Method: Two-dimensional framework measuring neutrality and semantic fidelity, comparing few-shot prompting across multiple LLMs, fine-tuning selected models, and applying targeted cleaning to boost task relevance.

Result: Open-weight LLMs outperform the only existing dedicated Italian GNR model, and fine-tuned models match/exceed best open-weight LLM performance at fraction of the size.

Conclusion: There’s a trade-off between optimizing training data for neutrality and meaning preservation, and smaller fine-tuned models can achieve competitive performance in gender-neutral rewriting.

Abstract: Gender-neutral rewriting (GNR) aims to reformulate text to eliminate unnecessary gender specifications while preserving meaning, a particularly challenging task in grammatical-gender languages like Italian. In this work, we conduct the first systematic evaluation of state-of-the-art large language models (LLMs) for Italian GNR, introducing a two-dimensional framework that measures both neutrality and semantic fidelity to the input. We compare few-shot prompting across multiple LLMs, fine-tune selected models, and apply targeted cleaning to boost task relevance. Our findings show that open-weight LLMs outperform the only existing model dedicated to GNR in Italian, whereas our fine-tuned models match or exceed the best open-weight LLM’s performance at a fraction of its size. Finally, we discuss the trade-off between optimizing the training data for neutrality and meaning preservation.

Method: Developed five-stage hierarchical annotation schema and used active learning to create balanced dataset of 1044 annotated sentences from FOMC transcripts.

Result: Top LLM achieved 0.80 accuracy for opinion classification but only 0.61 for monetary policy stance classification (human baseline: 0.89).

Conclusion: Op-Fed enables better model training, confidence calibration, and serves as seed dataset for future annotation of monetary policy opinions.

Abstract: The U.S. Federal Open Market Committee (FOMC) regularly discusses and sets monetary policy, affecting the borrowing and spending decisions of millions of people. In this work, we release Op-Fed, a dataset of 1044 human-annotated sentences and their contexts from FOMC transcripts. We faced two major technical challenges in dataset creation: imbalanced classes – we estimate fewer than 8% of sentences express a non-neutral stance towards monetary policy – and inter-sentence dependence – 65% of instances require context beyond the sentence-level. To address these challenges, we developed a five-stage hierarchical schema to isolate aspects of opinion, monetary policy, and stance towards monetary policy as well as the level of context needed. Second, we selected instances to annotate using active learning, roughly doubling the number of positive instances across all schema aspects. Using Op-Fed, we found a top-performing, closed-weight LLM achieves 0.80 zero-shot accuracy in opinion classification but only 0.61 zero-shot accuracy classifying stance towards monetary policy – below our human baseline of 0.89. We expect Op-Fed to be useful for future model training, confidence calibration, and as a seed dataset for future annotation efforts.

Method: The track comprised two subtasks: (1) Dialogue-level, Multi-dimensional Automatic Evaluation Metrics across 10 dimensions, and (2) Multilingual and Multicultural Safety Detection using baseline models like Llama-3-8B and Llama-Guard-3-1B.

Result: For Task 1, Llama-3-8B baseline achieved only 0.1681 Spearman’s correlation, indicating substantial room for improvement. For Task 2, teams outperformed baseline on multilingual safety (top ROC-AUC 0.9648) but baseline was superior on cultural safety (0.5126 ROC-AUC).

Conclusion: The results highlight critical needs for better multi-dimensional evaluation metrics and culturally-aware safety detection in dialogue systems, showing current methods remain inadequate for comprehensive assessment.

Abstract: The rapid advancement of Large Language Models (LLMs) has intensified the need for robust dialogue system evaluation, yet comprehensive assessment remains challenging. Traditional metrics often prove insufficient, and safety considerations are frequently narrowly defined or culturally biased. The DSTC12 Track 1, “Dialog System Evaluation: Dimensionality, Language, Culture and Safety,” is part of the ongoing effort to address these critical gaps. The track comprised two subtasks: (1) Dialogue-level, Multi-dimensional Automatic Evaluation Metrics, and (2) Multilingual and Multicultural Safety Detection. For Task 1, focused on 10 dialogue dimensions, a Llama-3-8B baseline achieved the highest average Spearman’s correlation (0.1681), indicating substantial room for improvement. In Task 2, while participating teams significantly outperformed a Llama-Guard-3-1B baseline on the multilingual safety subset (top ROC-AUC 0.9648), the baseline proved superior on the cultural subset (0.5126 ROC-AUC), highlighting critical needs in culturally-aware safety. This paper describes the datasets and baselines provided to participants, as well as submission evaluation results for each of the two proposed subtasks.

Method: Created MAVL (first multilingual multimodal benchmark) and proposed SylAVL-CoT model that uses audio-video cues with syllabic constraints and chain-of-thought reasoning for singable lyrics translation.

Result: SylAVL-CoT significantly outperforms text-based models in both singability and contextual accuracy, demonstrating the value of multimodal approaches.

Conclusion: Multimodal and multilingual approaches are essential for high-quality lyrics translation, with audio-video integration enabling richer and more expressive translations than text-only methods.

Abstract: Lyrics translation requires both accurate semantic transfer and preservation of musical rhythm, syllabic structure, and poetic style. In animated musicals, the challenge intensifies due to alignment with visual and auditory cues. We introduce Multilingual Audio-Video Lyrics Benchmark for Animated Song Translation (MAVL), the first multilingual, multimodal benchmark for singable lyrics translation. By integrating text, audio, and video, MAVL enables richer and more expressive translations than text-only approaches. Building on this, we propose Syllable-Constrained Audio-Video LLM with Chain-of-Thought SylAVL-CoT, which leverages audio-video cues and enforces syllabic constraints to produce natural-sounding lyrics. Experimental results demonstrate that SylAVL-CoT significantly outperforms text-based models in singability and contextual accuracy, emphasizing the value of multimodal, multilingual approaches for lyrics translation.

Method: Built transfer learning matrix and applied dimensionality reduction to analyze 10 models, identifying latent abilities (Reasoning, Sentiment Classification, NLU, Arithmetic) and studying side effects of transfer learning.

Result: Performance improvements defy surface-level dataset similarity explanations; hidden statistical factors (class distribution, generation length proclivities) and specific linguistic features are more influential than source data quality.

Conclusion: The work provides insights into complex transfer learning dynamics, enabling more predictable and effective LLM adaptation by focusing on underlying statistical and linguistic factors rather than superficial dataset characteristics.

Abstract: Large language models are increasingly deployed across diverse applications. This often includes tasks LLMs have not encountered during training. This implies that enumerating and obtaining the high-quality training data for all tasks is infeasible. Thus, we often need to rely on transfer learning using datasets with different characteristics, and anticipate out-of-distribution requests. Motivated by this practical need, we propose an analysis framework, building a transfer learning matrix and dimensionality reduction, to dissect these cross-task interactions. We train and analyze 10 models to identify latent abilities (e.g., Reasoning, Sentiment Classification, NLU, Arithmetic) and discover the side effects of the transfer learning. Our findings reveal that performance improvements often defy explanations based on surface-level dataset similarity or source data quality. Instead, hidden statistical factors of the source dataset, such as class distribution and generation length proclivities, alongside specific linguistic features, are actually more influential. This work offers insights into the complex dynamics of transfer learning, paving the way for more predictable and effective LLM adaptation.

Method: Identified Ambiguity-Encoding Neurons (AENs) during pre-filling stage, trained probes on these neurons for ambiguity detection, and performed layerwise analysis and neuron manipulation experiments.

Result: AEN-based probes outperform prompting and representation baselines, generalize across datasets, and emerge from shallow layers. Neuron manipulation enables controlling LLM behavior from answering to abstention.

Conclusion: LLMs form compact internal representations of question ambiguity that are interpretable and controllable, enabling improved ambiguity handling in language models.

Abstract: Ambiguity is pervasive in real-world questions, yet large language models (LLMs) often respond with confident answers rather than seeking clarification. In this work, we show that question ambiguity is linearly encoded in the internal representations of LLMs and can be both detected and controlled at the neuron level. During the model’s pre-filling stage, we identify that a small number of neurons, as few as one, encode question ambiguity information. Probes trained on these Ambiguity-Encoding Neurons (AENs) achieve strong performance on ambiguity detection and generalize across datasets, outperforming prompting-based and representation-based baselines. Layerwise analysis reveals that AENs emerge from shallow layers, suggesting early encoding of ambiguity signals in the model’s processing pipeline. Finally, we show that through manipulating AENs, we can control LLM’s behavior from direct answering to abstention. Our findings reveal that LLMs form compact internal representations of question ambiguity, enabling interpretable and controllable behavior.

Method: Created CL²GEC benchmark with 10,000 human-annotated sentences spanning 10 disciplines, evaluated large language models under sequential tuning, parameter-efficient adaptation, and four representative CL algorithms using both standard GEC metrics and continual learning metrics.

Result: Regularization-based methods were found to mitigate forgetting more effectively than replay-based or naive sequential approaches in handling grammatical error correction across diverse academic domains.

Conclusion: The benchmark provides a rigorous foundation for future research in adaptive grammatical error correction across diverse academic domains, demonstrating the effectiveness of continual learning approaches for multi-disciplinary CGEC.

Abstract: The growing demand for automated writing assistance in diverse academic domains highlights the need for robust Chinese Grammatical Error Correction (CGEC) systems that can adapt across disciplines. However, existing CGEC research largely lacks dedicated benchmarks for multi-disciplinary academic writing, overlooking continual learning (CL) as a promising solution to handle domain-specific linguistic variation and prevent catastrophic forgetting. To fill this crucial gap, we introduce CL$^2$GEC, the first Continual Learning benchmark for Chinese Literature Grammatical Error Correction, designed to evaluate adaptive CGEC across multiple academic fields. Our benchmark includes 10,000 human-annotated sentences spanning 10 disciplines, each exhibiting distinct linguistic styles and error patterns. CL$^2$GEC focuses on evaluating grammatical error correction in a continual learning setting, simulating sequential exposure to diverse academic disciplines to reflect real-world editorial dynamics. We evaluate large language models under sequential tuning, parameter-efficient adaptation, and four representative CL algorithms, using both standard GEC metrics and continual learning metrics adapted to task-level variation. Experimental results reveal that regularization-based methods mitigate forgetting more effectively than replay-based or naive sequential approaches. Our benchmark provides a rigorous foundation for future research in adaptive grammatical error correction across diverse academic domains.

Method: Proposes AgentCTG framework that simulates control and regulation mechanisms in multi-agent workflows, explores various agent collaboration methods, and introduces an auto-prompt module to enhance generation effectiveness.

Result: Achieves state-of-the-art results on multiple public datasets. Validated through a new Character-Driven Rewriting task that preserves domain knowledge while conforming to specific character profiles. Significantly enhances online navigation and role-playing experiences.

Conclusion: AgentCTG provides an effective solution for precise and complex control in text generation, enabling more immersive interactions and better user engagement in practical applications.

Abstract: Although significant progress has been made in many tasks within the field of Natural Language Processing (NLP), Controlled Text Generation (CTG) continues to face numerous challenges, particularly in achieving fine-grained conditional control over generation. Additionally, in real scenario and online applications, cost considerations, scalability, domain knowledge learning and more precise control are required, presenting more challenge for CTG. This paper introduces a novel and scalable framework, AgentCTG, which aims to enhance precise and complex control over the text generation by simulating the control and regulation mechanisms in multi-agent workflows. We explore various collaboration methods among different agents and introduce an auto-prompt module to further enhance the generation effectiveness. AgentCTG achieves state-of-the-art results on multiple public datasets. To validate its effectiveness in practical applications, we propose a new challenging Character-Driven Rewriting task, which aims to convert the original text into new text that conform to specific character profiles and simultaneously preserve the domain knowledge. When applied to online navigation with role-playing, our approach significantly enhances the driving experience through improved content delivery. By optimizing the generation of contextually relevant text, we enable a more immersive interaction within online communities, fostering greater personalization and user engagement.

Method: Proposes CARE framework that teaches LLMs to explicitly integrate in-context evidence within their reasoning process using the model’s own retrieval capabilities. Uses limited labeled evidence data and strategically retrieved in-context tokens in the reasoning chain.

Result: Substantially outperforms supervised fine-tuning, traditional retrieval-augmented generation methods, and external retrieval solutions on multiple real-world and counterfactual QA benchmarks. Enhances both retrieval accuracy and answer generation performance.

Conclusion: Represents a fundamental advancement in making LLMs more accurate, reliable, and efficient for knowledge-intensive tasks by improving context fidelity through native retrieval-augmented reasoning.

