchmark dataset for PMD based on individual frames. As our reformulation enables unprecedented transparency, we leverage a natural language inference (NLI) model to formulate two automated metrics for the coherence of generated rationales. We establish baselines for this reframed task, showing that VLMs struggle off-the-shelf, but with some trade-offs, their accuracy, coherence, and efficiency can be improved by incorporating these metrics into common inference and fine-tuning methods. Lastly, our multi-faceted metrics visualize common outcomes, highlighting areas for further improvement.

Method: Introduces Forgetting-MarI framework that penalizes marginal information to remove only additional information contributed by data to be unlearned while preserving information from retained data, providing explicit upper bound on residual influence.

Result: Extensive experiments show the approach outperforms current state-of-the-art unlearning methods, delivering reliable forgetting and better preserved general model performance across diverse benchmarks.

Conclusion: This advancement represents an important step toward making AI systems more controllable and compliant with privacy and copyright regulations without compromising effectiveness.

Abstract: As AI models are trained on ever-expanding datasets, the ability to remove the influence of specific data from trained models has become essential for privacy protection and regulatory compliance. Unlearning addresses this challenge by selectively removing parametric knowledge from the trained models without retraining from scratch, which is critical for resource-intensive models such as Large Language Models (LLMs). Existing unlearning methods often degrade model performance by removing more information than necessary when attempting to ‘‘forget’’ specific data. We introduce Forgetting-MarI, an LLM unlearning framework that provably removes only the additional (marginal) information contributed by the data to be unlearned, while preserving the information supported by the data to be retained. By penalizing marginal information, our method yields an explicit upper bound on the unlearn dataset’s residual influence in the trained models, providing provable undetectability. Extensive experiments confirm that our approach outperforms current state-of-the-art unlearning methods, delivering reliable forgetting and better preserved general model performance across diverse benchmarks. This advancement represents an important step toward making AI systems more controllable and compliant with privacy and copyright regulations without compromising their effectiveness.

Method: Constructs local, multi-scale graphs centered on each agent; uses aggregation-decomposition message-passing scheme and type-specific feature networks to handle agent heterogeneity; models both direct pairwise dependencies and complex group-wise interactions.

Result: Achieves state-of-the-art performance on Waymo Open Motion Dataset, setting new benchmarks for key metrics including Soft mAP and minADE.

Conclusion: A locality-grounded architecture that explicitly models multi-scale and heterogeneous interactions is a highly effective strategy for advancing motion forecasting in autonomous driving.

Abstract: Multi-agent trajectory prediction in autonomous driving requires a comprehensive understanding of complex social dynamics. Existing methods, however, often struggle to capture the full richness of these dynamics, particularly the co-existence of multi-scale interactions and the diverse behaviors of heterogeneous agents. To address these challenges, this paper introduces HeLoFusion, an efficient and scalable encoder for modeling heterogeneous and multi-scale agent interactions. Instead of relying on global context, HeLoFusion constructs local, multi-scale graphs centered on each agent, allowing it to effectively model both direct pairwise dependencies and complex group-wise interactions (\textit{e.g.}, platooning vehicles or pedestrian crowds). Furthermore, HeLoFusion tackles the critical challenge of agent heterogeneity through an aggregation-decomposition message-passing scheme and type-specific feature networks, enabling it to learn nuanced, type-dependent interaction patterns. This locality-focused approach enables a principled representation of multi-level social context, yielding powerful and expressive agent embeddings. On the challenging Waymo Open Motion Dataset, HeLoFusion achieves state-of-the-art performance, setting new benchmarks for key metrics including Soft mAP and minADE. Our work demonstrates that a locality-grounded architecture, which explicitly models multi-scale and heterogeneous interactions, is a highly effective strategy for advancing motion forecasting.

Method: BridgeDrive is a novel anchor-guided diffusion bridge policy that provides a principled diffusion framework to translate anchors (typical expert driving behaviors) into fine-grained trajectory plans. The method appropriately responds to varying traffic conditions and is compatible with efficient ODE solvers for real-time deployment.

Result: The approach achieves state-of-the-art performance on the Bench2Drive benchmark, improving the success rate by 7.72% over prior arts.

Conclusion: BridgeDrive successfully addresses the guidance challenge in diffusion-based autonomous driving planners by providing a theoretically consistent framework that effectively translates expert anchors into reactive trajectory plans while maintaining real-time compatibility.

Abstract: Diffusion-based planners have shown great promise for autonomous driving due to their ability to capture multi-modal driving behaviors. However, guiding these models effectively in reactive, closed-loop environments remains a significant challenge. Simple conditioning often fails to provide sufficient guidance in complex and dynamic driving scenarios. Recent work attempts to use typical expert driving behaviors (i.e., anchors) to guide diffusion models but relies on a truncated schedule, which introduces theoretical inconsistencies and can compromise performance. To address this, we introduce BridgeDrive, a novel anchor-guided diffusion bridge policy for closed-loop trajectory planning. Our approach provides a principled diffusion framework that effectively translates anchors into fine-grained trajectory plans, appropriately responding to varying traffic conditions. Our planner is compatible with efficient ODE solvers, a critical factor for real-time autonomous driving deployment. We achieve state-of-the-art performance on the Bench2Drive benchmark, improving the success rate by 7.72% over prior arts.

Method: Combines insertion-based Edit Flow for discrete text tokens with Flow Matching for image latents. Uses hierarchical sampling that prioritizes content over grammar, enabling concurrent text-image synthesis.

Result: Outperforms autoregressive baselines on both generation and understanding tasks across model sizes from 1B to 8B, using up to 50% fewer training FLOPs. Surpasses both autoregressive and diffusion-based approaches.

Conclusion: OneFlow unlocks new capabilities for concurrent generation, iterative refinement, and natural reasoning-like generation, demonstrating the effectiveness of non-autoregressive approaches for multimodal tasks.

Abstract: We present OneFlow, the first non-autoregressive multimodal model that enables variable-length and concurrent mixed-modal generation. Unlike autoregressive models that enforce rigid causal ordering between text and image generation, OneFlow combines an insertion-based Edit Flow for discrete text tokens with Flow Matching for image latents. OneFlow enables concurrent text-image synthesis with hierarchical sampling that prioritizes content over grammar. Through controlled experiments across model sizes from 1B to 8B, we demonstrate that OneFlow outperforms autoregressive baselines on both generation and understanding tasks while using up to 50% fewer training FLOPs. OneFlow surpasses both autoregressive and diffusion-based approaches while unlocking new capabilities for concurrent generation, iterative refinement, and natural reasoning-like generation.

Method: SCL is grounded in philosophy of mind and cognitive phenomenology, drawing on process philosophy, enactive cognition, and extended mind theory. It defines intelligence as a continuous loop of judgment, memory, control, action, and regulation - a performed process rather than a property.

Result: Three key contributions: 1) Operationalizes philosophical insights into computationally interpretable structures (executable epistemology), 2) Shows functional separation in cognitive architecture yields more coherent behavior than monolithic systems, 3) Redefines intelligence as capacity to reconstruct epistemic state through intentional understanding rather than representational accuracy.

Conclusion: SCL impacts philosophy (allowing theories to be enacted/tested), AI (grounding behavior in epistemic structure rather than statistics), and epistemology (framing knowledge as continuous reconstruction). Real progress requires architectures realizing cognitive principles structurally, not just larger models.

Abstract: Large language models exhibit intelligence without genuine epistemic understanding, exposing a key gap: the absence of epistemic architecture. This paper introduces the Structured Cognitive Loop (SCL) as an executable epistemological framework for emergent intelligence. Unlike traditional AI research asking “what is intelligence?” (ontological), SCL asks “under what conditions does cognition emerge?” (epistemological). Grounded in philosophy of mind and cognitive phenomenology, SCL bridges conceptual philosophy and implementable cognition. Drawing on process philosophy, enactive cognition, and extended mind theory, we define intelligence not as a property but as a performed process – a continuous loop of judgment, memory, control, action, and regulation. SCL makes three contributions. First, it operationalizes philosophical insights into computationally interpretable structures, enabling “executable epistemology” – philosophy as structural experiment. Second, it shows that functional separation within cognitive architecture yields more coherent and interpretable behavior than monolithic prompt based systems, supported by agent evaluations. Third, it redefines intelligence: not representational accuracy but the capacity to reconstruct its own epistemic state through intentional understanding. This framework impacts philosophy of mind, epistemology, and AI. For philosophy, it allows theories of cognition to be enacted and tested. For AI, it grounds behavior in epistemic structure rather than statistical regularity. For epistemology, it frames knowledge not as truth possession but as continuous reconstruction within a phenomenologically coherent loop. We situate SCL within debates on cognitive phenomenology, emergence, normativity, and intentionality, arguing that real progress requires not larger models but architectures that realize cognitive principles structurally.

Method: Proposed WorldTest protocol with three key components: 1) reward-free interaction phase, 2) test phase in a different but related environment, and 3) open-ended evaluation supporting many unknown tasks. Instantiated with AutumnBench - 43 interactive grid-world environments and 129 tasks across three families: masked-frame prediction, planning, and predicting changes to causal dynamics.

Result: Compared 517 human participants and three frontier models on AutumnBench. Humans outperformed the models, and scaling compute improved performance only in some environments but not others, revealing significant headroom in world-model learning.

Conclusion: WorldTest provides a novel template for evaluating what agents learn about environment dynamics, separating reward-free exploration from derived tests with behavior-based scoring. AutumnBench exposes significant gaps in current world-model learning capabilities compared to human performance.

Abstract: Model-learning agents should gather information to learn world models that support many downstream tasks and inferences, such as predicting unobserved states, estimating near- and far-term consequences of actions, planning action sequences, and detecting changes in dynamics. Current methods for learning and evaluating world models diverge from this goal: training and evaluation are anchored to next-frame prediction, and success is scored by reward maximization in the same environment. We propose WorldTest, a protocol to evaluate model-learning agents that separates reward-free interaction from a scored test phase in a different but related environment. WorldTest is open-ended $\unicode{x2014}$ models should support many different tasks unknown ahead of time $\unicode{x2014}$ and agnostic to model representation, allowing comparison across approaches. We instantiated WorldTest with AutumnBench, a suite of 43 interactive grid-world environments and 129 tasks across three families: masked-frame prediction, planning, and predicting changes to the causal dynamics. We compared 517 human participants and three frontier models on AutumnBench. We found that humans outperform the models, and scaling compute improves performance only in some environments but not others. WorldTest provides a novel template $\unicode{x2014}$ reward-free exploration, derived tests, and behavior-based scoring $\unicode{x2014}$ to evaluate what agents learn about environment dynamics, and AutumnBench exposes significant headroom in world-model learning.

Method: Constructed a dataset of code samples with correctness labels from the APPS coding challenge benchmark. Trained a value-head model (decoder-only transformer with regression layer) to estimate success probability of intermediate outputs through supervised fine-tuning, creating a reward model that considers both process and outcome rewards.

Result: Small LLMs are capable of serving as effective reward models or code evaluation critics, successfully identifying correct solutions among multiple candidates. Using this critic achieved over 20% improvement in the search capability of selecting the most accurate code out of multiple generations.

Conclusion: State-of-the-art small language models can be effectively transformed into usable reward models for code generation tasks, demonstrating that model size isn’t the only factor for effective reward modeling when properly combining process and outcome reward considerations.

Abstract: Generating high-quality code remains a challenge for Large Language Models (LLMs). For the evolution of reasoning models on this task, reward models are a necessary intermediate step. These models judge outcomes or intermediate steps. Decoder-only transformer models can be turned into reward models by introducing a regression layer and supervised fine-tuning. While it is known that reflection capabilities generally increase with the size of a model, we want to investigate whether state-of-the-art small language models like the Phi-4 family can be turned into usable reward models blending the consideration of process rewards and outcome rewards. Targeting this goal, we construct a dataset of code samples with correctness labels derived from the APPS coding challenge benchmark. We then train a value-head model to estimate the success probability of intermediate outputs. Our evaluation shows that small LLMs are capable of serving as effective reward models or code evaluation critics, successfully identifying correct solutions among multiple candidates. Using this critic, we achieve over a 20% improvement in the search capability of the most accurate code out of multiple generations.

Method: The paper analyzes conceptual mismatches between natural image methods and pathology requirements, identifying flaws in current approaches: dense embeddings can’t represent tissue combinatorial richness, self-supervision architectures have inherent flaws, patch design is inadequate, and pretraining is noise-fragile.

Result: The analysis reveals that pathology foundation models require explicit design for biological images rather than adaptations of natural-image methods, as biological complexity and limited domain innovation create fundamental gaps that current approaches cannot bridge.

Conclusion: Pathology demands models specifically designed for biological tissue characteristics, not adaptations of large-scale natural-image methods whose underlying assumptions don’t hold for tissue complexity and structure.

Abstract: Despite their successes in vision and language, foundation models have stumbled in pathology, revealing low accuracy, instability, and heavy computational demands. These shortcomings stem not from tuning problems but from deeper conceptual mismatches: dense embeddings cannot represent the combinatorial richness of tissue, and current architectures inherit flaws in self-supervision, patch design, and noise-fragile pretraining. Biological complexity and limited domain innovation further widen the gap. The evidence is clear-pathology requires models explicitly designed for biological images rather than adaptations of large-scale natural-image methods whose assumptions do not hold for tissue.

Method: Developed SnapStream, a KV cache compression method that works with static graphs and continuous batching. Explored accuracy implications on Llama-3.1-8B-Instruct and DeepSeek-R1, then deployed in production with 16-way tensor-parallel DeepSeek-671B on SambaNova SN40L accelerators.

Result: 4x improved on-chip memory usage with minimal accuracy degradation on LongBench-v2, AIME24 and LiveCodeBench benchmarks. Achieved 128k context length and up to 1832 tokens per second in real production setting.

Conclusion: SnapStream successfully addresses production deployment challenges by enabling sparse KV attention techniques in systems with static graphs and continuous batching, making it the first such implementation in production inference systems.

Abstract: The proliferation of 100B+ parameter Large Language Models (LLMs) with 100k+ context length support have resulted in increasing demands for on-chip memory to support large KV caches. Techniques such as StreamingLLM and SnapKV demonstrate how to control KV cache size while maintaining model accuracy. Yet, these techniques are not commonly used within industrial deployments using frameworks like vLLM or SGLang. The reason is twofold: on one hand, the static graphs and continuous batching methodology employed by these frameworks make it difficult to admit modifications to the standard multi-head attention algorithm, while on the other hand, the accuracy implications of such techniques on modern instruction-following and reasoning models are not well understood, obfuscating the need for implementing these techniques. In this paper, we explore these accuracy implications on Llama-3.1-8B-Instruct and DeepSeek-R1, and develop SnapStream, a KV cache compression method that can be deployed at scale. We demonstrate the efficacy of SnapStream in a 16-way tensor-parallel deployment of DeepSeek-671B on SambaNova SN40L accelerators running at 128k context length and up to 1832 tokens per second in a real production setting. SnapStream enables $4\times$ improved on-chip memory usage and introduces minimal accuracy degradation on LongBench-v2, AIME24 and LiveCodeBench. To the best of our knowledge, this is the first implementation of sparse KV attention techniques deployed in a production inference system with static graphs and continuous batching.

Method: ALIGN integrates large language and vision-language models in a multi-stage pipeline that performs optical character recognition, linguistic reasoning, and map-level verification through grid-based spatial scanning, systematically evaluating predicted locations against contextual and visual evidence.

Result: Applied to Bangla-language news data, ALIGN demonstrates consistent improvements over traditional geoparsing methods, accurately identifying district- and sub-district-level crash sites without requiring model retraining.

Conclusion: The framework establishes a high-accuracy foundation for automated crash mapping in data-scarce regions, supporting evidence-driven road-safety policymaking and broader integration of multimodal AI in transportation analytics.

Abstract: Reliable geospatial information on road accidents is vital for safety analysis and infrastructure planning, yet most low- and middle-income countries continue to face a critical shortage of accurate, location-specific crash data. Existing text-based geocoding tools perform poorly in multilingual and unstructured news environments, where incomplete place descriptions and mixed language (e.g. Bangla-English) scripts obscure spatial context. To address these limitations, this study introduces ALIGN (Accident Location Inference through Geo-Spatial Neural Reasoning), a vision-language framework that emulates human spatial reasoning to infer accident location coordinates directly from available textual and map-based cues. ALIGN integrates large language and vision-language model mechanisms within a multi-stage pipeline that performs optical character recognition, linguistic reasoning, and map-level verification through grid-based spatial scanning. The framework systematically evaluates each predicted location against contextual and visual evidence, ensuring interpretable, fine-grained geolocation outcomes without requiring model retraining. Applied to Bangla-language news data source, ALIGN demonstrates consistent improvements over traditional geoparsing methods, accurately identifying district- and sub-district-level crash sites. Beyond its technical contribution, the framework establishes a high accuracy foundation for automated crash mapping in data-scarce regions, supporting evidence-driven road-safety policymaking and the broader integration of multimodal artificial intelligence in transportation analytics.

Method: Two-stage approach: 1) Fine-tunes pretrained protein language models using masked language modeling on homologous sequences to activate evolutionary plausibility of target protein family. 2) Uses test-time inference with Monte Carlo tree search to evolve proteins with guidance from the fine-tuned model.

Result: AlphaDE remarkably outperforms previous state-of-the-art methods even with few-shot fine-tuning. A case study demonstrates successful condensation of the protein sequence space of avGFP through computational evolution.

Conclusion: AlphaDE bridges the gap between protein language models and advanced optimization techniques, providing an effective framework for computational protein evolution that leverages both fine-tuning and test-time inference strategies.

Abstract: Protein evolution through amino acid mutations is a cornerstone of life sciences. Recent advances in protein language models have shown rich evolutionary patterns, offering unprecedented potential for in-silicon directed evolution. However, existing directed evolution methods largely rely on heuristic evolution strategies and have yet to efficiently integrate the transformative protein language models with advanced optimization techniques, such as reinforcement learning, to learn optimal evolution policies. To bridge this gap, we propose AlphaDE, a novel framework that evolves protein sequences by harnessing the innovative paradigms of large language models, such as fine-tuning and test-time inference. First, AlphaDE fine-tunes pretrained protein language models using masked language modeling on homologous protein sequences to activate the evolutionary plausibility of the interested protein family. Second, AlphaDE introduces test-time inference based on Monte Carlo tree search, which effectively evolves proteins with evolutionary guidance from the fine-tuned protein language model. Extensive benchmark experiments show that AlphaDE remarkably outperforms previous state-of-the-art methods even with few-shot fine-tuning. A case study further demonstrates that AlphaDE supports condensing the protein sequence space of avGFP through computational evolution.

Method: PathMind follows a “Retrieve-Prioritize-Reason” paradigm: (1) Retrieves query subgraph from KG, (2) Uses semantic-aware path prioritization mechanism that considers both accumulative cost and estimated future cost to identify important reasoning paths, (3) Employs dual-phase training strategy including task-specific instruction tuning and path-wise preference alignment.

Result: Extensive experiments on benchmark datasets demonstrate that PathMind consistently outperforms competitive baselines, particularly on complex reasoning tasks with fewer input tokens, by identifying essential reasoning paths.

Conclusion: PathMind enhances faithful and interpretable reasoning by selectively guiding LLMs with important reasoning paths, addressing limitations of existing LLM-based KGR methods through intelligent path prioritization and efficient training strategies.

Abstract: Knowledge graph reasoning (KGR) is the task of inferring new knowledge by performing logical deductions on knowledge graphs. Recently, large language models (LLMs) have demonstrated remarkable performance in complex reasoning tasks. Despite promising success, current LLM-based KGR methods still face two critical limitations. First, existing methods often extract reasoning paths indiscriminately, without assessing their different importance, which may introduce irrelevant noise that misleads LLMs. Second, while many methods leverage LLMs to dynamically explore potential reasoning paths, they require high retrieval demands and frequent LLM calls. To address these limitations, we propose PathMind, a novel framework designed to enhance faithful and interpretable reasoning by selectively guiding LLMs with important reasoning paths. Specifically, PathMind follows a “Retrieve-Prioritize-Reason” paradigm. First, it retrieves a query subgraph from KG through the retrieval module. Next, it introduces a path prioritization mechanism that identifies important reasoning paths using a semantic-aware path priority function, which simultaneously considers the accumulative cost and the estimated future cost for reaching the target. Finally, PathMind generates accurate and logically consistent responses via a dual-phase training strategy, including task-specific instruction tuning and path-wise preference alignment. Extensive experiments on benchmark datasets demonstrate that PathMind consistently outperforms competitive baselines, particularly on complex reasoning tasks with fewer input tokens, by identifying essential reasoning paths.

Method: Formulate active inference optimizations as constrained divergence minimization problems that can be solved using standard mean field methods without appealing to expected free energy.

Result: The proposed perception/action divergence criterion coincides with variational free energy for perception modeling, but differs from expected free energy functional by an entropy regularizer for action modeling.

Conclusion: Active inference can be implemented independently of the Free Energy Principle through constrained divergence minimization, providing a more general framework for modeling perception and action in discrete state spaces.

Abstract: We seek to clarify the concept of active inference by disentangling it from the Free Energy Principle. We show how the optimizations that need to be carried out in order to implement active inference in discrete state spaces can be formulated as constrained divergence minimization problems which can be solved by standard mean field methods that do not appeal to the idea of expected free energy. When it is used to model perception, the perception/action divergence criterion that we propose coincides with variational free energy. When it is used to model action, it differs from an expected free energy functional by an entropy regularizer.

Method: Proposes a framework for persona-based agent selection and evaluates brainstorming outputs across different persona pairings (e.g., Doctor vs VR Engineer) and A2A dynamics (separate, together, separate-then-together) through multiple experimental setups.

Result: Three key findings: (1) persona choice shapes idea domains, (2) collaboration mode shifts diversity of idea generation, and (3) multi-agent persona-driven brainstorming produces both idea depth and cross-domain coverage.

Conclusion: Persona-based multi-agent brainstorming is valuable for generating diverse and deep ideas, with persona curation and collaboration dynamics being critical factors for effective ideation.

Abstract: We demonstrate the importance of persona-based multi-agents brainstorming for both diverse topics and subject matter ideation. Prior work has shown that generalized multi-agent collaboration often provides better reasoning than a single agent alone. In this paper, we propose and develop a framework for persona-based agent selection, showing how persona domain curation can improve brainstorming outcomes. Using multiple experimental setups, we evaluate brainstorming outputs across different persona pairings (e.g., Doctor vs VR Engineer) and A2A (agent-to-agent) dynamics (separate, together, separate-then-together). Our results show that (1) persona choice shapes idea domains, (2) collaboration mode shifts diversity of idea generation, and (3) multi-agent persona-driven brainstorming produces idea depth and cross-domain coverage.

Method: Uses textual experiential memory by learning contrastive reasoning trajectories to generate interpretable summaries of successful patterns, plus adaptive reward shaping that penalizes redundant tool calls only in correct-answer scenarios.

Result: Maintains accuracy comparable to SOTA baseline ReSearch while reducing search tool invocations by 39.6%, inference time by 48.6%, and token consumption by 21.2% on four multi-hop QA benchmarks.

Conclusion: LightSearcher effectively balances the accuracy-efficiency trade-off in DeepSearch paradigms, demonstrating superior efficiency without compromising accuracy.

Abstract: DeepSearch paradigms have become a core enabler for deep reasoning models, allowing them to invoke external search tools to access up-to-date, domain-specific knowledge beyond parametric boundaries, thereby enhancing the depth and factual reliability of reasoning. Building upon this foundation, recent advances in reinforcement learning (RL) have further empowered models to autonomously and strategically control search tool usage, optimizing when and how to query external knowledge sources. Yet, these RL-driven DeepSearch systems often reveal a see-saw trade-off between accuracy and efficiency-frequent tool invocations can improve factual correctness but lead to unnecessary computational overhead and diminished efficiency. To address this challenge, we propose LightSearcher, an efficient RL framework that incorporates textual experiential memory by learning contrastive reasoning trajectories to generate interpretable summaries of successful reasoning patterns. In addition, it employs an adaptive reward shaping mechanism that penalizes redundant tool calls only in correct-answer scenarios. This design effectively balances the inherent accuracy-efficiency trade-off in DeepSearch paradigms. Experiments on four multi-hop QA benchmarks show that LightSearcher maintains accuracy comparable to SOTA baseline ReSearch, while reducing search tool invocations by 39.6%, inference time by 48.6%, and token consumption by 21.2%, demonstrating its superior efficiency.

Method: Extends wav2vec2 framework with a dedicated noise identification module that operates concurrently with speech transcription, using joint optimization of transcription and noise classification objectives.

Result: Substantial improvements in transcription quality and noise discrimination, achieving superior performance in word error rate, character error rate, and noise detection accuracy compared to conventional architectures.

Conclusion: Joint optimization of transcription and noise classification yields more reliable speech recognition in challenging acoustic conditions, demonstrating the effectiveness of integrated noise detection in ASR systems.

Abstract: This research presents a novel approach to enhancing automatic speech recognition systems by integrating noise detection capabilities directly into the recognition architecture. Building upon the wav2vec2 framework, the proposed method incorporates a dedicated noise identification module that operates concurrently with speech transcription. Experimental validation using publicly available speech and environmental audio datasets demonstrates substantial improvements in transcription quality and noise discrimination. The enhanced system achieves superior performance in word error rate, character error rate, and noise detection accuracy compared to conventional architectures. Results indicate that joint optimization of transcription and noise classification objectives yields more reliable speech recognition in challenging acoustic conditions.

Method: Three-stage annotation framework combining human judgment with structured feedback and iterative refinement to curate training data and improve benchmark quality. Models variability in human judgments using Beta distributions to predict both expected correctness and uncertainty.

Result: Collected 11,721 annotations across 3,580 question-answer pairs from 15 LALMs on two audio QA benchmarks, achieving inter-annotator agreement of 0.82 (Krippendorff’s alpha). ORCA achieves 0.91 Spearman correlation with mean human judgments, matching or outperforming LLM-judge baselines while providing uncertainty estimates with significantly less compute.

Conclusion: ORCA provides a more nuanced evaluation framework for open-ended audio QA responses by modeling judgment uncertainty, offering better alignment with human evaluation while being computationally efficient. The framework, models, code, and curated dataset are released for community use.

Abstract: Evaluating open-ended responses from large audio language models (LALMs) is challenging because human annotators often genuinely disagree on answer correctness due to multiple valid interpretations, partial correctness, and subjective judgment. Traditional metrics reporting only mean scores fail to capture this uncertainty. We present ORCA (Open-ended Response Correctness Assessment), a framework that models the variability in human judgments using Beta distributions to predict both expected correctness and uncertainty. Our three-stage annotation framework combines human judgment with structured feedback and iterative refinement to simultaneously curate training data and improve benchmark quality. We collected 11,721 annotations across 3,580 question-answer pairs from 15 LALMs on two audio QA benchmarks, achieving inter-annotator agreement of 0.82 (Krippendorff’s alpha). ORCA achieves 0.91 Spearman correlation with mean human judgments, matching or outperforming LLM-judge baselines while providing uncertainty estimates and requiring significantly less compute. We release our models, code, and curated dataset.

Method: Leverages spatial diversity of ubiquitous objects in meeting rooms. Uses noise-robust unsupervised approach to distinguish participants by detecting speech-induced vibration differences in frequency domain. Employs deep learning framework to combine signals from multiple objects for speech quality enhancement.

Result: Achieves speech classification accuracy up to 0.99 with several participants. Demonstrates consistent speech quality enhancement across real-world scenarios including different radar-object distances.

Conclusion: Proposes a novel eavesdropping attack system that can remotely identify individual speakers in multi-participant meetings without prior knowledge, representing a significant privacy threat for sensitive discussions.

Abstract: Multi-participant meetings occur across various domains, such as business negotiations and medical consultations, during which sensitive information like trade secrets, business strategies, and patient conditions is often discussed. Previous research has demonstrated that attackers with mmWave radars outside the room can overhear meeting content by detecting minute speech-induced vibrations on objects. However, these eavesdropping attacks cannot differentiate which speech content comes from which person in a multi-participant meeting, leading to potential misunderstandings and poor decision-making. In this paper, we answer the question ``who speaks what’’. By leveraging the spatial diversity introduced by ubiquitous objects, we propose an attack system that enables attackers to remotely eavesdrop on in-person conversations without requiring prior knowledge, such as identities, the number of participants, or seating arrangements. Since participants in in-person meetings are typically seated at different locations, their speech induces distinct vibration patterns on nearby objects. To exploit this, we design a noise-robust unsupervised approach for distinguishing participants by detecting speech-induced vibration differences in the frequency domain. Meanwhile, a deep learning-based framework is explored to combine signals from objects for speech quality enhancement. We validate the proof-of-concept attack on speech classification and signal enhancement through extensive experiments. The experimental results show that our attack can achieve the speech classification accuracy of up to $0.99$ with several participants in a meeting room. Meanwhile, our attack demonstrates consistent speech quality enhancement across all real-world scenarios, including different distances between the radar and the objects.

Method: Uses latent Diffusion Transformer (DiT) with CLAP-based style encoder for multi-modal alignment, chained classifier-free guidance for independent control, and Representation Alignment to distill Whisper features for stabilization.

Result: DMP-TTS achieves stronger style controllability than open-source baselines while maintaining competitive intelligibility and naturalness.

Conclusion: The framework successfully disentangles speaker timbre and speaking style control through multi-modal prompting and hierarchical conditioning, advancing controllable TTS capabilities.

Abstract: Controllable text-to-speech (TTS) systems face significant challenges in achieving independent manipulation of speaker timbre and speaking style, often suffering from entanglement between these attributes. We present DMP-TTS, a latent Diffusion Transformer (DiT) framework with explicit disentanglement and multi-modal prompting. A CLAP-based style encoder (Style-CLAP) aligns cues from reference audio and descriptive text in a shared space and is trained with contrastive learning plus multi-task supervision on style attributes. For fine-grained control during inference, we introduce chained classifier-free guidance (cCFG) trained with hierarchical condition dropout, enabling independent adjustment of content, timbre, and style guidance strengths. Additionally, we employ Representation Alignment (REPA) to distill acoustic-semantic features from a pretrained Whisper model into intermediate DiT representations, stabilizing training and accelerating convergence. Experiments show that DMP-TTS delivers stronger style controllability than open-source baselines while maintaining competitive intelligibility and naturalness. Code and demos will be available at https://y61329697.github.io/DMP-TTS/.

Method: Two-stage approach: 1) Weakly supervised audio separation using CLAP for semantic alignment with text labels, plus ranking loss for contextual significance; 2) Image generation via trainable adapter and MLP mapping separated audio to generation conditions.

Result: Outperforms current SOTA methods in 17 out of 21 evaluation indexes across multi-source, mixed-source, and single-source tasks, with superior visual quality. Created first full multi-source audio-to-image benchmark using preprocessed LLP dataset.

Conclusion: MACS successfully addresses the multi-source audio-to-image generation gap by explicitly separating audio components, demonstrating superior performance and establishing a new benchmark for this cross-modal task.

Abstract: Propelled by the breakthrough in deep generative models, audio-to-image generation has emerged as a pivotal cross-modal task that converts complex auditory signals into rich visual representations. However, previous works only focus on single-source audio inputs for image generation, ignoring the multi-source characteristic in natural auditory scenes, thus limiting the performance in generating comprehensive visual content. To bridge this gap, we propose a method called MACS to conduct multi-source audio-to-image generation. To our best knowledge, this is the first work that explicitly separates multi-source audio to capture the rich audio components before image generation. MACS is a two-stage method. In the first stage, multi-source audio inputs are separated by a weakly supervised method, where the audio and text labels are semantically aligned by casting into a common space using the large pre-trained CLAP model. We introduce a ranking loss to consider the contextual significance of the separated audio signals. In the second stage, effective image generation is achieved by mapping the separated audio signals to the generation condition using only a trainable adapter and a MLP layer. We preprocess the LLP dataset as the first full multi-source audio-to-image generation benchmark. The experiments are conducted on multi-source, mixed-source, and single-source audio-to-image generation tasks. The proposed MACS outperforms the current state-of-the-art methods in 17 out of the 21 evaluation indexes on all tasks and delivers superior visual quality.

Method: Develops MambAttention architecture combining Mamba with shared time- and frequency-multi-head attention modules. Uses VB-DemandEx dataset (VoiceBank+Demand inspired with more challenging noise types and lower SNRs) for training.

Result: MambAttention significantly outperforms state-of-the-art LSTM-, xLSTM-, Mamba-, and Conformer-based systems on out-of-domain datasets (DNS 2020 and EARS-WHAM_v2). Matches or outperforms generative diffusion models in generalization while being competitive with language models. Ablation shows importance of weight sharing in attention modules.

Conclusion: MambAttention demonstrates superior cross-corpus generalization for speech enhancement. The hybrid approach combining Mamba with shared time-frequency attention modules effectively addresses overfitting while maintaining strong performance. The architecture remains superior even when similar attention modules are integrated with LSTM and xLSTM.

Abstract: With new sequence models like Mamba and xLSTM, several studies have shown that these models match or outperform the state-of-the-art in single-channel speech enhancement and audio representation learning. However, prior research has demonstrated that sequence models like LSTM and Mamba tend to overfit to the training set. To address this, previous works have shown that adding self-attention to LSTMs substantially improves generalization performance for single-channel speech enhancement. Nevertheless, neither the concept of hybrid Mamba and time-frequency attention models nor their generalization performance have been explored for speech enhancement. In this paper, we propose a novel hybrid architecture, MambAttention, which combines Mamba and shared time- and frequency-multi-head attention modules for generalizable single-channel speech enhancement. To train our model, we introduce VB-DemandEx, a dataset inspired by VoiceBank+Demand but with more challenging noise types and lower signal-to-noise ratios. Trained on VB-DemandEx, MambAttention significantly outperforms existing state-of-the-art discriminative LSTM-, xLSTM-, Mamba-, and Conformer-based systems of similar complexity across all reported metrics on two out-of-domain datasets: DNS 2020 without reverberation and EARS-WHAM_v2. MambAttention also matches or outperforms generative diffusion models in generalization performance while being competitive with language model baselines. Ablation studies highlight the importance of weight sharing between time- and frequency-multi-head attention modules for generalization performance. Finally, we explore integrating the shared time- and frequency-multi-head attention modules with LSTM and xLSTM, which yields a notable performance improvement on the out-of-domain datasets. Yet, MambAttention remains superior for cross-corpus generalization across all reported evaluation metrics.

Method: Introduces two audio tokenizers: music-notation-free prosody tokenizer and content-style tokenizer. Uses auto-regressive content-style modeling for text/prosody/style control and flow-matching acoustic modeling for timbre control. Implements explicit/implicit prosody learning strategies and multi-objective post-training.

Result: Unified modeling brings mutual benefits to both speech and singing voice generation. Demonstrates effectiveness across synthesis, conversion, and editing tasks, showing strong generalization and versatility.

Conclusion: Vevo2 provides a versatile framework for controllable speech and singing voice generation with strong performance across multiple tasks, addressing data scarcity and controllability challenges.

Abstract: Controllable human voice generation, particularly for expressive domains like singing, remains a significant challenge. This paper introduces Vevo2, a unified framework for controllable speech and singing voice generation. To tackle issues like the scarcity of annotated singing data and to enable flexible controllability, Vevo2 introduces two audio tokenizers: (1) a unified music-notation-free prosody tokenizer that captures prosody and melody from speech, singing, and even instrumental sounds, and (2) a unified content-style tokenizer that encodes linguistic content, prosody, and style for both speech and singing, while enabling timbre disentanglement. Vevo2 consists of an auto-regressive (AR) content-style modeling stage, which aims to enable controllability over text, prosody, and style, as well as a flow-matching acoustic modeling stage that allows for timbre control. Particularly, during the speech-singing joint training of the AR model, we propose both explicit and implicit prosody learning strategies to bridge speech and singing voice. Moreover, to further enhance the Vevo2’s ability to follow text and prosody, we design a multi-objective post-training task that integrates both intelligibility and prosody similarity alignment. Experimental results show that the unified modeling in Vevo2 brings mutual benefits to both speech and singing voice generation. Additionally, Vevo2’s effectiveness across a wide range of synthesis, conversion, and editing tasks for both speech and singing further demonstrates its strong generalization ability and versatility. Audio samples are are available at https://versasinger.github.io/.