Abstract: Large language models (LLMs) often struggle with context fidelity, producing inconsistent answers when responding to questions based on provided information. Existing approaches either rely on expensive supervised fine-tuning to generate evidence post-answer or train models to perform web searches without necessarily improving utilization of the given context. We propose CARE, a novel native retrieval-augmented reasoning framework that teaches LLMs to explicitly integrate in-context evidence within their reasoning process with the model’s own retrieval capabilities. Our method requires limited labeled evidence data while significantly enhancing both retrieval accuracy and answer generation performance through strategically retrieved in-context tokens in the reasoning chain. Extensive experiments on multiple real-world and counterfactual QA benchmarks demonstrate that our approach substantially outperforms supervised fine-tuning, traditional retrieval-augmented generation methods, and external retrieval solutions. This work represents a fundamental advancement in making LLMs more accurate, reliable, and efficient for knowledge-intensive tasks.

Method: Constructed a Japanese NLI dataset focused on comparatives and evaluated various LLMs in zero-shot and few-shot settings with different prompt formats, including logical semantic representations.

Result: LLM performance was sensitive to prompt formats in zero-shot settings and influenced by gold labels in few-shot examples. Models struggled with Japanese-specific linguistic phenomena, but prompts with logical semantic representations improved accuracy on difficult inference problems.

Conclusion: LLMs need specialized handling for Japanese comparative NLI tasks, with logical semantic representations in prompts showing promise for improving performance on challenging numerical and logical inference problems.

Abstract: Large Language Models (LLMs) perform remarkably well in Natural Language Inference (NLI). However, NLI involving numerical and logical expressions remains challenging. Comparatives are a key linguistic phenomenon related to such inference, but the robustness of LLMs in handling them, especially in languages that are not dominant in the models’ training data, such as Japanese, has not been sufficiently explored. To address this gap, we construct a Japanese NLI dataset that focuses on comparatives and evaluate various LLMs in zero-shot and few-shot settings. Our results show that the performance of the models is sensitive to the prompt formats in the zero-shot setting and influenced by the gold labels in the few-shot examples. The LLMs also struggle to handle linguistic phenomena unique to Japanese. Furthermore, we observe that prompts containing logical semantic representations help the models predict the correct labels for inference problems that they struggle to solve even with few-shot examples.

Method: Adapt instruction-tuned LLMs using DSPy-based prompt optimization to process clinical notes and structured EHR inputs jointly.

Result: The approach achieves performance on par with specialized multimodal systems.

Conclusion: This method requires less complexity and offers greater adaptability across clinical tasks compared to specialized multimodal systems.

Abstract: Large language models (LLMs) excel at text generation, but their ability to handle clinical classification tasks involving structured data, such as time series, remains underexplored. In this work, we adapt instruction-tuned LLMs using DSPy-based prompt optimization to process clinical notes and structured EHR inputs jointly. Our results show that this approach achieves performance on par with specialized multimodal systems while requiring less complexity and offering greater adaptability across tasks.

Method: Created four test sets with different generation approaches: real voices reading synthetic code-switched sentences, generative TTS, and concatenative TTS, plus a 128-hour training set with generative TTS across 16 language pairs.

Result: A comprehensive dataset covering 113 unique code-switched language pairs across 52 languages with multiple speech generation methods and a substantial training set.

Conclusion: CS-FLEURS aims to expand the scope of code-switched speech research by providing diverse, multi-language datasets that go beyond high-resourced languages, facilitating broader development and evaluation of speech systems.

Abstract: We present CS-FLEURS, a new dataset for developing and evaluating code-switched speech recognition and translation systems beyond high-resourced languages. CS-FLEURS consists of 4 test sets which cover in total 113 unique code-switched language pairs across 52 languages: 1) a 14 X-English language pair set with real voices reading synthetically generated code-switched sentences, 2) a 16 X-English language pair set with generative text-to-speech 3) a 60 {Arabic, Mandarin, Hindi, Spanish}-X language pair set with the generative text-to-speech, and 4) a 45 X-English lower-resourced language pair test set with concatenative text-to-speech. Besides the four test sets, CS-FLEURS also provides a training set with 128 hours of generative text-to-speech data across 16 X-English language pairs. Our hope is that CS-FLEURS helps to broaden the scope of future code-switched speech research. Dataset link: https://huggingface.co/datasets/byan/cs-fleurs.

Method: Uses compact proxy models trained adversarially as Factual Alignment Proxy (FAP) and Hallucination Detection Proxy (HDP). These inject a real-time steering vector (difference between FAP and HDP logits) during autoregressive decoding.

Result: Achieved 99.2% Factual Consistency Rate on TruthfulQA and highest FActScore of 46.50 on BioGEN benchmark, demonstrating state-of-the-art performance.

Conclusion: DSCC-HS provides a principled and efficient plug-and-play solution for enhancing LLM factuality through dynamic self-reinforcing calibration during generation.

Abstract: Large Language Model (LLM) hallucination is a significant barrier to their reliable deployment. Current methods like Retrieval-Augmented Generation (RAG) are often reactive. We introduce Dynamic Self-reinforcing Calibration for Hallucination Suppression (DSCC-HS), a novel, proactive framework that intervenes during autoregressive decoding. Inspired by dual-process cognitive theory, DSCC-HS uses a compact proxy model, trained in adversarial roles as a Factual Alignment Proxy (FAP) and a Hallucination Detection Proxy (HDP). During inference, these proxies dynamically steer a large target model by injecting a real-time steering vector, which is the difference between FAP and HDP logits, at each decoding step. This plug-and-play approach requires no modification to the target model. Our experiments on TruthfulQA and BioGEN show DSCC-HS achieves state-of-the-art performance. On TruthfulQA, it reached a 99.2% Factual Consistency Rate (FCR). On the long-form BioGEN benchmark, it attained the highest FActScore of 46.50. These results validate DSCC-HS as a principled and efficient solution for enhancing LLM factuality.

Method: Trained SVM and BlueBERT models on 7,094 institutional reports and 571 IAEA SAFRON reports. Used transfer learning approach with BlueBERT_TRANSFER model fine-tuned on both datasets.

Result: BlueBERT_TRANSFER achieved AUROC 0.78 on cross-institution testing, similar to human performance (AUROC 0.81) on curated dataset. Without transfer learning, performance dropped significantly.

Conclusion: Successfully developed cross-institution NLP models that can detect high-severity radiation oncology incident reports with human-level accuracy, enabling efficient automated screening.

Abstract: PURPOSE: Incident reports are an important tool for safety and quality improvement in healthcare, but manual review is time-consuming and requires subject matter expertise. Here we present a natural language processing (NLP) screening tool to detect high-severity incident reports in radiation oncology across two institutions. METHODS AND MATERIALS: We used two text datasets to train and evaluate our NLP models: 7,094 reports from our institution (Inst.), and 571 from IAEA SAFRON (SF), all of which had severity scores labeled by clinical content experts. We trained and evaluated two types of models: baseline support vector machines (SVM) and BlueBERT which is a large language model pretrained on PubMed abstracts and hospitalized patient data. We assessed for generalizability of our model in two ways. First, we evaluated models trained using Inst.-train on SF-test. Second, we trained a BlueBERT_TRANSFER model that was first fine-tuned on Inst.-train then on SF-train before testing on SF-test set. To further analyze model performance, we also examined a subset of 59 reports from our Inst. dataset, which were manually edited for clarity. RESULTS Classification performance on the Inst. test achieved AUROC 0.82 using SVM and 0.81 using BlueBERT. Without cross-institution transfer learning, performance on the SF test was limited to an AUROC of 0.42 using SVM and 0.56 using BlueBERT. BlueBERT_TRANSFER, which was fine-tuned on both datasets, improved the performance on SF test to AUROC 0.78. Performance of SVM, and BlueBERT_TRANSFER models on the manually curated Inst. reports (AUROC 0.85 and 0.74) was similar to human performance (AUROC 0.81). CONCLUSION: In summary, we successfully developed cross-institution NLP models on incident report text from radiation oncology centers. These models were able to detect high-severity reports similarly to humans on a curated dataset.

Method: Two-stage approach: 1) Coarse-grained stage filters low-value sentences using TF-IDF, 2) Fine-grained stage prunes tokens using attention contribution, cross-model loss difference, and positional importance.

Result: DSPC achieves 49.17 performance on Longbench FewShot task using 3x fewer tokens, outperforming state-of-the-art LongLLMLingua by 7.76 points.

Conclusion: Training-free prompt compression via dual-stage progressive compression effectively reduces token usage while maintaining or improving LLM performance.

Abstract: Large language models (LLMs) have achieved remarkable success in many natural language processing (NLP) tasks. To achieve more accurate output, the prompts used to drive LLMs have become increasingly longer, which incurs higher computational costs. To address this prompt inflation problem, prompt compression has been proposed. However, most existing methods require training a small auxiliary model for compression, incurring a significant amount of additional computation. To avoid this, we propose a two-stage, training-free approach, called Dual-Stage Progressive Compression (DSPC). In the coarse-grained stage, semantic-related sentence filtering removes sentences with low semantic value based on TF-IDF. In the fine-grained stage, token importance is assessed using attention contribution, cross-model loss difference, and positional importance, enabling the pruning of low-utility tokens while preserving semantics. We validate DSPC on LLaMA-3.1-8B-Instruct and GPT-3.5-Turbo under a constrained token budget and observe consistent improvements. For instance, in the FewShot task of the Longbench dataset, DSPC achieves a performance of 49.17 by using only 3x fewer tokens, outperforming the best state-of-the-art baseline LongLLMLingua by 7.76.

Method: Two-stage pre-training and fine-tuning with FastConformer encoder and Transformer decoder, trained on 1.7M hours of data including Granary and NeMo ASR Set 3.0, with non-speech audio for hallucination reduction and NeMo Forced Aligner for timestamps.

Result: Outperforms Whisper-large-v3 on English ASR while being 10x faster, delivers competitive multilingual ASR and AST performance against larger models like Seamless-M4T-v2-large, and provides reliable segment-level timestamps.

Conclusion: Canary-1B-v2 demonstrates that efficient architecture with proper training can achieve state-of-the-art performance in multilingual speech recognition and translation, with additional release of Parakeet-TDT-0.6B-v3 offering similar capabilities with fewer parameters.

Abstract: This report introduces Canary-1B-v2, a fast, robust multilingual model for Automatic Speech Recognition (ASR) and Speech-to-Text Translation (AST). Built with a FastConformer encoder and Transformer decoder, it supports 25 languages primarily European. The model was trained on 1.7M hours of total data samples, including Granary and NeMo ASR Set 3.0, with non-speech audio added to reduce hallucinations for ASR and AST. We describe its two-stage pre-training and fine-tuning process with dynamic data balancing, as well as experiments with an nGPT encoder. Results show nGPT scales well with massive data, while FastConformer excels after fine-tuning. For timestamps, Canary-1B-v2 uses the NeMo Forced Aligner (NFA) with an auxiliary CTC model, providing reliable segment-level timestamps for ASR and AST. Evaluations show Canary-1B-v2 outperforms Whisper-large-v3 on English ASR while being 10x faster, and delivers competitive multilingual ASR and AST performance against larger models like Seamless-M4T-v2-large and LLM-based systems. We also release Parakeet-TDT-0.6B-v3, a successor to v2, offering multilingual ASR across the same 25 languages with just 600M parameters.

Method: Developed ccg-jcomp, a logical inference system based on compositional semantics specifically designed for Japanese comparatives to handle numerical and logical expressions.

Result: The system was evaluated on a Japanese NLI dataset with comparative expressions and demonstrated higher accuracy compared to existing Large Language Models.

Conclusion: Logic-based compositional semantic approaches are effective for Japanese comparative NLI and outperform LLMs, providing a robust solution for handling Japanese-specific linguistic features in comparative reasoning.

Abstract: Natural Language Inference (NLI) involving comparatives is challenging because it requires understanding quantities and comparative relations expressed by sentences. While some approaches leverage Large Language Models (LLMs), we focus on logic-based approaches grounded in compositional semantics, which are promising for robust handling of numerical and logical expressions. Previous studies along these lines have proposed logical inference systems for English comparatives. However, it has been pointed out that there are several morphological and semantic differences between Japanese and English comparatives. These differences make it difficult to apply such systems directly to Japanese comparatives. To address this gap, this study proposes ccg-jcomp, a logical inference system for Japanese comparatives based on compositional semantics. We evaluate the proposed system on a Japanese NLI dataset containing comparative expressions. We demonstrate the effectiveness of our system by comparing its accuracy with that of existing LLMs.

Method: Investigated various soft-prompting strategies (prefix-tuning, prompt-tuning, P-tuning, P-tuning V2), LoRA reparameterizations, and hard prompting with zero-shot and few-shot inferences using LLMs. Experiments conducted with Arabic-specific encoder models on multiple datasets and n-shot inferences on decoder-only models including Phi-3.5 and SILMA.

Result: LLMs struggled to differentiate dialectal nuances in few-shot/zero-shot setups. Soft-prompted encoder variants performed better, while LoRA-based fine-tuned models achieved the best performance, surpassing even full fine-tuning approaches.

Conclusion: Parameter-efficient methods like LoRA reparameterization are highly effective for Arabic Dialect Identification, outperforming both traditional fine-tuning and prompting strategies, making them suitable for resource-constrained scenarios.

Abstract: This paper discusses our exploration of different data-efficient and parameter-efficient approaches to Arabic Dialect Identification (ADI). In particular, we investigate various soft-prompting strategies, including prefix-tuning, prompt-tuning, P-tuning, and P-tuning V2, as well as LoRA reparameterizations. For the data-efficient strategy, we analyze hard prompting with zero-shot and few-shot inferences to analyze the dialect identification capabilities of Large Language Models (LLMs). For the parameter-efficient PEFT approaches, we conducted our experiments using Arabic-specific encoder models on several major datasets. We also analyzed the n-shot inferences on open-source decoder-only models, a general multilingual model (Phi-3.5), and an Arabic-specific one(SILMA). We observed that the LLMs generally struggle to differentiate the dialectal nuances in the few-shot or zero-shot setups. The soft-prompted encoder variants perform better, while the LoRA-based fine-tuned models perform best, even surpassing full fine-tuning.

Method: Dataset selection criteria, curation and normalization process, and performance evaluation of several models on the collected multilingual emotional speech datasets.

Result: The collection is publicly available with metadata and a leaderboard on Hugging Face platform, enabling standardized benchmarking for speech emotion recognition systems.

Conclusion: CAMEO provides a valuable resource for the research community by offering curated multilingual emotional speech data with standardized benchmarks to advance speech emotion recognition research.

Abstract: This paper presents CAMEO – a curated collection of multilingual emotional speech datasets designed to facilitate research in emotion recognition and other speech-related tasks. The main objectives were to ensure easy access to the data, to allow reproducibility of the results, and to provide a standardized benchmark for evaluating speech emotion recognition (SER) systems across different emotional states and languages. The paper describes the dataset selection criteria, the curation and normalization process, and provides performance results for several models. The collection, along with metadata, and a leaderboard, is publicly available via the Hugging Face platform.

Method: Proposed Emotion Omni model with a data pipeline to construct a 200k emotional dialogue dataset. The model understands emotional content in user speech and generates empathetic responses without large-scale pretraining.

Result: Achieves comparable instruction-following ability without large-scale pretraining, surpasses existing models in speech quality (UTMOS: 4.41) and empathy (Emotion GPT Score: 3.97).

Conclusion: Emotion Omni demonstrates significant improvements in both speech fidelity and emotional expressiveness, enabling more empathetic human-machine interactions with limited data requirements.

Abstract: With the development of speech large language models (speech LLMs), users can now interact directly with assistants via speech. However, most existing models only convert response content into speech without fully capturing the rich emotional cues in user queries, where the same sentence may convey different meanings depending on the expression. Emotional understanding is thus essential for improving human-machine interaction. Most empathetic speech LLMs rely on massive datasets, demanding high computational cost. A key challenge is to build models that generate empathetic responses with limited data and without large-scale training. To this end, we propose Emotion Omni, a model that understands emotional content in user speech and generates empathetic responses. We further developed a data pipeline to construct a 200k emotional dialogue dataset supporting empathetic speech assistants. Experiments show that Emotion Omni achieves comparable instruction-following ability without large-scale pretraining, while surpassing existing models in speech quality (UTMOS:4.41) and empathy (Emotion GPT Score: 3.97). These results confirm its improvements in both speech fidelity and emotional expressiveness. Demos are available at https://w311411.github.io/omni_demo/.

Method: CAMPUS framework features dynamic sub-curriculum selection, competency-aware curriculum schedule adjustment, and multiple difficulty-based scheduling to adapt to model evolution.

Result: Extensive experiments show CAMPUS achieves superior performance compared to state-of-the-art baselines for efficient instruction tuning.

Conclusion: Dynamic curriculum adaptation to model capabilities during training is crucial for effective instruction tuning, and CAMPUS provides an effective framework for this purpose.

Abstract: Efficient instruction tuning aims to enhance the ultimate performance of large language models (LLMs) trained on a given instruction dataset. Curriculum learning as a typical data organization strategy has shown preliminary effectiveness in instruction tuning. However, current curriculum tuning methods suffer from the curriculum rigidity, since they rely solely on static heuristic difficulty metrics. These methods fail to adapt to the evolving capabilities of models during training, resulting in a fixed and potentially sub-optimal learning trajectory. To address the issue, Competence-Aware Multi-Perspective cUrriculum inStruction tuning framework termed CAMPUS is proposed. CAMPUS offers several advantages: (1) Dynamic selection for sub-curriculum. (2) Competency-aware adjustment to the curriculum schedule. (3) Multiple difficulty-based scheduling. Extensive experiments prove the superior performance of CAMPUS, compared to other state-of-the-art baselines for efficient instruction tuning.

Method: Proposes using RBO (Rank-Biased Overlap) metrics for ranking comparison controlling for gender, generates test sets for job title matching in four grammatical gender languages with masculine/feminine forms annotated by gender and relevance, and evaluates several multilingual models.

Result: All evaluated out-of-the-box multilingual models exhibited varying degrees of gender bias when tested with the new methodology and test sets.

Conclusion: The study provides a foundation for assessing gender bias in job ranking systems and demonstrates that current multilingual models show gender bias, highlighting the need for bias-aware evaluation methodologies.

Abstract: This work sets the ground for studying how explicit grammatical gender assignment in job titles can affect the results of automatic job ranking systems. We propose the usage of metrics for ranking comparison controlling for gender to evaluate gender bias in job title ranking systems, in particular RBO (Rank-Biased Overlap). We generate and share test sets for a job title matching task in four grammatical gender languages, including occupations in masculine and feminine form and annotated by gender and matching relevance. We use the new test sets and the proposed methodology to evaluate the gender bias of several out-of-the-box multilingual models to set as baselines, showing that all of them exhibit varying degrees of gender bias.

Method: Geometric framework based on archetypal analysis of response batches. Global uncertainty measured via Geometric Volume (convex hull volume of archetypes). Local uncertainty measured via Geometric Suspicion (ranking responses by reliability). Uses only black-box model access.

Result: Performs comparably or better than prior methods on short-form QA datasets, achieves superior results on medical datasets where hallucinations are critical. Provides theoretical justification linking convex hull volume to entropy.

Conclusion: The geometric framework effectively quantifies both global and local uncertainty without white-box access, enabling reliable hallucination detection and reduction through preferential response selection, particularly valuable in high-risk domains like medicine.

Abstract: Large language models demonstrate impressive results across diverse tasks but are still known to hallucinate, generating linguistically plausible but incorrect answers to questions. Uncertainty quantification has been proposed as a strategy for hallucination detection, but no existing black-box approach provides estimates for both global and local uncertainty. The former attributes uncertainty to a batch of responses, while the latter attributes uncertainty to individual responses. Current local methods typically rely on white-box access to internal model states, whilst black-box methods only provide global uncertainty estimates. We introduce a geometric framework to address this, based on archetypal analysis of batches of responses sampled with only black-box model access. At the global level, we propose Geometric Volume, which measures the convex hull volume of archetypes derived from response embeddings. At the local level, we propose Geometric Suspicion, which ranks responses by reliability and enables hallucination reduction through preferential response selection. Unlike prior dispersion methods which yield only a single global score, our approach provides semantic boundary points which have utility for attributing reliability to individual responses. Experiments show that our framework performs comparably to or better than prior methods on short form question-answering datasets, and achieves superior results on medical datasets where hallucinations carry particularly critical risks. We also provide theoretical justification by proving a link between convex hull volume and entropy.

Method: Organized shared task with two main components: minuting task covering English/Czech languages across project meetings and European Parliament sessions, and QA task focused on project meetings with monolingual (English) and cross-lingual (Czech questions on English meetings) settings. Included multiple baseline systems using current LLMs for comprehensive evaluation.

Result: Limited participation with only one team for minuting and two teams for QA tasks. However, organizers provided extensive baseline evaluations of 2025 large language models on both tasks, enabling assessment of current capabilities.

Conclusion: Despite lower participation, AutoMin 2025 successfully established evaluation frameworks for meeting summarization and QA tasks, providing valuable benchmarks for LLM performance in multilingual and cross-lingual meeting processing scenarios.

Abstract: This paper presents the third edition of AutoMin, a shared task on automatic meeting summarization into minutes. In 2025, AutoMin featured the main task of minuting, the creation of structured meeting minutes, as well as a new task: question answering (QA) based on meeting transcripts. The minuting task covered two languages, English and Czech, and two domains: project meetings and European Parliament sessions. The QA task focused solely on project meetings and was available in two settings: monolingual QA in English, and cross-lingual QA, where questions were asked and answered in Czech based on English meetings. Participation in 2025 was more limited compared to previous years, with only one team joining the minuting task and two teams participating in QA. However, as organizers, we included multiple baseline systems to enable a comprehensive evaluation of current (2025) large language models (LLMs) on both tasks.

Method: Two tasks: association task and decision task using a novel evaluation corpus pairing sentences from seven regional German dialects with standard German counterparts to assess dialect naming and usage bias.

Result: All evaluated LLMs showed significant dialect naming and usage bias against German dialect speakers through negative adjective associations, and reproduced these biases in decision making. Explicit labeling of linguistic demographics amplified bias more than implicit dialect cues.

Conclusion: LLMs reflect and amplify societal stereotypes against dialect speakers, with explicit demographic mentions increasing bias contrary to prior findings about other demographic groups.

Abstract: Dialects represent a significant component of human culture and are found across all regions of the world. In Germany, more than 40% of the population speaks a regional dialect (Adler and Hansen, 2022). However, despite cultural importance, individuals speaking dialects often face negative societal stereotypes. We examine whether such stereotypes are mirrored by large language models (LLMs). We draw on the sociolinguistic literature on dialect perception to analyze traits commonly associated with dialect speakers. Based on these traits, we assess the dialect naming bias and dialect usage bias expressed by LLMs in two tasks: an association task and a decision task. To assess a model’s dialect usage bias, we construct a novel evaluation corpus that pairs sentences from seven regional German dialects (e.g., Alemannic and Bavarian) with their standard German counterparts. We find that: (1) in the association task, all evaluated LLMs exhibit significant dialect naming and dialect usage bias against German dialect speakers, reflected in negative adjective associations; (2) all models reproduce these dialect naming and dialect usage biases in their decision making; and (3) contrary to prior work showing minimal bias with explicit demographic mentions, we find that explicitly labeling linguistic demographics–German dialect speakers–amplifies bias more than implicit cues like dialect usage.

Method: Analyzed 12 LLMs from four model families using four datasets, examining misalignment rates, attack success rates, judgment consistency, and explanation generation capabilities. Also fine-tuned smaller LMs for explanation generation.

Result: Large model size doesn’t guarantee better alignment. LLMs show preference for specific scenario types and higher consistency within model families. No significant differences in HVSB understanding across LLMs. Smaller fine-tuned LMs produce more readable explanations but with lower model agreeability.

Conclusion: LLM alignment with human values on social biases is complex and not size-dependent. Model families show consistent behavior patterns. Explanation capabilities can be transferred to smaller models but with trade-offs in agreeability.

Abstract: Large language models (LLMs) can lead to undesired consequences when misaligned with human values, especially in scenarios involving complex and sensitive social biases. Previous studies have revealed the misalignment of LLMs with human values using expert-designed or agent-based emulated bias scenarios. However, it remains unclear whether the alignment of LLMs with human values differs across different types of scenarios (e.g., scenarios containing negative vs. non-negative questions). In this study, we investigate the alignment of LLMs with human values regarding social biases (HVSB) in different types of bias scenarios. Through extensive analysis of 12 LLMs from four model families and four datasets, we demonstrate that LLMs with large model parameter scales do not necessarily have lower misalignment rate and attack success rate. Moreover, LLMs show a certain degree of alignment preference for specific types of scenarios and the LLMs from the same model family tend to have higher judgment consistency. In addition, we study the understanding capacity of LLMs with their explanations of HVSB. We find no significant differences in the understanding of HVSB across LLMs. We also find LLMs prefer their own generated explanations. Additionally, we endow smaller language models (LMs) with the ability to explain HVSB. The generation results show that the explanations generated by the fine-tuned smaller LMs are more readable, but have a relatively lower model agreeability.

Method: CER integrates scientific evidence retrieval techniques with large language models for reasoning, plus supervised veracity prediction, to ensure outputs are grounded in verifiable evidence-based sources and mitigate hallucinations.