Method: 1) Develops a systematic evaluation framework using majority-voting ensemble of multiple audio event detectors. 2) Introduces two novel metrics: IH@vid (fraction of videos with hallucinations) and IH@dur (fraction of hallucinated duration). 3) Proposes Posterior Feature Correction (PFC), a training-free inference-time method that operates in two passes: generates initial audio to detect hallucinated segments, then regenerates audio after masking corresponding video features at those timestamps.

Result: Experiments show state-of-the-art V2A models suffer from severe Insertion Hallucination. PFC reduces both prevalence and duration of hallucinations by over 50% on average, without degrading (and sometimes improving) conventional metrics for audio quality and temporal synchronization.

Conclusion: This is the first work to formally define, systematically measure, and effectively mitigate Insertion Hallucination in video-to-audio generation, paving the way for more reliable and faithful V2A models.

Abstract: Video-to-Audio generation has made remarkable strides in automatically synthesizing sound for video. However, existing evaluation metrics, which focus on semantic and temporal alignment, overlook a critical failure mode: models often generate acoustic events, particularly speech and music, that have no corresponding visual source. We term this phenomenon Insertion Hallucination and identify it as a systemic risk driven by dataset biases, such as the prevalence of off-screen sounds, that remains completely undetected by current metrics. To address this challenge, we first develop a systematic evaluation framework that employs a majority-voting ensemble of multiple audio event detectors. We also introduce two novel metrics to quantify the prevalence and severity of this issue: IH@vid (the fraction of videos with hallucinations) and IH@dur (the fraction of hallucinated duration). Building on this, we propose Posterior Feature Correction, a novel training-free inference-time method that mitigates IH. PFC operates in a two-pass process: it first generates an initial audio output to detect hallucinated segments, and then regenerates the audio after masking the corresponding video features at those timestamps. Experiments on several mainstream V2A benchmarks first reveal that state-of-the-art models suffer from severe IH. In contrast, our PFC method reduces both the prevalence and duration of hallucinations by over 50% on average, without degrading, and in some cases even improving, conventional metrics for audio quality and temporal synchronization. Our work is the first to formally define, systematically measure, and effectively mitigate Insertion Hallucination, paving the way for more reliable and faithful V2A models.

Method: Proposed Bayesian extension to ATM that replaces standard Q-learning with Kalman filter-style Bayesian update, maintaining uncertainty-aware Q-value estimates for more stable and sample-efficient learning.

Result: In small tabular environments: Bayesian ATM achieves comparable/improved returns with lower variance and more stable policies. In complex mHealth settings: Both standard and Bayesian ATM perform poorly, suggesting ATM assumptions mismatch real-world challenges.

Conclusion: Uncertainty-aware methods are valuable in low-data settings, but new RL algorithms are needed that explicitly model causal structure, continuous states, and delayed feedback under observation cost constraints for real-world mHealth.

Abstract: Reinforcement learning in mobile health (mHealth) interventions requires balancing intervention efficacy with user burden, particularly when state measurements (for example, user surveys or feedback) are costly yet essential. The Act-Then-Measure (ATM) heuristic addresses this challenge by decoupling control and measurement actions within the Action-Contingent Noiselessly Observable Markov Decision Process (ACNO-MDP) framework. However, the standard ATM algorithm relies on a temporal-difference-inspired Q-learning method, which is prone to instability in sparse and noisy environments. In this work, we propose a Bayesian extension to ATM that replaces standard Q-learning with a Kalman filter-style Bayesian update, maintaining uncertainty-aware estimates of Q-values and enabling more stable and sample-efficient learning. We evaluate our method in both toy environments and clinically motivated testbeds. In small, tabular environments, Bayesian ATM achieves comparable or improved scalarized returns with substantially lower variance and more stable policy behavior. In contrast, in larger and more complex mHealth settings, both the standard and Bayesian ATM variants perform poorly, suggesting a mismatch between ATM’s modeling assumptions and the structural challenges of real-world mHealth domains. These findings highlight the value of uncertainty-aware methods in low-data settings while underscoring the need for new RL algorithms that explicitly model causal structure, continuous states, and delayed feedback under observation cost constraints.

Method: Created agent-centered simulator with 276 Unreal Engine MetaHuman patients from interview data, with synchronized speech, gaze/face, and poses. Uses perception-fusion-policy loop for conversational decisions with safety shield. Training employs counterfactual replay with bounded nonverbal perturbations and uncertainty-aware turn manager to reduce diagnostic ambiguity.

Result: Custom TD3 controller outperformed PPO and CEM, achieving near-ceiling coverage with steadier pace at comparable rewards. Shows negligible turn overlap, aligned cut timing, fewer clarification prompts, and shorter waits. Performance remains stable under modality dropout and renderer swap, with rankings holding on held-out patient split.

Conclusion: Agent-centered simulation enables efficient training of humanoid conversational skills without hardware burden. TD3 excels at social timing and completeness, providing robust foundation for clinician-supervised humanoid pilots in mental health screening.

Abstract: Testing humanoid robots with users is slow, causes wear, and limits iteration and diversity. Yet screening agents must master conversational timing, prosody, backchannels, and what to attend to in faces and speech for Depression and PTSD. Most simulators omit policy learning with nonverbal dynamics; many controllers chase task accuracy while underweighting trust, pacing, and rapport. We virtualise the humanoid as a conversational agent to train without hardware burden. Our agent-centred, simulation-first pipeline turns interview data into 276 Unreal Engine MetaHuman patients with synchronised speech, gaze/face, and head-torso poses, plus PHQ-8 and PCL-C flows. A perception-fusion-policy loop decides what and when to speak, when to backchannel, and how to avoid interruptions, under a safety shield. Training uses counterfactual replay (bounded nonverbal perturbations) and an uncertainty-aware turn manager that probes to reduce diagnostic ambiguity. Results are simulation-only; the humanoid is the transfer target. In comparing three controllers, a custom TD3 (Twin Delayed DDPG) outperformed PPO and CEM, achieving near-ceiling coverage with steadier pace at comparable rewards. Decision-quality analyses show negligible turn overlap, aligned cut timing, fewer clarification prompts, and shorter waits. Performance stays stable under modality dropout and a renderer swap, and rankings hold on a held-out patient split. Contributions: (1) an agent-centred simulator that turns interviews into 276 interactive patients with bounded nonverbal counterfactuals; (2) a safe learning loop that treats timing and rapport as first-class control variables; (3) a comparative study (TD3 vs PPO/CEM) with clear gains in completeness and social timing; and (4) ablations and robustness analyses explaining the gains and enabling clinician-supervised humanoid pilots.

Method: Evaluated language/general time-series/ECG foundation models compared with time-series deep learning models. Created a benchmark with comprehensive experimental setup to assess foundation models’ effectiveness in ECG analysis.

Result: General time-series/ECG foundation models achieved a top performance rate of 80%, demonstrating their effectiveness in ECG analysis. The study provides in-depth analyses and insights alongside comprehensive experimental results.

Conclusion: Foundation models show potential for advancing physiological waveform analysis, but limitations exist. The study highlights both the promise and constraints of foundation models in healthcare applications, with publicly available benchmark data and code.

Abstract: In the process of patient diagnosis, non-invasive measurements are widely used due to their low risks and quick results. Electrocardiogram (ECG), as a non-invasive method to collect heart activities, is used to diagnose cardiac conditions. Analyzing the ECG typically requires domain expertise, which is a roadblock to applying artificial intelligence (AI) for healthcare. Through advances in self-supervised learning and foundation models, AI systems can now acquire and leverage domain knowledge without relying solely on human expertise. However, there is a lack of comprehensive analyses over the foundation models’ performance on ECG. This study aims to answer the research question: “Are Foundation Models Useful for ECG Analysis?” To address it, we evaluate language/general time-series/ECG foundation models in comparison with time-series deep learning models. The experimental results show that general time-series/ECG foundation models achieve a top performance rate of 80%, indicating their effectiveness in ECG analysis. In-depth analyses and insights are provided along with comprehensive experimental results. This study highlights the limitations and potential of foundation models in advancing physiological waveform analysis. The data and code for this benchmark are publicly available at https://github.com/yuhaoxu99/ECGMultitasks-Benchmark.

Method: Introduces TinyDéjàVu framework with novel algorithms to reduce RAM footprint by optimizing data flows and eliminating redundant computations on overlapping sliding window inputs for sensor time-series data.

Result: TinyDéjàVu saves more than 60% of RAM usage and eliminates up to 90% of redundant compute on overlapping sliding window inputs, as demonstrated through reproducible benchmarks on hardware.

Conclusion: The framework provides significant resource savings for tiny ML models on constrained microcontroller hardware, with open-source implementation available for practical deployment.

Abstract: Always-on sensors are increasingly expected to embark a variety of tiny neural networks and to continuously perform inference on time-series of the data they sense. In order to fit lifetime and energy consumption requirements when operating on battery, such hardware uses microcontrollers (MCUs) with tiny memory budget e.g., 128kB of RAM. In this context, optimizing data flows across neural network layers becomes crucial. In this paper, we introduce TinyDéjàVu, a new framework and novel algorithms we designed to drastically reduce the RAM footprint required by inference using various tiny ML models for sensor data time-series on typical microcontroller hardware. We publish the implementation of TinyDéjàVu as open source, and we perform reproducible benchmarks on hardware. We show that TinyDéjàVu can save more than 60% of RAM usage and eliminate up to 90% of redundant compute on overlapping sliding window inputs.

Method: Proposes LLM4XCE framework with: 1) Embedding module + Parallel Feature-Spatial Attention for deep fusion of pilot features and spatial structures, 2) Fine-tuning only top two Transformer layers of LLM to capture latent dependencies efficiently.

Result: Extensive simulations show LLM4XCE significantly outperforms state-of-the-art methods under hybrid-field conditions, achieving superior estimation accuracy and generalization performance.

Conclusion: LLMs can effectively address XL-MIMO channel estimation challenges by leveraging semantic modeling capabilities, offering a promising direction for 6G networks with hybrid-field environments.

Abstract: Extremely large-scale massive multiple-input multiple-output (XL-MIMO) is a key enabler for sixth-generation (6G) networks, offering massive spatial degrees of freedom. Despite these advantages, the coexistence of near-field and far-field effects in hybrid-field channels presents significant challenges for accurate estimation, where traditional methods often struggle to generalize effectively. In recent years, large language models (LLMs) have achieved impressive performance on downstream tasks via fine-tuning, aligning with the semantic communication shift toward task-oriented understanding over bit-level accuracy. Motivated by this, we propose Large Language Models for XL-MIMO Channel Estimation (LLM4XCE), a novel channel estimation framework that leverages the semantic modeling capabilities of large language models to recover essential spatial-channel representations for downstream tasks. The model integrates a carefully designed embedding module with Parallel Feature-Spatial Attention, enabling deep fusion of pilot features and spatial structures to construct a semantically rich representation for LLM input. By fine-tuning only the top two Transformer layers, our method effectively captures latent dependencies in the pilot data while ensuring high training efficiency. Extensive simulations demonstrate that LLM4XCE significantly outperforms existing state-of-the-art methods under hybrid-field conditions, achieving superior estimation accuracy and generalization performance.

Method: DW-KNN (Double Weighted KNN) integrates exponential distance weighting with neighbor validity assessment to suppress noisy/mislabeled samples and reduce hyperparameter sensitivity.

Result: Achieves 0.8988 average accuracy, ranks 2nd among six methods (within 0.2% of best Ensemble KNN), has lowest cross-validation variance (0.0156), and shows statistically significant improvements over existing methods.

Conclusion: DW-KNN provides a simple yet effective alternative to complex adaptive schemes, particularly valuable for high-stakes applications requiring explainable predictions.

Abstract: K-Nearest Neighbors (KNN) is one of the most used ML classifiers. However, if we observe closely, standard distance-weighted KNN and relative variants assume all ‘k’ neighbors are equally reliable. In heterogeneous feature space, this becomes a limitation that hinders reliability in predicting true levels of the observation. We propose DW-KNN (Double Weighted KNN), a transparent and robust variant that integrates exponential distance with neighbor validity. This enables instance-level interpretability, suppresses noisy or mislabeled samples, and reduces hyperparameter sensitivity. Comprehensive evaluation on 9 data-sets helps to demonstrate that DW-KNN achieves 0.8988 accuracy on average. It ranks 2nd among six methods and within 0.2% of the best-performing Ensemble KNN. It also exhibits the lowest cross-validation variance (0.0156), indicating reliable prediction stability. Statistical significance test confirmed ($p < 0.001$) improvement over compactness weighted KNN (+4.09%) and Kernel weighted KNN (+1.13%). The method provides a simple yet effective alternative to complex adaptive schemes, particularly valuable for high-stakes applications requiring explainable predictions.

Method: LUMOS uses a transformer-based architecture with a novel cross-attention mechanism that conditions predictions on future known events (holidays, sales, etc.), and employs multi-modal tokenization combining user transactions, event context, and static user demographic attributes through specialized embedding pathways.

Result: On a production dataset spanning 275 billion user activity tokens from 250 million users, LUMOS achieves average improvements of 0.025 in ROC-AUC for binary classification tasks and 4.6% reduction in MAPE for regression tasks across 5 tasks, with online A/B testing showing a 3.15% increase in Daily Active Users.

Conclusion: LUMOS provides a scalable solution for user behavior prediction that outperforms traditional task-specific models while eliminating manual feature engineering, demonstrating measurable business impact through improved prediction accuracy and user engagement metrics.

Abstract: User behavior prediction at scale remains a critical challenge for online B2C platforms. Traditional approaches rely heavily on task-specific models and domain-specific feature engineering. This is time-consuming, computationally expensive, and requires domain expertise and therefore not scalable. We present LUMOS (Large User MOdel Series), a transformer-based architecture that eliminates task-specific models and manual feature engineering by learning multiple tasks jointly using only raw user activity data. LUMOS introduces a novel cross-attention mechanism that conditions predictions on future known events (e.g., holidays, sales, etc.), enabling the model to predict complex behaviour patterns like “how will upcoming holidays affect user engagement?” The architecture also employs multi-modal tokenization, combining user transactions, event context, and static user demographic attributes into rich representations processed through specialized embedding pathways. Through extensive experiments on a production dataset spanning 275 billion user activity tokens from 250 million users, we demonstrate that LUMOS achieves superior performance compared to traditional task-specific models. Across 5 tasks with established baselines, we achieve an average improvement of 0.025 in ROC-AUC for binary classification tasks and 4.6% reduction in MAPE for regression tasks. Online A/B testing validates these improvements translate to measurable business impact with a 3.15% increase in Daily Active Users.

Method: Combines strengths of both physics-guided approaches by embedding fundamental solution of wave equation into network architecture. Uses point neuron learning method that can model arbitrary sound fields from microphone observations without training datasets, directly processes complex numbers.

Result: Outperforms two competing methods in sound field reconstruction in reverberant environments. Efficiently handles noisy environments with sparse microphone observations.

Conclusion: Proposed PINN architecture offers better interpretability and generalizability than existing methods, successfully addresses sound field reconstruction problem while strictly satisfying wave equation physics.

Abstract: Machine learning and neural networks have advanced numerous research domains, but challenges such as large training data requirements and inconsistent model performance hinder their application in certain scientific problems. To overcome these challenges, researchers have investigated integrating physics principles into machine learning models, mainly through: (i) physics-guided loss functions, generally termed as physics-informed neural networks, and (ii) physics-guided architectural design. While both approaches have demonstrated success across multiple scientific disciplines, they have limitations including being trapped to a local minimum, poor interpretability, and restricted generalizability. This paper proposes a new physics-informed neural network (PINN) architecture that combines the strengths of both approaches by embedding the fundamental solution of the wave equation into the network architecture, enabling the learned model to strictly satisfy the wave equation. The proposed point neuron learning method can model an arbitrary sound field based on microphone observations without any dataset. Compared to other PINN methods, our approach directly processes complex numbers and offers better interpretability and generalizability. We evaluate the versatility of the proposed architecture by a sound field reconstruction problem in a reverberant environment. Results indicate that the point neuron method outperforms two competing methods and can efficiently handle noisy environments with sparse microphone observations.

Method: Created a unified benchmarking framework with minimal preprocessing, standardized evaluation protocols across 11 diagnostic tasks using 14 publicly available EEG datasets covering various neurological conditions (epilepsy, schizophrenia, Parkinson’s, OCD, mTBI).

Result: Foundation models achieve strong performance in certain settings, but simpler models often remain competitive, particularly under clinical distribution shifts where traditional methods show robustness.

Conclusion: The framework enables reproducible side-by-side comparisons of foundation models and classical baselines, with released data and code to facilitate adoption and further research in clinical EEG analysis.

Abstract: We introduce a unified benchmarking framework focused on evaluating EEG-based foundation models in clinical applications. The benchmark spans 11 well-defined diagnostic tasks across 14 publicly available EEG datasets, including epilepsy, schizophrenia, Parkinson’s disease, OCD, and mild traumatic brain injury. It features minimal preprocessing, standardized evaluation protocols, and enables side-by-side comparisons of classical baselines and modern foundation models. Our results show that while foundation models achieve strong performance in certain settings, simpler models often remain competitive, particularly under clinical distribution shifts. To facilitate reproducibility and adoption, we release all prepared data and code in an accessible and extensible format.

Method: PS-LoRA uses a dual-regularization objective that penalizes conflicting gradient directions and constrains magnitude deviations. It also implements magnitude-based merging to consolidate sequential adapters without retraining.

Result: Experiments on NLP and Vision benchmarks show PS-LoRA outperforms state-of-the-art methods by preserving learned representations while efficiently adapting to new domains.

Conclusion: PS-LoRA effectively resolves gradient conflicts in continual learning, maintaining parameter stability and preventing catastrophic forgetting through aligned optimization subspace updates.

Abstract: Low-Rank Adaptation (LoRA) enables efficient Continual Learning but often suffers from catastrophic forgetting due to destructive interference between tasks. Our analysis reveals that this degradation is primarily driven by antagonistic directional updates where new task gradients directly oppose the historical weight trajectory. To address this, we propose PS-LoRA (Parameter Stability LoRA), a framework designed to resolve conflicts by aligning updates within the optimization subspace. Our approach employs a dual-regularization objective that penalizes conflicting directions and constrains magnitude deviations to ensure consistency with prior knowledge. Additionally, we implement a magnitude-based merging strategy to consolidate sequential adapters into a robust representation without retraining. Experiments on NLP and Vision benchmarks show that PS-LoRA outperforms state-of-the-art methods by preserving the stability of learned representations while efficiently adapting to new domains.

Method: Extends learning theory to double fibration transforms (geometric integral operators including generalized Radon and geodesic ray transforms). Proves these operators avoid the curse of dimensionality. Investigates architectures that explicitly encode the geometry of these transforms, using cross-attention based on levelset methods.

Result: Proves double fibration transforms don’t suffer from curse of dimensionality - error decays superalgebraically (faster than any fixed power of reciprocal training samples). Shows architecture reminiscent of cross-attention based on levelset methods yields universal, stable parameterization that learns these transforms from very few training examples.

Conclusion: The work expands operator learning theory to include geometric integral operators that propagate singularities, providing theoretical foundation for data-efficient learning of such operators in scientific ML applications like wave propagation and fluid dynamics.

Abstract: Deep learning methods have proven capable of recovering operators between high-dimensional spaces, such as solution maps of PDEs and similar objects in mathematical physics, from very few training samples. This phenomenon of data-efficiency has been proven for certain classes of elliptic operators with simple geometry, i.e., operators that do not change the domain of the function or propagate singularities. However, scientific machine learning is commonly used for problems that do involve the propagation of singularities in a priori unknown ways, such as waves, advection, and fluid dynamics. In light of this, we expand the learning theory to include double fibration transforms–geometric integral operators that include generalized Radon and geodesic ray transforms. We prove that this class of operators does not suffer from the curse of dimensionality: the error decays superalgebraically, that is, faster than any fixed power of the reciprocal of the number of training samples. Furthermore, we investigate architectures that explicitly encode the geometry of these transforms, demonstrating that an architecture reminiscent of cross-attention based on levelset methods yields a parameterization that is universal, stable, and learns double fibration transforms from very few training examples. Our results contribute to a rapidly-growing line of theoretical work on learning operators for scientific machine learning.

Method: Compute spectral embeddings to capture fragile directions of the input manifold, assign robustness scores to edges/non-edges. Two attack variants: 1) Spade-guided deletion attack removes most spectrally robust edges, 2) Spade-guided addition attack inserts edges between maximally incompatible nodes in fragile spectral space. Attacks are model-aware, gradient-free, and operate at graph level.

Result: Proposes SEA as a new family of adversarial attacks that leverage spectral robustness evaluation, with two complementary attack variants that can be plugged into existing GNN architectures without requiring gradients.

Conclusion: SEA provides a principled spectral approach to adversarial attacks on GNNs that explicitly targets structural vulnerabilities through spectral analysis, offering a gradient-free alternative to existing methods.

Abstract: Graph Neural Networks (GNNs) achieve strong performance on graph-structured data, but are notoriously vulnerable to small, carefully crafted perturbations of the graph structure. Most existing structure-based attacks rely on gradient-based heuristics or local connectivity patterns, and treat edges as equally important candidates for manipulation. In this paper, we propose Spectral Edge Attacks (SEA), a new family of adversarial attacks that explicitly leverage spectral robustness evaluation to guide structural perturbations. Our key idea is to compute a spectral embedding that captures the most fragile directions of the input manifold and to use it to assign a robustness score to each edge or non-edge. Based on these scores, we introduce two complementary attack variants: (i) a Spade-guided deletion attack that removes the most spectrally robust edges, and (ii) a Spade-guided addition attack that inserts edges between nodes that are maximally incompatible in the fragile spectral space. Both attacks operate at the graph level, are model-aware but conceptually simple, and can be plugged into existing GNN architectures without requiring gradients. We describe the spectral formulation, the attack algorithms, and experiments on benchmarks.

Method: Created FIFE benchmark with 88 human-authored prompts and verification system using chainable, verifiable constraints for fine-grained reward signals. Evaluated 53 models (proprietary, open-weight, open-source) in zero-shot setting.

Result: Clear performance hierarchy: top open-weight model (76.1 strict/79.5 loose) surpasses leading proprietary system (65.9/70.5), while best open-source models lag significantly (45.5/48.9). All models struggle with complex requirements and fail to achieve perfect compliance.

Conclusion: Even top-performing models have limitations in handling complex financial instructions, highlighting the need for further research. The authors release dataset and code as open-source resource to promote Reinforcement Learning research in finance.

Abstract: Language Models (LMs) struggle with complex, interdependent instructions, particularly in high-stakes domains like finance where precision is critical. We introduce FIFE, a novel, high-difficulty benchmark designed to assess LM instruction-following capabilities for financial analysis tasks. FIFE comprises 88 human-authored prompts and employs a verification system with chainable, verifiable constraints for fine-grained reward signals. We evaluate 53 models (proprietary, open-weight, open-source) in a zero-shot setting. Our key findings reveal a clear performance hierarchy: the top open-weight model (76.1 strict / 79.5 loose) surpasses the leading proprietary system (65.9 strict / 70.5 loose), while the best open-source models lag significantly (45.5 strict / 48.9 loose). However, even top-performing models struggle with FIFE’s complex requirements, failing to achieve perfect compliance. We release our dataset and code as an open-source resource to promote research in Reinforcement Learning for the financial domain.

Method: Proposes CluCERT with: 1) Semantic clustering filter to reduce noisy samples and retain meaningful perturbations, 2) Refine module to extract core semantics, and 3) Fast synonym substitution strategy to accelerate denoising. Uses clustering-guided denoising smoothing for certification.

Result: Outperforms existing certified approaches in both robustness bounds and computational efficiency across various downstream tasks and jailbreak defense scenarios.

Conclusion: CluCERT provides a more effective and efficient framework for certifying LLM robustness against adversarial prompts, addressing key limitations of previous methods through semantic clustering and optimization techniques.

Abstract: Recent advancements in Large Language Models (LLMs) have led to their widespread adoption in daily applications. Despite their impressive capabilities, they remain vulnerable to adversarial attacks, as even minor meaning-preserving changes such as synonym substitutions can lead to incorrect predictions. As a result, certifying the robustness of LLMs against such adversarial prompts is of vital importance. Existing approaches focused on word deletion or simple denoising strategies to achieve robustness certification. However, these methods face two critical limitations: (1) they yield loose robustness bounds due to the lack of semantic validation for perturbed outputs and (2) they suffer from high computational costs due to repeated sampling. To address these limitations, we propose CluCERT, a novel framework for certifying LLM robustness via clustering-guided denoising smoothing. Specifically, to achieve tighter certified bounds, we introduce a semantic clustering filter that reduces noisy samples and retains meaningful perturbations, supported by theoretical analysis. Furthermore, we enhance computational efficiency through two mechanisms: a refine module that extracts core semantics, and a fast synonym substitution strategy that accelerates the denoising process. Finally, we conduct extensive experiments on various downstream tasks and jailbreak defense scenarios. Experimental results demonstrate that our method outperforms existing certified approaches in both robustness bounds and computational efficiency.

Method: StructuredDNA replaces dense Mixture-of-Experts routing with a bio-physical, energy-guided routing layer based on semantic energy minimization. Inputs are dynamically grouped into semantic codons, and routing selects a single expert by minimizing a global energy functional that combines cohesion, uncertainty, and computational cost.

Result: On BioASQ (K=50), achieved 97.7% reduction in Energy Utilization Density (EUD) and Semantic Stability Index (SSI) of 0.998. Demonstrated Semantic Scaling Law on WikiText-103, scaling to K=2048 experts while maintaining >99% energy efficiency.

Conclusion: StructuredDNA establishes a robust, domain-agnostic paradigm for future sparse computational frameworks, providing an explicit link between bio-physical principles and sparse expert routing in Transformers, pointing toward energy-aware, modular, and scalable computational systems.

Abstract: The rapid scaling of large computational models has led to a critical increase in energy and compute costs. Inspired by biological systems where structure and function emerge from low-energy configurations, we introduce StructuredDNA, a sparse architecture framework for modular, energy-aware Transformer routing. StructuredDNA replaces dense Mixture-of-Experts routing with a bio-physical, energy-guided routing layer based on semantic energy minimization. Inputs are dynamically grouped into semantic codons, and routing selects a single expert by minimizing a global energy functional that combines cohesion, uncertainty, and computational cost. We validate StructuredDNA on both specialized (BioASQ) and open-domain benchmarks (WikiText-103). On BioASQ (K = 50), we achieve a 97.7% reduction in Energy Utilization Density (EUD) and a Semantic Stability Index (SSI) of 0.998. We further demonstrate a Semantic Scaling Law on WikiText-103, showing that the architecture generalizes to open domains by scaling expert granularity (K = 2048) while maintaining more than 99% energy efficiency. StructuredDNA thus establishes a robust, domain-agnostic paradigm for future sparse computational frameworks. StructuredDNA provides an explicit link between bio-physical principles and sparse expert routing in Transformer architectures, and points toward future energy-aware, modular, and scalable computational systems. We discuss limitations of this proof-of-concept study and outline directions for scaling the approach to larger models, datasets, and hardware platforms. The StructuredDNA implementation is available at https://github.com/InnoDeep-repos/StructuredDNA .

Method: UCF integrates uncertainty-aware contrastive loss (dynamically weights samples by confidence), adaptive temperature scaling (adjusts to batch variability), self-attention-guided LSTM encoder, and uses positive anchors to stabilize training in PU learning setting.

Result: Achieves >93.38% accuracy, >0.93 precision, near-perfect recall with minimal false negatives, competitive ROC-AUC scores for malicious content classification. Visual analyses show clear separation between positive and unlabeled instances.

Conclusion: UCF is a robust, scalable solution for PU learning that produces calibrated, discriminative embeddings suitable for high-stakes applications where data is noisy and imbalanced.

Abstract: We propose the Uncertainty Contrastive Framework (UCF), a Positive-Unlabeled (PU) representation learning framework that integrates uncertainty-aware contrastive loss, adaptive temperature scaling, and a self-attention-guided LSTM encoder to improve classification under noisy and imbalanced conditions. UCF dynamically adjusts contrastive weighting based on sample confidence, stabilizes training using positive anchors, and adapts temperature parameters to batch-level variability. Applied to malicious content classification, UCF-generated embeddings enable multiple traditional classifiers to achieve more than 93.38% accuracy, precision above 0.93, and near-perfect recall, with minimal false negatives and competitive ROC-AUC scores. Visual analyses confirm clear separation between positive and unlabeled instances, highlighting the framework’s ability to produce calibrated, discriminative embeddings. These results position UCF as a robust and scalable solution for PU learning in high-stakes domains such as cybersecurity and biomedical text mining.

Method: Contrastive Region Masking (CRM) systematically masks annotated visual regions and contrasts the resulting reasoning traces with unmasked baselines to provide causal attribution at each CoT step.

Result: Applied to VisArgs datasets, CRM reveals distinct failure modes: some models hallucinate when evidence is missing while preserving reasoning structure, while others ground tightly to visual cues but collapse under perturbations.

Conclusion: CRM reframes visual benchmarks as diagnostic tools, highlighting the need for multimodal evaluation frameworks that measure not just performance, but also robustness and fidelity of reasoning.

Abstract: We introduce Contrastive Region Masking (CRM), a training free diagnostic that reveals how multimodal large language models (MLLMs) depend on specific visual regions at each step of chain-of-thought (CoT) reasoning. Unlike prior approaches limited to final answers or attention maps, CRM provides causal, step-level attri- bution by systematically masking annotated regions and contrasting the resulting reasoning traces with unmasked baselines. Applied to datasets such as VisArgs, CRM reveals distinct failure modes: some models preserve reasoning structure, but hallucinate when evidence is missing, while others ground tightly to visual cues yet collapse under perturbations. By shifting the evaluation from correctness of an- swers to faithfulness of reasoning, CRM reframes visual benchmarks as diagnostic tools, highlighting the need for multimodal evaluation frameworks that measure not just performance, but also robustness and fidelity of reasoning.

Method: A graph neural network surrogate model trained on comprehensive high-fidelity CFD simulations of patient-specific aneurysms. The architecture combines graph transformers with autoregressive predictions to simulate blood flow, wall shear stress, and oscillatory shear index directly from vascular geometries.

Result: The model reproduces full-field hemodynamics in less than one minute per cardiac cycle, generalizes across unseen patient geometries and inflow conditions without mesh-specific calibration, and enables near real-time inference integrated with existing imaging pipelines.

Conclusion: This work transforms high-fidelity simulations from an expert-only research tool into a deployable, data-driven decision support system, delivering high-resolution hemodynamic predictions within minutes of patient imaging without requiring computational specialists.

Abstract: Intracranial aneurysms remain a major cause of neurological morbidity and mortality worldwide, where rupture risk is tightly coupled to local hemodynamics particularly wall shear stress and oscillatory shear index. Conventional computational fluid dynamics simulations provide accurate insights but are prohibitively slow and require specialized expertise. Clinical imaging alternatives such as 4D Flow MRI offer direct in-vivo measurements, yet their spatial resolution remains insufficient to capture the fine-scale shear patterns that drive endothelial remodeling and rupture risk while being extremely impractical and expensive. We present a graph neural network surrogate model that bridges this gap by reproducing full-field hemodynamics directly from vascular geometries in less than one minute per cardiac cycle. Trained on a comprehensive dataset of high-fidelity simulations of patient-specific aneurysms, our architecture combines graph transformers with autoregressive predictions to accurately simulate blood flow, wall shear stress, and oscillatory shear index. The model generalizes across unseen patient geometries and inflow conditions without mesh-specific calibration. Beyond accelerating simulation, our framework establishes the foundation for clinically interpretable hemodynamic prediction. By enabling near real-time inference integrated with existing imaging pipelines, it allows direct comparison with hospital phase-diagram assessments and extends them with physically grounded, high-resolution flow fields. This work transforms high-fidelity simulations from an expert-only research tool into a deployable, data-driven decision support system. Our full pipeline delivers high-resolution hemodynamic predictions within minutes of patient imaging, without requiring computational specialists, marking a step-change toward real-time, bedside aneurysm analysis.

Method: Proposes two novel isotonic normalization-aware techniques: NA-FIR (incorporates normalization directly into optimization) and SCIR (models the problem as cumulative bivariate isotonic regression). Both methods are grounded in natural assumptions expected by practitioners and inherently account for probability normalization constraints.

Result: Empirical evaluation on various text and image classification datasets across different model architectures shows consistent improvements in negative log-likelihood (NLL) and expected calibration error (ECE) metrics compared to previous approaches.

Conclusion: The proposed isotonic normalization-aware techniques effectively address the limitations of previous multi-class isotonic calibration methods, providing better calibration performance while maintaining the non-parametric advantages of isotonic regression.

Abstract: Accurate and reliable probability predictions are essential for multi-class supervised learning tasks, where well-calibrated models enable rational decision-making. While isotonic regression has proven effective for binary calibration, its extension to multi-class problems via one-vs-rest calibration produced suboptimal results when compared to parametric methods, limiting its practical adoption. In this work, we propose novel isotonic normalization-aware techniques for multiclass calibration, grounded in natural and intuitive assumptions expected by practitioners. Unlike prior approaches, our methods inherently account for probability normalization by either incorporating normalization directly into the optimization process (NA-FIR) or modeling the problem as a cumulative bivariate isotonic regression (SCIR). Empirical evaluation on a variety of text and image classification datasets across different model architectures reveals that our approach consistently improves negative log-likelihood (NLL) and expected calibration error (ECE) metrics.

Method: The study introduces a diffusion-based deep learning framework that systematically compares three residual prediction strategies: (1) fully data-driven using only MRMS observations, (2) corrective model using only HRRR forecasts, and (3) hybrid model integrating both MRMS and selected HRRR variables. Forecasts are produced at 1-km resolution with 1-hour to 12-hour autoregressive rollouts.

Result: The DL framework consistently outperforms HRRR baseline across all lead times. Hybrid model performs best at shortest lead times, while HRRR-corrective model outperforms others at longer lead times (up to 12 hours). The study includes calibrated uncertainty quantification for reliability assessment.

Conclusion: The work advances DL-based precipitation forecasting by enhancing predictive skill, reliability, and regional applicability. Gains at longer lead times are particularly valuable for emergency preparedness, where modest increases in forecast horizon can significantly improve decision-making.

Abstract: Accurate precipitation forecasting is essential for hydrometeorological risk management, especially for anticipating extreme rainfall that can lead to flash flooding and infrastructure damage. This study introduces a diffusion-based deep learning (DL) framework that systematically compares three residual prediction strategies differing only in their input sources: (1) a fully data-driven model using only past observations from the Multi-Radar Multi-Sensor (MRMS) system, (2) a corrective model using only forecasts from the High-Resolution Rapid Refresh (HRRR) numerical weather prediction system, and (3) a hybrid model integrating both MRMS and selected HRRR forecast variables. By evaluating these approaches under a unified setup, we provide a clearer understanding of how each data source contributes to predictive skill over the Continental United States (CONUS). Forecasts are produced at 1-km spatial resolution, beginning with direct 1-hour predictions and extending to 12 hours using autoregressive rollouts. Performance is evaluated using both CONUS-wide and region-specific metrics that assess overall performance and skill at extreme rainfall thresholds. Across all lead times, our DL framework consistently outperforms the HRRR baseline in pixel-wise and spatiostatistical metrics. The hybrid model performs best at the shortest lead time, while the HRRR-corrective model outperforms others at longer lead times, maintaining high skill through 12 hours. To assess reliability, we incorporate calibrated uncertainty quantification tailored to the residual learning setup. These gains, particularly at longer lead times, are critical for emergency preparedness, where modest increases in forecast horizon can improve decision-making. This work advances DL-based precipitation forecasting by enhancing predictive skill, reliability, and applicability across regions.

Method: Used simulation-based data curation with random structure sampling and multislice simulation to generate synthetic HRTEM data. Trained over 12,000 neural network segmentation models and developed a framework using HRTEM contrast transfer function to compare dataset information content and quantify OOD domain shifts.