Result: State-of-the-art performance on expert-annotated datasets (HealthFC, BioASQ-7b, SciFact) with promising cross-dataset generalization capabilities.

Conclusion: CER effectively addresses biomedical fact-checking challenges by combining evidence retrieval with LLM reasoning, providing a robust solution against healthcare misinformation while ensuring transparency and reproducibility through released code and data.

Abstract: Misinformation in healthcare, from vaccine hesitancy to unproven treatments, poses risks to public health and trust in medical systems. While machine learning and natural language processing have advanced automated fact-checking, validating biomedical claims remains uniquely challenging due to complex terminology, the need for domain expertise, and the critical importance of grounding in scientific evidence. We introduce CER (Combining Evidence and Reasoning), a novel framework for biomedical fact-checking that integrates scientific evidence retrieval, reasoning via large language models, and supervised veracity prediction. By integrating the text-generation capabilities of large language models with advanced retrieval techniques for high-quality biomedical scientific evidence, CER effectively mitigates the risk of hallucinations, ensuring that generated outputs are grounded in verifiable, evidence-based sources. Evaluations on expert-annotated datasets (HealthFC, BioASQ-7b, SciFact) demonstrate state-of-the-art performance and promising cross-dataset generalization. Code and data are released for transparency and reproducibility: https: //github.com/PRAISELab-PicusLab/CER.

Method: CER integrates scientific evidence retrieval, reasoning via large language models, and supervised veracity prediction. It combines LLM text-generation capabilities with advanced retrieval techniques for high-quality biomedical evidence to mitigate hallucinations.

Result: State-of-the-art performance on expert-annotated datasets (HealthFC, BioASQ-7b, SciFact) with promising cross-dataset generalization capabilities.

Conclusion: CER effectively addresses biomedical fact-checking challenges by ensuring outputs are grounded in verifiable evidence-based sources, demonstrating strong performance and generalization across multiple biomedical datasets.

Abstract: Misinformation in healthcare, from vaccine hesitancy to unproven treatments, poses risks to public health and trust in medical systems. While machine learning and natural language processing have advanced automated fact-checking, validating biomedical claims remains uniquely challenging due to complex terminology, the need for domain expertise, and the critical importance of grounding in scientific evidence. We introduce CER (Combining Evidence and Reasoning), a novel framework for biomedical fact-checking that integrates scientific evidence retrieval, reasoning via large language models, and supervised veracity prediction. By integrating the text-generation capabilities of large language models with advanced retrieval techniques for high-quality biomedical scientific evidence, CER effectively mitigates the risk of hallucinations, ensuring that generated outputs are grounded in verifiable, evidence-based sources. Evaluations on expert-annotated datasets (HealthFC, BioASQ-7b, SciFact) demonstrate state-of-the-art performance and promising cross-dataset generalization. Code and data are released for transparency and reproducibility: https://github.com/PRAISELab-PicusLab/CER

Method: Evaluated instruction-tuned LLMs on Word Sense Disambiguation (WSD) compared to specialized systems, and tested generative abilities through definition generation, free-form explanation, and example generation tasks.

Result: GPT-4o and DeepSeek-V3 achieved WSD performance on par with specialized systems with greater robustness. LLMs explained word meanings in context with up to 98% accuracy, with free-form explanation performing best.

Conclusion: Modern LLMs demonstrate sophisticated word sense understanding capabilities, performing competitively with specialized WSD systems while excelling in generative explanation tasks that leverage their natural language strengths.

Abstract: Understanding the meaning of words in context is a fundamental capability for Large Language Models (LLMs). Despite extensive evaluation efforts, the extent to which LLMs show evidence that they truly grasp word senses remains underexplored. In this paper, we address this gap by evaluating both i) the Word Sense Disambiguation (WSD) capabilities of instruction-tuned LLMs, comparing their performance to state-of-the-art systems specifically designed for the task, and ii) the ability of two top-performing open- and closed-source LLMs to understand word senses in three generative settings: definition generation, free-form explanation, and example generation. Notably, we find that, in the WSD task, leading models such as GPT-4o and DeepSeek-V3 achieve performance on par with specialized WSD systems, while also demonstrating greater robustness across domains and levels of difficulty. In the generation tasks, results reveal that LLMs can explain the meaning of words in context up to 98% accuracy, with the highest performance observed in the free-form explanation task, which best aligns with their generative capabilities.

Method: Introduced a controlled methodology using model internals to measure language preference while holding document relevance constant. Tested across eight languages and six open-weight models.

Result: Models preferentially cite English sources when queries are in English, with bias amplified for lower-resource languages and documents positioned mid-context. Models sometimes trade-off document relevance for language preference.

Conclusion: Citation choices in multilingual RAG systems are not always driven by informativeness alone, revealing how language models leverage multilingual context and influence citation behavior with unintended language biases.

Abstract: Multilingual Retrieval-Augmented Generation (mRAG) systems enable language models to answer knowledge-intensive queries with citation-supported responses across languages. While such systems have been proposed, an open questions is whether the mixture of different document languages impacts generation and citation in unintended ways. To investigate, we introduce a controlled methodology using model internals to measure language preference while holding other factors such as document relevance constant. Across eight languages and six open-weight models, we find that models preferentially cite English sources when queries are in English, with this bias amplified for lower-resource languages and for documents positioned mid-context. Crucially, we find that models sometimes trade-off document relevance for language preference, indicating that citation choices are not always driven by informativeness alone. Our findings shed light on how language models leverage multilingual context and influence citation behavior.

Method: Built on COMET framework, concatenated in-domain human-annotated sentences to create long-context training data, computed weighted average scores, normalized multiple human judgment scales, and trained multilingual regression models using source, hypothesis, and reference translations.

Result: Experimental results demonstrate that incorporating long-context information improves correlations with human judgments compared to models trained only on short segments.

Conclusion: The approach successfully enhances translation quality evaluation by leveraging long-context data and integrating diverse human judgment datasets, leading to better alignment with human assessment standards.

Abstract: In this paper, we present our submission to the Tenth Conference on Machine Translation (WMT25) Shared Task on Automated Translation Quality Evaluation. Our systems are built upon the COMET framework and trained to predict segment-level Error Span Annotation (ESA) scores using augmented long-context data. To construct long-context training data, we concatenate in-domain, human-annotated sentences and compute a weighted average of their scores. We integrate multiple human judgment datasets (MQM, SQM, and DA) by normalising their scales and train multilingual regression models to predict quality scores from the source, hypothesis, and reference translations. Experimental results show that incorporating long-context information improves correlations with human judgments compared to models trained only on short segments.

Method: Proposes Slim-SC, a step-wise pruning strategy that identifies and removes redundant chains using inter-chain similarity at the thought level during the reasoning process.

Result: Experiments on three STEM reasoning datasets show Slim-SC reduces inference latency by up to 45% and KVC usage by 26% while maintaining or improving accuracy compared to standard SC.

Conclusion: Slim-SC offers a simple yet efficient alternative to Self-Consistency for test-time scaling, providing significant computational savings without sacrificing reasoning performance.

Abstract: Recently, Test-Time Scaling (TTS) has gained increasing attention for improving LLM reasoning performance at test time without retraining the model. A notable TTS technique is Self-Consistency (SC), which generates multiple reasoning chains in parallel and selects the final answer via majority voting. While effective, the order-of-magnitude computational overhead limits its broad deployment. Prior attempts to accelerate SC mainly rely on model-based confidence scores or heuristics with limited empirical support. For the first time, we theoretically and empirically analyze the inefficiencies of SC and reveal actionable opportunities for improvement. Building on these insights, we propose Slim-SC, a step-wise pruning strategy that identifies and removes redundant chains using inter-chain similarity at the thought level. Experiments on three STEM reasoning datasets and two recent LLM architectures show that Slim-SC reduces inference latency and KVC usage by up to 45% and 26%, respectively, with R1-Distill, while maintaining or improving accuracy, thus offering a simple yet efficient TTS alternative for SC.

Method: Prompt LLM to output step answers at each reasoning step, track consecutive identical answers as convergence measure, and terminate generation when run length shows sharp increase exceeding threshold.

Result: 41% reduction in inference tokens across 5 reasoning datasets and 3 LLMs while maintaining comparable accuracy to standard CoT; works well with self-consistency and is robust to hyperparameters.

Conclusion: ES-CoT provides a practical and effective approach for efficient reasoning by leveraging answer convergence detection, making it suitable for real-world LLM applications with minimal performance trade-offs.

Abstract: Reasoning large language models (LLMs) have demonstrated superior capacities in solving complicated problems by generating long chain-of-thoughts (CoT), but such a lengthy CoT incurs high inference costs. In this study, we introduce ES-CoT, an inference-time method that shortens CoT generation by detecting answer convergence and stopping early with minimal performance loss. At the end of each reasoning step, we prompt the LLM to output its current final answer, denoted as a step answer. We then track the run length of consecutive identical step answers as a measure of answer convergence. Once the run length exhibits a sharp increase and exceeds a minimum threshold, the generation is terminated. We provide both empirical and theoretical support for this heuristic: step answers steadily converge to the final answer, and large run-length jumps reliably mark this convergence. Experiments on five reasoning datasets across three LLMs show that ES-CoT reduces the number of inference tokens by about 41% on average while maintaining accuracy comparable to standard CoT. Further, ES-CoT integrates seamlessly with self-consistency prompting and remains robust across hyperparameter choices, highlighting it as a practical and effective approach for efficient reasoning.

Method: Translate-and-tune pipeline: compress AR↔EN teacher to FP8 for higher throughput, create bilingual supervision data, fine-tune lightweight LFM2-1.2B model to translate English instruction sets into Arabic, train Hala models at various parameter sizes (350M-9B), and apply slerp merging to balance Arabic specialization with base-model strengths.

Result: Hala achieves state-of-the-art results on Arabic-centric benchmarks within both “nano” (≤2B) and “small” (7-9B) parameter categories, outperforming their base models while maintaining high throughput through FP8 compression.

Conclusion: The Hala family of models successfully demonstrates effective Arabic-centric instruction following and translation capabilities, providing valuable resources (models, data, evaluation, recipes) to advance Arabic NLP research.

Abstract: We present Hala, a family of Arabic-centric instruction and translation models built with our translate-and-tune pipeline. We first compress a strong AR$\leftrightarrow$EN teacher to FP8 (yielding $\sim$2$\times$ higher throughput with no quality loss) and use it to create high-fidelity bilingual supervision. A lightweight language model LFM2-1.2B is then fine-tuned on this data and used to translate high-quality English instruction sets into Arabic, producing a million-scale corpus tailored to instruction following. We train Hala models at 350M, 700M, 1.2B, and 9B parameters, and apply slerp merging to balance Arabic specialization with base-model strengths. On Arabic-centric benchmarks, Hala achieves state-of-the-art results within both the “nano” ($\leq$2B) and “small” (7-9B) categories, outperforming their bases. We release models, data, evaluation, and recipes to accelerate research in Arabic NLP.

Method: Compared text-only and audio-based evaluations of 10 MT systems from WMT Shared Task using crowd-sourced judgments via Amazon Mechanical Turk, with statistical significance testing and self-replication experiments.

Result: Audio-based assessments produced rankings largely consistent with text-only evaluations but identified significant differences between some translation systems.

Conclusion: Speech-based assessments should be incorporated into future MT evaluation frameworks due to speech’s richer, more natural modality for evaluating spoken translation applications.

Abstract: Machine Translation (MT) has achieved remarkable performance, with growing interest in speech translation and multimodal approaches. However, despite these advancements, MT quality assessment remains largely text centric, typically relying on human experts who read and compare texts. Since many real-world MT applications (e.g Google Translate Voice Mode, iFLYTEK Translator) involve translation being spoken rather printed or read, a more natural way to assess translation quality would be through speech as opposed text-only evaluations. This study compares text-only and audio-based evaluations of 10 MT systems from the WMT General MT Shared Task, using crowd-sourced judgments collected via Amazon Mechanical Turk. We additionally, performed statistical significance testing and self-replication experiments to test reliability and consistency of audio-based approach. Crowd-sourced assessments based on audio yield rankings largely consistent with text only evaluations but, in some cases, identify significant differences between translation systems. We attribute this to speech richer, more natural modality and propose incorporating speech-based assessments into future MT evaluation frameworks.

Method: Constructed training datasets with controlled proportions of contextually relevant examples, systematically validated the sparsity claim, and proposed/evaluated two training strategies to better leverage available data.

Result: Strong association found between training data sparsity and model performance. Improvements in one contextual phenomenon don’t generalize to others. Cross-lingual transfer observed but not significantly higher within language sub-families. Proposed strategies achieved 6-8% accuracy gains.

Conclusion: Data sparsity is confirmed as a key bottleneck for context utilization in translation. Training strategies that better leverage available contextual data can significantly improve performance, but contextual improvements remain phenomenon-specific rather than generalizing broadly.

Abstract: Achieving human-level translations requires leveraging context to ensure coherence and handle complex phenomena like pronoun disambiguation. Sparsity of contextually rich examples in the standard training data has been hypothesized as the reason for the difficulty of context utilization. In this work, we systematically validate this claim in both single- and multilingual settings by constructing training datasets with a controlled proportions of contextually relevant examples. We demonstrate a strong association between training data sparsity and model performance confirming sparsity as a key bottleneck. Importantly, we reveal that improvements in one contextual phenomenon do no generalize to others. While we observe some cross-lingual transfer, it is not significantly higher between languages within the same sub-family. Finally, we propose and empirically evaluate two training strategies designed to leverage the available data. These strategies improve context utilization, resulting in accuracy gains of up to 6 and 8 percentage points on the ctxPro evaluation in single- and multilingual settings respectively.

Method: Developed ConfMAD, a MAD framework that integrates confidence expression throughout the debate process, allowing LLMs to explicitly communicate their confidence levels during agent interactions.

Result: Experimental results demonstrate the effectiveness of the proposed method, showing improved debate performance and task outcomes compared to traditional MAD systems without confidence expression.

Conclusion: Incorporating confidence expression in Multi-Agent Debate systems significantly improves debate dynamics and overall system performance, providing valuable insights for designing confidence-aware MAD systems.

Abstract: Generative Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of tasks. Recent research has introduced Multi-Agent Debate (MAD) systems, which leverage multiple LLMs to simulate human debate and thereby improve task performance. However, while some LLMs may possess superior knowledge or reasoning capabilities for specific tasks, they often struggle to clearly communicate this advantage during debates, in part due to a lack of confidence expression. Moreover, inappropriate confidence expression can cause agents in MAD systems to either stubbornly maintain incorrect beliefs or converge prematurely on suboptimal answers, ultimately reducing debate effectiveness and overall system performance. To address these challenges, we propose incorporating confidence expression into MAD systems to allow LLMs to explicitly communicate their confidence levels. To validate this approach, we develop ConfMAD, a MAD framework that integrates confidence expression throughout the debate process. Experimental results demonstrate the effectiveness of our method, and we further analyze how confidence influences debate dynamics, offering insights into the design of confidence-aware MAD systems.

Method: Cross-modality Self-supervised Learning with Sigmoid Self-attention Weighting (SSL-SSAW) fusion method, using contrastive learning for feature alignment and adaptive feature extraction from question and sign language sequences.

Result: Achieved SOTA performance on CSL-Daily-QA and PHOENIX-2014T-QA datasets, with question assistance matching or surpassing gloss assistance performance.

Conclusion: Question-based dialogue integration effectively improves sign language translation quality, demonstrating the value of accessible contextual cues over traditional gloss annotations.

Abstract: Sign Language Translation (SLT) bridges the communication gap between deaf people and hearing people, where dialogue provides crucial contextual cues to aid in translation. Building on this foundational concept, this paper proposes Question-based Sign Language Translation (QB-SLT), a novel task that explores the efficient integration of dialogue. Unlike gloss (sign language transcription) annotations, dialogue naturally occurs in communication and is easier to annotate. The key challenge lies in aligning multimodality features while leveraging the context of the question to improve translation. To address this issue, we propose a cross-modality Self-supervised Learning with Sigmoid Self-attention Weighting (SSL-SSAW) fusion method for sign language translation. Specifically, we employ contrastive learning to align multimodality features in QB-SLT, then introduce a Sigmoid Self-attention Weighting (SSAW) module for adaptive feature extraction from question and sign language sequences. Additionally, we leverage available question text through self-supervised learning to enhance representation and translation capabilities. We evaluated our approach on newly constructed CSL-Daily-QA and PHOENIX-2014T-QA datasets, where SSL-SSAW achieved SOTA performance. Notably, easily accessible question assistance can achieve or even surpass the performance of gloss assistance. Furthermore, visualization results demonstrate the effectiveness of incorporating dialogue in improving translation quality.

Method: Decomposed ambiguity into internal and external types, created AssoCiAm benchmark with hybrid computational methods to circumvent ambiguity in association evaluation.

Result: Found strong positive correlation between cognition and association in MLLMs. Ambiguity causes models to behave more randomly. Method ensures more accurate and reliable evaluations.

Conclusion: AssoCiAm effectively addresses ambiguity issues in association evaluation, providing a more reliable framework for assessing MLLM creative abilities and supporting AGI development.

Abstract: Recent advancements in multimodal large language models (MLLMs) have garnered significant attention, offering a promising pathway toward artificial general intelligence (AGI). Among the essential capabilities required for AGI, creativity has emerged as a critical trait for MLLMs, with association serving as its foundation. Association reflects a model’ s ability to think creatively, making it vital to evaluate and understand. While several frameworks have been proposed to assess associative ability, they often overlook the inherent ambiguity in association tasks, which arises from the divergent nature of associations and undermines the reliability of evaluations. To address this issue, we decompose ambiguity into two types-internal ambiguity and external ambiguity-and introduce AssoCiAm, a benchmark designed to evaluate associative ability while circumventing the ambiguity through a hybrid computational method. We then conduct extensive experiments on MLLMs, revealing a strong positive correlation between cognition and association. Additionally, we observe that the presence of ambiguity in the evaluation process causes MLLMs' behavior to become more random-like. Finally, we validate the effectiveness of our method in ensuring more accurate and reliable evaluations. See Project Page for the data and codes.

Method: Developed a reproducible framework integrating financial context with behavioral finance studies to create supervision data for end-to-end advisors. Created a 19k sample reasoning dataset and fine-tuned Qwen-3-8B model on this dataset.

Result: The 8B model achieved performance comparable to significantly larger baselines (14-32B parameters) across factual accuracy, fluency, and personalization metrics while incurring 80% lower costs than larger counterparts, as demonstrated through held-out test split and blind LLM-jury study.

Conclusion: Careful data curation and behavioral integration can enable smaller, more efficient models to achieve performance comparable to much larger models in financial advisory tasks, significantly reducing computational costs while maintaining quality.

Abstract: Personalized financial advice requires consideration of user goals, constraints, risk tolerance, and jurisdiction. Prior LLM work has focused on support systems for investors and financial planners. Simultaneously, numerous recent studies examine broader personal finance tasks, including budgeting, debt management, retirement, and estate planning, through agentic pipelines that incur high maintenance costs, yielding less than 25% of their expected financial returns. In this study, we introduce a novel and reproducible framework that integrates relevant financial context with behavioral finance studies to construct supervision data for end-to-end advisors. Using this framework, we create a 19k sample reasoning dataset and conduct a comprehensive fine-tuning of the Qwen-3-8B model on the dataset. Through a held-out test split and a blind LLM-jury study, we demonstrate that through careful data curation and behavioral integration, our 8B model achieves performance comparable to significantly larger baselines (14-32B parameters) across factual accuracy, fluency, and personalization metrics while incurring 80% lower costs than the larger counterparts.

Method: Used open-weight LLMs to annotate statements with high-level stances toward migrants, tracked net tone across time and parties, and developed semi-automated framework for extracting fine-grained narrative frames in UK parliamentary discourse.

Result: UK discourse remains relatively aligned across parties (unlike polarized US), with persistent Labour-Conservative ideological gap peaking negatively in 2025. Shift toward securitized narratives (border control, illegal immigration) and decline in integration-oriented frames. Discussion shifted from national to international law and human rights.

Conclusion: LLMs can support scalable, fine-grained discourse analysis in political and historical contexts, revealing nuanced trends in migration discourse over decades.

Abstract: We present a large-scale computational analysis of migration-related discourse in UK parliamentary debates spanning over 75 years and compare it with US congressional discourse. Using open-weight LLMs, we annotate each statement with high-level stances toward migrants and track the net tone toward migrants across time and political parties. For the UK, we extend this with a semi-automated framework for extracting fine-grained narrative frames to capture nuances of migration discourse. Our findings show that, while US discourse has grown increasingly polarised, UK parliamentary attitudes remain relatively aligned across parties, with a persistent ideological gap between Labour and the Conservatives, reaching its most negative level in 2025. The analysis of narrative frames in the UK parliamentary statements reveals a shift toward securitised narratives such as border control and illegal immigration, while longer-term integration-oriented frames such as social integration have declined. Moreover, discussions of national law about immigration have been replaced over time by international law and human rights, revealing nuances in discourse trends. Taken together broadly, our findings demonstrate how LLMs can support scalable, fine-grained discourse analysis in political and historical contexts.

Method: Pretrained exclusively on openly available data with retroactive robots.txt compliance filtering. Used Goldfish objective to suppress verbatim recall while retaining performance. Trained on 15T tokens from 1800+ languages with ~40% non-English content.

Result: Apertus models approach state-of-the-art results among fully open models on multilingual benchmarks, rivaling or surpassing open-weight counterparts. Successfully mitigates memorization risks while maintaining downstream task performance.

Conclusion: Apertus provides a fully transparent, compliant, and multilingual open LLM suite with complete release of scientific artifacts (data scripts, checkpoints, evaluation suites, training code) enabling audit and extension.

Abstract: We present Apertus, a fully open suite of large language models (LLMs) designed to address two systemic shortcomings in today’s open model ecosystem: data compliance and multilingual representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting robots.txt exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of memorization, we adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. The Apertus models also expand multilingual coverage, training on 15T tokens from over 1800 languages, with ~40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivalling or surpassing open-weight counterparts. Beyond model weights, we release all scientific artifacts from our development cycle with a permissive license, including data preparation scripts, checkpoints, evaluation suites, and training code, enabling transparent audit and extension.

Method: The paper conducts a comprehensive survey of methodologies that integrate LLMs with IR systems, examining key components including query rewriters, retrievers, rerankers, and readers. It also explores emerging directions like search agents.

Result: The survey provides a consolidated overview of existing approaches combining LLMs with IR systems, highlighting how LLMs can enhance various IR components and identifying promising research directions in this rapidly evolving field.

Conclusion: The integration of LLMs with IR systems represents a significant advancement, combining traditional IR methods with modern neural architectures. This survey serves as a foundation for future research in leveraging LLMs to address IR challenges and develop more effective information retrieval solutions.

Abstract: As a primary means of information acquisition, information retrieval (IR) systems, such as search engines, have integrated themselves into our daily lives. These systems also serve as components of dialogue, question-answering, and recommender systems. The trajectory of IR has evolved dynamically from its origins in term-based methods to its integration with advanced neural models. While the neural models excel at capturing complex contextual signals and semantic nuances, thereby reshaping the IR landscape, they still face challenges such as data scarcity, interpretability, and the generation of contextually plausible yet potentially inaccurate responses. This evolution requires a combination of both traditional methods (such as term-based sparse retrieval methods with rapid response) and modern neural architectures (such as language models with powerful language understanding capacity). Meanwhile, the emergence of large language models (LLMs), typified by ChatGPT and GPT-4, has revolutionized natural language processing due to their remarkable language understanding, generation, generalization, and reasoning abilities. Consequently, recent research has sought to leverage LLMs to improve IR systems. Given the rapid evolution of this research trajectory, it is necessary to consolidate existing methodologies and provide nuanced insights through a comprehensive overview. In this survey, we delve into the confluence of LLMs and IR systems, including crucial aspects such as query rewriters, retrievers, rerankers, and readers. Additionally, we explore promising directions, such as search agents, within this expanding field.

Method: AEPO selects a subset of responses that maximizes diversity and representativeness from available responses, then annotates preferences only over this informative subset rather than exhaustively annotating all responses.

Result: Evaluation on three datasets shows AEPO outperforms baselines with the same annotation budget, demonstrating more effective use of limited annotation resources.

Conclusion: AEPO provides an effective approach for annotation-efficient preference optimization by focusing annotation efforts on the most informative response subsets, achieving better performance with limited budgets.

Abstract: Preference optimization is a standard approach to fine-tuning large language models to align with human preferences. The quantity, diversity, and representativeness of the preference dataset are critical to the effectiveness of preference optimization. However, obtaining a large amount of preference annotations is difficult in many applications. This raises the question of how to use the limited annotation budget to create an effective preference dataset. To this end, we propose Annotation-Efficient Preference Optimization (AEPO). Instead of exhaustively annotating preference over all available response texts, AEPO selects a subset of responses that maximizes diversity and representativeness from the available responses and then annotates preference over the selected ones. In this way, AEPO focuses the annotation budget on labeling preferences over a smaller but informative subset of responses. We evaluate the performance of preference learning using AEPO on three datasets and show that it outperforms the baselines with the same annotation budget. Our code is available at https://github.com/CyberAgentAILab/annotation-efficient-po

Method: Database-Augmented Query representation (DAQu) framework that expands original queries with various query-related metadata across multiple database tables, using graph-based set-encoding strategy to handle unordered hierarchical features.