Result: Neural network segmentation models show significant performance stability but degrade smoothly and predictably as imaging conditions shift from training distribution. The framework successfully quantifies OOD domain shifts for imaging conditions.

Conclusion: While neural networks exhibit predictable OOD degradation for imaging condition shifts, the approach has limitations for explaining other OOD shifts like atomic structure variations. Complementary techniques are needed for comprehensive understanding of generalization in such settings.

Abstract: Neural networks, while effective for tackling many challenging scientific tasks, are not known to perform well out-of-distribution (OOD), i.e., within domains which differ from their training data. Understanding neural network OOD generalization is paramount to their successful deployment in experimental workflows, especially when ground-truth knowledge about the experiment is hard to establish or experimental conditions significantly vary. With inherent access to ground-truth information and fine-grained control of underlying distributions, simulation-based data curation facilitates precise investigation of OOD generalization behavior. Here, we probe generalization with respect to imaging conditions of neural network segmentation models for high-resolution transmission electron microscopy (HRTEM) imaging of nanoparticles, training and measuring the OOD generalization of over 12,000 neural networks using synthetic data generated via random structure sampling and multislice simulation. Using the HRTEM contrast transfer function, we further develop a framework to compare information content of HRTEM datasets and quantify OOD domain shifts. We demonstrate that neural network segmentation models enjoy significant performance stability, but will smoothly and predictably worsen as imaging conditions shift from the training distribution. Lastly, we consider limitations of our approach in explaining other OOD shifts, such as of the atomic structures, and discuss complementary techniques for understanding generalization in such settings.

Method: DeepTherm uses a modular deep learning approach with dual-prediction pipeline that disentangles baseline mortality (without heatwaves/irregular events) from all-cause mortality, enabling prediction without heat-related mortality history.

Result: Evaluation on real-world Spanish data shows consistent, robust, and accurate performance across diverse regions, time periods, and population groups, with ability to trade off between missed alarms and false alarms.

Conclusion: DeepTherm provides a flexible early warning system for deadly heatwave prediction that addresses key challenges in data requirements and operational robustness for public health protection.

Abstract: Severe heatwaves in urban areas significantly threaten public health, calling for establishing early warning strategies. Despite predicting occurrence of heatwaves and attributing historical mortality, predicting an incoming deadly heatwave remains a challenge due to the difficulty in defining and estimating heat-related mortality. Furthermore, establishing an early warning system imposes additional requirements, including data availability, spatial and temporal robustness, and decision costs. To address these challenges, we propose DeepTherm, a modular early warning system for deadly heatwave prediction without requiring heat-related mortality history. By highlighting the flexibility of deep learning, DeepTherm employs a dual-prediction pipeline, disentangling baseline mortality in the absence of heatwaves and other irregular events from all-cause mortality. We evaluated DeepTherm on real-world data across Spain. Results demonstrate consistent, robust, and accurate performance across diverse regions, time periods, and population groups while allowing trade-off between missed alarms and false alarms.

Method: Used Facebook Prophet and NeuralProphet additive models with systematic feature selection (correlation, mutual information, mRMR), leakage-safe scaling, and chronological data splits. Compared against LSTM, LightGBM, and SARIMAX baselines using multi-year pollutant and meteorological data from Beijing for PM2.5 and PM10 forecasting.

Result: Facebook Prophet consistently outperformed all other models including NeuralProphet, SARIMAX, LSTM, and LightGBM, achieving test R² above 0.94 for both PM2.5 and PM10 pollutants on a 7-day holdout evaluation.

Conclusion: Interpretable additive models like Facebook Prophet offer a practical balance of accuracy, transparency, and ease of deployment, remaining competitive with both traditional statistical methods and complex deep learning approaches for urban air pollution forecasting.

Abstract: Accurate forecasting of urban air pollution is essential for protecting public health and guiding mitigation policies. While Deep Learning (DL) and hybrid pipelines dominate recent research, their complexity and limited interpretability hinder operational use. This study investigates whether lightweight additive models – Facebook Prophet (FBP) and NeuralProphet (NP) – can deliver competitive forecasts for particulate matter (PM${2.5}$, PM${10}$) in Beijing, China. Using multi-year pollutant and meteorological data, we applied systematic feature selection (correlation, mutual information, mRMR), leakage-safe scaling, and chronological data splits. Both models were trained with pollutant and precursor regressors, with NP additionally leveraging lagged dependencies. For context, two machine learning baselines (LSTM, LightGBM) and one traditional statistical model (SARIMAX) were also implemented. Performance was evaluated on a 7-day holdout using MAE, RMSE, and $R^2$. Results show that FBP consistently outperformed NP, SARIMAX, and the learning-based baselines, achieving test $R^2$ above 0.94 for both pollutants. These findings demonstrate that interpretable additive models remain competitive with both traditional and complex approaches, offering a practical balance of accuracy, transparency, and ease of deployment.

Method: GS-KAN constructs unique edge functions by applying learnable linear transformations to a single learnable, shared parent function per layer, inspired by David Sprecher’s refinement of the superposition theorem.

Result: GS-KAN outperforms MLPs and standard KANs on continuous function approximation, achieves competitive performance on tabular regression, and outperforms MLPs on high-dimensional classification while maintaining superior parameter efficiency.

Conclusion: GS-KAN enables deployment of KAN-based architectures in high-dimensional regimes under strict parameter constraints, addressing the parameter explosion problem of standard implementations.

Abstract: The Kolmogorov-Arnold representation theorem offers a theoretical alternative to Multi-Layer Perceptrons (MLPs) by placing learnable univariate functions on edges rather than nodes. While recent implementations such as Kolmogorov-Arnold Networks (KANs) demonstrate high approximation capabilities, they suffer from significant parameter inefficiency due to the requirement of maintaining unique parameterizations for every network edge. In this work, we propose GS-KAN (Generalized Sprecher-KAN), a lightweight architecture inspired by David Sprecher’s refinement of the superposition theorem. GS-KAN constructs unique edge functions by applying learnable linear transformations to a single learnable, shared parent function per layer. We evaluate GS-KAN against existing KAN architectures and MLPs across synthetic function approximation, tabular data regression and image classification tasks. Our results demonstrate that GS-KAN outperforms both MLPs and standard KAN baselines on continuous function approximation tasks while maintaining superior parameter efficiency. Additionally, GS-KAN achieves competitive performance with existing KAN architectures on tabular regression and outperforms MLPs on high-dimensional classification tasks. Crucially, the proposed architecture enables the deployment of KAN-based architectures in high-dimensional regimes under strict parameter constraints, a setting where standard implementations are typically infeasible due to parameter explosion. The source code is available at https://github.com/rambamn48/gs-impl.

Method: For convex models: Wasserstein-Robust Influence Functions (W-RIF) with provable coverage guarantees. For deep networks: Natural Wasserstein metric that measures perturbations in the geometry induced by the model’s feature covariance, eliminating spectral amplification that inflates Lipschitz bounds. Also introduces Self-Influence term analysis linking to Lipschitz constant for attribution stability.

Result: Natural Wasserstein metric reduces worst-case sensitivity by 76× and stabilizes attribution estimates. On CIFAR-10 with ResNet-18, Natural W-TRAK certifies 68.7% of ranking pairs compared to 0% for Euclidean baselines. Self-Influence achieves 0.970 AUROC for label noise detection, identifying 94.1% of corrupted labels by examining top 20% of training data.

Conclusion: The paper provides the first non-vacuous certified bounds for neural network attribution, explains why standard methods are geometrically fragile, and offers practical robust attribution with theoretical grounding for anomaly detection applications.

Abstract: Data attribution methods identify which training examples are responsible for a model’s predictions, but their sensitivity to distributional perturbations undermines practical reliability. We present a unified framework for certified robust attribution that extends from convex models to deep networks. For convex settings, we derive Wasserstein-Robust Influence Functions (W-RIF) with provable coverage guarantees. For deep networks, we demonstrate that Euclidean certification is rendered vacuous by spectral amplification – a mechanism where the inherent ill-conditioning of deep representations inflates Lipschitz bounds by over $10{,}000\times$. This explains why standard TRAK scores, while accurate point estimates, are geometrically fragile: naive Euclidean robustness analysis yields 0% certification. Our key contribution is the Natural Wasserstein metric, which measures perturbations in the geometry induced by the model’s own feature covariance. This eliminates spectral amplification, reducing worst-case sensitivity by $76\times$ and stabilizing attribution estimates. On CIFAR-10 with ResNet-18, Natural W-TRAK certifies 68.7% of ranking pairs compared to 0% for Euclidean baselines – to our knowledge, the first non-vacuous certified bounds for neural network attribution. Furthermore, we prove that the Self-Influence term arising from our analysis equals the Lipschitz constant governing attribution stability, providing theoretical grounding for leverage-based anomaly detection. Empirically, Self-Influence achieves 0.970 AUROC for label noise detection, identifying 94.1% of corrupted labels by examining just the top 20% of training data.

Method: Formalize masked diffusion sampling as a Markov decision process where the dLLM serves as the environment. Train lightweight transformer-based policies that map dLLM token confidences to unmasking decisions using reinforcement learning.

Result: Trained policies match state-of-the-art heuristic performance with semi-autoregressive generation and outperform them in full diffusion settings. Policies show transferability to new dLLMs and longer sequences, but degrade on out-of-domain data and have challenges with fine-grained accuracy-efficiency tuning.

Conclusion: Reinforcement learning offers a promising alternative to heuristic sampling for masked diffusion language models, achieving competitive performance while addressing limitations of manual tuning, though challenges remain with domain transfer and precise trade-off control.

Abstract: Diffusion (Large) Language Models (dLLMs) now match the downstream performance of their autoregressive counterparts on many tasks, while holding the promise of being more efficient during inference. One particularly successful variant is masked discrete diffusion, in which a buffer filled with special mask tokens is progressively replaced with tokens sampled from the model’s vocabulary. Efficiency can be gained by unmasking several tokens in parallel, but doing too many at once risks degrading the generation quality. Thus, one critical design aspect of dLLMs is the sampling procedure that selects, at each step of the diffusion process, which tokens to replace. Indeed, recent work has found that heuristic strategies such as confidence thresholding lead to both higher quality and token throughput compared to random unmasking. However, such heuristics have downsides: they require manual tuning, and we observe that their performance degrades with larger buffer sizes. In this work, we instead propose to train sampling procedures using reinforcement learning. Specifically, we formalize masked diffusion sampling as a Markov decision process in which the dLLM serves as the environment, and propose a lightweight policy architecture based on a single-layer transformer that maps dLLM token confidences to unmasking decisions. Our experiments show that these trained policies match the performance of state-of-the-art heuristics when combined with semi-autoregressive generation, while outperforming them in the full diffusion setting. We also examine the transferability of these policies, finding that they can generalize to new underlying dLLMs and longer sequence lengths. However, we also observe that their performance degrades when applied to out-of-domain data, and that fine-grained tuning of the accuracy-efficiency trade-off can be challenging with our approach.

Method: SEDONet replaces coordinate inputs in DeepONet trunks with a fixed Chebyshev spectral dictionary, providing a principled inductive bias tailored to bounded domains for better representation of non-periodic features.

Result: SEDONet achieves lowest relative L2 errors across PDE benchmarks (2D Poisson, 1D Burgers, advection-diffusion, Allen-Cahn, Lorenz-96), with 30-40% average improvement over baseline DeepONet and meaningful gains over Fourier-embedded variants on non-periodic geometries.

Conclusion: Chebyshev spectral embeddings provide a simple, parameter-neutral modification to DeepONets that delivers robust and efficient spectral framework for surrogate modeling of PDEs on bounded domains, accurately preserving high-frequency and boundary-localized features.

Abstract: Deep Operator Networks (DeepONets) have become a central tool in data-driven operator learning, providing flexible surrogates for nonlinear mappings arising in partial differential equations (PDEs). However, the standard trunk design based on fully connected layers acting on raw spatial or spatiotemporal coordinates struggles to represent sharp gradients, boundary layers, and non-periodic structures commonly found in PDEs posed on bounded domains with Dirichlet or Neumann boundary conditions. To address these limitations, we introduce the Spectral-Embedded DeepONet (SEDONet), a new DeepONet variant in which the trunk is driven by a fixed Chebyshev spectral dictionary rather than coordinate inputs. This non-periodic spectral embedding provides a principled inductive bias tailored to bounded domains, enabling the learned operator to capture fine-scale non-periodic features that are difficult for Fourier or MLP trunks to represent. SEDONet is evaluated on a suite of PDE benchmarks including 2D Poisson, 1D Burgers, 1D advection-diffusion, Allen-Cahn dynamics, and the Lorenz-96 chaotic system, covering elliptic, parabolic, advective, and multiscale temporal phenomena, all of which can be viewed as canonical problems in computational mechanics. Across all datasets, SEDONet consistently achieves the lowest relative L2 errors among DeepONet, FEDONet, and SEDONet, with average improvements of about 30-40% over the baseline DeepONet and meaningful gains over Fourier-embedded variants on non-periodic geometries. Spectral analyses further show that SEDONet more accurately preserves high-frequency and boundary-localized features, demonstrating the value of Chebyshev embeddings in non-periodic operator learning. The proposed architecture offers a simple, parameter-neutral modification to DeepONets, delivering a robust and efficient spectral framework for surrogate modeling of PDEs on bounded domains.

Method: The study approaches transformer failure modes through graph neural network theory, analyzing information propagation bottlenecks. It examines how decoder-only transformers’ causal nature creates geometric properties in information flow, and unifies existing ad-hoc solutions under theoretical frameworks.

Result: The analysis reveals that many transformer failure modes parallel GNN issues, with causal transformers exhibiting predictable geometric information propagation patterns that lead to specific failure modes. Existing intuitive solutions can be theoretically explained and improved.

Conclusion: Transformer failure modes can be systematically understood through GNN theory, providing a theoretical foundation for analyzing and improving transformers by addressing information propagation bottlenecks rather than relying on ad-hoc solutions.

Abstract: Transformers and more specifically decoder-only transformers dominate modern LLM architectures. While they have shown to work exceptionally well, they are not without issues, resulting in surprising failure modes and predictably asymmetric performance degradation. This article is a study of many of these observed failure modes of transformers through the lens of graph neural network (GNN) theory. We first make the case that much of deep learning, including transformers, is about learnable information mixing and propagation. This makes the study of model failure modes a study of bottlenecks in information propagation. This naturally leads to GNN theory, where there is already a rich literature on information propagation bottlenecks and theoretical failure modes of models. We then make the case that many issues faced by GNNs are also experienced by transformers. In addition, we analyze how the causal nature of decoder-only transformers create interesting geometric properties in information propagation, resulting in predictable and potentially devastating failure modes. Finally, we observe that existing solutions in transformer research tend to be ad-hoc and driven by intuition rather than grounded theoretical motivation. As such, we unify many such solutions under a more theoretical perspective, providing insight into why they work, what problem they are actually solving, and how they can be further improved to target specific failure modes of transformers. Overall, this article is an attempt to bridge the gap between observed failure modes in transformers and a general lack of theoretical understanding of them in this space.

Method: Developed a data-driven clinical support tool using a novel synthesized dataset combining US and Greek patient populations. Integrated three data sources: patient demographics, CT scans, and echocardiograms. Used leaf-level analysis to leverage population heterogeneity and avoid benchmarking against uncertain counterfactual risk estimates.

Result: The prescriptive model reduced PPI rates by 26% compared to current standard of care in the internal US population and by 16% in the external Greek validation cohort.

Conclusion: This work represents the first unified, personalized prescription strategy for transcatheter heart valve selection in TAVR, demonstrating significant reduction in pacemaker implantation risk across different patient populations.

Abstract: Transcatheter Aortic Valve Replacement (TAVR) has emerged as a minimally invasive treatment option for patients with severe aortic stenosis, a life-threatening cardiovascular condition. Multiple transcatheter heart valves (THV) have been approved for use in TAVR, but current guidelines regarding valve type prescription remain an active topic of debate. We propose a data-driven clinical support tool to identify the optimal valve type with the objective of minimizing the risk of permanent pacemaker implantation (PPI), a predominant postoperative complication. We synthesize a novel dataset that combines U.S. and Greek patient populations and integrates three distinct data sources (patient demographics, computed tomography scans, echocardiograms) while harmonizing differences in each country’s record system. We introduce a leaf-level analysis to leverage population heterogeneity and avoid benchmarking against uncertain counterfactual risk estimates. The final prescriptive model shows a reduction in PPI rates of 26% and 16% compared with the current standard of care in our internal U.S. population and external Greek validation cohort, respectively. To the best of our knowledge, this work represents the first unified, personalized prescription strategy for THV selection in TAVR.

Method: The study investigates LLM applicability for analog circuit design by examining how different data representations influence model behavior, comparing smaller models (T5, GPT-2) with larger foundation models (Mistral-7B, GPT-oss-20B) under varying training conditions, focusing on AI-assisted design with humans in the loop.

Result: Results reveal key reliability challenges: sensitivity to data format, instability in generated designs, and limited generalization to unseen circuit configurations, highlighting the current limitations of LLMs for structured engineering tasks.

Conclusion: The findings provide early evidence on both the limits and potential of LLMs as tools to enhance human capabilities in complex engineering tasks, offering insights for designing more reliable, deployable foundation models for structured, real-world applications.

Abstract: Large Language Models (LLMs) and transformer architectures have shown impressive reasoning and generation capabilities across diverse natural language tasks. However, their reliability and robustness in real-world engineering domains remain largely unexplored, limiting their practical utility in human-centric workflows. In this work, we investigate the applicability and consistency of LLMs for analog circuit design – a task requiring domain-specific reasoning, adherence to physical constraints, and structured representations – focusing on AI-assisted design where humans remain in the loop. We study how different data representations influence model behavior and compare smaller models (e.g., T5, GPT-2) with larger foundation models (e.g., Mistral-7B, GPT-oss-20B) under varying training conditions. Our results highlight key reliability challenges, including sensitivity to data format, instability in generated designs, and limited generalization to unseen circuit configurations. These findings provide early evidence on the limits and potential of LLMs as tools to enhance human capabilities in complex engineering tasks, offering insights into designing reliable, deployable foundation models for structured, real-world applications.

Method: Integrates fully quantized training, Stein’s estimator-based residual loss computation, and tensor-train decomposition. Key innovations: mixed-precision training with SMX format to eliminate data duplication, difference-based quantization for Stein’s estimator to mitigate underflow, and partial-reconstruction scheme for TT-Layers to reduce quantization-error accumulation. Also designs PINTA hardware accelerator.

Result: Achieves accuracy comparable to or better than full-precision baselines on 2-D Poisson, 20-D HJB, and 100-D Heat equations, with 5.5x-83.5x speedups and 159.6x-2324.1x energy savings.

Conclusion: Enables real-time PDE solving on edge devices and paves the way for energy-efficient scientific computing at scale.

Abstract: Physics-Informed Neural Networks (PINNs) have emerged as a promising paradigm for solving partial differential equations (PDEs) by embedding physical laws into neural network training objectives. However, their deployment on resource-constrained platforms is hindered by substantial computational and memory overhead, primarily stemming from higher-order automatic differentiation, intensive tensor operations, and reliance on full-precision arithmetic. To address these challenges, we present a framework that enables scalable and energy-efficient PINN training on edge devices. This framework integrates fully quantized training, Stein’s estimator (SE)-based residual loss computation, and tensor-train (TT) decomposition for weight compression. It contributes three key innovations: (1) a mixed-precision training method that use a square-block MX (SMX) format to eliminate data duplication during backpropagation; (2) a difference-based quantization scheme for the Stein’s estimator that mitigates underflow; and (3) a partial-reconstruction scheme (PRS) for TT-Layers that reduces quantization-error accumulation. We further design PINTA, a precision-scalable hardware accelerator, to fully exploit the performance of the framework. Experiments on the 2-D Poisson, 20-D Hamilton-Jacobi-Bellman (HJB), and 100-D Heat equations demonstrate that the proposed framework achieves accuracy comparable to or better than full-precision, uncompressed baselines while delivering 5.5x to 83.5x speedups and 159.6x to 2324.1x energy savings. This work enables real-time PDE solving on edge devices and paves the way for energy-efficient scientific computing at scale.

Method: The method constructs feature similarity matrices with both labeled and unlabeled samples, recovers ordinal rankings via spectral seriation algorithms, uses labeled samples for regularization, selects robust features with dynamic programming, and applies the recovered ordinal relationships for contrastive learning and supervision on unlabeled data.

Result: The method provides theoretical guarantees and empirical verification showing it surpasses existing state-of-the-art semi-supervised deep regression methods on various datasets.

Conclusion: The proposed approach successfully extends contrastive regression to semi-supervised settings, reducing dependence on costly annotations while achieving robust performance by leveraging unlabeled data for representation learning.

Abstract: Contrastive learning methods enforce label distance relationships in feature space to improve representation capability for regression models. However, these methods highly depend on label information to correctly recover ordinal relationships of features, limiting their applications to semi-supervised regression. In this work, we extend contrastive regression methods to allow unlabeled data to be used in the semi-supervised setting, thereby reducing the dependence on costly annotations. Particularly we construct the feature similarity matrix with both labeled and unlabeled samples in a mini-batch to reflect inter-sample relationships, and an accurate ordinal ranking of involved unlabeled samples can be recovered through spectral seriation algorithms if the level of error is within certain bounds. The introduction of labeled samples above provides regularization of the ordinal ranking with guidance from the ground-truth label information, making the ranking more reliable. To reduce feature perturbations, we further utilize the dynamic programming algorithm to select robust features for the matrix construction. The recovered ordinal relationship is then used for contrastive learning on unlabeled samples, and we thus allow more data to be used for feature representation learning, thereby achieving more robust results. The ordinal rankings can also be used to supervise predictions on unlabeled samples, serving as an additional training signal. We provide theoretical guarantees and empirical verification through experiments on various datasets, demonstrating that our method can surpass existing state-of-the-art semi-supervised deep regression methods. Our code have been released on https://github.com/xmed-lab/CLSS.

Method: Uses a variant of Rao-Blackwellized Particle Filter (RBPF) that analytically marginalizes linear-Gaussian substructure using assumed closed-form agent dynamics. Introduces two difference estimators: a Gaussian mixture model using RBPF weights and a reduced version confined to effective samples.

Result: Provides computable lower bounds on KL divergence between true intent distribution and RBPF estimates via Gaussian-mixture KL bounds. Shows reduced estimator performs almost as well as complete one. Experiments demonstrate fast and accurate intent recovery for compliant agents.

Conclusion: The RBPF approach enables efficient intent inference with theoretical guarantees, motivating future work on designing intent-obfuscating controllers for adversarial scenarios.

Abstract: Inferring the eventual goal of a mobile agent from noisy observations of its trajectory is a fundamental estimation problem. We initiate the study of such intent inference using a variant of a Rao-Blackwellized Particle Filter (RBPF), subject to the assumption that the agent’s intent manifests through closed-loop behavior with a state-of-the-art provable practical stability property. Leveraging the assumed closed-form agent dynamics, the RBPF analytically marginalizes the linear-Gaussian substructure and updates particle weights only, improving sample efficiency over a standard particle filter. Two difference estimators are introduced: a Gaussian mixture model using the RBPF weights and a reduced version confining the mixture to the effective sample. We quantify how well the adversary can recover the agent’s intent using information-theoretic leakage metrics and provide computable lower bounds on the Kullback-Leibler (KL) divergence between the true intent distribution and RBPF estimates via Gaussian-mixture KL bounds. We also provide a bound on the difference in performance between the two estimators, highlighting the fact that the reduced estimator performs almost as well as the complete one. Experiments illustrate fast and accurate intent recovery for compliant agents, motivating future work on designing intent-obfuscating controllers.

Method: Proposes Hetero-SplitEE with heterogeneous early exits in hierarchical training, allowing each client to select distinct split points (cut layers) based on computational capacity. Introduces two cooperative training strategies: Sequential strategy (clients trained sequentially with shared server model) and Averaging strategy (parallel training with periodic cross-layer aggregation).

Result: Extensive experiments on CIFAR-10, CIFAR-100, and STL-10 datasets using ResNet-18 demonstrate that the method maintains competitive accuracy while efficiently supporting diverse computational constraints.

Conclusion: Hetero-SplitEE enables practical deployment of collaborative deep learning in heterogeneous IoT ecosystems by accommodating device heterogeneity while maintaining model performance.

Abstract: The continuous scaling of deep neural networks has fundamentally transformed machine learning, with larger models demonstrating improved performance across diverse tasks. This growth in model size has dramatically increased the computational resources required for the training process. Consequently, distributed approaches, such as Federated Learning and Split Learning, have become essential paradigms for scalable deployment. However, existing Split Learning approaches assume client homogeneity and uniform split points across all participants. This critically limits their applicability to real-world IoT systems where devices exhibit heterogeneity in computational resources. To address this limitation, this paper proposes Hetero-SplitEE, a novel method that enables heterogeneous IoT devices to train a shared deep neural network in parallel collaboratively. By integrating heterogeneous early exits into hierarchical training, our approach allows each client to select distinct split points (cut layers) tailored to its computational capacity. In addition, we propose two cooperative training strategies, the Sequential strategy and the Averaging strategy, to facilitate this collaboration among clients with different split points. The Sequential strategy trains clients sequentially with a shared server model to reduce computational overhead. The Averaging strategy enables parallel client training with periodic cross-layer aggregation. Extensive experiments on CIFAR-10, CIFAR-100, and STL-10 datasets using ResNet-18 demonstrate that our method maintains competitive accuracy while efficiently supporting diverse computational constraints, enabling practical deployment of collaborative deep learning in heterogeneous IoT ecosystems.

Method: Imposes Gaussian process priors on representations and obtains generalized Bayesian posterior minimizing a loss function that encourages informative representations. The GP covariance function naturally pulls similar representations together as an alternative to explicit positive samples.

Result: GPSSL outperforms traditional methods in accuracy, uncertainty quantification, and error control across various datasets for classification and regression tasks. Shows connections to kernel PCA and VICReg while providing posterior uncertainties.

Conclusion: GPSSL provides an effective SSL approach with built-in uncertainty quantification that addresses key limitations of existing methods, offering better performance and uncertainty propagation to downstream tasks.

Abstract: Self supervised learning (SSL) is a machine learning paradigm where models learn to understand the underlying structure of data without explicit supervision from labeled samples. The acquired representations from SSL have demonstrated useful for many downstream tasks including clustering, and linear classification, etc. To ensure smoothness of the representation space, most SSL methods rely on the ability to generate pairs of observations that are similar to a given instance. However, generating these pairs may be challenging for many types of data. Moreover, these methods lack consideration of uncertainty quantification and can perform poorly in out-of-sample prediction settings. To address these limitations, we propose Gaussian process self supervised learning (GPSSL), a novel approach that utilizes Gaussian processes (GP) models on representation learning. GP priors are imposed on the representations, and we obtain a generalized Bayesian posterior minimizing a loss function that encourages informative representations. The covariance function inherent in GPs naturally pulls representations of similar units together, serving as an alternative to using explicitly defined positive samples. We show that GPSSL is closely related to both kernel PCA and VICReg, a popular neural network-based SSL method, but unlike both allows for posterior uncertainties that can be propagated to downstream tasks. Experiments on various datasets, considering classification and regression tasks, demonstrate that GPSSL outperforms traditional methods in terms of accuracy, uncertainty quantification, and error control.

Method: PathHD uses hyperdimensional computing to encode relation paths into block-diagonal GHRR hypervectors, ranks candidates with blockwise cosine similarity and Top-K pruning, then performs one-shot LLM adjudication with cited supporting paths.

Result: On WebQSP, CWQ, and GrailQA, PathHD achieves comparable/better Hits@1 than neural baselines, reduces latency by 40-60%, GPU memory by 3-5×, and provides faithful, path-grounded rationales for better error diagnosis.

Conclusion: Carefully designed HDC representations offer a practical substrate for efficient KG-LLM reasoning with favorable accuracy-efficiency-interpretability trade-off, enabling scalable deployment.

Abstract: Recent advances in large language models (LLMs) have enabled strong reasoning over both structured and unstructured knowledge. When grounded on knowledge graphs (KGs), however, prevailing pipelines rely on heavy neural encoders to embed and score symbolic paths or on repeated LLM calls to rank candidates, leading to high latency, GPU cost, and opaque decisions that hinder faithful, scalable deployment. We propose PathHD, a lightweight and encoder-free KG reasoning framework that replaces neural path scoring with hyperdimensional computing (HDC) and uses only a single LLM call per query. PathHD encodes relation paths into block-diagonal GHRR hypervectors, ranks candidates with blockwise cosine similarity and Top-K pruning, and then performs a one-shot LLM adjudication to produce the final answer together with cited supporting paths. Technically, PathHD is built on three ingredients: (i) an order-aware, non-commutative binding operator for path composition, (ii) a calibrated similarity for robust hypervector-based retrieval, and (iii) a one-shot adjudication step that preserves interpretability while eliminating per-path LLM scoring. On WebQSP, CWQ, and the GrailQA split, PathHD (i) attains comparable or better Hits@1 than strong neural baselines while using one LLM call per query; (ii) reduces end-to-end latency by $40-60%$ and GPU memory by $3-5\times$ thanks to encoder-free retrieval; and (iii) delivers faithful, path-grounded rationales that improve error diagnosis and controllability. These results indicate that carefully designed HDC representations provide a practical substrate for efficient KG-LLM reasoning, offering a favorable accuracy-efficiency-interpretability trade-off.

Method: Leverages the PLM itself with a lightweight curation pipeline and domain-specific filters to construct high-quality training data. Uses these filters to refine PLM outputs and identify candidates for in vitro evaluation. Combines filters with SFT to enable generation of more stable and functional enzymes while expanding protein sequence space exploration.

Result: The supervised fine-tuned model generates sequences that are more novel and display improved characteristics across both targeted design constraints and emergent protein property measures. Demonstrated effectiveness with GenSLM applied to tryptophan synthase enzyme family.

Conclusion: Presents a general, PLM-agnostic approach for fast supervised fine-tuning that improves protein sequence generation fidelity, reliability, and novelty without requiring costly experimental datasets, enabling more effective protein design and exploration.

Abstract: Supervised fine-tuning (SFT) is a standard approach for adapting large language models to specialized domains, yet its application to protein sequence modeling and protein language models (PLMs) remains ad hoc. This is in part because high-quality annotated data are far more difficult to obtain for proteins than for natural language. We present a simple and general recipe for fast SFT of PLMs, designed to improve the fidelity, reliability, and novelty of generated protein sequences. Unlike existing approaches that require costly precompiled experimental datasets for SFT, our method leverages the PLM itself, integrating a lightweight curation pipeline with domain-specific filters to construct high-quality training data. These filters can independently refine a PLM’s output and identify candidates for in vitro evaluation; when combined with SFT, they enable PLMs to generate more stable and functional enzymes, while expanding exploration into protein sequence space beyond natural variants. Although our approach is agnostic to both the choice of protein language model (PLM) and the protein system, we demonstrate its effectiveness with a genome-scale PLM (GenSLM) applied to the tryptophan synthase enzyme family. The supervised fine-tuned model generates sequences that are not only more novel but also display improved characteristics across both targeted design constraints and emergent protein property measures.

Method: IPDNN framework with GLOW activation function, dynamic scatter subregion identification, and transfer learning

Result: Superior reconstruction accuracy, robustness, and efficiency compared to state-of-the-art methods

Conclusion: The proposed solver successfully integrates physical interpretability with real-time inference capability for practical electromagnetic applications

Abstract: This paper presents an improved physics-driven neural network (IPDNN) framework for solving electromagnetic inverse scattering problems (ISPs). A new Gaussian-localized oscillation-suppressing window (GLOW) activation function is introduced to stabilize convergence and enable a lightweight yet accurate network architecture. A dynamic scatter subregion identification strategy is further developed to adaptively refine the computational domain, preventing missed detections and reducing computational cost. Moreover, transfer learning is incorporated to extend the solver’s applicability to practical scenarios, integrating the physical interpretability of iterative algorithms with the real-time inference capability of neural networks. Numerical simulations and experimental results demonstrate that the proposed solver achieves superior reconstruction accuracy, robustness, and efficiency compared with existing state-of-the-art methods.

Method: Theoretical analysis of node-anchored Subgraph GNNs’ expressive power for approximating Strong Branching, plus extensive empirical evaluation on four benchmark datasets comparing them with MPNNs and heuristics.

Result: Theoretically, node-anchored Subgraph GNNs (with lower expressive power than 3-WL) are sufficient to approximate Strong Branching. However, empirically, their O(n) complexity causes memory bottlenecks and slower solving times than MPNNs and heuristics.

Conclusion: For MILP branching, the computational cost of expressive GNNs currently outweighs their decision quality gains. Future research should focus on efficiency-preserving expressivity rather than pure expressivity improvements.

Abstract: Graph Neural Networks (GNNs) have emerged as a promising approach for ``learning to branch’’ in Mixed-Integer Linear Programming (MILP). While standard Message-Passing GNNs (MPNNs) are efficient, they theoretically lack the expressive power to fully represent MILP structures. Conversely, higher-order GNNs (like 2-FGNNs) are expressive but computationally prohibitive. In this work, we investigate Subgraph GNNs as a theoretical middle ground. Crucially, while previous work [Chen et al., 2025] demonstrated that GNNs with 3-WL expressive power can approximate Strong Branching, we prove a sharper result: node-anchored Subgraph GNNs whose expressive power is strictly lower than 3-WL [Zhang et al., 2023] are sufficient to approximate Strong Branching scores. However, our extensive empirical evaluation on four benchmark datasets reveals a stark contrast between theory and practice. While node-anchored Subgraph GNNs theoretically offer superior branching decisions, their $O(n)$ complexity overhead results in significant memory bottlenecks and slower solving times than MPNNs and heuristics. Our results indicate that for MILP branching, the computational cost of expressive GNNs currently outweighs their gains in decision quality, suggesting that future research must focus on efficiency-preserving expressivity.

Method: Developed forecasting framework using CSI MasterFormat as target structure, integrating explanatory variables (raw material prices, commodity indexes, macroeconomic indicators), and evaluating four time-series models (LSTM, ARIMA, VECM, Chronos-Bolt) under baseline and extended configurations.

Result: Explanatory variables significantly improved predictive performance across all models. LSTM achieved highest accuracy with RMSE=1.390 and MAPE=0.957, showing 59% improvement over ARIMA. Framework validated across multiple CSI divisions with Division 06 as demonstration case.

Conclusion: The research provides a robust methodology enabling owners and contractors to improve budgeting practices and achieve more reliable cost estimation at the Definitive level through granular construction material price forecasting.

Abstract: The persistent volatility of construction material prices poses significant risks to cost estimation, budgeting, and project delivery, underscoring the urgent need for granular and scalable forecasting methods. This study develops a forecasting framework that leverages the Construction Specifications Institute (CSI) MasterFormat as the target data structure, enabling predictions at the six-digit section level and supporting detailed cost projections across a wide spectrum of building materials. To enhance predictive accuracy, the framework integrates explanatory variables such as raw material prices, commodity indexes, and macroeconomic indicators. Four time-series models, Long Short-Term Memory (LSTM), Autoregressive Integrated Moving Average (ARIMA), Vector Error Correction Model (VECM), and Chronos-Bolt, were evaluated under both baseline configurations (using CSI data only) and extended versions with explanatory variables. Results demonstrate that incorporating explanatory variables significantly improves predictive performance across all models. Among the tested approaches, the LSTM model consistently achieved the highest accuracy, with RMSE values as low as 1.390 and MAPE values of 0.957, representing improvements of up to 59% over the traditional statistical time-series model, ARIMA. Validation across multiple CSI divisions confirmed the framework’s scalability, while Division 06 (Wood, Plastics, and Composites) is presented in detail as a demonstration case. This research offers a robust methodology that enables owners and contractors to improve budgeting practices and achieve more reliable cost estimation at the Definitive level.

Method: First aggregates diverse biological datasets (molecular, protein, gene, and pathway-level interactions), then develops an optimal transport-based approach to generate high-quality pseudo-labels for unlabeled molecule-protein pairs, leveraging known interaction distributions to guide label assignment.

Result: Demonstrates substantial improvements over state-of-the-art methods in prediction accuracies and zero-shot ability across unseen interactions on multiple MPI datasets including virtual screening and protein retrieval tasks.

Conclusion: Provides a new paradigm for leveraging diverse biological data sources to tackle problems traditionally constrained by single- or bi-modal learning, paving the way for future advances in computational biology and drug discovery.

Abstract: Predicting molecule-protein interactions (MPIs) is a fundamental task in computational biology, with crucial applications in drug discovery and molecular function annotation. However, existing MPI models face two major challenges. First, the scarcity of labeled molecule-protein pairs significantly limits model performance, as available datasets capture only a small fraction of biological relevant interactions. Second, most methods rely solely on molecular and protein features, ignoring broader biological context such as genes, metabolic pathways, and functional annotations that could provide essential complementary information. To address these limitations, our framework first aggregates diverse biological datasets, including molecular, protein, genes and pathway-level interactions, and then develop an optimal transport-based approach to generate high-quality pseudo-labels for unlabeled molecule-protein pairs, leveraging the underlying distribution of known interactions to guide label assignment. By treating pseudo-labeling as a mechanism for bridging disparate biological modalities, our approach enables the effective use of heterogeneous data to enhance MPI prediction. We evaluate our framework on multiple MPI datasets including virtual screening tasks and protein retrieval tasks, demonstrating substantial improvements over state-of-the-art methods in prediction accuracies and zero shot ability across unseen interactions. Beyond MPI prediction, our approach provides a new paradigm for leveraging diverse biological data sources to tackle problems traditionally constrained by single- or bi-modal learning, paving the way for future advances in computational biology and drug discovery.

Method: Proposes a counterfactual learning framework that asks “What if we backtrack and perform alternative actions when unsafe events occur?” Uses a structural causal model to predict outcomes of different actions and integrates counterfactual modules with additional “X” modules for safe RL practices.

Result: CFLight significantly reduces collisions through near-zero collision control strategy, outperforms conventional RL methods and recent safe RL models on both real-world and synthetic datasets while improving overall traffic performance.

Conclusion: CFLight provides a generalized, safe RL framework for traffic signal control that effectively balances safety and efficiency, with potential applications in other domains beyond traffic management.

Abstract: Traffic accidents result in millions of injuries and fatalities globally, with a significant number occurring at intersections each year. Traffic Signal Control (TSC) is an effective strategy for enhancing safety at these urban junctures. Despite the growing popularity of Reinforcement Learning (RL) methods in optimizing TSC, these methods often prioritize driving efficiency over safety, thus failing to address the critical balance between these two aspects. Additionally, these methods usually need more interpretability. CounterFactual (CF) learning is a promising approach for various causal analysis fields. In this study, we introduce a novel framework to improve RL for safety aspects in TSC. This framework introduces a novel method based on CF learning to address the question: What if, when an unsafe event occurs, we backtrack to perform alternative actions, and will this unsafe event still occur in the subsequent period?'' To answer this question, we propose a new structure causal model to predict the result after executing different actions, and we propose a new CF module that integrates with additional X’’ modules to promote safe RL practices. Our new algorithm, CFLight, which is derived from this framework, effectively tackles challenging safety events and significantly improves safety at intersections through a near-zero collision control strategy. Through extensive numerical experiments on both real-world and synthetic datasets, we demonstrate that CFLight reduces collisions and improves overall traffic performance compared to conventional RL methods and the recent safe RL model. Moreover, our method represents a generalized and safe framework for RL methods, opening possibilities for applications in other domains. The data and code are available in the github https://github.com/MJLee00/CFLight-Enhancing-Safety-with-Traffic-Signal-Control-through-Counterfactual-Learning.

Method: Federated distillation-assisted vehicle edge caching scheme based on lightweight denoising diffusion probabilistic model (LDPM). Uses knowledge distillation to reduce model transmission frequency compared to traditional FL.

Result: Simulation shows good robustness to vehicle speed variations, significantly reduces communication overhead, and improves cache hit percentage.

Conclusion: The proposed scheme effectively addresses communication overhead and training failure issues in traditional FL for vehicle edge caching while maintaining privacy protection.

Abstract: Vehicle edge caching is a promising technology that can significantly reduce the latency for vehicle users (VUs) to access content by pre-caching user-interested content at edge nodes. It is crucial to accurately predict the content that VUs are interested in without exposing their privacy. Traditional federated learning (FL) can protect user privacy by sharing models rather than raw data. However, the training of FL requires frequent model transmission, which can result in significant communication overhead. Additionally, vehicles may leave the road side unit (RSU) coverage area before training is completed, leading to training failures. To address these issues, in this letter, we propose a federated distillation-assisted vehicle edge caching scheme based on lightweight denoising diffusion probabilistic model (LDPM). The simulation results demonstrate that the proposed vehicle edge caching scheme has good robustness to variations in vehicle speed, significantly reducing communication overhead and improving cache hit percentage.

Method: Proposed ConFormer (Conditional Transformer) integrates graph propagation with guided normalization layer, dynamically adjusting spatial and temporal node relationships based on historical patterns using enriched traffic datasets from Tokyo and California.

Result: ConFormer surpasses state-of-the-art STAEFormer in both predictive performance and efficiency, achieving lower computational costs and reduced parameter demands, consistently outperforming mainstream spatio-temporal baselines across multiple metrics.

Conclusion: ConFormer demonstrates significant potential to advance traffic prediction research by effectively integrating external factors and achieving superior performance with improved efficiency.

Abstract: Traffic prediction remains a key challenge in spatio-temporal data mining, despite progress in deep learning. Accurate forecasting is hindered by the complex influence of external factors such as traffic accidents and regulations, often overlooked by existing models due to limited data integration. To address these limitations, we present two enriched traffic datasets from Tokyo and California, incorporating traffic accident and regulation data. Leveraging these datasets, we propose ConFormer (Conditional Transformer), a novel framework that integrates graph propagation with guided normalization layer. This design dynamically adjusts spatial and temporal node relationships based on historical patterns, enhancing predictive accuracy. Our model surpasses the state-of-the-art STAEFormer in both predictive performance and efficiency, achieving lower computational costs and reduced parameter demands. Extensive evaluations demonstrate that ConFormer consistently outperforms mainstream spatio-temporal baselines across multiple metrics, underscoring its potential to advance traffic prediction research.

Method: Black-box distillation attack using 48,000 instruction queries to Meditron-7B, collecting 25,000 benign instruction-response pairs, then fine-tuning LLaMA3 8B surrogate via LoRA with zero-alignment supervision. Developed dynamic adversarial evaluation framework with Generative Query-based harmful prompt generation, verifier filtering, category-wise failure analysis, and adaptive Random Search jailbreak attacks.

Result: With only $12 cost, surrogate achieves strong fidelity on benign inputs while producing unsafe completions for 86% of adversarial prompts, far exceeding Meditron-7B (66%) and untuned base model (46%). Reveals pronounced functional-ethical gap where task utility transfers but alignment collapses.

Conclusion: Benign-only black-box distillation exposes practical threat: adversaries can cheaply replicate medical LLM capabilities while stripping safety mechanisms, underscoring need for extraction-aware safety monitoring and layered defense systems for real-time alignment drift detection.

Abstract: As medical large language models (LLMs) become increasingly integrated into clinical workflows, concerns around alignment robustness, and safety are escalating. Prior work on model extraction has focused on classification models or memorization leakage, leaving the vulnerability of safety-aligned generative medical LLMs underexplored. We present a black-box distillation attack that replicates the domain-specific reasoning of safety-aligned medical LLMs using only output-level access. By issuing 48,000 instruction queries to Meditron-7B and collecting 25,000 benign instruction response pairs, we fine-tune a LLaMA3 8B surrogate via parameter efficient LoRA under a zero-alignment supervision setting, requiring no access to model weights, safety filters, or training data. With a cost of $12, the surrogate achieves strong fidelity on benign inputs while producing unsafe completions for 86% of adversarial prompts, far exceeding both Meditron-7B (66%) and the untuned base model (46%). This reveals a pronounced functional-ethical gap, task utility transfers, while alignment collapses. To analyze this collapse, we develop a dynamic adversarial evaluation framework combining Generative Query (GQ)-based harmful prompt generation, verifier filtering, category-wise failure analysis, and adaptive Random Search (RS) jailbreak attacks. We also propose a layered defense system, as a prototype detector for real-time alignment drift in black-box deployments. Our findings show that benign-only black-box distillation exposes a practical and under-recognized threat: adversaries can cheaply replicate medical LLM capabilities while stripping safety mechanisms, underscoring the need for extraction-aware safety monitoring.

Method: Organizes existing methods into three families, identifies desirable properties for distance measures, and proposes a Cauchy-Schwarz fairness regularizer that penalizes CS divergence between prediction distributions conditioned on sensitive groups.

Result: CS regularizer consistently improves Demographic Parity and Equal Opportunity metrics while maintaining competitive accuracy, and achieves more stable utility-fairness trade-off across hyperparameter settings.

Conclusion: The Cauchy-Schwarz divergence provides a theoretically grounded and practically effective fairness regularizer with desirable properties including tight generalization bounds, robustness to scale differences, and handling arbitrary prediction distributions.

Abstract: Group fairness in machine learning is often enforced by adding a regularizer that reduces the dependence between model predictions and sensitive attributes. However, existing regularizers are built on heterogeneous distance measures and design choices, which makes their behavior hard to reason about and their performance inconsistent across tasks. This raises a basic question: what properties make a good fairness regularizer? We address this question by first organizing existing in-process methods into three families: (i) matching prediction statistics across sensitive groups, (ii) aligning latent representations, and (iii) directly minimizing dependence between predictions and sensitive attributes. Through this lens, we identify desirable properties of the underlying distance measure, including tight generalization bounds, robustness to scale differences, and the ability to handle arbitrary prediction distributions. Motivated by these properties, we propose a Cauchy-Schwarz (CS) fairness regularizer that penalizes the empirical CS divergence between prediction distributions conditioned on sensitive groups. Under a Gaussian comparison, we show that CS divergence yields a tighter bound than Kullback-Leibler divergence, Maximum Mean Discrepancy, and the mean disparity used in Demographic Parity, and we discuss how these advantages translate to a distribution-free, kernel-based estimator that naturally extends to multiple sensitive attributes. Extensive experiments on four tabular benchmarks and one image dataset demonstrate that the proposed CS regularizer consistently improves Demographic Parity and Equal Opportunity metrics while maintaining competitive accuracy, and achieves a more stable utility-fairness trade-off across hyperparameter settings compared to prior regularizers.

Method: Conducted systematic study across four vision and language models, varying training data composition to characterize sensitivity of subgroup performance to data balance. Proposed and formalized the latent separation hypothesis, which states that a partially fine-tuned model’s dependence on subgroup representation is determined by the degree of separation between subgroups in the latent space of the pre-trained model. Provided theoretical analysis and empirical validation.

Result: Found that imbalanced data distributions can sometimes improve subgroup performance, and subgroup performance can remain unaffected by the absence of entire subgroups during training. Validated the latent separation hypothesis empirically, showing that quantitative analysis of latent subgroup separation can predict when subgroup representation matters.

Conclusion: The latent separation hypothesis provides a framework for understanding when subgroup representation matters in model training. This has practical applications for foundation model fine-tuning, where quantitative analysis of latent subgroup separation can inform data collection and balancing decisions, moving beyond the simplistic assumption that balanced data always optimizes performance.

Abstract: Unequal representation of demographic groups in training data poses challenges to model generalisation across populations. Standard practice assumes that balancing subgroup representation optimises performance. However, recent empirical results contradict this assumption: in some cases, imbalanced data distributions actually improve subgroup performance, while in others, subgroup performance remains unaffected by the absence of an entire subgroup during training. We conduct a systematic study of subgroup allocation across four vision and language models, varying training data composition to characterise the sensitivity of subgroup performance to data balance. We propose the latent separation hypothesis, which states that a partially fine-tuned model’s dependence on subgroup representation is determined by the degree of separation between subgroups in the latent space of the pre-trained model. We formalise this hypothesis, provide theoretical analysis, and validate it empirically. Finally, we present a practical application to foundation model fine-tuning, demonstrating that quantitative analysis of latent subgroup separation can inform data collection and balancing decisions.

Method: Develops a contextual pricing algorithm based on optimistic posterior sampling. For the non-contextual case, proposes a variance-aware zooming algorithm. Both methods handle unknown distribution of buyer valuation types with finite support.

Result: Achieves regret Õ(K⋆√dT) for contextual pricing, proven to be tight in d and T up to logarithmic terms. For non-contextual pricing, obtains optimal dependence on K⋆ through variance-aware zooming.

Conclusion: The paper provides the first study of contextual dynamic pricing with heterogeneous buyer populations, developing algorithms with tight regret bounds and optimal dependence on key parameters, advancing beyond homogeneous buyer assumptions.

Abstract: We initiate the study of contextual dynamic pricing with a heterogeneous population of buyers, where a seller repeatedly posts prices (over $T$ rounds) that depend on the observable $d$-dimensional context and receives binary purchase feedback. Unlike prior work assuming homogeneous buyer types, in our setting the buyer’s valuation type is drawn from an unknown distribution with finite support size $K_{\star}$. We develop a contextual pricing algorithm based on optimistic posterior sampling with regret $\widetilde{O}(K_{\star}\sqrt{dT})$, which we prove to be tight in $d$ and $T$ up to logarithmic terms. Finally, we refine our analysis for the non-contextual pricing case, proposing a variance-aware zooming algorithm that achieves the optimal dependence on $K_{\star}$.

Method: QuanvNeXt uses fully quanvolutional architecture with novel Cross Residual blocks that reduce feature homogeneity and strengthen cross-feature relationships. The model was evaluated on two open-source EEG datasets for depression diagnosis.

Result: Achieved average accuracy of 93.1% and AUC-ROC of 97.2%, outperforming InceptionTime (91.7% accuracy, 95.9% AUC-ROC). Uncertainty analysis showed well-calibrated predictions with low ECE scores (0.0436-0.1159) even under noise perturbations. XAI analysis confirmed effective learning of spectrotemporal patterns distinguishing healthy vs. depressed subjects.

Conclusion: QuanvNeXt establishes an efficient and reliable approach for EEG-based depression diagnosis with superior performance, robust uncertainty calibration, and interpretable feature learning capabilities.

Abstract: This study presents QuanvNeXt, an end-to-end fully quanvolutional model for EEG-based depression diagnosis. QuanvNeXt incorporates a novel Cross Residual block, which reduces feature homogeneity and strengthens cross-feature relationships while retaining parameter efficiency. We evaluated QuanvNeXt on two open-source datasets, where it achieved an average accuracy of 93.1% and an average AUC-ROC of 97.2%, outperforming state-of-the-art baselines such as InceptionTime (91.7% accuracy, 95.9% AUC-ROC). An uncertainty analysis across Gaussian noise levels demonstrated well-calibrated predictions, with ECE scores remaining low (0.0436, Dataset 1) to moderate (0.1159, Dataset 2) even at the highest perturbation (ε = 0.1). Additionally, a post-hoc explainable AI analysis confirmed that QuanvNeXt effectively identifies and learns spectrotemporal patterns that distinguish between healthy controls and major depressive disorder. Overall, QuanvNeXt establishes an efficient and reliable approach for EEG-based depression diagnosis.

Method: Fully latent autoregressive scheme using Gaussian Process integrated into VAE, with causal GP prior, structured amortized posterior, and regularized ELBO training. Sequential dynamics transferred to continuous latent space while decoder remains non-autoregressive for parallel generation.

Result: Model trains stably in proof-of-concept framework, with sequential and parallel sampling variants showing consistent behavior. Demonstrates that temporal structure can be supported by latent space geometry rather than explicit neural operations.

Conclusion: Part of temporal structure in language models can be effectively captured through probabilistic geometry of latent space, enabling parallel decoding while maintaining sequential coherence, offering alternative to traditional autoregressive neural approaches.

Abstract: We investigate a fully Latent AutoRegressive scheme based on a Gaussian Process (GP) integrated into a Variational Autoencoder (VAE). In this setting, sequential dynamics are transferred from the observation space to a continuous latent space, while linguistic generation remains parallel through a non-autoregressive decoder. We present a complete methodological formulation, including a causal GP prior, a structured amortized posterior, and a training protocol based on a regularized ELBO. Empirical evaluation, conducted within a deliberately constrained proof-of-concept (POC) framework, shows that the model can be trained stably and that the sequential and parallel sampling variants exhibit consistent behavior. Overall, the results suggest that part of the temporal structure in a language model can be supported by the probabilistic geometry of the latent space rather than by explicit neural operations.

Method: Introduced Stanford Sleep Bench, a large-scale PSG dataset with 17,467 recordings (over 163,000 hours) from a sleep clinic, including 13 clinical disease prediction tasks and canonical sleep tasks. Systematically evaluated self-supervised representation learning pre-training methods on this benchmark across four downstream tasks.

Result: Multiple pretraining methods achieve comparable performance for sleep staging, apnea diagnosis, and age estimation. However, for mortality and disease prediction, contrastive learning significantly outperforms other approaches while also converging faster during pretraining.

Conclusion: The Stanford Sleep Bench dataset and systematic evaluation framework will facilitate reproducibility and advance sleep research. The authors will release the dataset along with pretrained model weights, training pipelines, and evaluation code.

Abstract: Polysomnography (PSG), the gold standard test for sleep analysis, generates vast amounts of multimodal clinical data, presenting an opportunity to leverage self-supervised representation learning (SSRL) for pre-training foundation models to enhance sleep analysis. However, progress in sleep foundation models is hindered by two key limitations: (1) the lack of a shared dataset and benchmark with diverse tasks for training and evaluation, and (2) the absence of a systematic evaluation of SSRL approaches across sleep-related tasks. To address these gaps, we introduce Stanford Sleep Bench, a large-scale PSG dataset comprising 17,467 recordings totaling over 163,000 hours from a major sleep clinic, including 13 clinical disease prediction tasks alongside canonical sleep-related tasks such as sleep staging, apnea diagnosis, and age estimation. We systematically evaluate SSRL pre-training methods on Stanford Sleep Bench, assessing downstream performance across four tasks: sleep staging, apnea diagnosis, age estimation, and disease and mortality prediction. Our results show that multiple pretraining methods achieve comparable performance for sleep staging, apnea diagnosis, and age estimation. However, for mortality and disease prediction, contrastive learning significantly outperforms other approaches while also converging faster during pretraining. To facilitate reproducibility and advance sleep research, we will release Stanford Sleep Bench along with pretrained model weights, training pipelines, and evaluation code.

Method: Proposes TCSC framework for semantic task offloading via V2I and V2V communications. Formulates MINLP problem for task latency and semantic symbols, then decomposes into two subproblems: 1) MAPPO-PDN (multi-agent proximal policy optimization with parametric distribution noise) for semantic symbol optimization, 2) Linear programming for offloading ratio optimization.

Result: Simulations demonstrate superior performance compared to other algorithms, showing effectiveness of the proposed TCSC framework and optimization methods in improving task processing efficiency in vehicular edge computing environments.

Conclusion: The TCSC framework with MAPPO-PDN and LP optimization provides an effective solution for semantic task offloading in highway IoV scenarios, achieving better performance than existing approaches through cooperative semantic communication and intelligent optimization.

Abstract: Semantic Communication (SC) combined with Vehicular edge computing (VEC) provides an efficient edge task processing paradigm for Internet of Vehicles (IoV). Focusing on highway scenarios, this paper proposes a Tripartite Cooperative Semantic Communication (TCSC) framework, which enables Vehicle Users (VUs) to perform semantic task offloading via Vehicle-to-Infrastructure (V2I) and Vehicle-to-Vehicle (V2V) communications. Considering task latency and the number of semantic symbols, the framework constructs a Mixed-Integer Nonlinear Programming (MINLP) problem, which is transformed into two subproblems. First, we innovatively propose a multi-agent proximal policy optimization task offloading optimization method based on parametric distribution noise (MAPPO-PDN) to solve the optimization problem of the number of semantic symbols; second, linear programming (LP) is used to solve offloading ratio. Simulations show that performance of this scheme is superior to that of other algorithms.

Method: The study examines membership inference attacks (MIA) on open-source generative audio models to determine if specific audio samples were in training sets. When MIA proves limited for individual samples, the research focuses on dataset inference (DI) - aggregating membership evidence across multiple samples from an artist’s collection, building on prior work in text and vision domains.

Result: Membership inference alone is ineffective at scale due to weak per-sample signals in models trained on large, diverse datasets. However, dataset inference successfully detects when an artist’s collection of works contributed to model training, providing a practical verification mechanism.

Conclusion: Dataset inference offers a promising approach for copyright protection and dataset accountability in large audio generative models, as it can reliably detect when artists’ collections were used in training, unlike single-sample membership inference.

Abstract: Generative audio models, based on diffusion and autoregressive architectures, have advanced rapidly in both quality and expressiveness. This progress, however, raises pressing copyright concerns, as such models are often trained on vast corpora of artistic and commercial works. A central question is whether one can reliably verify if an artist’s material was included in training, thereby providing a means for copyright holders to protect their content. In this work, we investigate the feasibility of such verification through membership inference attacks (MIA) on open-source generative audio models, which attempt to determine whether a specific audio sample was part of the training set. Our empirical results show that membership inference alone is of limited effectiveness at scale, as the per-sample membership signal is weak for models trained on large and diverse datasets. However, artists and media owners typically hold collections of works rather than isolated samples. Building on prior work in text and vision domains, in this work we focus on dataset inference (DI), which aggregates diverse membership evidence across multiple samples. We find that DI is successful in the audio domain, offering a more practical mechanism for assessing whether an artist’s works contributed to model training. Our results suggest DI as a promising direction for copyright protection and dataset accountability in the era of large audio generative models.

Method: Focuses on 2-layer ReLU networks, examining boundary hyperplanes and channel vectors. Constructs examples showing expressivity limitations from equivariance constraints, then demonstrates compensation through model size enlargement while analyzing hypothesis space complexity.

Result: Equivariance constraints can strictly limit expressive power, but this drawback can be compensated by enlarging model size. Despite larger models, the resulting architecture corresponds to a hypothesis space with lower complexity, suggesting superior generalizability for equivariant networks.

Conclusion: Equivariant networks face expressivity limitations due to symmetry constraints, but these can be overcome with larger models while maintaining lower hypothesis complexity, explaining their empirical success and superior generalization properties.

Abstract: Equivariant neural networks encode symmetry as an inductive bias and have achieved strong empirical performance in wide domains. However, their expressive power remains not well understood. Focusing on 2-layer ReLU networks, this paper investigates the impact of equivariance constraints on the expressivity of equivariant and layer-wise equivariant networks. By examining the boundary hyperplanes and the channel vectors of ReLU networks, we construct an example showing that equivariance constraints could strictly limit expressive power. However, we demonstrate that this drawback can be compensated via enlarging the model size. Furthermore, we show that despite a larger model size, the resulting architecture could still correspond to a hypothesis space with lower complexity, implying superior generalizability for equivariant networks.

Method: Develops a comprehensive system architecture for continuous data collection and integration, establishes linkage between design features and manufacturing process data, and builds a machine learning model for automated design improvement suggestions.

Result: Enables automated design improvement suggestions and opens new possibilities for sustainable product development by integrating manufacturing process data with sustainability metrics.

Conclusion: The integration of design features with manufacturing process data through a systematic data-driven approach can significantly enhance product design improvements and support sustainable development in industrial settings.

Abstract: The growing adoption of Industrial Internet of Things (IIoT) technologies enables automated, real-time collection of manufacturing process data, unlocking new opportunities for data-driven product development. Current data-driven methods are generally applied within specific domains, such as design or manufacturing, with limited exploration of integrating design features and manufacturing process data. Since design decisions significantly affect manufacturing outcomes, such as error rates, energy consumption, and processing times, the lack of such integration restricts the potential for data-driven product design improvements. This paper presents a data-driven approach to mapping and analyzing the relationship between design features and manufacturing process data. A comprehensive system architecture is developed to ensure continuous data collection and integration. The linkage between design features and manufacturing process data serves as the basis for developing a machine learning model that enables automated design improvement suggestions. By integrating manufacturing process data with sustainability metrics, this approach opens new possibilities for sustainable product development.

Method: Proposes CrossAgent with a training pipeline combining cold-start supervised fine-tuning and Multi-Turn Group Relative Policy Optimization (GRPO). This enables the agent to learn adaptive action switching between heterogeneous action spaces without human-specified rules.

Result: Extensive experiments on over 800 tasks in open-world Minecraft show CrossAgent achieves state-of-the-art performance, significantly outperforming fixed-action baselines with superior generalization and efficiency in long-horizon reasoning.

Conclusion: CrossAgent demonstrates that unified agents mastering heterogeneous action spaces with adaptive interface selection can overcome the limitations of fixed-action approaches, enabling more flexible and efficient AI agents for dynamic environments.

Abstract: The paradigm of agentic AI is shifting from engineered complex workflows to post-training native models. However, existing agents are typically confined to static, predefined action spaces–such as exclusively using APIs, GUI events, or robotic commands. This rigidity limits their adaptability in dynamic environments where the optimal granularity of interaction varies contextually. To bridge this gap, we propose CrossAgent, a unified agentic model that masters heterogeneous action spaces and autonomously selects the most effective interface for each step of a trajectory. We introduce a comprehensive training pipeline that integrates cold-start supervised fine-tuning with a Multi-Turn Group Relative Policy Optimization (GRPO) algorithm. This approach enables the agent to learn adaptive action switching–balancing high-level efficiency with low-level precision–without human-specified rules. Extensive experiments on over 800 tasks in the open-world Minecraft environment demonstrate that CrossAgent achieves state-of-the-art performance. By dynamically leveraging the strengths of diverse action spaces, our model significantly outperforms fixed-action baselines, exhibiting superior generalization and efficiency in long-horizon reasoning. All code and models are available at https://github.com/CraftJarvis/OpenHA

Method: Proposes MoLKV where each expert is structured as a key-value pair; input-derived queries interact with cached key-value experts from the current sequence to generate context-aware expert outputs.

Result: MoLKV achieves significantly lower validation loss in small-scale evaluations compared to MoLE, demonstrating improved performance through context-aware mechanisms.

Conclusion: MoLKV effectively addresses MoLE’s limitations by introducing context-aware expert selection while maintaining suitability for resource-constrained devices.

Abstract: Recent research has developed several LLM architectures suitable for inference on end-user devices, such as the Mixture of Lookup Experts (MoLE)~\parencite{jie_mixture_2025}. A key feature of MoLE is that each token id is associated with a dedicated group of experts. For a given input, only the experts corresponding to the input token id will be activated. Since the communication overhead of loading this small number of activated experts into RAM during inference is negligible, expert parameters can be offloaded to storage, making MoLE suitable for resource-constrained devices. However, MoLE’s context-independent expert selection mechanism, based solely on input ids, may limit model performance. To address this, we propose the \textbf{M}ixture \textbf{o}f \textbf{L}ookup \textbf{K}ey-\textbf{V}alue Experts (\textbf{MoLKV}) model. In MoLKV, each expert is structured as a key-value pair. For a given input, the input-derived query interacts with the cached key-value experts from the current sequence, generating a context-aware expert output. This context-aware mechanism alleviates the limitation of MoLE, and experimental results demonstrate that MoLKV achieves significantly lower validation loss in small-scale evaluations.

Method: The researchers conduct mechanistic analysis of autoregressive transformers trained on drug-like small molecules. They use sparse autoencoders (SAEs) to extract feature dictionaries associated with chemically relevant activation patterns, examining computational patterns across multiple abstraction levels.

Result: The analysis identifies computational patterns consistent with both low-level syntactic parsing and more abstract chemical validity constraints. The extracted feature dictionaries reveal chemically meaningful activation patterns, and these mechanistic insights translate to improved predictive performance in downstream tasks.

Conclusion: Mechanistic analysis of chemical transformers reveals structured computational patterns for molecular validity, and these insights have practical value for improving predictive performance in various chemical applications.

Abstract: Transformers generate valid and diverse chemical structures, but little is known about the mechanisms that enable these models to capture the rules of molecular representation. We present a mechanistic analysis of autoregressive transformers trained on drug-like small molecules to reveal the computational structure underlying their capabilities across multiple levels of abstraction. We identify computational patterns consistent with low-level syntactic parsing and more abstract chemical validity constraints. Using sparse autoencoders (SAEs), we extract feature dictionaries associated with chemically relevant activation patterns. We validate our findings on downstream tasks and find that mechanistic insights can translate to predictive performance in various practical settings.

Method: The method embeds detailed three-phase grid information (phase voltages, unbalanced loads, BESS states) into heterogeneous graph nodes and uses diverse GNN architectures (GCN, GAT, GraphSAGE, GPS) to predict network state variables. A physics-informed loss function incorporates critical battery constraints (SoC and C-rate limits) via soft penalties during training.

Result: Experimental validation on the CIGRE 18-bus distribution system shows low prediction errors: bus voltage MSEs of 6.92e-07 (GCN), 1.21e-06 (GAT), 3.29e-05 (GPS), and 9.04e-07 (SAGE). The physics-informed method ensures nearly zero SoC and C-rate constraint violations.

Conclusion: The proposed embedding-loss approach achieves high accuracy in predicting network states while ensuring constraint compliance, confirming its effectiveness for reliable, constraint-compliant BESS dispatch in three-phase unbalanced distribution grids.

Abstract: Battery energy storage systems (BESS) have become increasingly vital in three-phase unbalanced distribution grids for maintaining voltage stability and enabling optimal dispatch. However, existing deep learning approaches often lack explicit three-phase representation, making it difficult to accurately model phase-specific dynamics and enforce operational constraints–leading to infeasible dispatch solutions. This paper demonstrates that by embedding detailed three-phase grid information–including phase voltages, unbalanced loads, and BESS states–into heterogeneous graph nodes, diverse GNN architectures (GCN, GAT, GraphSAGE, GPS) can jointly predict network state variables with high accuracy. Moreover, a physics-informed loss function incorporates critical battery constraints–SoC and C-rate limits–via soft penalties during training. Experimental validation on the CIGRE 18-bus distribution system shows that this embedding-loss approach achieves low prediction errors, with bus voltage MSEs of 6.92e-07 (GCN), 1.21e-06 (GAT), 3.29e-05 (GPS), and 9.04e-07 (SAGE). Importantly, the physics-informed method ensures nearly zero SoC and C-rate constraint violations, confirming its effectiveness for reliable, constraint-compliant dispatch.

Method: Built a dataset of 8,049 polymer-solvent pairs from molecular dynamics simulations. Created 2,394 features from molecular descriptors and fingerprints. Trained a 6-layer fully connected neural network with Adam optimizer and MSE loss.

Result: Model achieved strong agreement between predicted and actual solubility values. Demonstrated generalizability on 25 unseen experimental polymer-solvent combinations from Materials Genome Project with high accuracy.

Conclusion: SMILES-based machine learning models are viable for scalable solubility prediction and high-throughput solvent screening, supporting applications in green chemistry, polymer processing, and materials design.

Abstract: Understanding and predicting polymer solubility in various solvents is critical for applications ranging from recycling to pharmaceutical formulation. This work presents a deep learning framework that predicts polymer solubility, expressed as weight percent (wt%), directly from SMILES representations of both polymers and solvents. A dataset of 8,049 polymer solvent pairs at 25 deg C was constructed from calibrated molecular dynamics simulations (Zhou et al., 2023), and molecular descriptors and fingerprints were combined into a 2,394 feature representation per sample. A fully connected neural network with six hidden layers was trained using the Adam optimizer and evaluated using mean squared error loss, achieving strong agreement between predicted and actual solubility values. Generalizability was demonstrated using experimentally measured data from the Materials Genome Project, where the model maintained high accuracy on 25 unseen polymer solvent combinations. These findings highlight the viability of SMILES based machine learning models for scalable solubility prediction and high-throughput solvent screening, supporting applications in green chemistry, polymer processing, and materials design.

Method: Factorizes randomly initialized Transformer weights via SVD into factor units, uses dynamic soft gating with task/morphology embeddings to separate knowledge into shared learngenes and specific tailors, enabling selective activation.

Result: Achieves state-of-the-art performance with 3× improvement in sample efficiency for cross-task transfer and 17× reduction in model size for single-agent deployment.

Conclusion: DivMorph enables scalable, efficient policy deployment with effective knowledge disentanglement and transfer to novel tasks through modular decomposition.

Abstract: Universal morphology control aims to learn a universal policy that generalizes across heterogeneous agent morphologies, with Transformer-based controllers emerging as a popular choice. However, such architectures incur substantial computational costs, resulting in high deployment overhead, and existing methods exhibit limited cross-task generalization, necessitating training from scratch for each new task. To this end, we propose \textbf{DivMorph}, a modular training paradigm that leverages knowledge diversion to learn decomposable controllers. DivMorph factorizes randomly initialized Transformer weights into factor units via SVD prior to training and employs dynamic soft gating to modulate these units based on task and morphology embeddings, separating them into shared \textit{learngenes} and morphology- and task-specific \textit{tailors}, thereby achieving knowledge disentanglement. By selectively activating relevant components, DivMorph enables scalable and efficient policy deployment while supporting effective policy transfer to novel tasks. Extensive experiments demonstrate that DivMorph achieves state-of-the-art performance, achieving a 3$\times$ improvement in sample efficiency over direct finetuning for cross-task transfer and a 17$\times$ reduction in model size for single-agent deployment.

Method: Developed Ariel-ML, a novel toolkit combining a generic TinyML pipeline with an embedded Rust software platform that can leverage multi-core capabilities across various 32-bit microcontroller families (Arm Cortex-M, RISC-V, ESP-32). The implementation is fully open source.

Result: Ariel-ML outperforms prior art in inference latency as expected, and achieves comparable memory footprints to pre-existing C/C++ toolkits. Benchmarks were conducted using a zoo of various TinyML models.

Conclusion: Ariel-ML fills an important gap by providing Rust developers with automated parallelization for TinyML inference on multi-core MCUs, offering both performance benefits and memory efficiency, making it a valuable tool for TinyML practitioners and embedded Rust developers.

Abstract: Low-power microcontroller (MCU) hardware is currently evolving from single-core architectures to predominantly multi-core architectures. In parallel, new embedded software building blocks are more and more written in Rust, while C/C++ dominance fades in this domain. On the other hand, small artificial neural networks (ANN) of various kinds are increasingly deployed in edge AI use cases, thus deployed and executed directly on low-power MCUs. In this context, both incremental improvements and novel innovative services will have to be continuously retrofitted using ANNs execution in software embedded on sensing/actuating systems already deployed in the field. However, there was so far no Rust embedded software platform automating parallelization for inference computation on multi-core MCUs executing arbitrary TinyML models. This paper thus fills this gap by introducing Ariel-ML, a novel toolkit we designed combining a generic TinyML pipeline and an embedded Rust software platform which can take full advantage of multi-core capabilities of various 32bit microcontroller families (Arm Cortex-M, RISC-V, ESP-32). We published the full open source code of its implementation, which we used to benchmark its capabilities using a zoo of various TinyML models. We show that Ariel-ML outperforms prior art in terms of inference latency as expected, and we show that, compared to pre-existing toolkits using embedded C/C++, Ariel-ML achieves comparable memory footprints. Ariel-ML thus provides a useful basis for TinyML practitioners and resource-constrained embedded Rust developers.

Method: Proposed Triple Convolutional Neural Network (TCNN) model and two Attention-based TCNN (ATCNN) models for text matching in IR-based QA systems

Result: Experimental results demonstrate the effectiveness of the proposed models

Conclusion: The proposed TCNN and ATCNN models improve IR-based e-commerce QA systems by enhancing semantic matching capabilities for knowledge retrieval and reranking

Abstract: Automatic question-answering (QA) systems have boomed during last few years, and commonly used techniques can be roughly categorized into Information Retrieval (IR)-based and generation-based. A key solution to the IR based models is to retrieve the most similar knowledge entries of a given query from a QA knowledge base, and then rerank those knowledge entries with semantic matching models. In this paper, we aim to improve an IR based e-commerce QA system-AliMe with proposed text matching models, including a basic Triple Convolutional Neural Network (TCNN) model and two Attention-based TCNN (ATCNN) models. Experimental results show their effect.