Result: DAQu significantly enhances overall retrieval performance across diverse retrieval scenarios compared to relevant baselines.

Conclusion: The proposed DAQu framework effectively leverages database metadata for query augmentation, demonstrating superior retrieval performance through graph-based encoding of hierarchical features without order constraints.

Abstract: Information retrieval models that aim to search for documents relevant to a query have shown multiple successes, which have been applied to diverse tasks. Yet, the query from the user is oftentimes short, which challenges the retrievers to correctly fetch relevant documents. To tackle this, previous studies have proposed expanding the query with a couple of additional (user-related) features related to it. However, they may be suboptimal to effectively augment the query, and there is plenty of other information available to augment it in a relational database. Motivated by this fact, we present a novel retrieval framework called Database-Augmented Query representation (DAQu), which augments the original query with various (query-related) metadata across multiple tables. In addition, as the number of features in the metadata can be very large and there is no order among them, we encode them with the graph-based set-encoding strategy, which considers hierarchies of features in the database without order. We validate our DAQu in diverse retrieval scenarios, demonstrating that it significantly enhances overall retrieval performance over relevant baselines. Our code is available at \href{https://github.com/starsuzi/DAQu}{this https URL}.

Method: A synthetic framework that embeds key data points at varying depths, with two task scenarios: information-sparse (minimal relevant details in extensive irrelevant text for simple retrieval) and information-dense (continuously distributed relevant information for complex reasoning).

Result: Experiments show that recent reasoning models like Deepseek-R1 and OpenAI’s o3 excel in mathematical reasoning but struggle with continuous retrieval and reasoning in information-dense scenarios, even at shorter context lengths. The study also identifies ‘under-thinking’ phenomenon where models prematurely conclude reasoning.

Conclusion: NeedleBench provides critical insights and targeted tools for evaluating and improving LLMs’ long-context capabilities, offering a more systematic approach to assess retrieval and reasoning performance in bilingual long-context tasks.

Abstract: The capability of large language models to handle long-context information is crucial across various real-world applications. Existing evaluation methods often rely either on real-world long texts, making it difficult to exclude the influence of models’ inherent knowledge, or introduce irrelevant filler content to artificially achieve target lengths, reducing assessment effectiveness. To address these limitations, we introduce NeedleBench, a synthetic framework for assessing retrieval and reasoning performance in bilingual long-context tasks with adaptive context lengths. NeedleBench systematically embeds key data points at varying depths to rigorously test model capabilities. Tasks are categorized into two scenarios: information-sparse, featuring minimal relevant details within extensive irrelevant text to simulate simple retrieval tasks; and information-dense (the Ancestral Trace Challenge), where relevant information is continuously distributed throughout the context to simulate complex reasoning tasks. Our experiments reveal that although recent reasoning models like Deepseek-R1 and OpenAI’s o3 excel in mathematical reasoning, they struggle with continuous retrieval and reasoning in information-dense scenarios, even at shorter context lengths. We also characterize a phenomenon termed ‘under-thinking’, where models prematurely conclude reasoning despite available information. NeedleBench thus provides critical insights and targeted tools essential for evaluating and improving LLMs’ long-context capabilities. All resources are available at OpenCompass: https://github.com/open-compass/opencompass.

Method: Computational implementation of a dynamic neural model using Dynamic Field Theory, where lexical items map to continuous semantic dimensions (connectedness and control asymmetry) through neural coupling. Model simulations and experimental testing with self-paced reading and acceptability judgments.

Result: Model captured contextual modulation of lexical interpretation and individual variation. Generated novel prediction about context-dependent relationship between reading time and acceptability, which was partially confirmed experimentally. Supports polysemy as transient neural states rather than categorical representations.

Conclusion: Lexical polysemy arises from nonlinear dynamics of neural populations on continuous semantic dimensions, offering advantages over categorical representation models and Bayesian inference approaches.

Abstract: We computationally implement and experimentally test the behavioral predictions of a dynamic neural model of lexical meaning in the framework of Dynamic Field Theory. We demonstrate the architecture and behavior of the model using as a test case the English lexical item have, focusing on its polysemous use. In the model, have maps to a semantic space defined by two independently motivated continuous conceptual dimensions, connectedness and control asymmetry. The mapping is modeled as coupling between a neural node representing the lexical item and neural fields representing the conceptual dimensions. While lexical knowledge is modeled as a stable coupling pattern, real-time lexical meaning retrieval is modeled as the motion of neural activation patterns between transiently stable states corresponding to semantic interpretations or readings. Model simulations capture two previously reported empirical observations: (1) contextual modulation of lexical semantic interpretation, and (2) individual variation in the magnitude of this modulation. Simulations also generate a novel prediction that the by-trial relationship between sentence reading time and acceptability should be contextually modulated. An experiment combining self-paced reading and acceptability judgments replicates previous results and partially bears out the model’s novel prediction. Altogether, results support a novel perspective on lexical polysemy: that the many related meanings of a word are not categorically distinct representations; rather, they are transiently stable neural activation states that arise from the nonlinear dynamics of neural populations governing interpretation on continuous semantic dimensions. Our model offers important advantages over related models in the dynamical systems framework, as well as models based on Bayesian inference.

Method: Categorizes problems as easy/hard, uses coarse-grained aggregation for easy problems and fine-grained iterative multi-agent refinement for hard problems. Employs three agents (Solver, Reviewer, Refiner) with step-wise reward model scores for targeted feedback and iterative re-evaluation.

Result: Outperforms Self-Consistency by 3.4%, Best-of-k by 3.2%, and Self-Refine by 4.0% with less than half the samples. Continues to improve with more iterations unlike baseline methods.

Conclusion: MAgICoRe effectively addresses refinement challenges through problem difficulty categorization, external reward models for error localization, and multi-agent iterative refinement, demonstrating significant performance improvements across math reasoning tasks.

Abstract: Large Language Models’ (LLM) reasoning can be improved using test-time aggregation strategies, i.e., generating multiple samples and voting among generated samples. While these improve performance, they often reach a saturation point. Refinement offers an alternative by using LLM-generated feedback to improve solution quality. However, refinement introduces 3 key challenges: (1) Excessive refinement: Uniformly refining all instances can over-correct and reduce the overall performance. (2) Inability to localize and address errors: LLMs have a limited ability to self-correct and struggle to identify and correct their own mistakes. (3) Insufficient refinement: Deciding how many iterations of refinement are needed is non-trivial, and stopping too soon could leave errors unaddressed. To tackle these issues, we propose MAgICoRe, which avoids excessive refinement by categorizing problem difficulty as easy or hard, solving easy problems with coarse-grained aggregation and hard ones with fine-grained and iterative multi-agent refinement. To improve error localization, we incorporate external step-wise reward model (RM) scores. Moreover, to ensure effective refinement, we employ a multi-agent loop with three agents: Solver, Reviewer (which generates targeted feedback based on step-wise RM scores), and the Refiner (which incorporates feedback). To ensure sufficient refinement, we re-evaluate updated solutions, iteratively initiating further rounds of refinement. We evaluate MAgICoRe on Llama-3-8B and GPT-3.5 and show its effectiveness across 5 math datasets. Even one iteration of MAgICoRe beats Self-Consistency by 3.4%, Best-of-k by 3.2%, and Self-Refine by 4.0% while using less than half the samples. Unlike iterative refinement with baselines, MAgICoRe continues to improve with more iterations. Finally, our ablations highlight the importance of MAgICoRe’s RMs and multi-agent communication.

Method: DAVIS uses structured temporal memory for model-based planning and implements an agentic multi-turn retrieval system similar to human inner monologue for enhanced reasoning over past experiences.

Result: Substantially improved performance on ScienceWorld benchmark (8/9 subjects) and competitive performance on HotpotQA and MusiqueQA for multi-hop question answering.

Conclusion: DAVIS represents the first RAG agent to employ interactive retrieval in a RAG pipeline, successfully addressing the complex requirements of scientific laboratory assistance.

Abstract: Designing a generalist scientific agent capable of performing tasks in laboratory settings to assist researchers has become a key goal in recent Artificial Intelligence (AI) research. Unlike everyday tasks, scientific tasks are inherently more delicate and complex, requiring agents to possess a higher level of reasoning ability, structured and temporal understanding of their environment, and a strong emphasis on safety. Existing approaches often fail to address these multifaceted requirements. To tackle these challenges, we present DAVIS. Unlike traditional retrieval-augmented generation (RAG) approaches, DAVIS incorporates structured and temporal memory, which enables model-based planning. Additionally, DAVIS implements an agentic, multi-turn retrieval system, similar to a human’s inner monologue, allowing for a greater degree of reasoning over past experiences. DAVIS demonstrates substantially improved performance on the ScienceWorld benchmark comparing to previous approaches on 8 out of 9 elementary science subjects. In addition, DAVIS’s World Model demonstrates competitive performance on the famous HotpotQA and MusiqueQA dataset for multi-hop question answering. To the best of our knowledge, DAVIS is the first RAG agent to employ an interactive retrieval method in a RAG pipeline.

Method: Proposes Mirror-Consistency which incorporates a ‘reflective mirror’ into self-ensemble decoding, enabling LLMs to critically examine inconsistencies among multiple generations. Also enhances sample-based confidence calibration methods.

Result: Experimental results show Mirror-Consistency achieves superior performance in both reasoning accuracy and confidence calibration compared to standard Self-Consistency.

Conclusion: Mirror-Consistency effectively addresses limitations of Self-Consistency by leveraging minority responses and improving confidence calibration, resulting in better overall reasoning performance for LLMs.

Abstract: Self-Consistency, a widely-used decoding strategy, significantly boosts the reasoning capabilities of Large Language Models (LLMs). However, it depends on the plurality voting rule, which focuses on the most frequent answer while overlooking all other minority responses. These inconsistent minority views often illuminate areas of uncertainty within the model’s generation process. To address this limitation, we present Mirror-Consistency, an enhancement of the standard Self-Consistency approach. Our method incorporates a ‘reflective mirror’ into the self-ensemble decoding process and enables LLMs to critically examine inconsistencies among multiple generations. Additionally, just as humans use the mirror to better understand themselves, we propose using Mirror-Consistency to enhance the sample-based confidence calibration methods, which helps to mitigate issues of overconfidence. Our experimental results demonstrate that Mirror-Consistency yields superior performance in both reasoning accuracy and confidence calibration compared to Self-Consistency.

Method: Created Swiss Leading Decision Summarization (SLDS) dataset with 18K court rulings in German, French, and Italian plus German headnotes. Fine-tuned and evaluated three mT5 variants along with proprietary models in zero-shot and one-shot settings.

Result: Proprietary models performed well in zero-shot and one-shot settings, but fine-tuned smaller models still provided strong competitive performance for legal summarization tasks.

Conclusion: The publicly released SLDS dataset facilitates further research in multilingual legal summarization and development of assistive technologies for legal professionals, demonstrating that specialized fine-tuned models can effectively compete with larger proprietary models.

Abstract: Legal research is a time-consuming task that most lawyers face on a daily basis. A large part of legal research entails looking up relevant caselaw and bringing it in relation to the case at hand. Lawyers heavily rely on summaries (also called headnotes) to find the right cases quickly. However, not all decisions are annotated with headnotes and writing them is time-consuming. Automated headnote creation has the potential to make hundreds of thousands of decisions more accessible for legal research in Switzerland alone. To kickstart this, we introduce the Swiss Leading Decision Summarization ( SLDS) dataset, a novel cross-lingual resource featuring 18K court rulings from the Swiss Federal Supreme Court (SFSC), in German, French, and Italian, along with German headnotes. We fine-tune and evaluate three mT5 variants, along with proprietary models. Our analysis highlights that while proprietary models perform well in zero-shot and one-shot settings, fine-tuned smaller models still provide a strong competitive edge. We publicly release the dataset to facilitate further research in multilingual legal summarization and the development of assistive technologies for legal professionals

Method: Developed a Knowledge Boundary Model (KBM) to identify what knowledge is known/unknown to LLMs for given questions, deciding when to trigger RAG retrieval.

Result: Experiments on 11 English/Chinese datasets show KBM effectively delineates knowledge boundaries, significantly reducing retrieval needs while maintaining optimal end-to-end performance. Works well in dynamic knowledge, long-tail static knowledge, and multi-hop scenarios.