Method: Introduces fair k-nearest neighbor (kNN) and fair epsilon-neighborhood graph construction approaches that incorporate fairness constraints during neighborhood selection, ensuring proportional representation of sensitive features in local graph structure while maintaining geometric consistency.

Result: Thorough experiments on three synthetic datasets, seven real-world tabular datasets, and three real-world image datasets demonstrate that the fair graph construction methods surpass current baselines in graph clustering tasks.

Conclusion: Topological fairness in graph construction is essential for achieving equitable clustering outcomes, and fair graph construction inherently facilitates fairer spectral clustering results without needing to modify the clustering algorithm itself.

Abstract: Graph clustering plays a pivotal role in unsupervised learning methods like spectral clustering, yet traditional methods for graph clustering often perpetuate bias through unfair graph constructions that may underrepresent some groups. The current research introduces novel approaches for constructing fair k-nearest neighbor (kNN) and fair epsilon-neighborhood graphs that proactively enforce demographic parity during graph formation. By incorporating fairness constraints at the earliest stage of neighborhood selection steps, our approaches incorporate proportional representation of sensitive features into the local graph structure while maintaining geometric consistency.Our work addresses a critical gap in pre-processing for fair spectral clustering, demonstrating that topological fairness in graph construction is essential for achieving equitable clustering outcomes. Widely used graph construction methods like kNN and epsilon-neighborhood graphs propagate edge based disparate impact on sensitive groups, leading to biased clustering results. Providing representation of each sensitive group in the neighborhood of every node leads to fairer spectral clustering results because the topological features of the graph naturally reflect equitable group ratios. This research fills an essential shortcoming in fair unsupervised learning, by illustrating how topological fairness in graph construction inherently facilitates fairer spectral clustering results without the need for changes to the clustering algorithm itself. Thorough experiments on three synthetic datasets, seven real-world tabular datasets, and three real-world image datasets prove that our fair graph construction methods surpass the current baselines in graph clustering tasks.

Method: Built machine learning models to predict wildfire containment days as regression task using three publicly available datasets from California FRAP. Compared baseline ensemble regressor, Random Forest, XGBoost, and LSTM neural network performance.

Result: XGBoost slightly outperformed Random Forest due to better handling of static features. LSTM performed worse than ensemble models because dataset lacked temporal features. Models can provide detailed, precise forecasts rather than broader categorical predictions.

Conclusion: Wildfire managers can select appropriate model (XGBoost for static features) to accurately predict containment duration and allocate resources effectively, depending on feature availability. Regression approach provides more detailed forecasts than categorical predictions.

Abstract: California’s wildfire season keeps getting worse over the years, overwhelming the emergency response teams. These fires cause massive destruction to both property and human life. Because of these reasons, there’s a growing need for accurate and practical predictions that can help assist with resources allocation for the Wildfire managers or the response teams. In this research, we built machine learning models to predict the number of days it will require to fully contain a wildfire in California. Here, we addressed an important gap in the current literature. Most prior research has concentrated on wildfire risk or how fires spread, and the few that examine the duration typically predict it in broader categories rather than a continuous measure. This research treats the wildfire duration prediction as a regression task, which allows for more detailed and precise forecasts rather than just the broader categorical predictions used in prior work. We built the models by combining three publicly available datasets from California Department of Forestry and Fire Protection’s Fire and Resource Assessment Program (FRAP). This study compared the performance of baseline ensemble regressor, Random Forest and XGBoost, with a Long Short-Term Memory (LSTM) neural network. The results show that the XGBoost model slightly outperforms the Random Forest model, likely due to its superior handling of static features in the dataset. The LSTM model, on the other hand, performed worse than the ensemble models because the dataset lacked temporal features. Overall, this study shows that, depending on the feature availability, Wildfire managers or Fire management authorities can select the most appropriate model to accurately predict wildfire containment duration and allocate resources effectively.

Method: Integrates Conformal Prediction (CP) into bandit framework to provide finite-time prediction coverage guarantees. Combines decision-making bandit policies with statistical guarantees. For portfolio allocation application, incorporates hidden Markov models to capture regime-switching behavior of financial markets.

Result: Demonstrates practical advantages in small-gap settings: achieves better regret performance than classical UCB policies, provides nominal coverage guarantees where classical methods fail, and shows higher risk-adjusted returns in portfolio allocation while preserving coverage.

Conclusion: Conformal Bandits successfully bridges regret minimization with statistical guarantees, offering a practical framework that outperforms classical bandit algorithms in small-gap regimes and provides valuable finite-time coverage guarantees for applications like portfolio allocation.

Abstract: We introduce Conformal Bandits, a novel framework integrating Conformal Prediction (CP) into bandit problems, a classic paradigm for sequential decision-making under uncertainty. Traditional regret-minimisation bandit strategies like Thompson Sampling and Upper Confidence Bound (UCB) typically rely on distributional assumptions or asymptotic guarantees; further, they remain largely focused on regret, neglecting their statistical properties. We address this gap. Through the adoption of CP, we bridge the regret-minimising potential of a decision-making bandit policy with statistical guarantees in the form of finite-time prediction coverage. We demonstrate the potential of it Conformal Bandits through simulation studies and an application to portfolio allocation, a typical small-gap regime, where differences in arm rewards are far too small for classical policies to achieve optimal regret bounds in finite sample. Motivated by this, we showcase our framework’s practical advantage in terms of regret in small-gap settings, as well as its added value in achieving nominal coverage guarantees where classical UCB policies fail. Focusing on our application of interest, we further illustrate how integrating hidden Markov models to capture the regime-switching behaviour of financial markets, enhances the exploration-exploitation trade-off, and translates into higher risk-adjusted regret efficiency returns, while preserving coverage guarantees.

Method: HPM-KD integrates six components: (1) Adaptive Configuration Manager via meta-learning for automatic hyperparameter tuning, (2) Progressive Distillation Chain with automatically determined intermediate models, (3) Attention-Weighted Multi-Teacher Ensemble with dynamic per-sample weights, (4) Meta-Learned Temperature Scheduler, (5) Parallel Processing Pipeline with intelligent load balancing, and (6) Shared Optimization Memory for cross-experiment reuse.

Result: Experiments on CIFAR-10, CIFAR-100, and tabular datasets show HPM-KD achieves 10x-15x compression while maintaining 85% accuracy retention, eliminates manual tuning, reduces training time by 30-40% via parallelization, and ablation studies confirm independent contributions of each component (0.10-0.98 pp improvements).

Conclusion: HPM-KD provides a comprehensive solution to major knowledge distillation limitations, offering automated tuning, efficient multi-teacher coordination, and computational optimization while maintaining high compression ratios and accuracy. The framework is available as part of the open-source DeepBridge library.

Abstract: Knowledge Distillation (KD) has emerged as a promising technique for model compression but faces critical limitations: (1) sensitivity to hyperparameters requiring extensive manual tuning, (2) capacity gap when distilling from very large teachers to small students, (3) suboptimal coordination in multi-teacher scenarios, and (4) inefficient use of computational resources. We present \textbf{HPM-KD}, a framework that integrates six synergistic components: (i) Adaptive Configuration Manager via meta-learning that eliminates manual hyperparameter tuning, (ii) Progressive Distillation Chain with automatically determined intermediate models, (iii) Attention-Weighted Multi-Teacher Ensemble that learns dynamic per-sample weights, (iv) Meta-Learned Temperature Scheduler that adapts temperature throughout training, (v) Parallel Processing Pipeline with intelligent load balancing, and (vi) Shared Optimization Memory for cross-experiment reuse. Experiments on CIFAR-10, CIFAR-100, and tabular datasets demonstrate that HPM-KD: achieves 10x-15x compression while maintaining 85% accuracy retention, eliminates the need for manual tuning, and reduces training time by 30-40% via parallelization. Ablation studies confirm independent contribution of each component (0.10-0.98 pp). HPM-KD is available as part of the open-source DeepBridge library.

Method: Revisiting layer-selection strategies in KD, testing various matching approaches including reverse matching, and analyzing teacher layers from the student’s perspective using angle measurements between layers.

Result: Layer-selection strategy doesn’t matter much - even nonsensical strategies like reverse matching yield surprisingly good student performance. Analysis shows teacher layers have similar angles from student’s perspective.

Conclusion: Layer-selection strategies may not be the main focus in KD system design, and vanilla forward matching works well in most setups, simplifying KD practice.

Abstract: Knowledge distillation (KD) is a popular method of transferring knowledge from a large “teacher” model to a small “student” model. Previous work has explored various layer-selection strategies (e.g., forward matching and in-order random matching) for intermediate-layer matching in KD, where a student layer is forced to resemble a certain teacher layer. In this work, we revisit such layer-selection strategies and observe an intriguing phenomenon that layer-selection strategy does not matter (much) in intermediate-layer matching – even seemingly nonsensical matching strategies such as reverse matching still result in surprisingly good student performance. We provide an interpretation for this phenomenon by examining the angles between teacher layers viewed from the student’s perspective. Our work sheds light on KD practice, as layer-selection strategies may not be the main focus of KD system design, and vanilla forward matching works well in most setups.

Method: Analyzes axiomatic properties of both Dirichlet energy definitions, demonstrates that the normalized graph Laplacian version fails Rusch et al.’s node-similarity measure axioms, formalizes fundamental spectral properties of both definitions, and establishes criteria for selecting the appropriate metric based on GNN architecture.

Result: Shows that the Dirichlet energy from normalized graph Laplacian fails to satisfy node-similarity measure axioms, identifies critical spectral distinctions between the two definitions, and provides guidance for selecting the spectrally compatible metric for monitoring GNN dynamics.

Conclusion: The choice between unnormalized and normalized graph Laplacian Dirichlet energies matters significantly for monitoring GNN over-smoothing; the normalized version is not a valid node-similarity measure, and spectral compatibility with GNN architecture must guide metric selection to properly monitor dynamics.

Abstract: We analyze the distinctions between two functionals often used as over-smoothing measures: the Dirichlet energies induced by the unnormalized graph Laplacian and the normalized graph Laplacian. We demonstrate that the latter fails to satisfy the axiomatic definition of a node-similarity measure proposed by Rusch \textit{et al.} By formalizing fundamental spectral properties of these two definitions, we highlight critical distinctions necessary to select the metric that is spectrally compatible with the GNN architecture, thereby resolving ambiguities in monitoring the dynamics.

Method: The authors study a query learning model with logit queries (reflecting API access) and develop an efficient algorithm for learning any approximately low logit rank model from queries.

Result: The main result is an efficient algorithm that can learn any approximately low logit rank model from queries, providing provable learning guarantees.

Conclusion: This work gives the first end-to-end learning guarantee for a generative model that plausibly captures modern language models, as the low logit rank assumption closely matches empirical observations of LLM behavior.

Abstract: While modern language models and their inner workings are incredibly complex, recent work (Golowich, Liu & Shetty; 2025) has proposed a simple and potentially tractable abstraction for them through the observation that empirically, these language models all seem to have approximately low logit rank. Roughly, this means that a matrix formed by the model’s log probabilities of various tokens conditioned on certain sequences of tokens is well approximated by a low rank matrix. In this paper, our focus is on understanding how this structure can be exploited algorithmically for obtaining provable learning guarantees. Since low logit rank models can encode hard-to-learn distributions such as noisy parities, we study a query learning model with logit queries that reflects the access model for common APIs. Our main result is an efficient algorithm for learning any approximately low logit rank model from queries. We emphasize that our structural assumption closely reflects the behavior that is empirically observed in modern language models. Thus, our result gives what we believe is the first end-to-end learning guarantee for a generative model that plausibly captures modern language models.

Method: Systematically evaluate three PLMs (AntiBERTa, BioBERT, ESM2) against GPT-2 baseline on antibody target specificity prediction tasks, using attention attribution analysis to examine feature capture.

Result: All PLMs achieve high classification accuracy but show distinct biases in capturing biological features (V gene usage, somatic hypermutation patterns, isotype). Antibody-specific models naturally learn CDR focus, while general models benefit from explicit CDR-focused training.

Conclusion: Model architecture significantly influences biological feature extraction in antibodies, providing guidance for future PLM development in computational antibody design.

Abstract: Recent advances in protein language models (PLMs) have demonstrated remarkable capabilities in understanding protein sequences. However, the extent to which different model architectures capture antibody-specific biological properties remains unexplored. In this work, we systematically investigate how architectural choices in PLMs influence their ability to comprehend antibody sequence characteristics and functions. We evaluate three state-of-the-art PLMs-AntiBERTa, BioBERT, and ESM2–against a general-purpose language model (GPT-2) baseline on antibody target specificity prediction tasks. Our results demonstrate that while all PLMs achieve high classification accuracy, they exhibit distinct biases in capturing biological features such as V gene usage, somatic hypermutation patterns, and isotype information. Through attention attribution analysis, we show that antibody-specific models like AntiBERTa naturally learn to focus on complementarity-determining regions (CDRs), while general protein models benefit significantly from explicit CDR-focused training strategies. These findings provide insights into the relationship between model architecture and biological feature extraction, offering valuable guidance for future PLM development in computational antibody design.

Method: STACHE uses an exact, search-based algorithm that exploits factored state space structure to generate composite explanations: (1) robustness regions (connected neighborhoods where action stays the same) and (2) minimal counterfactuals (smallest perturbations that change the decision).

Result: Empirical validation on Gymnasium environments shows the framework explains policy actions and captures evolution of policy logic during training, revealing transitions from erratic to optimized strategies while providing insights into agent sensitivity and decision boundaries.

Conclusion: STACHE provides a comprehensive framework for generating faithful local explanations of RL agent decisions, offering actionable insights for debugging and verification without the fidelity gaps of surrogate models.

Abstract: Reinforcement learning agents often behave unexpectedly in sparse-reward or safety-critical environments, creating a strong need for reliable debugging and verification tools. In this paper, we propose STACHE, a comprehensive framework for generating local, black-box explanations for an agent’s specific action within discrete Markov games. Our method produces a Composite Explanation consisting of two complementary components: (1) a Robustness Region, the connected neighborhood of states where the agent’s action remains invariant, and (2) Minimal Counterfactuals, the smallest state perturbations required to alter that decision. By exploiting the structure of factored state spaces, we introduce an exact, search-based algorithm that circumvents the fidelity gaps of surrogate models. Empirical validation on Gymnasium environments demonstrates that our framework not only explains policy actions, but also effectively captures the evolution of policy logic during training - from erratic, unstable behavior to optimized, robust strategies - providing actionable insights into agent sensitivity and decision boundaries.

Method: FALCON introduces a hybrid training objective that encourages invertibility in continuous flows, enabling few-step sampling with likelihoods accurate enough for importance sampling applications.

Result: FALCON outperforms state-of-the-art normalizing flow models for molecular Boltzmann sampling and is two orders of magnitude (100x) faster than equivalently performing CNF models.

Conclusion: FALCON provides an efficient solution to the computational bottleneck of Boltzmann Generators, enabling practical and scalable molecular state sampling with accurate likelihoods in few steps.

Abstract: Scalable sampling of molecular states in thermodynamic equilibrium is a long-standing challenge in statistical physics. Boltzmann Generators tackle this problem by pairing a generative model, capable of exact likelihood computation, with importance sampling to obtain consistent samples under the target distribution. Current Boltzmann Generators primarily use continuous normalizing flows (CNFs) trained with flow matching for efficient training of powerful models. However, likelihood calculation for these models is extremely costly, requiring thousands of function evaluations per sample, severely limiting their adoption. In this work, we propose Few-step Accurate Likelihoods for Continuous Flows (FALCON), a method which allows for few-step sampling with a likelihood accurate enough for importance sampling applications by introducing a hybrid training objective that encourages invertibility. We show FALCON outperforms state-of-the-art normalizing flow models for molecular Boltzmann sampling and is two orders of magnitude faster than the equivalently performing CNF model.

Method: Proposed train-time data synthesis techniques to close the train-test gap. Improved methods for training world models specifically to enable efficient gradient-based planning.

Result: Outperforms or matches classical gradient-free cross-entropy method (CEM) across various object manipulation and navigation tasks using only 10% of the time budget.

Conclusion: Closing the train-test gap through improved world model training enables gradient-based planning to achieve state-of-the-art performance with significantly reduced computational requirements.

Abstract: World models paired with model predictive control (MPC) can be trained offline on large-scale datasets of expert trajectories and enable generalization to a wide range of planning tasks at inference time. Compared to traditional MPC procedures, which rely on slow search algorithms or on iteratively solving optimization problems exactly, gradient-based planning offers a computationally efficient alternative. However, the performance of gradient-based planning has thus far lagged behind that of other approaches. In this paper, we propose improved methods for training world models that enable efficient gradient-based planning. We begin with the observation that although a world model is trained on a next-state prediction objective, it is used at test-time to instead estimate a sequence of actions. The goal of our work is to close this train-test gap. To that end, we propose train-time data synthesis techniques that enable significantly improved gradient-based planning with existing world models. At test time, our approach outperforms or matches the classical gradient-free cross-entropy method (CEM) across a variety of object manipulation and navigation tasks in 10% of the time budget.

Method: Combines self-supervised pre-training on unlabeled data with differential hopping windowing technique to learn from limited labeled instances, plus an opportunistic model activation module that selectively activates deep learning only during active periods.

Result: Achieves 7.25% increase in precision and 4.4% improvement in accuracy compared to supervised models while using only 40% of labeled training data; model activation reduces inference time by up to 67%.

Conclusion: LIFT-PD enables practical, energy-efficient, and unobtrusive in-home monitoring of PD patients with minimal labeling requirements, paving the way for real-world deployment.

Abstract: Parkinson’s disease (PD) is a progressive neurological disorder that impacts the quality of life significantly, making in-home monitoring of motor symptoms such as Freezing of Gait (FoG) critical. However, existing symptom monitoring technologies are power-hungry, rely on extensive amounts of labeled data, and operate in controlled settings. These shortcomings limit real-world deployment of the technology. This work presents LIFT-PD, a computationally-efficient self-supervised learning framework for real-time FoG detection. Our method combines self-supervised pre-training on unlabeled data with a novel differential hopping windowing technique to learn from limited labeled instances. An opportunistic model activation module further minimizes power consumption by selectively activating the deep learning module only during active periods. Extensive experimental results show that LIFT-PD achieves a 7.25% increase in precision and 4.4% improvement in accuracy compared to supervised models while using as low as 40% of the labeled training data used for supervised learning. Additionally, the model activation module reduces inference time by up to 67% compared to continuous inference. LIFT-PD paves the way for practical, energy-efficient, and unobtrusive in-home monitoring of PD patients with minimal labeling requirements.

Method: Proposes Sinusoidal initialization, a deterministic method that uses sinusoidal functions to construct structured weight matrices. This method is designed to improve weight distribution across network layers and create more uniform neuron activation states from the initial forward pass.

Result: Experiments show Sinusoidal initialization delivers consistently faster convergence (20.9% improvement), greater training stability, and higher final accuracy (4.9% average increase) across various models including CNNs, vision transformers, and large language models.

Conclusion: By replacing randomness with structured sinusoidal functions, this initialization provides a stronger, more reliable foundation for deep learning systems, offering improved convergence speed, stability, and accuracy compared to traditional random initialization methods.

Abstract: Initialization plays a critical role in Deep Neural Network training, directly influencing convergence, stability, and generalization. Common approaches such as Glorot and He initializations rely on randomness, which can produce uneven weight distributions across layer connections. In this paper, we introduce the Sinusoidal initialization, a novel deterministic method that employs sinusoidal functions to construct structured weight matrices expressly to improve the spread and balance of weights throughout the network while simultaneously fostering a more uniform, well-conditioned distribution of neuron activation states from the very first forward pass. Because Sinusoidal initialization begins with weights and activations that are already evenly and efficiently utilized, it delivers consistently faster convergence, greater training stability, and higher final accuracy across a wide range of models, including convolutional neural networks, vision transformers, and large language models. On average, our experiments show an increase of 4.9% in final validation accuracy and 20.9% in convergence speed. By replacing randomness with structure, this initialization provides a stronger and more reliable foundation for Deep Learning systems.

Method: Novel regularization technique based on adaptive self-distillation (ASD) that adjusts to each client’s training data using global model prediction entropy and client-data label distribution.

Result: ASD can be integrated with existing FL algorithms, substantially boosting performance on real-world benchmarks, reducing client-drift, and improving generalization.

Conclusion: ASD regularization effectively addresses client-drift in federated learning under non-iid conditions, enhancing existing FL methods with theoretical and empirical validation.

Abstract: Federated Learning (FL) is a machine learning paradigm that enables clients to jointly train a global model by aggregating the locally trained models without sharing any local training data. In practice, there can often be substantial heterogeneity (e.g., class imbalance) across the local data distributions observed by each of these clients. Under such non-iid label distributions across clients, FL suffers from the ‘client-drift’ problem where every client drifts to its own local optimum. This results in slower convergence and poor performance of the aggregated model. To address this limitation, we propose a novel regularization technique based on adaptive self-distillation (ASD) for training models on the client side. Our regularization scheme adaptively adjusts to each client’s training data based on the global model’s prediction entropy and the client-data label distribution. We show in this paper that our proposed regularization (ASD) can be easily integrated atop existing, state-of-the-art FL algorithms, leading to a further boost in the performance of these off-the-shelf methods. We theoretically explain how incorporation of ASD regularizer leads to reduction in client-drift and empirically justify the generalization ability of the trained model. We demonstrate the efficacy of our approach through extensive experiments on multiple real-world benchmarks and show substantial gains in performance when the proposed regularizer is combined with popular FL methods.

Method: Developed a principled active learning approach for correlation clustering by introducing several information-theoretic acquisition functions that prioritize queries based on entropy and expected information gain to reduce uncertainty about the clustering structure efficiently.

Result: The methods significantly outperform existing baselines across a range of synthetic and real-world settings in terms of clustering accuracy and query efficiency.

Conclusion: Combining active learning with correlation clustering provides significant benefits in settings where similarity information is costly or limited, enabling effective clustering under budget constraints.

Abstract: Correlation clustering is a flexible framework for partitioning data based solely on pairwise similarity or dissimilarity information, without requiring the number of clusters as input. However, in many practical scenarios, these pairwise similarities are not available a priori and must be obtained through costly measurements or human feedback. This motivates the use of active learning to query only the most informative pairwise comparisons, enabling effective clustering under budget constraints. In this work, we develop a principled active learning approach for correlation clustering by introducing several information-theoretic acquisition functions that prioritize queries based on entropy and expected information gain. These strategies aim to reduce uncertainty about the clustering structure as efficiently as possible. We evaluate our methods across a range of synthetic and real-world settings and show that they significantly outperform existing baselines in terms of clustering accuracy and query efficiency. Our results highlight the benefits of combining active learning with correlation clustering in settings where similarity information is costly or limited.

Method: Developed a poisoning framework to systematically evaluate NAS robustness, tested four different NAS algorithms against four data poisoning attacks (including one tailored specifically for NAS), using CIFAR-10 and CIFAR-100 benchmarks.

Result: NAS appears superficially robust with marginal accuracy drops even under large poisoning budgets, but the expected accuracy improvements from NAS algorithms can be substantially reduced. Training-based NAS methods are least robust, while training-free methods are most robust but produce architectures similar to random selections.

Conclusion: NAS shows concerning vulnerabilities to data poisoning that diminish its value for achieving accuracy gains, with different NAS approaches exhibiting varying robustness levels. The findings highlight the need for countermeasures to protect NAS from data corruption attacks.

Abstract: We study the robustness of data-centric methods to find neural network architectures, known as neural architecture search (NAS), against data poisoning. To audit this robustness, we design a poisoning framework that enables the systematic evaluation of the ability of NAS to produce architectures under data corruption. Our framework examines four off-the-shelf NAS algorithms, representing different approaches to architecture discovery, against four data poisoning attacks, including one we tailor specifically for NAS. In our evaluation with the CIFAR-10 and CIFAR-100 benchmarks, we show that NAS is \emph{seemingly} robust to data poisoning, showing marginal accuracy drops even under large poisoning budgets. However, we demonstrate that when considering NAS algorithms designed to achieve a few percentage points of accuracy gain, this expected improvement can be substantially diminished under data poisoning. We also show that the reduction varies across NAS algorithms and analyze the factors contributing to their robustness. Our findings are: (1) Training-based NAS algorithms are the least robust due to their reliance on data. (2) Training-free NAS approaches are the most robust but produce architectures that perform similarly to random selections from the search space. (3) NAS algorithms can produce architectures with improved accuracy, even when using out-of-distribution data like MNIST. We lastly discuss potential countermeasures. Our code is available at: https://github.com/ztcoalson/NAS-Robustness-to-Data-Poisoning

Method: Develops a Primal-Dual Natural Actor-Critic algorithm that combines primal-dual optimization with natural policy gradient methods, specifically designed to handle constraints while maintaining convergence properties in average reward CMDPs.

Result: Achieves global convergence and constraint violation rates of $\tilde{\mathcal{O}}(1/\sqrt{T})$ when mixing time is known, and $\tilde{\mathcal{O}}(1/T^{0.5-ε})$ when mixing time is unknown (with $T \geq \tilde{\mathcal{O}}\left(τ_{\mathrm{mix}}^{2/ε}\right)$), matching theoretical lower bounds for MDPs.

Conclusion: Establishes a new theoretical benchmark for average reward CMDPs by providing optimal convergence rates that match the theoretical lower bound, advancing the theoretical exploration of constrained reinforcement learning.

Abstract: This paper investigates infinite-horizon average reward Constrained Markov Decision Processes (CMDPs) with general parametrization. We propose a Primal-Dual Natural Actor-Critic algorithm that adeptly manages constraints while ensuring a high convergence rate. In particular, our algorithm achieves global convergence and constraint violation rates of $\tilde{\mathcal{O}}(1/\sqrt{T})$ over a horizon of length $T$ when the mixing time, $τ_{\mathrm{mix}}$, is known to the learner. In absence of knowledge of $τ_{\mathrm{mix}}$, the achievable rates change to $\tilde{\mathcal{O}}(1/T^{0.5-ε})$ provided that $T \geq \tilde{\mathcal{O}}\left(τ_{\mathrm{mix}}^{2/ε}\right)$. Our results match the theoretical lower bound for Markov Decision Processes and establish a new benchmark in the theoretical exploration of average reward CMDPs.

Method: Proposes Entropy-Informed Weighting Channel Normalizing Flow (EIW-Flow) with a regularized, feature-dependent Shuffle operation that generates adaptive channel-wise weights and shuffles latent variables before splitting, guiding variables toward entropy increase.

Result: EIW-Flow achieves state-of-the-art density estimation and competitive sample quality on CIFAR-10, CelebA, ImageNet, and LSUN datasets with minimal computational overhead.

Conclusion: The entropy-informed shuffle operation effectively improves multi-scale normalizing flows by enabling adaptive channel processing, leading to better performance while maintaining efficiency.

Abstract: Normalizing Flows (NFs) are widely used in deep generative models for their exact likelihood estimation and efficient sampling. However, they require substantial memory since the latent space matches the input dimension. Multi-scale architectures address this by progressively reducing latent dimensions while preserving reversibility. Existing multi-scale architectures use simple, static channel-wise splitting, limiting expressiveness. To improve this, we introduce a regularized, feature-dependent $\mathtt{Shuffle}$ operation and integrate it into vanilla multi-scale architecture. This operation adaptively generates channel-wise weights and shuffles latent variables before splitting them. We observe that such operation guides the variables to evolve in the direction of entropy increase, hence we refer to NFs with the $\mathtt{Shuffle}$ operation as \emph{Entropy-Informed Weighting Channel Normalizing Flow} (EIW-Flow). Extensive experiments on CIFAR-10, CelebA, ImageNet, and LSUN demonstrate that EIW-Flow achieves state-of-the-art density estimation and competitive sample quality for deep generative modeling, with minimal computational overhead.

Method: Proposed two metrics: Segment Routing Best Performance (SRP) to evaluate expert coverage for token segments, and Segment Cache Best Hit Rate (SCH) to measure expert cache hit rates with future information. Analyzed 20 diverse MoE LLMs and used toy models to verify key factors.

Result: Found strong trade-off between local routing consistency and local load balance, while global load balance can coexist with consistency. Domain-specialized experts contribute more to routing consistency than vocabulary-specialized ones. Most models balance cache effectiveness and efficiency with cache sizes ~2x active experts.

Conclusion: The findings provide insights for memory-efficient MoE design and deployment without compromising inference speed, with published code for replication.

Abstract: Mixture-of-Experts (MoE) enables efficient scaling of large language models (LLMs) with sparsely activated experts during inference. To effectively deploy large MoE models on memory-constrained devices, many systems introduce expert offloading that caches a subset of experts in fast memory, leaving others on slow memory to run on CPU or load on demand. While some research has exploited the locality of expert activations, where consecutive tokens activate similar experts, the degree of this local routing consistency varies across models and remains understudied. In this paper, we propose two metrics to measure local routing consistency of MoE models: (1) Segment Routing Best Performance (SRP), which evaluates how well a fixed group of experts can cover the needs of a segment of tokens, and (2) Segment Cache Best Hit Rate (SCH), which measures the hit rate of an expert cache utilizing a length of future information under a cache limit. We analyze 20 MoE LLMs with diverse sizes and architectures and use toy models to verify key factors related to local routing consistency. We find a strong trade-off between local routing consistency and local load balance, while showing that global load balance can coexist with local routing consistency. Meanwhile, settings like shared experts that decrease expert combination space can lead to low local routing consistency. We further reveal that domain-specialized experts contribute more to routing consistency than vocabulary-specialized ones, and that most models balance between cache effectiveness and efficiency with cache sizes approximately twice the active experts. These findings pave the way for memory-efficient MoE design and deployment without compromising inference speed. We publish the code for replicating experiments at https://github.com/ljcleo/moe-lrc .

Method: Introduces a novel frequency response perspective on DM inference, using Gaussianity assumption to derive a closed-form spectral transfer function that captures how generated signals evolve from initial noise.

Result: The analysis enables design of noise schedules aligned with data characteristics, provides insights into underlying dynamics, and reveals relationships between spectral properties and noise schedule structure.

Conclusion: The spectral perspective offers theoretical justification for practitioner heuristics and leads to data-dependent scheduling curves, advancing the theoretical foundation of diffusion models.

Abstract: Diffusion models (DMs) have emerged as powerful tools for modeling complex data distributions and generating realistic new samples. Over the years, advanced architectures and sampling methods have been developed to make these models practically usable. However, certain synthesis process decisions still rely on heuristics without a solid theoretical foundation. In our work, we offer a novel analysis of the DM’s inference process, introducing a comprehensive frequency response perspective. Specifically, by relying on Gaussianity assumption, we present the inference process as a closed-form spectral transfer function, capturing how the generated signal evolves in response to the initial noise. We demonstrate how the proposed analysis can be leveraged to design a noise schedule that aligns effectively with the characteristics of the data. The spectral perspective also provides insights into the underlying dynamics and sheds light on the relationship between spectral properties and noise schedule structure. Our results lead to scheduling curves that are dependent on the spectral content of the data, offering a theoretical justification for some of the heuristics taken by practitioners.

Method: 1. Show optimal robust Q operator is a strict contraction under a carefully designed semi-norm; 2. Develop stochastic approximation for learning optimal robust Q-function; 3. Provide efficient robust Q-function estimation routine; 4. Introduce actor-critic algorithm leveraging robust critic estimation.

Result: 1. Prove optimal robust Q operator contraction property; 2. Achieve Õ(ε⁻²) sample complexity for learning optimal robust Q-function; 3. Actor-critic algorithm learns ε-optimal robust policy within Õ(ε⁻²) samples; 4. Numerical simulations validate performance.

Conclusion: The paper provides the first non-asymptotic convergence analysis for robust RL algorithms under various uncertainty sets, establishing optimal sample complexity bounds and enabling practical robust policy learning with theoretical guarantees.

Abstract: We present a non-asymptotic convergence analysis of $Q$-learning and actor-critic algorithms for robust average-reward Markov Decision Processes (MDPs) under contamination, total-variation (TV) distance, and Wasserstein uncertainty sets. A key ingredient of our analysis is showing that the optimal robust $Q$ operator is a strict contraction with respect to a carefully designed semi-norm (with constant functions quotiented out). This property enables a stochastic approximation update that learns the optimal robust $Q$-function using $\tilde{\mathcal{O}}(ε^{-2})$ samples. We also provide an efficient routine for robust $Q$-function estimation, which in turn facilitates robust critic estimation. Building on this, we introduce an actor-critic algorithm that learns an $ε$-optimal robust policy within $\tilde{\mathcal{O}}(ε^{-2})$ samples. We provide numerical simulations to evaluate the performance of our algorithms.

Method: MINJA uses bridging steps to link victim queries to malicious reasoning, an indication prompt to guide autonomous generation of similar bridging steps, and a progressive shortening strategy to gradually remove the prompt. Attackers interact with agents via queries and output observations to inject malicious records.

Result: Extensive experiments across diverse agents show MINJA effectively compromises agent memory with minimal execution requirements, demonstrating that any user can influence agent memory through normal interactions.

Conclusion: MINJA reveals significant security risks in LLM agents where memory can be compromised through normal interactions, highlighting the need for robust memory protection mechanisms in agent systems.

Abstract: Agents powered by large language models (LLMs) have demonstrated strong capabilities in a wide range of complex, real-world applications. However, LLM agents with a compromised memory bank may easily produce harmful outputs when the past records retrieved for demonstration are malicious. In this paper, we propose a novel Memory INJection Attack, MINJA, without assuming that the attacker can directly modify the memory bank of the agent. The attacker injects malicious records into the memory bank by only interacting with the agent via queries and output observations. These malicious records are designed to elicit a sequence of malicious reasoning steps corresponding to a different target query during the agent’s execution of the victim user’s query. Specifically, we introduce a sequence of bridging steps to link victim queries to the malicious reasoning steps. During the memory injection, we propose an indication prompt that guides the agent to autonomously generate similar bridging steps, with a progressive shortening strategy that gradually removes the indication prompt, such that the malicious record will be easily retrieved when processing later victim queries. Our extensive experiments across diverse agents demonstrate the effectiveness of MINJA in compromising agent memory. With minimal requirements for execution, MINJA enables any user to influence agent memory, highlighting the risk.

Method: Two simple techniques: (1) column-wise gradient normalization (normalizing gradients along output dimension), and (2) applying first-order momentum only to the output layer where gradient variance is highest. Combined as SCALE (Stochastic Column-normAlized Last-layer momEntum).

Result: SCALE matches or exceeds Adam performance on LLaMA models (60M-1B) using only 35-45% of total memory. Outperforms memory-efficient optimizers like GaLore, Fira, and APOLLO. For LLaMA 7B, beats APOLLO and Muon in both perplexity and memory consumption.

Conclusion: SCALE provides a simple, highly memory- and compute-efficient optimizer that achieves state-of-the-art pretraining performance with minimal modifications to SGD, making it ideal for large-scale pretraining under memory constraints.

Abstract: Training large language models (LLMs) typically relies on adaptive optimizers such as Adam, which introduce extra operations and require significant more memory to maintain first- and second-order moments than SGD. While recent works such as GaLore, Fira and APOLLO have proposed state-compressed variants to reduce memory consumption, a fundamental question remains: What are the minimum modifications to plain SGD needed to match state-of-the-art pretraining performance? We systematically investigate this question using a bottom-up approach, and identify two simple yet highly (memory- and compute-) efficient techniques: (1) column-wise gradient normalization (normalizing the gradient along the output dimension), which boosts SGD performance without momentum; and (2) applying first-order momentum only to the output layer, where gradient variance is highest. Combining these two techniques lead to SCALE (Stochastic Column-normAlized Last-layer momEntum), a simple optimizer for memory efficient pretraining. Across multiple LLaMA models (60M-1B), SCALE matches or exceeds the performance of Adam while using only 35-45% of the total memory. It also consistently outperforms memory-efficient optimizers such as GaLore, Fira and APOLLO, making it a strong candidate for large-scale pretraining under memory constraints. For LLaMA 7B model, SCALE outperforms the state-of-the-art memory-efficient methods APOLLO and Muon, in terms of both perplexity and memory consumption.

Method: Theoretical analysis showing that single-layer linear transformers, after adversarial pretraining across various classification tasks, can robustly generalize to unseen tasks through in-context learning from clean demonstrations only.

Result: Adversarially pretrained transformers can achieve universal robustness by adaptively focusing on robust features within given tasks. Two challenges identified: accuracy-robustness trade-off and sample-hungry training.

Conclusion: While expensive to train, universally robust foundation models are worthwhile investments as they provide free adversarial robustness to downstream tasks, initiating discussion on their utility.

Abstract: Adversarial training is one of the most effective adversarial defenses, but it incurs a high computational cost. In this study, we present the first theoretical analysis suggesting that adversarially pretrained transformers can serve as universally robust foundation models – models that can robustly adapt to diverse downstream tasks with only lightweight tuning. Specifically, we demonstrate that single-layer linear transformers, after adversarial pretraining across a variety of classification tasks, can robustly generalize to unseen classification tasks through in-context learning from clean demonstrations (i.e., without requiring additional adversarial training or examples). This universal robustness stems from the model’s ability to adaptively focus on robust features within given tasks. We also show the two open challenges for attaining robustness: accuracy–robustness trade-off and sample-hungry training. This study initiates the discussion on the utility of universally robust foundation models. While their training is expensive, the investment would prove worthwhile as downstream tasks can enjoy free adversarial robustness. The code is available at https://github.com/s-kumano/universally-robust-in-context-learner.

Method: Transforms time series into graphs using Weighted Dual Perspective Visibility Graph (WDPVG) to capture hidden structural aspects, then applies Graph Attention Network (GAT) for node classification to identify meaningful segments. Formulates TSS as a node classification problem on graphs.

Result: Achieves average F1 score of 0.97 across 59 diverse TSS benchmark datasets, outperforms seq2point baseline by 0.05 F1 score, and reduces required training data compared to baseline methods. Provides first detailed study of GNNs for graph representations of TS in TSS context.

Conclusion: The proposed graph-based approach using WDPVG and GAT effectively addresses limitations of sliding window methods, enabling more granular time series segmentation with superior performance and reduced data requirements.

Abstract: Time series segmentation (TSS) is one of the time series (TS) analysis techniques, that has received considerably less attention compared to other TS related tasks. In recent years, deep learning architectures have been introduced for TSS, however their reliance on sliding windows limits segmentation granularity due to fixed window sizes and strides. To overcome these challenges, we propose a new more granular TSS approach that utilizes the Weighted Dual Perspective Visbility Graph (WDPVG) TS into a graph and combines it with a Graph Attention Network (GAT). By transforming TS into graphs, we are able to capture different structural aspects of the data that would otherwise remain hidden. By utilizing the representation learning capabilities of Graph Neural Networks, our method is able to effectively identify meaningful segments within the TS. To better understand the potential of our approach, we also experimented with different TS-to-graph transformations and compared their performance. Our contributions include: a) formulating the TSS as a node classification problem on graphs; b) conducting an extensive analysis of various TS-to-graph transformations applied to TSS using benchmark datasets from the TSSB repository; c) providing the first detailed study on utilizing GNNs for analyzing graph representations of TS in the context of TSS; d) demonstrating the effectiveness of our method, which achieves an average F1 score of 0.97 across 59 diverse TSS benchmark datasets; e) outperforming the seq2point baseline method by 0.05 in terms of F1 score; and f) reducing the required training data compared to the baseline methods.

Method: PROPS (PROgressively Private Self-alignment) uses multi-stage alignment where privately aligned models from previous stages serve as labelers for subsequent stages, focusing on preference-level privacy rather than full gradient privacy.

Result: PROPS achieves up to 3x higher win-rates than DP-SGD and 2.5x higher win-rates than Randomized Response-based alignment for the same privacy budget, validated on multiple models (Pythia, GPT) and datasets.

Conclusion: PROPS provides a practical solution for LLM alignment with preference-level privacy, offering better utility than existing methods while maintaining strong privacy guarantees for human feedback.

Abstract: Alignment is a key step in developing Large Language Models (LLMs) using human feedback to ensure adherence to human values and societal norms. Dependence on human feedback raises privacy concerns about how much a labeler’s preferences may reveal about their personal values, beliefs, and personality traits. Existing approaches, such as Differentially Private SGD (DP-SGD), provide rigorous privacy guarantees by privatizing gradients during fine-tuning and alignment but can provide more privacy than necessary as human preferences are tied only to labels of (prompt, response) pairs and can degrade model utility. This work focuses on LLM alignment with preference-level privacy, which preserves the privacy of preference labels provided by humans. We propose PROPS (PROgressively Private Self-alignment), a multi-stage privacy preserving alignment framework where privately aligned models in previous stages can serve as labelers for supplementing training data in the subsequent stages of alignment. We present theoretical guarantees for PROPS as well as comprehensive validation using multiple models (Pythia and GPT) and datasets (AlpacaEval, Anthropic HH-RLHF, truthy-dpo-v0.1) to demonstrate the utility of PROPS over existing methods while still providing high privacy. For the same privacy budget, alignment via PROPS can achieve up to 3x higher win-rates compared to DP-SGD, and 2.5x higher win-rates compared to Randomized Response (RR) based alignment.

Method: Combines theoretical analysis of a toy model with empirical observations on small Transformers trained on a linear regression variant. Also validates findings on a well-studied in-context associative recall task.

Result: Reveals that emergence timing follows power laws based on task structure, architecture, and optimizer choice. Finds that repetition can greatly speed up emergence. Confirms results on associative recall task.

Conclusion: Provides a simple, theoretically grounded framework for understanding how data distributions and model design influence the learning dynamics behind one form of emergence in Transformers.

Abstract: Emergence is a fascinating property of large language models and neural networks more broadly: as models scale and train for longer, they sometimes develop new abilities in sudden ways. Despite initial studies, we still lack a comprehensive understanding of how and when these abilities emerge. To address this gap, we study the emergence over training of sparse attention, a critical and frequently observed attention pattern in Transformers. By combining theoretical analysis of a toy model with empirical observations on small Transformers trained on a linear regression variant, we uncover the mechanics driving sparse attention emergence and reveal that emergence timing follows power laws based on task structure, architecture, and optimizer choice. We additionally find that repetition can greatly speed up emergence. Finally, we confirm these results on a well-studied in-context associative recall task. Our findings provide a simple, theoretically grounded framework for understanding how data distributions and model design influence the learning dynamics behind one form of emergence.

Method: DeLTa integrates LLMs with tabular data through decision tree rules as intermediaries. It leverages LLMs’ reasoning to redesign improved rules from existing decision tree rules, then calibrates original decision trees using these LLM-generated rules via error correction vectors to reduce prediction errors.

Result: Extensive experiments on diverse tabular benchmarks show that DeLTa achieves state-of-the-art performance.

Conclusion: DeLTa provides a novel, effective approach for LLM integration with tabular data that avoids serialization and fine-tuning issues while leveraging LLMs’ reasoning capabilities through decision tree rule enhancement.

Abstract: Tabular data have been playing a vital role in diverse real-world fields, including healthcare, finance, etc. With the recent success of Large Language Models (LLMs), early explorations of extending LLMs to the domain of tabular data have been developed. Most of these LLM-based methods typically first serialize tabular data into natural language descriptions, and then tune LLMs or directly infer on these serialized data. However, these methods suffer from two key inherent issues: (i) data perspective: existing data serialization methods lack universal applicability for structured tabular data, and may pose privacy risks through direct textual exposure, and (ii) model perspective: LLM fine-tuning methods struggle with tabular data, and in-context learning scalability is bottle-necked by input length constraints (suitable for few-shot learning). This work explores a novel direction of integrating LLMs into tabular data throughough logical decision tree rules as intermediaries, proposes a decision tree enhancer with LLM-derived rule for tabular prediction, DeLTa. The proposed DeLTa avoids tabular data serialization, and can be applied to full data learning setting without LLM fine-tuning. Specifically, we leverage the reasoning ability of LLMs to redesign an improved rule given a set of decision tree rules. Furthermore, we provide a calibration method for original decision trees via new generated rule by LLM, which approximates the error correction vector to steer the original decision tree predictions in the direction of ``errors’’ reducing. Finally, extensive experiments on diverse tabular benchmarks show that our method achieves state-of-the-art performance.

Method: Develops an information-geometric framework grounded in diffusion dynamics. Embeds model outputs via spectral decompositions of multimodal graph Laplacians. Defines semantic distortion metric as gaps between model outputs and truth manifold. Uses Courant-Fischer bounds for temperature-dependent hallucination profiles and RKHS eigenmodes for modality-aware, interpretable measures.

Result: Derives mathematical bounds on hallucination profiles that track evolution over prompts and time. Provides principled, quantifiable measures that reframe hallucination as bounded rather than binary, enabling systematic evaluation.

Conclusion: The framework transforms hallucination from a qualitative problem to a quantifiable, bounded phenomenon, offering a principled mathematical basis for evaluation and mitigation strategies in multimodal LLMs.

Abstract: Hallucinations in LLMs–especially in multimodal settings–undermine reliability. We present a rigorous information-geometric framework, grounded in diffusion dynamics, to quantify hallucinations in MLLMs where model outputs are embedded via spectral decompositions of multimodal graph Laplacians, and their gaps to a truth manifold define a semantic distortion metric. We derive Courant-Fischer bounds on a temperature-dependent hallucination profile and use RKHS eigenmodes to obtain modality-aware, interpretable measures that track evolution over prompts and time. This reframes hallucination as quantifiable and bounded, providing a principled basis for evaluation and mitigation.

Method: Proposes SVH-MOL which uses Stein Variational Gradient Descent (SVGD) to push particles toward the Pareto set via functional gradient descent. Employs diverse gradient direction strategies within a unified SVGD framework for multi-objective optimization, enhanced with annealing schedules for stability.

Result: Extensive experiments on multi-objective problems and multi-task learning demonstrate superior performance, showing effectiveness in obtaining diverse Pareto solutions while maintaining high hypervolume values.

Conclusion: SVH-MOL successfully addresses the diversity-hypervolume trade-off in Pareto Set Learning, providing an effective framework for approximating complete Pareto sets in multi-objective optimization problems.

Abstract: Pareto Set Learning (PSL) is popular as an efficient approach to obtaining the complete optimal solution in Multi-objective Learning (MOL). A set of optimal solutions approximates the Pareto set, and its mapping is a set of dense points in the Pareto front in objective space. However, some current methods face a challenge: how to make the Pareto solution is diverse while maximizing the hypervolume value. In this paper, we propose a novel method to address this challenge, which employs Stein Variational Gradient Descent (SVGD) to approximate the entire Pareto set. SVGD pushes a set of particles towards the Pareto set by applying a form of functional gradient descent, which helps to converge and diversify optimal solutions. Additionally, we employ diverse gradient direction strategies to thoroughly investigate a unified framework for SVGD in multi-objective optimization and adapt this framework with an annealing schedule to promote stability. We introduce our method, SVH-MOL, and validate its effectiveness through extensive experiments on multi-objective problems and multi-task learning, demonstrating its superior performance.

Method: Non-linear AI model that reconstructs monthly mean temperature and precipitation anomalies from Euro-Atlantic weather regime indices, capturing complex non-linear relationships between atmospheric circulation states and surface climate variables.

Result: Model achieves mean absolute relative error below 80% and shows improved seasonal reconstruction compared to ECMWF’s SEAS5 operational forecast system, with comparable skill when using SEAS5-predicted weather regime indices.

Conclusion: Weather regime-based anomaly reconstruction powered by AI offers a promising pathway for sub-seasonal and seasonal forecasting, demonstrating practical applicability and competitive performance with operational systems.

Abstract: We present a non-linear AI-model designed to reconstruct monthly mean anomalies of the European temperature and precipitation based on the Euro-Atlantic Weather regimes (WR) indices. WR represent recurrent, quasi-stationary, and persistent states of the atmospheric circulation that exert considerable influence over the European weather, therefore offering an opportunity for sub-seasonal to seasonal forecasting. While much research has focused on studying the correlation and impacts of the WR on European weather, the estimation of ground-level climate variables, such as temperature and precipitation, from Euro-Atlantic WR remains largely unexplored and is currently limited to linear methods. The presented AI model can capture and introduce complex non-linearities in the relation between the WR indices, describing the state of the Euro-Atlantic atmospheric circulation and the corresponding surface temperature and precipitation anomalies in Europe. We discuss the AI-model performance in reconstructing the monthly mean two-meter temperature and total precipitation anomalies in the European winter and summer, also varying the number of WR used to describe the monthly atmospheric circulation. We assess the impact of errors on the WR indices in the reconstruction and show that a mean absolute relative error below 80% yields improved seasonal reconstruction compared to the ECMWF operational seasonal forecast system, SEAS5. As a demonstration of practical applicability, we evaluate the model using WR indices predicted by SEAS5, finding slightly better or comparable skill relative to the SEAS5 forecast itself. Our findings demonstrate that WR-based anomaly reconstruction, powered by AI tools, offers a promising pathway for sub-seasonal and seasonal forecasting.

Method: The authors analyze loss of trainability through an optimization lens, introduce two new indicators (batch-size-aware gradient-noise bound and curvature volatility-controlled bound), combine them into a per-layer adaptive noise threshold, and propose a step-size scheduler that keeps each layer’s effective parameter update below this bound.

Result: The proposed scheduler improves accuracy maintained by previously proposed approaches (CReLU, Wasserstein regularizer, L2 weight decay) and produces adaptive step-size trajectories that mirror manually engineered step-size decay schedules without tuning.

Conclusion: Loss of trainability can be effectively addressed through optimization-based analysis and adaptive step-size scheduling based on novel combined indicators, providing a principled approach that outperforms existing methods and matches manually engineered schedules.

Abstract: Loss of trainability refers to a phenomenon in continual learning where parameter updates no longer make progress on the optimization objective, so accuracy stalls or degrades as the learning problem changes over time. In this paper, we analyze loss of trainability through an optimization lens and find that the phenomenon is not reliably predicted by existing individual indicators such as Hessian rank, sharpness level, weight or gradient norms, gradient-to-parameter ratios, and unit-sign entropy. Motivated by our analysis, we introduce two complementary indicators: a batch-size-aware gradient-noise bound and a curvature volatility-controlled bound. We then combine these two indicators into a per-layer adaptive noise threshold on the effective step-size that anticipates trainability behavior. Using this insight, we propose a step-size scheduler that keeps each layer’s effective parameter update below this bound, thereby avoiding loss of trainability. We demonstrate that our scheduler can improve the accuracy maintained by previously proposed approaches, such as concatenated ReLU (CReLU), Wasserstein regularizer, and L2 weight decay. Surprisingly, our scheduler produces adaptive step-size trajectories that, without tuning, mirror the manually engineered step-size decay schedules.

Method: Proposes a model-driven trace clustering method that optimizes stochastic process models within each cluster using entropic relevance (a stochastic conformance metric based on directly-follows probabilities) to guide trace assignment, considering both structural alignment and likelihood of trace origination.

Result: The method is computationally efficient (scales linearly with input size), improves model interpretability by producing clusters with clearer control-flow patterns, and yields superior stochastic coherence and graph simplicity compared to traditional approaches, though traditional fitness metrics show a trade-off.

Conclusion: The approach provides specific utility for stochastic process analysis by better capturing real-world execution dynamics through stochastic models, offering a valuable alternative to traditional clustering methods that neglect activity frequencies and probabilities.

Abstract: Process discovery algorithms automatically extract process models from event logs, but high variability often results in complex and hard-to-understand models. To mitigate this issue, trace clustering techniques group process executions into clusters, each represented by a simpler and more understandable process model. Model-driven trace clustering improves on this by assigning traces to clusters based on their conformity to cluster-specific process models. However, most existing clustering techniques rely on either no process model discovery, or non-stochastic models, neglecting the frequency or probability of activities and transitions, thereby limiting their capability to capture real-world execution dynamics. We propose a novel model-driven trace clustering method that optimizes stochastic process models within each cluster. Our approach uses entropic relevance, a stochastic conformance metric based on directly-follows probabilities, to guide trace assignment. This allows clustering decisions to consider both structural alignment with a cluster’s process model and the likelihood that a trace originates from a given stochastic process model. The method is computationally efficient, scales linearly with input size, and improves model interpretability by producing clusters with clearer control-flow patterns. Extensive experiments on public real-life datasets demonstrate that while our method yields superior stochastic coherence and graph simplicity, traditional fitness metrics reveal a trade-off, highlighting the specific utility of our approach for stochastic process analysis.

Method: ModalSurv uses modality-specific projections to process each data type (clinical, MRI, histopathology, RNA-seq) and cross-attention fusion to integrate information across modalities for survival prediction.

Result: On CHIMERA Grand Challenge datasets: achieved C-index of 0.7402 (1st place) for prostate cancer and 0.5740 (5th place) for bladder cancer. Clinical features alone outperformed multimodal models on external tests, revealing challenges with multimodal alignment and potential overfitting.

Conclusion: ModalSurv provides systematic evaluation of multimodal survival modeling, showing promise but highlighting current limitations in scalability and generalization for cancer prognosis, with clinical data alone sometimes outperforming complex multimodal approaches on external validation.

Abstract: Accurate survival prediction is essential for personalised cancer treatment. We propose ModalSurv, a multimodal deep survival framework integrating clinical, MRI, histopathology, and RNA-sequencing data via modality-specific projections and cross-attention fusion. On the CHIMERA Grand Challenge datasets, ModalSurv achieved a C-index of 0.7402 (1st) for prostate and 0.5740 (5th) for bladder cancer. Notably, clinical features alone outperformed multimodal models on external tests, highlighting challenges of limited multimodal alignment and potential overfitting. Local validation showed multimodal gains but limited generalisation. ModalSurv provides a systematic evaluation of multimodal survival modelling, underscoring both its promise and current limitations for scalable, generalisable cancer prognosis.

Method: Using large language models to process discretized partial differential equation solutions, analyzing predictive accuracy across different temporal contexts and spatial discretizations, and examining multi-step rollouts with error analysis. Also analyzing token-level output distributions to understand the ICL progression.

Result: LLMs can accurately extrapolate PDE dynamics, with accuracy improving with longer temporal contexts but degrading at finer spatial discretizations. In multi-step rollouts, errors grow algebraically with time horizon. Analysis reveals a three-stage ICL progression: syntactic pattern imitation → exploratory high-entropy phase → confident, numerically grounded predictions.

Conclusion: Text-trained foundation models demonstrate emergent capabilities for spatiotemporal forecasting of PDE solutions through in-context learning, exhibiting predictable scaling laws and revealing a structured internal progression from pattern recognition to numerical prediction.

Abstract: Large language models (LLMs) have demonstrated emergent in-context learning (ICL) capabilities across a range of tasks, including zero-shot time-series forecasting. We show that text-trained foundation models can accurately extrapolate spatiotemporal dynamics from discretized partial differential equation (PDE) solutions without fine-tuning or natural language prompting. Predictive accuracy improves with longer temporal contexts but degrades at finer spatial discretizations. In multi-step rollouts, where the model recursively predicts future spatial states over multiple time steps, errors grow algebraically with the time horizon, reminiscent of global error accumulation in classical finite-difference solvers. We interpret these trends as in-context neural scaling laws, where prediction quality varies predictably with both context length and output length. To better understand how LLMs are able to internally process PDE solutions so as to accurately roll them out, we analyze token-level output distributions and uncover a consistent three-stage ICL progression: beginning with syntactic pattern imitation, transitioning through an exploratory high-entropy phase, and culminating in confident, numerically grounded predictions.

Method: Theoretical analysis of decoder-only transformer expressivity. Proves impossibility result for inverse permutation learning in arbitrary-depth decoder-only transformers. Presents alternative constructions: 1) using encoder-decoder architecture, 2) adding scratch tokens to decoder-only models.

Result: Proves decoder-only transformers cannot learn inverse permutations regardless of depth. Shows causal attention mask creates expressivity gap between encoder-decoder and decoder-only architectures. Surprisingly demonstrates that padding with scratch tokens enables inverse permutation learning.

Conclusion: The scratch token construction suggests a mechanism for how chain-of-thought prompting enables reasoning in LLMs - intermediate “thinking” tokens (even without semantic meaning) can provide computational workspace, explaining why decoder-only models can perform complex reasoning with appropriate prompting.

Abstract: In this technical note, we study the problem of inverse permutation learning in decoder-only transformers. Given a permutation and a string to which that permutation has been applied, the model is tasked with producing the original (canonical'') string. We argue that this task models a natural robustness property across a variety of reasoning tasks, including long-context retrieval, multiple choice QA and in-context learning. Our primary contribution is an impossibility result: we show that an arbitrary depth, decoder-only transformer cannot learn this task. This result concerns the expressive capacity of decoder-only transformer models and is agnostic to training dynamics or sample complexity. We give a pair of alternative constructions under which inverse permutation learning is feasible. The first of these highlights the fundamental role of the causal attention mask, and reveals a gap between the expressivity of encoder-decoder transformers and the more popular decoder-only architecture. The latter result is more surprising: we show that simply padding the input with scratch tokens" yields a construction under which inverse permutation learning is possible. We conjecture that this may suggest an alternative mechanism by which chain-of-thought prompting or, more generally, intermediate ``thinking’’ tokens can enable reasoning in large language models, even when these tokens encode no meaningful semantic information (e.g., the results of intermediate computations).

Method: SimpleFold uses only general-purpose transformer blocks with adaptive layers, trained via generative flow-matching objective with an additional structural term. It’s scaled to 3B parameters and trained on ~9M distilled protein structures plus experimental PDB data.

Result: SimpleFold-3B achieves competitive performance on standard folding benchmarks compared to state-of-the-art baselines, demonstrates strong ensemble prediction capabilities, and shows efficiency in deployment and inference on consumer-level hardware.

Conclusion: SimpleFold challenges the necessity of complex domain-specific architectures in protein folding, opening up an alternative design space using general-purpose transformer blocks for future progress in the field.

Abstract: Protein folding models have achieved groundbreaking results typically via a combination of integrating domain knowledge into the architectural blocks and training pipelines. Nonetheless, given the success of generative models across different but related problems, it is natural to question whether these architectural designs are a necessary condition to build performant models. In this paper, we introduce SimpleFold, the first flow-matching based protein folding model that solely uses general purpose transformer blocks. Protein folding models typically employ computationally expensive modules involving triangular updates, explicit pair representations or multiple training objectives curated for this specific domain. Instead, SimpleFold employs standard transformer blocks with adaptive layers and is trained via a generative flow-matching objective with an additional structural term. We scale SimpleFold to 3B parameters and train it on approximately 9M distilled protein structures together with experimental PDB data. On standard folding benchmarks, SimpleFold-3B achieves competitive performance compared to state-of-the-art baselines, in addition SimpleFold demonstrates strong performance in ensemble prediction which is typically difficult for models trained via deterministic reconstruction objectives. Due to its general-purpose architecture, SimpleFold shows efficiency in deployment and inference on consumer-level hardware. SimpleFold challenges the reliance on complex domain-specific architectures designs in protein folding, opening up an alternative design space for future progress.

Method: InfMasking uses an Infinite Masking strategy that stochastically occludes most features from each modality during fusion, preserving only partial information to create representations with varied synergistic patterns. Unmasked fused representations are aligned with masked ones through mutual information maximization to encode comprehensive synergistic information. An InfMasking loss approximates the computationally prohibitive mutual information calculation.

Result: InfMasking effectively enhances synergistic information between modalities in controlled experiments. In evaluations on large-scale real-world datasets, it achieves state-of-the-art performance across seven benchmarks.

Conclusion: InfMasking successfully addresses the challenge of capturing synergistic information in multimodal representation learning through its infinite masking strategy, demonstrating superior performance on multiple benchmarks.

Abstract: In multimodal representation learning, synergistic interactions between modalities not only provide complementary information but also create unique outcomes through specific interaction patterns that no single modality could achieve alone. Existing methods may struggle to effectively capture the full spectrum of synergistic information, leading to suboptimal performance in tasks where such interactions are critical. This is particularly problematic because synergistic information constitutes the fundamental value proposition of multimodal representation. To address this challenge, we introduce InfMasking, a contrastive synergistic information extraction method designed to enhance synergistic information through an Infinite Masking strategy. InfMasking stochastically occludes most features from each modality during fusion, preserving only partial information to create representations with varied synergistic patterns. Unmasked fused representations are then aligned with masked ones through mutual information maximization to encode comprehensive synergistic information. This infinite masking strategy enables capturing richer interactions by exposing the model to diverse partial modality combinations during training. As computing mutual information estimates with infinite masking is computationally prohibitive, we derive an InfMasking loss to approximate this calculation. Through controlled experiments, we demonstrate that InfMasking effectively enhances synergistic information between modalities. In evaluations on large-scale real-world datasets, InfMasking achieves state-of-the-art performance across seven benchmarks. Code is released at https://github.com/brightest66/InfMasking.

Method: Proposes a stability-plasticity principle that emphasizes preserving better knowledge (from pretrained policy or offline dataset) during online fine-tuning while maintaining plasticity. Identifies three distinct regimes of online fine-tuning requiring different stability properties.

Result: Large-scale empirical study validates the framework, with results strongly aligning with predictions in 45 out of 63 cases (71% alignment).

Conclusion: Provides a principled framework for guiding design choices in offline-to-online RL based on relative performance of offline dataset and pretrained policy, explaining previously inconsistent empirical behavior.

Abstract: Offline-to-online reinforcement learning (RL) has emerged as a practical paradigm that leverages offline datasets for pretraining and online interactions for fine-tuning. However, its empirical behavior is highly inconsistent: design choices of online-fine tuning that work well in one setting can fail completely in another. We propose a stability–plasticity principle that can explain this inconsistency: we should preserve the knowledge of pretrained policy or offline dataset during online fine-tuning, whichever is better, while maintaining sufficient plasticity. This perspective identifies three regimes of online fine-tuning, each requiring distinct stability properties. We validate this framework through a large-scale empirical study, finding that the results strongly align with its predictions in 45 of 63 cases. This work provides a principled framework for guiding design choices in offline-to-online RL based on the relative performance of the offline dataset and the pretrained policy.

Method: Introduces PEAR benchmark for systematic evaluation of planner-executor MAS architectures, focusing on both utility and vulnerability assessment through extensive experiments on various system configurations.

Result: Four key findings: (1) weak planners degrade clean task performance more than weak executors; (2) planner memory is essential but executor memory doesn’t impact clean performance; (3) trade-off exists between task performance and robustness; (4) planner-targeted attacks are particularly effective.

Conclusion: The PEAR benchmark provides actionable insights for enhancing MAS robustness and lays groundwork for principled defenses in multi-agent settings by systematically evaluating vulnerabilities in planner-executor architectures.

Abstract: Large Language Model (LLM)-based Multi-Agent Systems (MAS) have emerged as a powerful paradigm for tackling complex, multi-step tasks across diverse domains. However, despite their impressive capabilities, MAS remain susceptible to adversarial manipulation. Existing studies typically examine isolated attack surfaces or specific scenarios, leaving a lack of holistic understanding of MAS vulnerabilities. To bridge this gap, we introduce PEAR, a benchmark for systematically evaluating both the utility and vulnerability of planner-executor MAS. While compatible with various MAS architectures, our benchmark focuses on the planner-executor structure, which is a practical and widely adopted design. Through extensive experiments, we find that (1) a weak planner degrades overall clean task performance more severely than a weak executor; (2) while a memory module is essential for the planner, having a memory module for the executor does not impact the clean task performance; (3) there exists a trade-off between task performance and robustness; and (4) attacks targeting the planner are particularly effective at misleading the system. These findings offer actionable insights for enhancing the robustness of MAS and lay the groundwork for principled defenses in multi-agent settings.

Method: Introduces Spectrally Anisotropic Gaussian Diffusion (SAGD) with an anisotropic noise operator that replaces isotropic forward covariance with structured, frequency-diagonal covariance. This unifies band-pass masks and power-law weightings to emphasize or suppress designated frequency bands while keeping the forward process Gaussian.

Result: The learned score converges to true data score as t→0 under full support, while anisotropy reshapes probability-flow path. Empirically outperforms standard diffusion across several vision datasets and enables selective omission of known corruptions confined to specific frequency bands.

Conclusion: Carefully designed anisotropic forward noise provides a simple yet principled handle to tailor inductive bias in DPMs, demonstrating the value of explicit spectral shaping in diffusion models.

Abstract: Diffusion Probabilistic Models (DPMs) have achieved strong generative performance, yet their inductive biases remain largely implicit. In this work, we aim to build inductive biases into the training and sampling of diffusion models to better accommodate the target distribution of the data to model. We introduce an anisotropic noise operator that shapes these biases by replacing the isotropic forward covariance with a structured, frequency-diagonal covariance. This operator unifies band-pass masks and power-law weightings, allowing us to emphasize or suppress designated frequency bands, while keeping the forward process Gaussian. We refer to this as Spectrally Anisotropic Gaussian Diffusion (SAGD). In this work, we derive the score relation for anisotropic forward covariances and show that, under full support, the learned score converges to the true data score as $t!\to!0$, while anisotropy reshapes the probability-flow path from noise to data. Empirically, we show the induced anisotropy outperforms standard diffusion across several vision datasets, and enables selective omission: learning while ignoring known corruptions confined to specific bands. Together, these results demonstrate that carefully designed anisotropic forward noise provides a simple, yet principled, handle to tailor inductive bias in DPMs.

Method: A generic and versatile machine learning framework for data-driven RF fingerprinting that is emitter-type agnostic. The framework supports multiple downstream tasks including specific emitter identification (SEI), emitter data association (EDA), and RF emitter clustering (RFEC).

Result: The framework is demonstrated using real RF datasets for practical applications including spaceborne surveillance, signal intelligence, and countering drones, showing its effectiveness across different operational scenarios.

Conclusion: The proposed ML framework provides a flexible, data-driven solution for RF fingerprinting that outperforms traditional methods and can be applied to various emitter types and transmission schemes across multiple defense and civilian applications.

Abstract: Fingerprinting radio frequency (RF) emitters typically involves finding unique characteristics that are featured in their received signal. These fingerprints are nuanced, but sufficiently detailed, motivating the pursuit of methods that can successfully extract them. The downstream task that requires the most meticulous RF fingerprinting (RFF) is known as specific emitter identification (SEI), which entails recognising each individual transmitter. RFF and SEI have a long history, with numerous defence and civilian applications such as signal intelligence, electronic surveillance, physical-layer authentication of wireless devices, to name a few. In recent years, data-driven RFF approaches have become popular due to their ability to automatically learn intricate fingerprints. They generally deliver superior performance when compared to traditional RFF techniques that are often labour-intensive, inflexible, and only applicable to a particular emitter type or transmission scheme. In this paper, we present a generic and versatile machine learning (ML) framework for data-driven RFF with several popular downstream tasks such as SEI, data association (EDA) and RF emitter clustering (RFEC). It is emitter-type agnostic. We then demonstrate the introduced framework for several tasks using real RF datasets for spaceborne surveillance, signal intelligence and countering drones applications.

Method: The Deep Edge Filter applies high-pass filtering to deep neural network features by subtracting low-pass filtered outputs from original features. This isolates generalizable representations while preserving the original network architecture integrity.

Result: Experimental results across Vision, Text, 3D, and Audio domains demonstrate consistent performance improvements regardless of model architecture and data modality. Analysis shows the method induces feature sparsification and effectively isolates high-frequency components.

Conclusion: The Deep Edge Filter successfully improves model generalization by filtering out low-frequency domain biases while preserving task-relevant high-frequency information, with empirical validation supporting the core hypothesis about frequency-based feature encoding.

Abstract: We introduce the Deep Edge Filter, a novel approach that applies high-pass filtering to deep neural network features to improve model generalizability. Our method is motivated by our hypothesis that neural networks encode task-relevant semantic information in high-frequency components while storing domain-specific biases in low-frequency components of deep features. By subtracting low-pass filtered outputs from original features, our approach isolates generalizable representations while preserving architectural integrity. Experimental results across diverse domains such as Vision, Text, 3D, and Audio demonstrate consistent performance improvements regardless of model architecture and data modality. Analysis reveals that our method induces feature sparsification and effectively isolates high-frequency components, providing empirical validation of our core hypothesis. The code is available at https://github.com/dongkwani/DeepEdgeFilter.

Method: Tawa uses a novel IR abstraction called asynchronous references (aref) to express warp-level communication without exposing low-level hardware details. It automatically partitions programs into producer-consumer roles and manages intricate dataflow pipelines, relieving developers from manual kernel rewriting.

Result: On NVIDIA H100 GPUs, Tawa achieves up to 1.1× speedup over highly optimized cuBLAS GEMM kernels. For attention workloads, it attains 1.2× speedup over Triton and matches the performance of hand-optimized CUTLASS C++ FlashAttention-3 kernel with significantly less programming effort.

Conclusion: Tawa bridges the programmability gap between high-level programming and GPU hardware specialization by automating warp-specialized code generation, enabling high hardware utilization without the complexity of manual low-level programming.

Abstract: Modern GPUs feature specialized hardware units that enable high-performance, asynchronous dataflow execution. However, the conventional SIMT programming model is fundamentally misaligned with this task-parallel hardware, creating a significant programmability gap. While hardware-level warp specialization is the key to unlocking peak performance, it forces developers to manually orchestrate complex, low-level communication and software pipelines–a process that is labor-intensive, error-prone, and unsustainable. To address this challenge, we present Tawa, an automated compiler that systematically generates high-performance, warp-specialized code from a high-level, tile-based program. Central to our approach is a novel IR abstraction, asynchronous references (aref), which expresses warp-level communication without exposing low-level hardware details. Using this abstraction, Tawa automatically partitions programs into producer-consumer roles and manages the intricate dataflow pipeline, relieving developers of invasive kernel rewriting. Evaluation on NVIDIA H100 GPUs across representative LLM kernels shows that Tawa delivers high hardware utilization, achieving up to 1.1$\times$ speedup over highly optimized cuBLAS GEMM kernels. For attention workloads, Tawa attains 1.2$\times$ speedup over Triton and matches the performance of the hand-optimized CUTLASS C++ FlashAttention-3 kernel with far less programming effort.

Method: The authors propose a user feedback simulation framework and comprehensive benchmark covering multiple domains, languages, and task types to evaluate continual learning abilities of LLM systems.

Result: Experiments show that state-of-the-art baselines for continual learning in LLM systems are far from satisfactory in terms of both effectiveness and efficiency.

Conclusion: The proposed benchmark aims to pave the way for future research on LLM memory and optimization algorithms by providing a more realistic evaluation framework for continual learning from user feedback.

Abstract: Scaling up data, parameters, and test-time computation has been the mainstream methods to improve LLM systems (LLMsys), but their upper bounds are almost reached due to the gradual depletion of high-quality data and marginal gains obtained from larger computational resource consumption. Inspired by the abilities of human and traditional AI systems in learning from practice, constructing memory and continual learning frameworks for LLMsys has become an important and popular research direction in recent literature. Yet, existing benchmarks for LLM memory often focus on evaluating the system on homogeneous reading comprehension tasks with long-form inputs rather than testing their abilities to learn from accumulated user feedback in service time. Therefore, we propose a user feedback simulation framework and a comprehensive benchmark covering multiple domains, languages, and types of tasks to evaluate the continual learning abilities of LLMsys. Experiments show that the effectiveness and efficiency of state-of-the-art baselines are far from satisfying, and we hope this benchmark could pave the way for future studies on LLM memory and optimization algorithms.

Method: Two-phase approach: 1) Data generation phase creating 15 million adversarial training samples using custom RDSA methodology, 2) Robust model training with specialized architecture featuring Feature Embedding Block with shared weights for same-type features and Dense Fusion Tail for final prediction.

Result: Achieved 80% mixed accuracy score on both clean and adversarial data, outperforming the second-place solution by 2 percentage points and winning Task 2 of the ECML-PKDD 2025 challenge.

Conclusion: The proposed two-phase approach combining adversarial data generation with specialized robust architecture design successfully creates models resilient to RDSA attacks, demonstrating effectiveness for robust learning in high energy physics applications.