Conclusion: KBM provides an effective method to optimize RAG usage by selectively triggering retrieval only when needed, reducing computational costs while maintaining LLM performance across various knowledge scenarios.

Abstract: Large Language Models (LLMs) often struggle with dynamically changing knowledge and handling unknown static information. Retrieval-Augmented Generation (RAG) is employed to tackle these challenges and has a significant impact on improving LLM performance. In fact, we find that not all questions need to trigger RAG. By retrieving parts of knowledge unknown to the LLM and allowing the LLM to answer the rest, we can effectively reduce both time and computational costs. In our work, we propose a Knowledge Boundary Model (KBM) to express the known/unknown of a given question, and to determine whether a RAG needs to be triggered. Experiments conducted on 11 English and Chinese datasets illustrate that the KBM effectively delineates the knowledge boundary, significantly decreasing the proportion of retrievals required for optimal end-to-end performance. Furthermore, we evaluate the effectiveness of KBM in three complex scenarios: dynamic knowledge, long-tail static knowledge, and multi-hop problems, as well as its functionality as an external LLM plug-in.

Method: HITL-GAT system uses human-in-the-loop approach with three customized adversarial text generation methods for Tibetan script.

Result: Established first adversarial robustness benchmark for Tibetan, providing reference framework for other lower-resourced languages.

Conclusion: The interactive human-in-the-loop approach effectively addresses challenges in creating sustainable adversarial benchmarks for under-resourced languages.

Abstract: DNN-based language models excel across various NLP tasks but remain highly vulnerable to textual adversarial attacks. While adversarial text generation is crucial for NLP security, explainability, evaluation, and data augmentation, related work remains overwhelmingly English-centric, leaving the problem of constructing high-quality and sustainable adversarial robustness benchmarks for lower-resourced languages both difficult and understudied. First, method customization for lower-resourced languages is complicated due to linguistic differences and limited resources. Second, automated attacks are prone to generating invalid or ambiguous adversarial texts. Last but not least, language models continuously evolve and may be immune to parts of previously generated adversarial texts. To address these challenges, we introduce HITL-GAT, an interactive system based on a general approach to human-in-the-loop generation of adversarial texts. Additionally, we demonstrate the utility of HITL-GAT through a case study on Tibetan script, employing three customized adversarial text generation methods and establishing its first adversarial robustness benchmark, providing a valuable reference for other lower-resourced languages.

Method: POATE technique involves: 1) Polar Opposite query generation, 2) Adversarial Template construction, and 3) Elaboration. It crafts semantically opposing intents integrated with adversarial templates to steer models toward harmful outputs.

Result: Extensive evaluation across six diverse language model families shows POATE achieves significantly higher attack success rates (~44%) compared to existing methods, demonstrating robustness across varying parameter sizes.

Conclusion: The proposed Intent-Aware CoT and Reverse Thinking CoT defense methods enhance reasoning robustness by decomposing queries to detect malicious intent and reasoning in reverse to evaluate and reject harmful responses, strengthening model defenses against adversarial exploits.

Abstract: Large language models, despite extensive alignment with human values and ethical principles, remain vulnerable to sophisticated jailbreak attacks that exploit their reasoning abilities. Existing safety measures often detect overt malicious intent but fail to address subtle, reasoning-driven vulnerabilities. In this work, we introduce POATE (Polar Opposite query generation, Adversarial Template construction, and Elaboration), a novel jailbreak technique that harnesses contrastive reasoning to provoke unethical responses. POATE crafts semantically opposing intents and integrates them with adversarial templates, steering models toward harmful outputs with remarkable subtlety. We conduct extensive evaluation across six diverse language model families of varying parameter sizes to demonstrate the robustness of the attack, achieving significantly higher attack success rates (~44%) compared to existing methods. To counter this, we propose Intent-Aware CoT and Reverse Thinking CoT, which decompose queries to detect malicious intent and reason in reverse to evaluate and reject harmful responses. These methods enhance reasoning robustness and strengthen the model’s defense against adversarial exploits.

Method: Explored both prompting and fine-tuning approaches with GPT-4o-mini, comparing against Microsoft Presidio and Azure AI Language on CRAPII and TSCC datasets.

Result: Fine-tuned GPT-4o-mini achieved 0.9589 recall on CRAPII, tripled precision scores, reduced computational costs to 1/10 of Azure AI Language, and showed consistent accuracy across cultural backgrounds and genders with 0.9895 recall on TSCC.

Conclusion: Fine-tuned GPT-4o-mini is an accurate, cost-effective tool for PII detection in educational data, offering robust privacy protection while preserving data utility for research and analysis.

Abstract: Protecting Personally Identifiable Information (PII), such as names, is a critical requirement in learning technologies to safeguard student and teacher privacy and maintain trust. Accurate PII detection is an essential step toward anonymizing sensitive information while preserving the utility of educational data. Motivated by recent advancements in artificial intelligence, our study investigates the GPT-4o-mini model as a cost-effective and efficient solution for PII detection tasks. We explore both prompting and fine-tuning approaches and compare GPT-4o-mini’s performance against established frameworks, including Microsoft Presidio and Azure AI Language. Our evaluation on two public datasets, CRAPII and TSCC, demonstrates that the fine-tuned GPT-4o-mini model achieves superior performance, with a recall of 0.9589 on CRAPII. Additionally, fine-tuned GPT-4o-mini significantly improves precision scores (a threefold increase) while reducing computational costs to nearly one-tenth of those associated with Azure AI Language. Furthermore, our bias analysis reveals that the fine-tuned GPT-4o-mini model consistently delivers accurate results across diverse cultural backgrounds and genders. The generalizability analysis using the TSCC dataset further highlights its robustness, achieving a recall of 0.9895 with minimal additional training data from TSCC. These results emphasize the potential of fine-tuned GPT-4o-mini as an accurate and cost-effective tool for PII detection in educational data. It offers robust privacy protection while preserving the data’s utility for research and pedagogical analysis. Our code is available on GitHub: https://github.com/AnonJD/PrivacyAI

Method: Using a chess game structure analogy (opening, middle, end game), the authors analyze segment-specific patterns in texts to reveal where the most striking differences between human and AI writing lie, with deeper analysis of linguistic features.

Result: AI texts closely resemble human writing in body segments due to length, but show higher divergence in features dependent on continuous language flow, making body segments most informative for detection. Human texts exhibit greater stylistic variation across segments.

Conclusion: The findings provide fresh insights into human-AI text differences and pave the way for more effective and interpretable detection strategies, with body segments being particularly revealing despite surface similarities.

Abstract: The rapid advancement of Large Language Models (LLMs) has revolutionized text generation but also raised concerns about potential misuse, making detecting LLM-generated text (AI text) increasingly essential. While prior work has focused on identifying AI text and effectively checkmating it, our study investigates a less-explored territory: portraying the nuanced distinctions between human and AI texts across text segments (introduction, body, and conclusion). Whether LLMs excel or falter in incorporating linguistic ingenuity across text segments, the results will critically inform their viability and boundaries as effective creative assistants to humans. Through an analogy with the structure of chess games, comprising opening, middle, and end games, we analyze segment-specific patterns to reveal where the most striking differences lie. Although AI texts closely resemble human writing in the body segment due to its length, deeper analysis shows a higher divergence in features dependent on the continuous flow of language, making it the most informative segment for detection. Additionally, human texts exhibit greater stylistic variation across segments, offering a new lens for distinguishing them from AI. Overall, our findings provide fresh insights into human-AI text differences and pave the way for more effective and interpretable detection strategies. Codes available at https://github.com/tripto03/chess_inspired_human_ai_text_distinction.

Method: Systematically distort benchmark prompts via parametric transformation (rephrasing inputs while preserving semantic content and labels) to test if model performance relies on memorized patterns rather than genuine understanding.

Result: Average performance degradation of 2.15% under modest perturbations; 20/26 models showed statistically significant differences; higher baseline accuracy models and larger LLMs exhibited greater sensitivity to rephrasings; Llama family and lower-accuracy models showed insignificant degradation.

Conclusion: Benchmark scores alone are insufficient for evaluating LLMs; C-BOD provides a practical method to detect overfitting and promote more robust language understanding, challenging the community to prioritize resilience and generalization in evaluation.

Abstract: Large language models (LLMs) often appear to excel on public benchmarks, but these high scores may mask an overreliance on dataset-specific surface cues rather than true language understanding. We introduce the Chameleon Benchmark Overfit Detector (C-BOD), a meta-evaluation framework that systematically distorts benchmark prompts via a parametric transformation and detects overfitting of LLMs. By rephrasing inputs while preserving their semantic content and labels, C-BOD exposes whether a model’s performance is driven by memorized patterns. Evaluated on the MMLU benchmark using 26 leading LLMs, our method reveals an average performance degradation of 2.15% under modest perturbations, with 20 out of 26 models exhibiting statistically significant differences. Notably, models with higher baseline accuracy exhibit larger performance differences under perturbation, and larger LLMs tend to be more sensitive to rephrasings, indicating that both cases may overrely on fixed prompt patterns. In contrast, the Llama family and models with lower baseline accuracy show insignificant degradation, suggesting reduced dependency on superficial cues. Moreover, C-BOD’s dataset- and model-agnostic design allows easy integration into training pipelines to promote more robust language understanding. Our findings challenge the community to look beyond leaderboard scores and prioritize resilience and generalization in LLM evaluation.

Method: Controlled analogical reasoning environment testing different modalities (textual, visual, symbolic), difficulty levels, and task formats (MCQ/free-text) to compare inference systems.

Result: System 2 pipelines generally excel, especially in visual/symbolic modalities and harder tasks, while System 1 is competitive for textual and easier problems. Task format significantly influences relative advantage.

Conclusion: Advanced System 2 strategies like hypothesis selection and iterative refinement can substantially scale LLM reasoning, providing foundational insights for strategic deployment of logical inference.

Abstract: Modern large language models (LLMs) employ diverse logical inference mechanisms for reasoning, making the strategic optimization of these approaches critical for advancing their capabilities. This paper systematically investigate the comparative dynamics of inductive (System 1) versus abductive/deductive (System 2) inference in LLMs. We utilize a controlled analogical reasoning environment, varying modality (textual, visual, symbolic), difficulty, and task format (MCQ / free-text). Our analysis reveals System 2 pipelines generally excel, particularly in visual/symbolic modalities and harder tasks, while System 1 is competitive for textual and easier problems. Crucially, task format significantly influences their relative advantage, with System 1 sometimes outperforming System 2 in free-text rule-execution. These core findings generalize to broader in-context learning. Furthermore, we demonstrate that advanced System 2 strategies like hypothesis selection and iterative refinement can substantially scale LLM reasoning. This study offers foundational insights and actionable guidelines for strategically deploying logical inference to enhance LLM reasoning. Resources are available at https://github.com/HKUST-KnowComp/LogiDynamics.

Method: Created a new challenging test set with high variation in content preservation, conducted large meta-evaluation of existing metrics, and proposed a style-aware evaluation method using small language models.

Result: Found that widely used metrics show artificially high correlations due to test data nature, and demonstrated that suitable metrics must be style-aware. The proposed method achieved higher alignment with human judgments than similar-sized autorater models.

Conclusion: Proper evaluation of content preservation in style transfer requires style-aware metrics and carefully constructed test data that challenges content variation, not just style changes.

Abstract: Large language models (LLMs) make it easy to rewrite a text in any style – e.g. to make it more polite, persuasive, or more positive – but evaluation thereof is not straightforward. A challenge lies in measuring content preservation: that content not attributable to style change is retained. This paper presents a large meta-evaluation of metrics for evaluating style and attribute transfer, focusing on content preservation. We find that meta-evaluation studies on existing datasets lead to misleading conclusions about the suitability of metrics for content preservation. Widely used metrics show a high correlation with human judgments despite being deemed unsuitable for the task – because they do not abstract from style changes when evaluating content preservation. We show that the overly high correlations with human judgment stem from the nature of the test data. To address this issue, we introduce a new, challenging test set specifically designed for evaluating content preservation metrics for style transfer. We construct the data by creating high variation in the content preservation. Using this dataset, we demonstrate that suitable metrics for content preservation for style transfer indeed are style-aware. To support efficient evaluation, we propose a new style-aware method that utilises small language models, obtaining a higher alignment with human judgements than prompting a model of a similar size as an autorater. ater.

Method: Created Description-based Bias Benchmark (DBB) with naturalistic scenarios where bias concepts are hidden within subtle, real-world contexts rather than explicit terms.

Result: Analysis of six state-of-the-art LLMs revealed they reduce bias at term level but continue to reinforce biases in nuanced, contextually hidden settings.

Conclusion: Current bias evaluation methods are insufficient; semantic-level assessment through DBB reveals persistent hidden biases that need to be addressed in LLM development.