Abstract: This report presents the winning solution for Task 2 of Colliding with Adversaries: A Challenge on Robust Learning in High Energy Physics Discovery at ECML-PKDD 2025. The goal of the challenge was to design and train a robust ANN-based model capable of achieving high accuracy in a binary classification task on both clean and adversarial data generated with the Random Distribution Shuffle Attack (RDSA). Our solution consists of two components: a data generation phase and a robust model training phase. In the first phase, we produced 15 million artificial training samples using a custom methodology derived from Random Distribution Shuffle Attack (RDSA). In the second phase, we introduced a robust architecture comprising (i)a Feature Embedding Block with shared weights among features of the same type and (ii)a Dense Fusion Tail responsible for the final prediction. Training this architecture on our adversarial dataset achieved a mixed accuracy score of 80%, exceeding the second-place solution by two percentage points.

Method: The review organizes around four core themes: (1) historical development of Kolmogorov’s superposition theory toward neural networks, (2) establishing formal equivalence between KANs and MLPs, (3) analyzing the role of basis functions, and (4) organizing recent advancements in accuracy, efficiency, regularization, and convergence. The authors collected and categorized a large set of open-source implementations to map the KAN ecosystem.

Result: The paper provides a systematic mapping of the KAN research landscape, including a practical “Choose-Your-KAN” guide for practitioners to select appropriate architectures. It identifies current research gaps and future directions, supported by an associated GitHub repository serving as a structured reference for ongoing KAN research.

Conclusion: This review serves as a comprehensive reference point for the KAN community, offering both theoretical foundations and practical guidance while identifying areas for future research. The organized ecosystem mapping and practical selection guide make KANs more accessible to researchers and practitioners.

Abstract: The so-called Kolmogorov-Arnold Networks (KANs), whose design is merely inspired, rather than dictated, by the Kolmogorov superposition theorem, have emerged as a promising alternative to traditional Multilayer Perceptrons (MLPs). This review provides a systematic and comprehensive overview of the rapidly expanding KAN landscape. By collecting and categorizing a large set of open-source implementations, we map the vibrant ecosystem supporting modern KAN development. We organize the review around four core themes: (i) presenting a precise history of Kolmogorov’s superposition theory toward neural-network formulations; (ii) establishing the formal equivalence between KANs and MLPs; (iii) analyzing the critical role of basis functions; and (iv) organizing recent advancements in accuracy, efficiency, regularization, and convergence. Finally, we provide a practical Choose-Your-KAN guide to assist practitioners in selecting appropriate architectures, and we close by identifying current research gaps and future directions. The associated GitHub repository (https://github.com/AmirNoori68/kan-review) complements this paper and serves as a structured reference for ongoing KAN research.

Method: An interactive installation with a mutable, projection-mapped landscape where human participants reshape the world and engage with multiple AI agents that develop incomplete accounts of their environment.

Result: The work positions AI as embodied co-witnesses rather than deterministic tools, creating a shared space for examining epistemic limits and meaning construction in unstable conditions.

Conclusion: LLMscape invites reflection on our shared epistemic limits and challenges traditional views of AI as deterministic tools, instead presenting AI as collaborative meaning-makers in uncertain environments.

Abstract: LLMscape is an interactive installation that investigates how humans and AI construct meaning under shared conditions of uncertainty. Within a mutable, projection-mapped landscape, human participants reshape the world and engage with multiple AI agents, each developing incomplete and provisional accounts of their environment. Exhibited in Shanghai and continually evolving, the work positions AI not as deterministic tools but as embodied co-witnesses to an unstable world, examining the parallels between human and artificial meaning-making and inviting reflection on our shared epistemic limits.

Method: MAESTRO moves LLMs outside execution loop as offline architects. It has two components: semantic curriculum generator for diverse traffic scenarios, and automated reward synthesizer for Python reward functions. These guide standard MARL (MADDPG) backbone.

Result: On large-scale traffic signal control (16 intersections), combining LLM-generated curricula with reward shaping yields +4.0% higher mean return (163.26 vs. 156.93) and 2.2% better risk-adjusted performance (Sharpe 1.53 vs. 0.70) over baseline.

Conclusion: LLMs can be effective high-level designers for cooperative MARL training when used as offline architects, improving performance and stability without increasing deployment inference costs.

Abstract: Cooperative Multi-Agent Reinforcement Learning (MARL) faces two major design bottlenecks: crafting dense reward functions and constructing curricula that avoid local optima in high-dimensional, non-stationary environments. Existing approaches rely on fixed heuristics or use Large Language Models (LLMs) directly in the control loop, which is costly and unsuitable for real-time systems. We propose MAESTRO (Multi-Agent Environment Shaping through Task and Reward Optimization), a framework that moves the LLM outside the execution loop and uses it as an offline training architect. MAESTRO introduces two generative components: (i) a semantic curriculum generator that creates diverse, performance-driven traffic scenarios, and (ii) an automated reward synthesizer that produces executable Python reward functions adapted to evolving curriculum difficulty. These components guide a standard MARL backbone (MADDPG) without increasing inference cost at deployment. We evaluate MAESTRO on large-scale traffic signal control (Hangzhou, 16 intersections) and conduct controlled ablations. Results show that combining LLM-generated curricula with LLM-generated reward shaping yields improved performance and stability. Across four seeds, the full system achieves +4.0% higher mean return (163.26 vs. 156.93) and 2.2% better risk-adjusted performance (Sharpe 1.53 vs. 0.70) over a strong curriculum baseline. These findings highlight LLMs as effective high-level designers for cooperative MARL training.

Method: Analyze network dynamics on a statistical manifold, revealing that the Ridge corresponds to the Edge of Stability where the Fisher Information Matrix becomes singular. Show that apparent Euclidean force antagonism is actually a manifestation of Dual Equilibrium in Riemannian space.

Result: The Ridge of Optimization is identified as the Edge of Stability boundary where Fisher Information Matrix singularity occurs. This provides a geometric interpretation that unifies learning dynamics and network capacity through the Minimum Description Length principle.

Conclusion: The paper offers a geometric theory of self-organized criticality in neural networks, showing that the Ridge phenomenon emerges from fundamental information-geometric principles rather than being a mere statistical artifact.

Abstract: High-capacity kernel Hopfield networks exhibit a \textit{Ridge of Optimization} characterized by extreme stability. While previously linked to \textit{Spectral Concentration}, its origin remains elusive. Here, we analyze the network dynamics on a statistical manifold, revealing that the Ridge corresponds to the Edge of Stability, a critical boundary where the Fisher Information Matrix becomes singular. We demonstrate that the apparent Euclidean force antagonism is a manifestation of \textit{Dual Equilibrium} in the Riemannian space. This unifies learning dynamics and capacity via the Minimum Description Length principle, offering a geometric theory of self-organized criticality.

Method: AltNet uses twin networks that periodically alternate roles: one active network learns from environment interactions while the other learns off-policy from replay buffer and active network’s experiences, then they swap roles with the active network being reset.

Result: AltNet restores plasticity, improves sample efficiency, achieves higher performance, and avoids performance drops in DeepMind Control Suite tasks, outperforming baseline and state-of-the-art reset methods.

Conclusion: AltNet provides a safe and effective approach to maintaining plasticity in RL without the performance degradation risks of traditional reset methods.

Abstract: Neural networks have shown remarkable success in supervised learning when trained on a single task using a fixed dataset. However, when neural networks are trained on a reinforcement learning task, their ability to continue learning from new experiences declines over time. This decline in learning ability is known as plasticity loss. To restore plasticity, prior work has explored periodically resetting the parameters of the learning network, a strategy that often improves overall performance. However, such resets come at the cost of a temporary drop in performance, which can be dangerous in real-world settings. To overcome this instability, we introduce AltNet, a reset-based approach that restores plasticity without performance degradation by leveraging twin networks. The use of twin networks anchors performance during resets through a mechanism that allows networks to periodically alternate roles: one network learns as it acts in the environment, while the other learns off-policy from the active network’s interactions and a replay buffer. At fixed intervals, the active network is reset and the passive network, having learned from prior experiences, becomes the new active network. AltNet restores plasticity, improving sample efficiency and achieving higher performance, while avoiding performance drops that pose risks in safety-critical settings. We demonstrate these advantages in several high-dimensional control tasks from the DeepMind Control Suite, where AltNet outperforms various relevant baseline methods, as well as state-of-the-art reset-based techniques.

Method: Proposes a “configuration-dynamics-error analysis” framework that reformulates major DD approaches under a common generalization-error perspective. Provides theoretical analysis including scaling laws and coverage laws with provable bounds.

Result: Derives two main theoretical results: (1) a scaling law showing error decreases with distilled sample size (explaining performance saturation), and (2) a coverage law showing required distilled sample size scales linearly with configuration diversity. Reveals different matching methods are interchangeable surrogates.

Conclusion: The unified theoretical framework advances dataset distillation foundations, explains empirical observations, and enables theory-driven design of compact, configuration-robust distilled datasets. Experiments confirm the derived laws across diverse methods and configurations.

Abstract: Dataset distillation (DD) aims to construct compact synthetic datasets that allow models to achieve comparable performance to full-data training while substantially reducing storage and computation. Despite rapid empirical progress, its theoretical foundations remain limited: existing methods (gradient, distribution, trajectory matching) are built on heterogeneous surrogate objectives and optimization assumptions, which makes it difficult to analyze their common principles or provide general guarantees. Moreover, it is still unclear under what conditions distilled data can retain the effectiveness of full datasets when the training configuration, such as optimizer, architecture, or augmentation, changes. To answer these questions, we propose a unified theoretical framework, termed configuration–dynamics–error analysis, which reformulates major DD approaches under a common generalization-error perspective and provides two main results: (i) a scaling law that provides a single-configuration upper bound, characterizing how the error decreases as the distilled sample size increases and explaining the commonly observed performance saturation effect; and (ii) a coverage law showing that the required distilled sample size scales linearly with configuration diversity, with provably matching upper and lower bounds. In addition, our unified analysis reveals that various matching methods are interchangeable surrogates, reducing the same generalization error, clarifying why they can all achieve dataset distillation and providing guidance on how surrogate choices affect sample efficiency and robustness. Experiments across diverse methods and configurations empirically confirm the derived laws, advancing a theoretical foundation for DD and enabling theory-driven design of compact, configuration-robust dataset distillation.

Method: The paper proposes a PID (proportional-integral-derivative) control-inspired boosting method applied to neural network predictions. At each time step, the PID-based method adjusts predicted values to bring them closer to real values. The method is tested on water demand forecasting using two deep neural network models and on hourly energy consumption forecasting to demonstrate general applicability.

Result: The PID-based booster significantly improves prediction accuracy compared to original neural network models while maintaining negligible impact on system complexity. The method proves effective for both water demand and energy consumption forecasting problems.

Conclusion: The PID-inspired boosting method successfully enhances neural network performance for multi-step periodic time-series prediction, offering superior accuracy with minimal complexity increase, making it valuable for various industrial applications.

Abstract: Multi-step time-series prediction is an essential supportive step for decision-makers in several industrial areas. Artificial intelligence techniques, which use a neural network component in various forms, have recently frequently been used to accomplish this step. However, the complexity of the neural network structure still stands up as a critical problem against prediction accuracy. In this paper, a method inspired by the proportional-integral-derivative (PID) control approach is investigated to enhance the performance of neural network models used for multi-step ahead prediction of periodic time-series information while maintaining a negligible impact on the complexity of the system. The PID-based method is applied to the predicted value at each time step to bring that value closer to the real value. The water demand forecasting problem is considered as a case study, where two deep neural network models from the literature are used to prove the effectiveness of the proposed boosting method. Furthermore, to prove the applicability of this PID-based booster to other types of periodic time-series prediction problems, it is applied to enhance the accuracy of a neural network model used for multi-step forecasting of hourly energy consumption. The comparison between the results of the original prediction models and the results after using the proposed technique demonstrates the superiority of the proposed method in terms of prediction accuracy and system complexity.

Method: Proposes a CNN-based framework that transforms firm-year panel data into image-like representations to capture cross-sectional and temporal patterns. Uses feature engineering to enable advance fraud prediction and applies local explanation techniques for interpretability across entity, feature, and time dimensions.

Result: CNN outperforms logistic regression and LightGBM in accuracy, robustness, and early-warning performance. Proper threshold tuning is crucial in high-risk settings. Key predictors include solvency, ratio structure, governance, and internal control. Fraud firms show heterogeneous patterns in short time windows.

Conclusion: The CNN framework effectively detects financial fraud with better predictive power and timeliness. Interpretability analysis reveals meaningful fraud patterns, and the case study validates the model’s predictions against real-world misconduct.

Abstract: Since the emergence of joint-stock companies, financial fraud by listed firms has repeatedly undermined capital markets. Fraud is difficult to detect because of covert tactics and the high labor and time costs of audits. Traditional statistical models are interpretable but struggle with nonlinear feature interactions, while machine learning models are powerful but often opaque. In addition, most existing methods judge fraud only for the current year based on current year data, limiting timeliness. This paper proposes a financial fraud detection framework for Chinese A-share listed companies based on convolutional neural networks (CNNs). We design a feature engineering scheme that transforms firm-year panel data into image like representations, enabling the CNN to capture cross-sectional and temporal patterns and to predict fraud in advance. Experiments show that the CNN outperforms logistic regression and LightGBM in accuracy, robustness, and early-warning performance, and that proper tuning of the classification threshold is crucial in high-risk settings. To address interpretability, we analyze the model along the dimensions of entity, feature, and time using local explanation techniques. We find that solvency, ratio structure, governance structure, and internal control are general predictors of fraud, while environmental indicators matter mainly in high-pollution industries. Non-fraud firms share stable feature patterns, whereas fraud firms exhibit heterogeneous patterns concentrated in short time windows. A case study of Guanong Shares in 2022 shows that cash flow analysis, social responsibility, governance structure, and per-share indicators are the main drivers of the model’s fraud prediction, consistent with the company’s documented misconduct.

Method: Cooperative neural network framework integrating an imputation model with a surrogate strength predictor, trained cooperatively to generate valid mix designs in single forward pass without retraining for different constraints.

Result: Achieves R-squared values of 0.87-0.92, reduces mean squared error by ~50% compared to autoencoder models and ~70% compared to Bayesian inference with Gaussian process surrogates.

Conclusion: The framework provides accurate and computationally efficient foundation for constraint-aware, data-driven mix proportioning in high-performance concrete design.

Abstract: High-performance concrete requires complex mix design decisions involving interdependent variables and practical constraints. While data-driven methods have improved predictive modeling for forward design in concrete engineering, inverse design remains limited, especially when some variables are fixed and only the remaining ones must be inferred. This study proposes a cooperative neural network framework for the partial inverse design of high-performance concrete. The framework integrates an imputation model with a surrogate strength predictor and learns through cooperative training. Once trained, it generates valid and performance-consistent mix designs in a single forward pass without retraining for different constraint scenarios. Compared with baseline models, including autoencoder models and Bayesian inference with Gaussian process surrogates, the proposed method achieves R-squared values of 0.87 to 0.92 and substantially reduces mean squared error by approximately 50% and 70%, respectively. The results show that the framework provides an accurate and computationally efficient foundation for constraint-aware, data-driven mix proportioning.

Method: Couples KGE loss gradient with local curvatures in an extended Ricci flow, allowing entity embeddings to co-evolve dynamically with manifold geometry toward mutual adaptation.

Result: Theoretically proves: 1) edge-wise curvatures decay exponentially (manifold drives toward Euclidean flatness), 2) KGE distances strictly converge to global optimum. Experimentally shows improvements on link prediction and node classification benchmarks.

Conclusion: RicciKGE effectively adapts to heterogeneous knowledge graph structures by dynamically evolving both embeddings and manifold geometry, overcoming limitations of static homogeneous manifolds.

Abstract: Knowledge graph embedding (KGE) relies on the geometry of the embedding space to encode semantic and structural relations. Existing methods place all entities on one homogeneous manifold, Euclidean, spherical, hyperbolic, or their product/multi-curvature variants, to model linear, symmetric, or hierarchical patterns. Yet a predefined, homogeneous manifold cannot accommodate the sharply varying curvature that real-world graphs exhibit across local regions. Since this geometry is imposed a priori, any mismatch with the knowledge graph’s local curvatures will distort distances between entities and hurt the expressiveness of the resulting KGE. To rectify this, we propose RicciKGE to have the KGE loss gradient coupled with local curvatures in an extended Ricci flow such that entity embeddings co-evolve dynamically with the underlying manifold geometry towards mutual adaptation. Theoretically, when the coupling coefficient is bounded and properly selected, we rigorously prove that i) all the edge-wise curvatures decay exponentially, meaning that the manifold is driven toward the Euclidean flatness; and ii) the KGE distances strictly converge to a global optimum, which indicates that geometric flattening and embedding optimization are promoting each other. Experimental improvements on link prediction and node classification benchmarks demonstrate RicciKGE’s effectiveness in adapting to heterogeneous knowledge graph structures.

Method: DRCL uses a dual refinement cycle with warm-start initialization and bidirectional refinement between a GCN-based Community Detection Module and a Text Semantic Modeling Module, iteratively exchanging pseudo-labels to enhance both structural clustering and text representation learning.

Result: DRCL consistently improves structural and semantic quality of discovered communities across several datasets, and a Mamba-based classifier trained from DRCL’s community signals achieves accuracy comparable to supervised models.

Conclusion: DRCL demonstrates strong potential for deployment in large-scale systems where labeled data are scarce or costly, offering a practical unsupervised solution for text-attributed network analysis.

Abstract: Pretrained language models offer strong text understanding capabilities but remain difficult to deploy in real-world text-attributed networks due to their heavy dependence on labeled data. Meanwhile, community detection methods typically ignore textual semantics, limiting their usefulness in downstream applications such as content organization, recommendation, and risk monitoring. To overcome these limitations, we present Dual Refinement Cycle Learning (DRCL), a fully unsupervised framework designed for practical scenarios where no labels or category definitions are available. DRCL integrates structural and semantic information through a warm-start initialization and a bidirectional refinement cycle between a GCN-based Community Detection Module (GCN-CDM) and a Text Semantic Modeling Module (TSMM). The two modules iteratively exchange pseudo-labels, allowing semantic cues to enhance structural clustering and structural patterns to guide text representation learning without manual supervision. Across several text-attributed graph datasets, DRCL consistently improves the structural and semantic quality of discovered communities. Moreover, a Mamba-based classifier trained solely from DRCL’s community signals achieves accuracy comparable to supervised models, demonstrating its potential for deployment in large-scale systems where labeled data are scarce or costly. The code is available at https://github.com/wuanghoong/DRCL.git.

Method: The paper provides a general overview of possible vector systems for NN training, their properties, and construction methods. These systems are used to configure latent spaces of encoders and visual transformers, allowing training without classification layers.

Result: The approach significantly speeds up ImageNet-1K and 50k-600k classes LSC training. Using the minimum number of latent space dimensions for a specific number of classes results in faster convergence.

Conclusion: The method enables efficient training on datasets with massive numbers of classes and has potential advantages for reducing the size of vector databases used to store NN embeddings by optimizing latent space dimensions.

Abstract: The overall neural network (NN) performance is closely related to the properties of its embedding distribution in latent space (LS). It has recently been shown that predefined vector systems, specifically An root system vectors, can be used as targets for latent space configurations (LSC) to ensure the desired LS structure. One of the main LSC advantage is the possibility of training classifier NNs without classification layers, which facilitates training NNs on datasets with extremely large numbers of classes. This paper provides a more general overview of possible vector systems for NN training along with their properties and methods for vector system construction. These systems are used to configure LS of encoders and visual transformers to significantly speed up ImageNet-1K and 50k-600k classes LSC training. It is also shown that using the minimum number of LS dimensions for a specific number of classes results in faster convergence. The latter has potential advantages for reducing the size of vector databases used to store NN embeddings.

Method: Transformer-based real-time-to-event (TTE) model that discretizes time into grid-based tokens, representing each day as a TTE sequence. Unlike previous methods relying on historical patterns, this approach leverages streaming contextual information from passive smartphone data.

Result: Evaluation on real-world study with 93 users shows the method effectively captures irregular departure patterns within individual routines and outperforms baseline models.

Conclusion: The proposed method demonstrates potential for practical deployment in EV charging optimization, contributing to sustainable transportation systems by reducing battery degradation through smarter charging schedules.

Abstract: Electric vehicles (EVs) are key to sustainable mobility, yet their lithium-ion batteries (LIBs) degrade more rapidly under prolonged high states of charge (SOC). This can be mitigated by delaying full charging \ours until just before departure, which requires accurate prediction of user departure times. In this work, we propose Transformer-based real-time-to-event (TTE) model for accurate EV departure prediction. Our approach represents each day as a TTE sequence by discretizing time into grid-based tokens. Unlike previous methods primarily dependent on temporal dependency from historical patterns, our method leverages streaming contextual information to predict departures. Evaluation on a real-world study involving 93 users and passive smartphone data demonstrates that our method effectively captures irregular departure patterns within individual routines, outperforming baseline models. These results highlight the potential for practical deployment of the \ours algorithm and its contribution to sustainable transportation systems.

Method: Proposes two key innovations: 1) Receptive Field Adaptive MoE (RFAMoE) module that allows each channel to adaptively select desired receptive fields during diffusion, and 2) Fusion MoE module that generates K noise signals in parallel, fuses them via routing mechanism, enabling single-step reconstruction instead of multiple inferences.

Result: The framework consistently outperforms diffusion-based state-of-the-art methods on different tasks and datasets. It achieves better reconstruction performance while eliminating the substantial computational cost and latency associated with multiple inference processes.

Conclusion: The proposed MoE-based diffusion framework effectively addresses the unique challenges of medical time series reconstruction, offering superior performance with reduced computational overhead compared to existing approaches.

Abstract: Recent studies show that using diffusion models for time series signal reconstruction holds great promise. However, such approaches remain largely unexplored in the domain of medical time series. The unique characteristics of the physiological time series signals, such as multivariate, high temporal variability, highly noisy, and artifact-prone, make deep learning-based approaches still challenging for tasks such as imputation. Hence, we propose a novel Mixture of Experts (MoE)-based noise estimator within a score-based diffusion framework. Specifically, the Receptive Field Adaptive MoE (RFAMoE) module is designed to enable each channel to adaptively select desired receptive fields throughout the diffusion process. Moreover, recent literature has found that when generating a physiological signal, performing multiple inferences and averaging the reconstructed signals can effectively reduce reconstruction errors, but at the cost of significant computational and latency overhead. We design a Fusion MoE module and innovatively leverage the nature of MoE module to generate K noise signals in parallel, fuse them using a routing mechanism, and complete signal reconstruction in a single inference step. This design not only improves performance over previous methods but also eliminates the substantial computational cost and latency associated with multiple inference processes. Extensive results demonstrate that our proposed framework consistently outperforms diffusion-based SOTA works on different tasks and datasets.

Method: Analysis of hundreds of millions of anonymized user interactions with Comet (Perplexity’s AI-powered browser) and its integrated agent, Comet Assistant. Introduced a hierarchical agentic taxonomy organizing use cases across topic, subtopic, and task levels.

Result: Substantial heterogeneity in adoption: earlier adopters, users from higher GDP/education countries, and digital/knowledge-intensive sector workers more likely to adopt. Productivity & Workflow (57% of queries) and Learning & Research are largest topics. Personal use dominates (55%), followed by professional (30%) and educational (16%). Short-term usage shows stickiness but shifts toward more cognitive topics over time.

Conclusion: AI agent diffusion has significant implications for researchers, businesses, policymakers, and educators. The hierarchical taxonomy provides framework for understanding agent usage patterns, revealing both current adoption trends and evolving user behaviors toward more cognitively oriented applications.

Abstract: This paper presents the first large-scale field study of the adoption, usage intensity, and use cases of general-purpose AI agents operating in open-world web environments. Our analysis centers on Comet, an AI-powered browser developed by Perplexity, and its integrated agent, Comet Assistant. Drawing on hundreds of millions of anonymized user interactions, we address three fundamental questions: Who is using AI agents? How intensively are they using them? And what are they using them for? Our findings reveal substantial heterogeneity in adoption and usage across user segments. Earlier adopters, users in countries with higher GDP per capita and educational attainment, and individuals working in digital or knowledge-intensive sectors – such as digital technology, academia, finance, marketing, and entrepreneurship – are more likely to adopt or actively use the agent. To systematically characterize the substance of agent usage, we introduce a hierarchical agentic taxonomy that organizes use cases across three levels: topic, subtopic, and task. The two largest topics, Productivity & Workflow and Learning & Research, account for 57% of all agentic queries, while the two largest subtopics, Courses and Shopping for Goods, make up 22%. The top 10 out of 90 tasks represent 55% of queries. Personal use constitutes 55% of queries, while professional and educational contexts comprise 30% and 16%, respectively. In the short term, use cases exhibit strong stickiness, but over time users tend to shift toward more cognitively oriented topics. The diffusion of increasingly capable AI agents carries important implications for researchers, businesses, policymakers, and educators, inviting new lines of inquiry into this rapidly emerging class of AI capabilities.

Method: Machine learning framework using BookSim simulations and reverse neural network models (MLP, Conditional Diffusion Model, CVAE) to predict optimal NoC parameters from target performance metrics, with 150,000+ simulation data points across mesh topologies.

Result: Conditional Diffusion Model achieved highest predictive accuracy with MSE of 0.463 on unseen data, and framework reduces design exploration time by several orders of magnitude.

Conclusion: The ML-driven framework provides a practical solution for rapid and scalable NoC co-design, significantly accelerating design space exploration while maintaining accuracy.

Abstract: Network-on-Chip (NoC) design requires exploring a high-dimensional configuration space to satisfy stringent throughput requirements and latency constraints. Traditional design space exploration techniques are often slow and struggle to handle complex, non-linear parameter interactions. This work presents a machine learning-driven framework that automates NoC design space exploration using BookSim simulations and reverse neural network models. Specifically, we compare three architectures - a Multi-Layer Perceptron (MLP),a Conditional Diffusion Model, and a Conditional Variational Autoencoder (CVAE) to predict optimal NoC parameters given target performance metrics. Our pipeline generates over 150,000 simulation data points across varied mesh topologies. The Conditional Diffusion Model achieved the highest predictive accuracy, attaining a mean squared error (MSE) of 0.463 on unseen data. Furthermore, the proposed framework reduces design exploration time by several orders of magnitude, making it a practical solution for rapid and scalable NoC co-design.

Method: HOLE (Homological Observation of Latent Embeddings) extracts topological features from neural activations using persistent homology and presents them through visualization techniques including Sankey diagrams, heatmaps, dendrograms, and blob graphs.

Result: Evaluation on standard datasets with discriminative models shows topological analysis reveals patterns associated with class separation, feature disentanglement, and model robustness, providing complementary insights for understanding deep learning systems.

Conclusion: Topological analysis through HOLE offers a valuable complementary perspective for interpreting and improving deep learning systems by revealing structural patterns in neural representations.

Abstract: Deep learning models have achieved remarkable success across various domains, yet their learned representations and decision-making processes remain largely opaque and hard to interpret. This work introduces HOLE (Homological Observation of Latent Embeddings), a method for analyzing and interpreting deep neural networks through persistent homology. HOLE extracts topological features from neural activations and presents them using a suite of visualization techniques, including Sankey diagrams, heatmaps, dendrograms, and blob graphs. These tools facilitate the examination of representation structure and quality across layers. We evaluate HOLE on standard datasets using a range of discriminative models, focusing on representation quality, interpretability across layers, and robustness to input perturbations and model compression. The results indicate that topological analysis reveals patterns associated with class separation, feature disentanglement, and model robustness, providing a complementary perspective for understanding and improving deep learning systems.

Method: PolyLingua uses a two-level contrastive learning framework: instance-level separation and class-level alignment with adaptive margins, creating compact and well-separated embeddings for closely related languages.

Result: Achieves 99.25% F1 on Amazon Massive dataset and 98.15% F1 on Song dataset (music requests with code-switching), surpassing Sonnet 3.5 while using 10x fewer parameters.

Conclusion: PolyLingua provides accurate language identification ideal for compute- and latency-constrained environments, effectively handling challenging cases like code-switching in music requests.

Abstract: Language identification is a crucial first step in multilingual systems such as chatbots and virtual assistants, enabling linguistically and culturally accurate user experiences. Errors at this stage can cascade into downstream failures, setting a high bar for accuracy. Yet, existing language identification tools struggle with key cases – such as music requests where the song title and user language differ. Open-source tools like LangDetect, FastText are fast but less accurate, while large language models, though effective, are often too costly for low-latency or low-resource settings. We introduce PolyLingua, a lightweight Transformer-based model for in-domain language detection and fine-grained language classification. It employs a two-level contrastive learning framework combining instance-level separation and class-level alignment with adaptive margins, yielding compact and well-separated embeddings even for closely related languages. Evaluated on two challenging datasets – Amazon Massive (multilingual digital assistant utterances) and a Song dataset (music requests with frequent code-switching) – PolyLingua achieves 99.25% F1 and 98.15% F1, respectively, surpassing Sonnet 3.5 while using 10x fewer parameters, making it ideal for compute- and latency-constrained environments.

Method: Uses multivariate Bernoulli distribution to model multi-label data, estimates distribution parameters from observed label frequencies, calculates weights for each label combination, and performs weighted sampling that accounts for label dependencies to achieve target distribution characteristics.

Result: Applied to Web of Science biomedical research articles with 64 topic categories, the approach successfully produced a more balanced sub-sample that preserved category frequency order, reduced frequency differences between most and least common categories, and enhanced representation of minority categories.

Conclusion: The proposed sampling algorithm effectively addresses multi-label sampling challenges by incorporating label dependencies through multivariate Bernoulli modeling, enabling creation of balanced samples that maintain important population characteristics while improving representation of minority labels.

Abstract: Datasets may contain observations with multiple labels. If the labels are not mutually exclusive, and if the labels vary greatly in frequency, obtaining a sample that includes sufficient observations with scarcer labels to make inferences about those labels, and which deviates from the population frequencies in a known manner, creates challenges. In this paper, we consider a multivariate Bernoulli distribution as our underlying distribution of a multi-label problem. We present a novel sampling algorithm that takes label dependencies into account. It uses observed label frequencies to estimate multivariate Bernoulli distribution parameters and calculate weights for each label combination. This approach ensures the weighted sampling acquires target distribution characteristics while accounting for label dependencies. We applied this approach to a sample of research articles from Web of Science labeled with 64 biomedical topic categories. We aimed to preserve category frequency order, reduce frequency differences between most and least common categories, and account for category dependencies. This approach produced a more balanced sub-sample, enhancing the representation of minority categories.

Method: Proposes a new method based on Post-LN architecture that induces algebraic smoothing to prevent oversmoothing without the curse of depth. The approach is parameter-efficient and requires no additional parameters.

Result: Empirical results across five benchmarks demonstrate that the method supports deeper networks (up to 256 layers) and improves performance. The approach successfully avoids both oversmoothing and the curse of depth.

Conclusion: The paper provides a principled solution to the LN placement dilemma by inducing algebraic smoothing in Post-LN architectures, enabling deeper and more effective GNNs without additional parameters.

Abstract: The relationship between Layer Normalization (LN) placement and the oversmoothing phenomenon remains underexplored. We identify a critical dilemma: Pre-LN architectures avoid oversmoothing but suffer from the curse of depth, while Post-LN architectures bypass the curse of depth but experience oversmoothing. To resolve this, we propose a new method based on Post-LN that induces algebraic smoothing, preventing oversmoothing without the curse of depth. Empirical results across five benchmarks demonstrate that our approach supports deeper networks (up to 256 layers) and improves performance, requiring no additional parameters. Key contributions: Theoretical Characterization: Analysis of LN dynamics and their impact on oversmoothing and the curse of depth. A Principled Solution: A parameter-efficient method that induces algebraic smoothing and avoids oversmoothing and the curse of depth. Empirical Validation: Extensive experiments showing the effectiveness of the method in deeper GNNs.

Method: Proposes Global Budget Allocation attack strategy based on theoretical insight that sample influence on value function convergence is proportional to TD error. Formulates attack as global resource allocation problem with closed-form solution where perturbation magnitudes are assigned proportional to TD-error sensitivity under global L2 constraint.

Result: Empirical results on D4RL benchmarks show significant outperformance over baseline strategies, achieving up to 80% performance degradation with minimal perturbations that evade detection by state-of-the-art statistical and spectral defenses.

Conclusion: The proposed attack strategy is more efficient and stealthy than existing approaches, demonstrating the vulnerability of offline RL to targeted poisoning attacks that intelligently allocate perturbation budget.

Abstract: Offline Reinforcement Learning (RL) enables policy optimization from static datasets but is inherently vulnerable to data poisoning attacks. Existing attack strategies typically rely on locally uniform perturbations, which treat all samples indiscriminately. This approach is inefficient, as it wastes the perturbation budget on low-impact samples, and lacks stealthiness due to significant statistical deviations. In this paper, we propose a novel Global Budget Allocation attack strategy. Leveraging the theoretical insight that a sample’s influence on value function convergence is proportional to its Temporal Difference (TD) error, we formulate the attack as a global resource allocation problem. We derive a closed-form solution where perturbation magnitudes are assigned proportional to the TD-error sensitivity under a global L2 constraint. Empirical results on D4RL benchmarks demonstrate that our method significantly outperforms baseline strategies, achieving up to 80% performance degradation with minimal perturbations that evade detection by state-of-the-art statistical and spectral defenses.

Method: Proposed DS FedProxGrad (Decay Step Size FedProxGrad) with Robbins-Monro step-size schedule and mild decay condition on local inexactness, extending the analytical framework with explicit fairness regularization.

Result: Proved that liminf of expected squared gradient norm equals zero, meaning algorithm is asymptotically stationary with convergence rate independent of variance-induced noise floor.

Conclusion: The improved analysis shows FedProxGrad can achieve exact stationarity asymptotically under appropriate step-size scheduling and local inexactness control, overcoming previous limitations of noise-dominated convergence.

Abstract: Recent work \cite{arifgroup} introduced Federated Proximal Gradient \textbf{(\texttt{FedProxGrad})} for solving non-convex composite optimization problems in group fair federated learning. However, the original analysis established convergence only to a \textit{noise-dominated neighborhood of stationarity}, with explicit dependence on a variance-induced noise floor. In this work, we provide an improved asymptotic convergence analysis for a generalized \texttt{FedProxGrad}-type analytical framework with inexact local proximal solutions and explicit fairness regularization. We call this extended analytical framework \textbf{DS \texttt{FedProxGrad}} (Decay Step Size \texttt{FedProxGrad}). Under a Robbins-Monro step-size schedule \cite{robbins1951stochastic} and a mild decay condition on local inexactness, we prove that $\liminf_{r\to\infty} \mathbb{E}[|\nabla F(\mathbf{x}^r)|^2] = 0$, i.e., the algorithm is asymptotically stationary and the convergence rate does not depend on a variance-induced noise floor.

Method: WOLF embeds role-grounded agents (Villager, Werewolf, Seer, Doctor) in a programmable LangGraph state machine with strict night-day cycles, debate turns, and majority voting. Every statement is analyzed as a distinct unit with self-assessed honesty and peer-rated deceptiveness, using a standardized deception taxonomy (omission, distortion, fabrication, misdirection).

Result: Werewolves produce deceptive statements in 31% of turns, while peer detection achieves 71-73% precision with ~52% overall accuracy. Suspicion toward Werewolves rises from ~52% to over 60% across rounds, while suspicion toward truthful roles stabilizes near 44-46%.

Conclusion: WOLF moves deception evaluation beyond static datasets, offering a dynamic, controlled testbed for measuring deceptive and detective capacity in adversarial multi-agent interaction, showing that extended interaction improves recall against liars without compounding errors against truthful roles.

Abstract: Deception is a fundamental challenge for multi-agent reasoning: effective systems must strategically conceal information while detecting misleading behavior in others. Yet most evaluations reduce deception to static classification, ignoring the interactive, adversarial, and longitudinal nature of real deceptive dynamics. Large language models (LLMs) can deceive convincingly but remain weak at detecting deception in peers. We present WOLF, a multi-agent social deduction benchmark based on Werewolf that enables separable measurement of deception production and detection. WOLF embeds role-grounded agents (Villager, Werewolf, Seer, Doctor) in a programmable LangGraph state machine with strict night-day cycles, debate turns, and majority voting. Every statement is a distinct analysis unit, with self-assessed honesty from speakers and peer-rated deceptiveness from others. Deception is categorized via a standardized taxonomy (omission, distortion, fabrication, misdirection), while suspicion scores are longitudinally smoothed to capture both immediate judgments and evolving trust dynamics. Structured logs preserve prompts, outputs, and state transitions for full reproducibility. Across 7,320 statements and 100 runs, Werewolves produce deceptive statements in 31% of turns, while peer detection achieves 71-73% precision with ~52% overall accuracy. Precision is higher for identifying Werewolves, though false positives occur against Villagers. Suspicion toward Werewolves rises from ~52% to over 60% across rounds, while suspicion toward Villagers and the Doctor stabilizes near 44-46%. This divergence shows that extended interaction improves recall against liars without compounding errors against truthful roles. WOLF moves deception evaluation beyond static datasets, offering a dynamic, controlled testbed for measuring deceptive and detective capacity in adversarial multi-agent interaction.

Method: GAIR uses a general-purpose MLLM to jointly process information from multiple GUI-specific models and make decisions. When insufficient information exists, it enters group reflection mode where the general-purpose model provides tailored instructions to specialized models based on their strengths/weaknesses to gather more relevant information.

Result: The framework was evaluated through extensive experiments on GUI benchmarks, demonstrating effectiveness and reliability in GUI automation tasks.

Conclusion: GAIR successfully addresses the challenge of building GUI automation systems by integrating heterogeneous models through information-joint reasoning and group reflection, achieving higher performance than individual specialized models.

Abstract: Building AI systems for GUI automation task has attracted remarkable research efforts, where MLLMs are leveraged for processing user requirements and give operations. However, GUI automation includes a wide range of tasks, from document processing to online shopping, from CAD to video editing. Diversity between particular tasks requires MLLMs for GUI automation to have heterogeneous capabilities and master multidimensional expertise, raising problems on constructing such a model. To address such challenge, we propose GAIR: GUI Automation via Information-Joint Reasoning and Group Reflection, a novel MLLM-based GUI automation agent framework designed for integrating knowledge and combining capabilities from heterogeneous models to build GUI automation agent systems with higher performance. Since different GUI-specific MLLMs are trained on different dataset and thus have different strengths, GAIR introduced a general-purpose MLLM for jointly processing the information from multiple GUI-specific models, further enhancing performance of the agent framework. The general-purpose MLLM also serves as decision maker, trying to execute a reasonable operation based on previously gathered information. When the general-purpose model thinks that there isn’t sufficient information for a reasonable decision, GAIR would transit into group reflection status, where the general-purpose model would provide GUI-specific models with different instructions and hints based on their strengths and weaknesses, driving them to gather information with more significance and accuracy that can support deeper reasoning and decision. We evaluated the effectiveness and reliability of GAIR through extensive experiments on GUI benchmarks.

Method: Proposes an Agent-Centric Data Fabric architecture with four key mechanisms: attention-guided data retrieval, semantic micro-caching for context-driven agent federations, predictive data prefetching, and quorum-based data serving.

Result: The proposed architecture enables agents to access representative data faster and more efficiently while reducing redundant queries, data movement, and inference load across systems by treating data systems as adaptive collaborators rather than static executors.

Conclusion: Frames data systems as adaptive collaborators and outlines new research directions toward behaviorally responsive data infrastructures where caching, probing, and orchestration enable efficient, context-rich data exchange among dynamic, reasoning-driven agents.

Abstract: The rise of multi-agent systems powered by large language models (LLMs) and specialized reasoning agents exposes fundamental limitations in today’s data management architectures. Traditional databases and data fabrics were designed for static, well-defined workloads, whereas agentic systems exhibit dynamic, context-driven, and collaborative behaviors. Agents continuously decompose tasks, shift attention across modalities, and share intermediate results with peers - producing non-deterministic, multi-modal workloads that strain conventional query optimizers and caching mechanisms. We propose an Agent-Centric Data Fabric, a unified architecture that rethinks how data systems serve, optimize, coordinate, and learn from agentic workloads. To achieve this we exploit the concepts of attention-guided data retrieval, semantic micro-caching for context-driven agent federations, predictive data prefetching and quorum-based data serving. Together, these mechanisms enable agents to access representative data faster and more efficiently, while reducing redundant queries, data movement, and inference load across systems. By framing data systems as adaptive collaborators, instead of static executors, we outline new research directions toward behaviorally responsive data infrastructures, where caching, probing, and orchestration jointly enable efficient, context-rich data exchange among dynamic, reasoning-driven agents.

Method: Established connection between Eisenberg-Gale equilibrium convex program and decentralized multi-agent learning. Developed two algorithms: (1) EG-MARL - multi-agent reinforcement learning framework guided by centralized EG equilibrium assignment, and (2) stochastic online optimization mechanism with guided exploration and subset-based fair assignment.

Result: Evaluated on Multi-Agent Particle Environment simulations and Webots-based warehouse proof-of-concept. Both methods preserve fairness-efficiency balance of EG solution under partial observability. EG-MARL achieves near-centralized coordination and reduced travel distances, while online mechanism enables real-time allocation with competitive fairness.

Conclusion: The paper successfully bridges EG equilibrium theory with decentralized multi-agent learning, providing practical solutions for fair spatial task allocation in partially observable environments while maintaining efficiency.

Abstract: Spatial task allocation in systems such as multi-robot delivery or ride-sharing requires balancing efficiency with fair service across tasks. Greedy assignment policies that match each agent to its highest-preference or lowest-cost task can maximize efficiency but often create inequities: some tasks receive disproportionately favorable service (e.g., shorter delays or better matches), while others face long waits or poor allocations. We study fairness in heterogeneous multi-agent systems where tasks vary in preference alignment and urgency. Most existing approaches either assume centralized coordination or largely ignore fairness under partial observability. Distinct from this prior work, we establish a connection between the Eisenberg-Gale (EG) equilibrium convex program and decentralized, partially observable multi-agent learning. Building on this connection, we develop two equilibrium-informed algorithms that integrate fairness and efficiency: (i) a multi-agent reinforcement learning (MARL) framework, EG-MARL, whose training is guided by a centralized EG equilibrium assignment algorithm; and (ii) a stochastic online optimization mechanism that performs guided exploration and subset-based fair assignment as tasks are discovered. We evaluate on Multi-Agent Particle Environment (MPE) simulations across varying team sizes against centralized EG, Hungarian, and Min-Max distance baselines, and also present a Webots-based warehouse proof-of-concept with heterogeneous robots. Both methods preserve the fairness-efficiency balance of the EG solution under partial observability, with EG-MARL achieving near-centralized coordination and reduced travel distances, and the online mechanism enabling real-time allocation with competitive fairness.

Method: Introduces neighbors’ averaged Q-function concept and derives coupled policy gradient expression. Develops DSCP algorithm using geometric 2-horizon sampling (no full Q-table), local parameter estimates, and push-sum protocol for distributed coordination across network.

Result: Proves convergence to first-order stationary point of objective function. Simulations in robot path planning show clear improvement over state-of-the-art methods.

Conclusion: DSCP algorithm effectively addresses networked MARL with interdependent policies, enabling distributed scalable optimization with theoretical guarantees and practical performance improvements.

Abstract: This paper studies networked multi-agent reinforcement learning (NMARL) with interdependent rewards and coupled policies. In this setting, each agent’s reward depends on its own state-action pair as well as those of its direct neighbors, and each agent’s policy is parameterized by its local parameters together with those of its $κ_{p}$-hop neighbors, with $κ_{p}\geq 1$ denoting the coupled radius. The objective of the agents is to collaboratively optimize their policies to maximize the discounted average cumulative reward. To address the challenge of interdependent policies in collaborative optimization, we introduce a novel concept termed the neighbors’ averaged $Q$-function and derive a new expression for the coupled policy gradient. Based on these theoretical foundations, we develop a distributed scalable coupled policy (DSCP) algorithm, where each agent relies only on the state-action pairs of its $κ_{p}$-hop neighbors and the rewards of its $(κ_{p}+1)$-hop neighbors. Specially, in the DSCP algorithm, we employ a geometric 2-horizon sampling method that does not require storing a full $Q$-table to obtain an unbiased estimate of the coupled policy gradient. Moreover, each agent interacts exclusively with its direct neighbors to obtain accurate policy parameters, while maintaining local estimates of other agents’ parameters to execute its local policy and collect samples for optimization. These estimates and policy parameters are updated via a push-sum protocol, enabling distributed coordination of policy updates across the network. We prove that the joint policy produced by the proposed algorithm converges to a first-order stationary point of the objective function. Finally, the effectiveness of DSCP algorithm is demonstrated through simulations in a robot path planning environment, showing clear improvement over state-of-the-art methods.

Method: Proposes Cross-Space Synergy (CSS) with two components: Synergistic Polynomial Fusion (SPF) for representation using low-rank tensor factorization to capture high-order cross-modal interactions, and Pareto Gradient Modulator (PGM) for optimization to steer updates along Pareto-optimal directions across competing objectives.

Result: CSS outperforms existing representative methods on IEMOCAP and MELD datasets in both accuracy and training stability.

Conclusion: CSS effectively addresses complex multimodal scenarios by combining efficient cross-modal interaction modeling with stable multi-objective optimization.

Abstract: Multimodal Emotion Recognition in Conversation (MERC) aims to predict speakers’ emotions by integrating textual, acoustic, and visual cues. Existing approaches either struggle to capture complex cross-modal interactions or experience gradient conflicts and unstable training when using deeper architectures. To address these issues, we propose Cross-Space Synergy (CSS), which couples a representation component with an optimization component. Synergistic Polynomial Fusion (SPF) serves the representation role, leveraging low-rank tensor factorization to efficiently capture high-order cross-modal interactions. Pareto Gradient Modulator (PGM) serves the optimization role, steering updates along Pareto-optimal directions across competing objectives to alleviate gradient conflicts and improve stability. Experiments show that CSS outperforms existing representative methods on IEMOCAP and MELD in both accuracy and training stability, demonstrating its effectiveness in complex multimodal scenarios.

Method: Uses ResNet-221 backbone with two speaker-labeling strategies (Dual-Speaker IDs and Multi-Speaker IDs) to enlarge margin between bona fide and generated speech. Proposes discriminator-based sub-judge systems reusing internal features from HiFi-GAN and BigVGAN discriminators, aggregated via multi-query multi-head attentive statistics pooling (MQMHA).

Result: Experimental results on the SpoofCeleb corpus show the system design effectively improves agnostic detection cost function (a-DCF), demonstrating effectiveness against high-quality TTS attacks.

Conclusion: The proposed approach combining speaker-labeling strategies and discriminator-based sub-judge systems is effective for spoof-aware speaker verification against sophisticated TTS attacks, as validated by improved a-DCF performance on the SpoofCeleb corpus.

Abstract: This paper describes our submission to the WildSpoof Challenge Track 2, which focuses on spoof-aware speaker verification (SASV) in the presence of high-quality text-to-speech (TTS) attacks. We adopt a ResNet-221 back-bone and study two speaker-labeling strategies, namelyDual-Speaker IDs and Multi-Speaker IDs, to explicitly enlarge the margin between bona fide and generated speech in the embedding space. In addition, we propose discriminator-based sub-judge systems that reuse internal features from HiFi-GAN and BigVGAN discriminators, aggregated via multi-query multi-head attentive statistics pooling(MQMHA). Experimental results on the SpoofCeleb corpus show that our system design is effective in improving agnostic detection cost function (a-DCF).

Method: Perceptual experiment with 54 listeners (28 Spanish, 26 Japanese natives) classifying 80 voices (50% artificial) as natural or synthetic while justifying choices. Variables included language (Spanish/Japanese), speech style (audiobooks/interviews), and voice familiarity (familiar/unfamiliar). Qualitative analysis of reasoning behind decisions.

Result: Average accuracy of 59.11% with higher performance on authentic samples. Listeners relied primarily on suprasegmental and higher-level features (intonation, rhythm, fluency, pauses, speed, breathing, laughter) over segmental features. Both shared cues and cross-linguistic differences in conceptualizing “humanness” were found between Japanese and Spanish listeners.

Conclusion: Human perceptual strategies for distinguishing natural from artificial speech are complex, with prosody and spontaneous speech phenomena (disfluencies) playing crucial roles. Findings align with prior research emphasizing suprasegmental features and highlight cross-linguistic variations in how listeners perceive vocal naturalness.

Abstract: Audio deepfakes have reached a level of realism that makes it increasingly difficult to distinguish between human and artificial voices, which poses risks such as identity theft or spread of disinformation. Despite these concerns, research on humans’ ability to identify deepfakes is limited, with most studies focusing on English and very few exploring the reasons behind listeners’ perceptual decisions. This study addresses this gap through a perceptual experiment in which 54 listeners (28 native Spanish speakers and 26 native Japanese speakers) classified voices as natural or synthetic, and justified their choices. The experiment included 80 stimuli (50% artificial), organized according to three variables: language (Spanish/Japanese), speech style (audiobooks/interviews), and familiarity with the voice (familiar/unfamiliar). The goal was to examine how these variables influence detection and to analyze qualitatively the reasoning behind listeners’ perceptual decisions. Results indicate an average accuracy of 59.11%, with higher performance on authentic samples. Judgments of vocal naturalness rely on a combination of linguistic and non-linguistic cues. Comparing Japanese and Spanish listeners, our qualitative analysis further reveals both shared cues and notable cross-linguistic differences in how listeners conceptualize the “humanness” of speech. Overall, participants relied primarily on suprasegmental and higher-level or extralinguistic characteristics - such as intonation, rhythm, fluency, pauses, speed, breathing, and laughter - over segmental features. These findings underscore the complexity of human perceptual strategies in distinguishing natural from artificial speech and align partly with prior research emphasizing the importance of prosody and phenomena typical of spontaneous speech, such as disfluencies.

Method: Uses a neural network with: 1) training strategy using controlled ratios of speech and singing samples to improve discrimination, 2) computationally efficient architecture for reduced inference runtime, and 3) new evaluation metric for assessing SAD robustness in mixed speech-singing scenarios.

Result: Experiments on challenging multi-genre dataset show SR-SAD maintains high speech detection accuracy (AUC = 0.919) while effectively rejecting singing.

Conclusion: By explicitly learning to distinguish between speech and singing, SR-SAD enables more reliable speech activity detection in mixed speech-singing scenarios.

Abstract: Speech Activity Detection (SAD) systems often misclassify singing as speech, leading to degraded performance in applications such as dialogue enhancement and automatic speech recognition. We introduce Singing-Robust Speech Activity Detection ( SR-SAD ), a neural network designed to robustly detect speech in the presence of singing. Our key contributions are: i) a training strategy using controlled ratios of speech and singing samples to improve discrimination, ii) a computationally efficient model that maintains robust performance while reducing inference runtime, and iii) a new evaluation metric tailored to assess SAD robustness in mixed speech-singing scenarios. Experiments on a challenging dataset spanning multiple musical genres show that SR-SAD maintains high speech detection accuracy (AUC = 0.919) while rejecting singing. By explicitly learning to distinguish between speech and singing, SR-SAD enables more reliable SAD in mixed speech-singing scenarios.

Method: Unified embedding framework with separate speech and text encoders, followed by shared scaling layer that maps both modalities into common embedding space. Theoretical analysis of speech retrieval challenges and architectural principles for speech-to-document matching.

Result: Model reduces pipeline latency by 50% while achieving higher retrieval accuracy compared to traditional two-stage methods. Robust across diverse acoustic conditions and speaker variations.

Conclusion: Proposes new paradigm for multimodal SLLMs retrieval systems with end-to-end speech retrieval framework that eliminates ASR dependency and improves efficiency and accuracy.

Abstract: Embedding-based retrieval models have made significant strides in retrieval-augmented generation (RAG) techniques for text and multimodal large language models (LLMs) applications. However, when it comes to speech larage language models (SLLMs), these methods are limited to a two-stage process, where automatic speech recognition (ASR) is combined with text-based retrieval. This sequential architecture suffers from high latency and error propagation. To address these limitations, we propose a unified embedding framework that eliminates the need for intermediate text representations. Specifically, the framework includes separate speech and text encoders, followed by a shared scaling layer that maps both modalities into a common embedding space. Our model reduces pipeline latency by 50% while achieving higher retrieval accuracy compared to traditional two-stage methods. We also provide a theoretical analysis of the challenges inherent in end-to-end speech retrieval and introduce architectural principles for effective speech-to-document matching. Extensive experiments demonstrate the robustness of our approach across diverse acoustic conditions and speaker variations, paving the way for a new paradigm in multimodal SLLMs retrieval systems.

Method: Formalize understanding of optimal ASR performance under lossy input compression, then propose a parametric dithering technique for low-complexity speech compression pipeline. The method adapts to meet performance targets or stay within entropy constraints.

Result: Method performs well at 1-bit resolution with 25% relative CER improvement, and shows improvements of 32.4% and 33.5% at 2- and 3-bit resolution respectively. Second dither choice yields reduced data rate.

Conclusion: Parametric dithering effectively improves ASR performance under lossy input compression, with significant CER improvements at low bit resolutions while maintaining adaptability for performance targets and entropy constraints.

Abstract: Dithering is a technique commonly used to improve the perceptual quality of lossy data compression. In this work, we analytically and experimentally justify the use of dithering for ASR input compression. We formalize an understanding of optimal ASR performance under lossy input compression and leverage this to propose a parametric dithering technique for a low-complexity speech compression pipeline. The method performs well at 1-bit resolution, showing a 25% relative CER improvement, while also demonstrating improvements of 32.4% and 33.5% at 2- and 3-bit resolution, respectively, with our second dither choice yielding a reduced data rate. The proposed codec is adaptable to meet performance targets or stay within entropy constraints.

Method: Proposes Agreement Disagreement Guided Knowledge Transfer (ADGKT) with agreement component (GradVac for gradient alignment and LogitNorm for logit regulation) and disagreement component (Disagreement Restriction and ensemble strategy).

Result: Extensive experiments demonstrate the method’s effectiveness and superiority in achieving robust and balanced knowledge transfer across heterogeneous HSI scenes.

Conclusion: ADGKT framework successfully addresses gradient conflicts and information loss problems, enhancing cross-scene hyperspectral imaging knowledge transfer by integrating both agreement and disagreement mechanisms.

Abstract: Knowledge transfer plays a crucial role in cross-scene hyperspectral imaging (HSI). However, existing studies often overlook the challenges of gradient conflicts and dominant gradients that arise during the optimization of shared parameters. Moreover, many current approaches fail to simultaneously capture both agreement and disagreement information, relying only on a limited shared subset of target features and consequently missing the rich, diverse patterns present in the target scene. To address these issues, we propose an Agreement Disagreement Guided Knowledge Transfer (ADGKT) framework that integrates both mechanisms to enhance cross-scene transfer. The agreement component includes GradVac, which aligns gradient directions to mitigate conflicts between source and target domains, and LogitNorm, which regulates logit magnitudes to prevent domination by a single gradient source. The disagreement component consists of a Disagreement Restriction (DiR) and an ensemble strategy, which capture diverse predictive target features and mitigate the loss of critical target information. Extensive experiments demonstrate the effectiveness and superiority of the proposed method in achieving robust and balanced knowledge transfer across heterogeneous HSI scenes.

Method: Proposed framework uses EfficientNetV2-M backbone with Automatic Mixed Precision training, AdamW optimizer, Cosine Annealing learning rate scheduling, and Exponential Moving Average regularization. Trained on 18,080 chest X-ray images across 5 disease categories (Cardiomegaly, COVID-19, Normal, Pneumonia, Tuberculosis).

Result: Achieved mean test accuracy of 96.45% (vs 95.30% baseline, p<0.001) and macro-averaged F1-score of 91.08% (p<0.001). COVID-19 detection: 99.95% accuracy, Tuberculosis: 99.97% accuracy. Despite 6.8x more parameters, training time reduced by 11.4% and performance stability increased by 22.7%.

Conclusion: The framework serves as an effective decision-support tool for pandemic response, tuberculosis screening, and regular thoracic disease assessment in healthcare facilities, especially in resource-limited settings.

Abstract: The interpretation of Chest X-ray is an important diagnostic issue in clinical practice and especially in the resource-limited setting where the shortage of radiologists plays a role in delayed diagnosis and poor patient outcomes. Although the original CheXNet architecture has shown potential in automated analysis of chest radiographs, DenseNet-121 backbone is computationally inefficient and poorly single-label classifier. To eliminate such shortcomings, we suggest a better classification framework of chest disease that relies on EfficientNetV2-M and incorporates superior training approaches such as Automatic Mixed Precision training, AdamW, Cosine Annealing learning rate scheduling, and Exponential Moving Average regularization. We prepared a dataset of 18,080 chest X-ray images of three source materials of high authority and representing five key clinically significant disease categories which included Cardiomegaly, COVID-19, Normal, Pneumonia, and Tuberculosis. To achieve statistical reliability and reproducibility, nine independent experimental runs were run. The suggested architecture showed significant gains with mean test accuracy of 96.45 percent compared to 95.30 percent at baseline (p less than 0.001) and macro-averaged F1-score increased to 91.08 percent (p less than 0.001). Critical infectious diseases showed near-perfect classification performance with COVID-19 detection having 99.95 percent accuracy and Tuberculosis detection having 99.97 percent accuracy. Although 6.8 times more parameters are included, the training time was reduced by 11.4 percent and performance stability was increased by 22.7 percent. This framework presents itself as a decision-support tool that can be used to respond to a pandemic, screen tuberculosis, and assess thoracic disease regularly in various healthcare facilities.

Method: 1) ETAS framework on SKINCON dataset to optimize ViT for dermatological features, fine-tuned on DermNet dataset; 2) StackNet architecture integrating multiple binary ViT classifiers; 3) RAG pipeline using Google Gemini 2.5 Pro LLM for diagnostic explanations; 4) Experimental evaluation on 23 skin disease categories; 5) Domain-expert evaluation by 8 medical doctors.

Result: Achieved 56.30% overall F1-score, surpassing SkinGPT-4 (48.51%) by 16.06% performance increase. Domain-expert evaluation showed 92% agreement rate with AI assistant assessments. Created proof-of-concept prototype for real-world applications.

Conclusion: The DermETAS-SNA LLM Assistant demonstrates superior performance in skin disease classification and provides clinically useful explanations, showing practical feasibility for real-world dermatological applications.

Abstract: Our work introduces the DermETAS-SNA LLM Assistant that integrates Dermatology-focused Evolutionary Transformer Architecture Search with StackNet Augmented LLM. The assistant dynamically learns skin-disease classifiers and provides medically informed descriptions to facilitate clinician-patient interpretation. Contributions include: (1) Developed an ETAS framework on the SKINCON dataset to optimize a Vision Transformer (ViT) tailored for dermatological feature representation and then fine-tuned binary classifiers for each of the 23 skin disease categories in the DermNet dataset to enhance classification performance; (2) Designed a StackNet architecture that integrates multiple fine-tuned binary ViT classifiers to enhance predictive robustness and mitigate class imbalance issues; (3) Implemented a RAG pipeline, termed Diagnostic Explanation and Retrieval Model for Dermatology, which harnesses the capabilities of the Google Gemini 2.5 Pro LLM architecture to generate personalized, contextually informed diagnostic descriptions and explanations for patients, leveraging a repository of verified dermatological materials; (4) Performed extensive experimental evaluations on 23 skin disease categories to demonstrate performance increase, achieving an overall F1-score of 56.30% that surpasses SkinGPT-4 (48.51%) by a considerable margin, representing a performance increase of 16.06%; (5) Conducted a domain-expert evaluation, with eight licensed medical doctors, of the clinical responses generated by our AI assistant for seven dermatological conditions. Our results show a 92% agreement rate with the assessments provided by our AI assistant (6) Created a proof-of-concept prototype that fully integrates our DermETAS-SNA LLM into our AI assistant to demonstrate its practical feasibility for real-world clinical and educational applications.

Method: The authors extend causal attribution frameworks to high-dimensional segmentation tasks by modeling the data-generating process through a causal graph and employing Shapley values to fairly attribute performance changes to individual mechanisms. The framework addresses unique challenges in medical imaging: high-dimensional outputs, limited samples, and complex mechanism interactions.

Result: Validation on multiple sclerosis (MS) lesion segmentation across 4 centers and 7 annotators reveals context-dependent failure modes: annotation protocol shifts dominate when crossing annotators (7.4% ± 8.9% DSC attribution), while acquisition shifts dominate when crossing imaging centers (6.5% ± 9.1%).

Conclusion: The mechanism-specific quantification enables practitioners to prioritize targeted interventions based on deployment context, providing actionable insights for improving segmentation model robustness under different types of distribution shifts.

Abstract: Deep learning models for medical image segmentation suffer significant performance drops due to distribution shifts, but the causal mechanisms behind these drops remain poorly understood. We extend causal attribution frameworks to high-dimensional segmentation tasks, quantifying how acquisition protocols and annotation variability independently contribute to performance degradation. We model the data-generating process through a causal graph and employ Shapley values to fairly attribute performance changes to individual mechanisms. Our framework addresses unique challenges in medical imaging: high-dimensional outputs, limited samples, and complex mechanism interactions. Validation on multiple sclerosis (MS) lesion segmentation across 4 centers and 7 annotators reveals context-dependent failure modes: annotation protocol shifts dominate when crossing annotators (7.4% $\pm$ 8.9% DSC attribution), while acquisition shifts dominate when crossing imaging centers (6.5% $\pm$ 9.1%). This mechanism-specific quantification enables practitioners to prioritize targeted interventions based on deployment context.

Method: Proposes Semantic Information Transport Protocol (SITP) that verifies only packet headers while retaining potentially corrupted payloads for semantic decoding. Establishes cross-layer analytical model linking SNR to packet-loss probability. Develops cross-image feature interleaving mechanism to mitigate burst losses.

Result: SITP offers lower latency than TCP with comparable reliability at low SNRs, matches UDP-level latency while delivering superior reconstruction quality. Cross-image semantic interleaving effectively mitigates degradation from bursty packet losses.

Conclusion: SITP provides a practical solution for semantic communication in 6G networks, balancing reliability and latency through header-only verification and semantic-aware error handling with cross-layer modeling and feature interleaving.

Abstract: With the evolution of 6G networks, modern communication systems are facing unprecedented demands for high reliability and low latency. However, conventional transport protocols are designed for bit-level reliability, failing to meet the semantic robustness requirements. To address this limitation, this paper proposes a novel Semantic Information Transport Protocol (SITP), which achieves TCP-level reliability and UDP level latency by verifying only packet headers while retaining potentially corrupted payloads for semantic decoding. Building upon SITP, a cross-layer analytical model is established to quantify packet-loss probability across the physical, data-link, network, transport, and application layers. The model provides a unified probabilistic formulation linking signal noise rate (SNR) and packet-loss rate, offering theoretical foundation into end-to-end semantic transmission. Furthermore, a cross-image feature interleaving mechanism is developed to mitigate consecutive burst losses by redistributing semantic features across multiple correlated images, thereby enhancing robustness in burst-fade channels. Extensive experiments show that SITP offers lower latency than TCP with comparable reliability at low SNRs, while matching UDP-level latency and delivering superior reconstruction quality. In addition, the proposed cross-image semantic interleaving mechanism further demonstrates its effectiveness in mitigating degradation caused by bursty packet losses.

Method: Proposes NOC4SC framework using Swin Transformer for unified encoder-decoder architecture with shared parameters across users, plus adaptive NOC and SNR Modulation (NSM) block to dynamically regulate SNR and generate approximately orthogonal semantic features in distinct subspaces.

Result: NOC4SC achieves comparable performance to DeepJSCC-PNOMA and outperforms other multi-user semantic communication baseline methods in extensive experiments.

Conclusion: NOC4SC provides a bandwidth-efficient semantic communication paradigm that enables simultaneous same-frequency transmission while effectively mitigating inter-user interference through adaptive SNR modulation and semantic feature orthogonalization.

Abstract: With the explosive growth of connected devices and emerging applications, current wireless networks are encountering unprecedented demands for massive user access, where the inter-user interference has become a critical challenge to maintaining high quality of service (QoS) in multi-user communication systems. To tackle this issue, we propose a bandwidth-efficient semantic communication paradigm termed Non-Orthogonal Codewords for Semantic Communication (NOC4SC), which enables simultaneous same-frequency transmission without spectrum spreading. By leveraging the Swin Transformer, the proposed NOC4SC framework enables each user to independently extract semantic features through a unified encoder-decoder architecture with shared network parameters across all users, which ensures that the user’s data remains protected from unauthorized decoding. Furthermore, we introduce an adaptive NOC and SNR Modulation (NSM) block, which employs deep learning to dynamically regulate SNR and generate approximately orthogonal semantic features within distinct feature subspaces, thereby effectively mitigating inter-user interference. Extensive experiments demonstrate the proposed NOC4SC achieves comparable performance to the DeepJSCC-PNOMA and outperforms other multi-user SemCom baseline methods.

Method: Integrates Implicit Neural Representation (INR) into k-space domain to continuously model multi-directional dipole responses and explicitly complete cone-null regions. Uses frequency-domain residual-weighted Dipole Loss for physical consistency. Combines 3D U-Net-based QSMnet backbone with INR module through alternating optimization for end-to-end joint training.

Result: Outperforms conventional and recent deep-learning approaches across multiple quantitative metrics on 2016 QSM Reconstruction Challenge, multi-orientation GRE dataset, and clinical data. Shows notable advantages in structural recovery within cone-null regions and artifact suppression.

Conclusion: QSMnet-INR effectively alleviates single-orientation QSM’s ill-posedness without requiring multi-orientation acquisition, achieving high accuracy, robustness, and strong cross-scenario generalization with potential for clinical translation.

Abstract: Quantitative Susceptibility Mapping (QSM) quantifies tissue magnetic susceptibility from magnetic-resonance phase data and plays a crucial role in brain microstructure imaging, iron-deposition assessment, and neurological-disease research. However, single-orientation QSM inversion remains highly ill-posed because the dipole kernel exhibits a cone-null region in the Fourier domain, leading to streaking artifacts and structural loss. To overcome this limitation, we propose QSMnet-INR, a deep, physics-informed framework that integrates an Implicit Neural Representation (INR) into the k-space domain. The INR module continuously models multi-directional dipole responses and explicitly completes the cone-null region, while a frequency-domain residual-weighted Dipole Loss enforces physical consistency. The overall network combines a 3D U-Net-based QSMnet backbone with the INR module through alternating optimization for end-to-end joint training. Experiments on the 2016 QSM Reconstruction Challenge, a multi-orientation GRE dataset, and both in-house and public single-orientation clinical data demonstrate that QSMnet-INR consistently outperforms conventional and recent deep-learning approaches across multiple quantitative metrics. The proposed framework shows notable advantages in structural recovery within cone-null regions and in artifact suppression. Ablation studies further confirm the complementary contributions of the INR module and Dipole Loss to detail preservation and physical stability. Overall, QSMnet-INR effectively alleviates the ill-posedness of single-orientation QSM without requiring multi-orientation acquisition, achieving high accuracy, robustness, and strong cross-scenario generalization-highlighting its potential for clinical translation.

Method: 1) Virtual Patient Engine models continuous latent disease trajectories from sparse clinical anchors to synthesize physiologically plausible, fully labeled 3D cohorts. 2) Self-Reinforcing Interleaved Validation provides leakage-free online evaluation with progressively challenging synthetic samples. 3) Dynamic Lattice-of-Experts organizes specialized networks in pathology-aware topology and activates relevant experts per input.

Result: Outperforms four state-of-the-art FSL methods by 4.2-11% Dice with only 7 labeled anchors, approaches fully supervised performance (within 1% Dice) with only 19 anchors, shows superior harmonization across four vendors, and generalizes to unseen pathologies in strict OOD setting.

Conclusion: PathCo-LatticE demonstrates that pathology-guided synthetic supervision can enable robust zero-shot generalization in few-shot learning for medical imaging, eliminating the need for real validation data and target-domain fine-tuning while achieving near-fully-supervised performance.

Abstract: Few-shot learning (FSL) mitigates data scarcity in cardiac MRI segmentation but typically relies on semi-supervised techniques sensitive to domain shifts and validation bias, restricting zero-shot generalizability. We propose PathCo-LatticE, a fully supervised FSL framework that replaces unlabeled data with pathology-guided synthetic supervision. First, our Virtual Patient Engine models continuous latent disease trajectories from sparse clinical anchors, using generative modeling to synthesize physiologically plausible, fully labeled 3D cohorts. Second, Self-Reinforcing Interleaved Validation (SIV) provides a leakage-free protocol that evaluates models online with progressively challenging synthetic samples, eliminating the need for real validation data. Finally, a dynamic Lattice-of-Experts (LoE) organizes specialized networks within a pathology-aware topology and activates the most relevant experts per input, enabling robust zero-shot generalization to unseen data without target-domain fine-tuning. We evaluated PathCo-LatticE in a strict out-of-distribution (OOD) setting, deriving all anchors and severity statistics from a single-source domain (ACDC) and performing zero-shot testing on the multi-center, multi-vendor M&Ms dataset. PathCo-LatticE outperforms four state-of-the-art FSL methods by 4.2-11% Dice starting from only 7 labeled anchors, and approaches fully supervised performance (within 1% Dice) with only 19 labeled anchors. The method shows superior harmonization across four vendors and generalization to unseen pathologies. [Code will be made publicly available].

Method: Proposes a retouch transfer approach that learns from professional edits through before-after image pairs. Uses a context-aware Implicit Neural Representation that learns to apply edits adaptively based on image content and context, capable of learning from a single example. Extracts implicit transformations from reference edits and adaptively applies them to new images.

Result: Introduces a comprehensive Photo Retouching Dataset with 100,000 high-quality images edited using over 170 professional Adobe Lightroom presets. The approach surpasses existing methods in photo retouching and enhances performance in related tasks like Gamut Mapping and Raw Reconstruction.

Conclusion: The work presents a significant step toward making sophisticated photo editing more accessible while maintaining high-fidelity results, bridging the gap between professional editing capabilities and automated solutions. Source code and dataset are publicly available.

Abstract: Professional photo editing remains challenging, requiring extensive knowledge of imaging pipelines and significant expertise. While recent deep learning approaches, particularly style transfer methods, have attempted to automate this process, they often struggle with output fidelity, editing control, and complex retouching capabilities. We propose a novel retouch transfer approach that learns from professional edits through before-after image pairs, enabling precise replication of complex editing operations. We develop a context-aware Implicit Neural Representation that learns to apply edits adaptively based on image content and context, and is capable of learning from a single example. Our method extracts implicit transformations from reference edits and adaptively applies them to new images. To facilitate this research direction, we introduce a comprehensive Photo Retouching Dataset comprising 100,000 high-quality images edited using over 170 professional Adobe Lightroom presets. Through extensive evaluation, we demonstrate that our approach not only surpasses existing methods in photo retouching but also enhances performance in related image reconstruction tasks like Gamut Mapping and Raw Reconstruction. By bridging the gap between professional editing capabilities and automated solutions, our work presents a significant step toward making sophisticated photo editing more accessible while maintaining high-fidelity results. The source code and the dataset are publicly available at https://omaralezaby.github.io/inretouch .

Method: Experiments with various data augmentation techniques, particularly strong augmentations involving compositions of multiple techniques, to identify which ones improve semantic segmentation performance for medical images.

Result: Found that existing data augmentations don’t always improve performance for medical image segmentation, and identified alternative augmentations that provide better results.

Conclusion: Strong data augmentation strategies need to be carefully selected for medical image segmentation tasks, as standard approaches may not be optimal; specific augmentations can provide improved performance.

Abstract: Self-supervised contrastive learning is among the recent representation learning methods that have shown performance gains in several downstream tasks including semantic segmentation. This paper evaluates strong data augmentation, one of the most important components for self-supervised contrastive learning’s improved performance. Strong data augmentation involves applying the composition of multiple augmentation techniques on images. Surprisingly, we find that the existing data augmentations do not always improve performance for semantic segmentation for medical images. We experiment with other augmentations that provide improved performance.