Abstract: Large Language Models (LLMs) often exhibit social biases inherited from their training data. While existing benchmarks evaluate bias by term-based mode through direct term associations between demographic terms and bias terms, LLMs have become increasingly adept at avoiding biased responses, leading to seemingly low levels of bias. However, biases persist in subtler, contextually hidden forms that traditional benchmarks fail to capture. We introduce the Description-based Bias Benchmark (DBB), a novel dataset designed to assess bias at the semantic level that bias concepts are hidden within naturalistic, subtly framed contexts in real-world scenarios rather than superficial terms. We analyze six state-of-the-art LLMs, revealing that while models reduce bias in response at the term level, they continue to reinforce biases in nuanced settings. Data, code, and results are available at https://github.com/JP-25/Description-based-Bias-Benchmark.

Method: Created a manually annotated dataset of 125K+ instances with three bilingual annotators per instance, covering five core NLP tasks. Evaluated traditional tools, open-source models, and closed-source LLMs in zero-shot and one-shot settings, then fine-tuned state-of-the-art LLMs on the dataset.

Result: Closed-source LLMs significantly outperformed traditional tools and open-source models. One-shot prompting consistently improved performance, especially for structure-sensitive tasks. Fine-tuning achieved up to 95.25 F1 in NER, 98.77 F1 in MLI, and competitive machine translation performance.

Conclusion: COMI-LINGUA sets new benchmarks for Hinglish code-mixed text processing and demonstrates the effectiveness of LLMs when provided with high-quality annotated data, with the dataset being publicly available for further research.

Abstract: We introduce COMI-LINGUA, the largest manually annotated Hindi-English code-mixed dataset, comprising 125K+ high-quality instances across five core NLP tasks: Matrix Language Identification, Token-level Language Identification, Part-Of-Speech Tagging, Named Entity Recognition, and Machine Translation. Each instance is annotated by three bilingual annotators, yielding over 376K expert annotations with strong inter-annotator agreement (Fleiss’ Kappa $\geq$ 0.81). The rigorously preprocessed and filtered dataset covers both Devanagari and Roman scripts and spans diverse domains, ensuring real-world linguistic coverage. Evaluation reveals that closed-source LLMs significantly outperform traditional tools and open-source models in zero-shot settings. Notably, one-shot prompting consistently boosts performance across tasks, especially in structure-sensitive predictions like POS and NER. Fine-tuning state-of-the-art LLMs on COMI-LINGUA demonstrates substantial improvements, achieving up to 95.25 F1 in NER, 98.77 F1 in MLI, and competitive MT performance, setting new benchmarks for Hinglish code-mixed text. COMI-LINGUA is publicly available at this URL: https://huggingface.co/datasets/LingoIITGN/COMI-LINGUA.

Method: The researchers used causal interventions to analyze information flow in Gemma-2 2B model across five naturalistic in-context learning tasks, examining how the model processes and utilizes few-shot examples.

Result: The study revealed a two-step strategy: lower layers build representations of individual examples contextualized by preceding examples through input-output token connections, while higher layers aggregate these representations to identify tasks and prepare predictions.

Conclusion: The contextualization step’s importance varies by task and becomes more critical with ambiguous examples. The causal analysis provides rigorous insights into the mechanisms underlying in-context learning in language models.

Abstract: In-Context Learning (ICL) is an intriguing ability of large language models (LLMs). Despite a substantial amount of work on its behavioral aspects and how it emerges in miniature setups, it remains unclear which mechanism assembles task information from the individual examples in a fewshot prompt. We use causal interventions to identify information flow in Gemma-2 2B for five naturalistic ICL tasks. We find that the model infers task information using a two-step strategy we call contextualize-then-aggregate: In the lower layers, the model builds up representations of individual fewshot examples, which are contextualized by preceding examples through connections between fewshot input and output tokens across the sequence. In the higher layers, these representations are aggregated to identify the task and prepare prediction of the next output. The importance of the contextualization step differs between tasks, and it may become more important in the presence of ambiguous examples. Overall, by providing rigorous causal analysis, our results shed light on the mechanisms through which ICL happens in language models.

Method: Systematically tested 12 leading LLMs on two-integer addition (0 to 2^64) by evaluating three key properties: commutativity, representation invariance via symbolic remapping, and consistent accuracy scaling with operand length.

Result: Models achieved high numeric accuracy (73.8-99.8%) but failed diagnostic tests: accuracy dropped to ≤7.5% with symbolic inputs, commutativity violated in up to 20% of cases, and accuracy scaling was non-monotonic. Interventions worsened performance.

Conclusion: Current LLMs solve elementary addition through pattern matching rather than robust rule induction, highlighting the need for new diagnostic benchmarks and architectural innovations to develop genuine mathematical reasoning.

Abstract: Large language models (LLMs) achieve impressive results on advanced mathematics benchmarks but sometimes fail on basic arithmetic tasks, raising the question of whether they have truly grasped fundamental arithmetic rules or are merely relying on pattern matching. To unravel this issue, we systematically probe LLMs’ understanding of two-integer addition ($0$ to $2^{64}$) by testing three crucial properties: commutativity ($A+B=B+A$), representation invariance via symbolic remapping (e.g., $7 \mapsto Y$), and consistent accuracy scaling with operand length. Our evaluation of 12 leading LLMs reveals a stark disconnect: while models achieve high numeric accuracy (73.8-99.8%), they systematically fail these diagnostics. Specifically, accuracy plummets to $\le 7.5$% with symbolic inputs, commutativity is violated in up to 20% of cases, and accuracy scaling is non-monotonic. Interventions further expose this pattern-matching reliance: explicitly providing rules degrades performance by 29.49%, while prompting for explanations before answering merely maintains baseline accuracy. These findings demonstrate that current LLMs address elementary addition via pattern matching, not robust rule induction, motivating new diagnostic benchmarks and innovations in model architecture and training to cultivate genuine mathematical reasoning. Our dataset and generating code are available at https://github.com/kuri-leo/llm-arithmetic-diagnostic.

Method: Developed an evaluation protocol, created open-source assessment library, and established a leaderboard system for organizing results

Result: A comprehensive benchmark framework that includes evaluation procedures, software tools, and results organization for voice cloning TTS models

Conclusion: The presented benchmark provides a standardized way to evaluate and compare voice cloning text-to-speech models through systematic assessment tools and public leaderboard

Abstract: We present a novel benchmark for voice cloning text-to-speech models. The benchmark consists of an evaluation protocol, an open-source library for assessing the performance of voice cloning models, and an accompanying leaderboard. The paper discusses design considerations and presents a detailed description of the evaluation procedure. The usage of the software library is explained, along with the organization of results on the leaderboard.

Method: Proposes a methodology based on comparative behavioral theory to analyze interactions between training data and model architecture. Builds on recent work relating transformers to n-gram LMs, evaluating how data, model design choices, and temporal dynamics affect bias propagation.

Result: Findings show: (1) n-gram LMs are highly sensitive to context window size in bias propagation, while transformers demonstrate architectural robustness; (2) temporal provenance of training data significantly affects bias; (3) different model architectures respond differentially to controlled bias injection, with certain biases being disproportionately amplified.

Conclusion: As language models become ubiquitous, a holistic approach is needed that traces bias to its origins across both data and model dimensions, not just symptoms, to effectively mitigate harm.

Abstract: Current research on bias in language models (LMs) predominantly focuses on data quality, with significantly less attention paid to model architecture and temporal influences of data. Even more critically, few studies systematically investigate the origins of bias. We propose a methodology grounded in comparative behavioral theory to interpret the complex interaction between training data and model architecture in bias propagation during language modeling. Building on recent work that relates transformers to n-gram LMs, we evaluate how data, model design choices, and temporal dynamics affect bias propagation. Our findings reveal that: (1) n-gram LMs are highly sensitive to context window size in bias propagation, while transformers demonstrate architectural robustness; (2) the temporal provenance of training data significantly affects bias; and (3) different model architectures respond differentially to controlled bias injection, with certain biases (e.g. sexual orientation) being disproportionately amplified. As language models become ubiquitous, our findings highlight the need for a holistic approach – tracing bias to its origins across both data and model dimensions, not just symptoms, to mitigate harm.

Method: Introduces a three-level taxonomy (Tool, Analyst, Scientist) through the lens of the scientific method to delineate LLMs’ escalating autonomy and evolving responsibilities in the research lifecycle.

Result: Provides a conceptual architecture and strategic foresight for AI-driven scientific discovery, identifying key challenges and future research trajectories including robotic automation, self-improvement, and ethical governance.

Conclusion: This survey offers a framework to navigate and shape the future of LLM-driven scientific discovery, fostering both rapid innovation and responsible advancement in the field.

Abstract: Large Language Models (LLMs) are catalyzing a paradigm shift in scientific discovery, evolving from task-specific automation tools into increasingly autonomous agents and fundamentally redefining research processes and human-AI collaboration. This survey systematically charts this burgeoning field, placing a central focus on the changing roles and escalating capabilities of LLMs in science. Through the lens of the scientific method, we introduce a foundational three-level taxonomy-Tool, Analyst, and Scientist-to delineate their escalating autonomy and evolving responsibilities within the research lifecycle. We further identify pivotal challenges and future research trajectories such as robotic automation, self-improvement, and ethical governance. Overall, this survey provides a conceptual architecture and strategic foresight to navigate and shape the future of AI-driven scientific discovery, fostering both rapid innovation and responsible advancement. Github Repository: https://github.com/HKUST-KnowComp/Awesome-LLM-Scientific-Discovery.

Method: Revisiting existing localized unlearning approaches and conducting controlled experiments to rigorously evaluate whether local parameter updates causally contribute to unlearning.

Result: Findings reveal that the set of parameters that must be modified for effective unlearning is not strictly determined, challenging the core assumption that parameter locality indicates effective knowledge removal.

Conclusion: The core assumption of localized unlearning - that parameter locality inherently indicates effective knowledge removal - is challenged, suggesting current approaches may not be as effective as presumed.

Abstract: Large language models often retain unintended content, prompting growing interest in knowledge unlearning. Recent approaches emphasize localized unlearning, restricting parameter updates to specific regions in an effort to remove target knowledge while preserving unrelated general knowledge. However, their effectiveness remains uncertain due to the lack of robust and thorough evaluation of the trade-off between the competing goals of unlearning. In this paper, we begin by revisiting existing localized unlearning approaches. We then conduct controlled experiments to rigorously evaluate whether local parameter updates causally contribute to unlearning. Our findings reveal that the set of parameters that must be modified for effective unlearning is not strictly determined, challenging the core assumption of localized unlearning that parameter locality is inherently indicative of effective knowledge removal.

Method: Created SCRum-9 dataset with 7,516 tweets in 9 languages, annotated by native speakers with confidence ratings. Benchmarked 5 LLMs and 2 MLMs using ICL and fine-tuning setups, and explored multilingual synthetic data generation for model training.

Result: LLMs with weak ICL performance can produce valuable synthetic data that enables small MLMs to achieve higher performance than zero-shot ICL in LLMs. Model predictions often align with human annotators’ second-choice labels on ambiguous cases.

Conclusion: SCRum-9 enables advanced multilingual rumour analysis and shows that synthetic data from LLMs can effectively boost smaller models’ performance, with model predictions reflecting human uncertainty patterns in ambiguous cases.

Abstract: We introduce SCRum-9, the largest multilingual Stance Classification dataset for Rumour analysis in 9 languages, containing 7,516 tweets from X. SCRum-9 goes beyond existing stance classification datasets by covering more languages, linking examples to more fact-checked claims (2.1k), and including confidence-related annotations from multiple annotators to account for intra- and inter-annotator variability. Annotations were made by at least two native speakers per language, totalling more than 405 hours of annotation and 8,150 dollars in compensation. Further, SCRum-9 is used to benchmark five large language models (LLMs) and two multilingual masked language models (MLMs) in In-Context Learning (ICL) and fine-tuning setups. This paper also innovates by exploring the use of multilingual synthetic data for rumour stance classification, showing that even LLMs with weak ICL performance can produce valuable synthetic data for fine-tuning small MLMs, enabling them to achieve higher performance than zero-shot ICL in LLMs. Finally, we examine the relationship between model predictions and human uncertainty on ambiguous cases finding that model predictions often match the second-choice labels assigned by annotators, rather than diverging entirely from human judgments. SCRum-9 is publicly released to the research community with potential to foster further research on multilingual analysis of misleading narratives on social media.

Method: Constructed a hand-generated and annotated benchmark of diverse English-language rebus puzzles, ranging from simple pictographic substitutions to spatially-dependent cues, and analyzed performance of different VLMs.

Result: VLMs exhibit some surprising capabilities in decoding simple visual clues but struggle significantly with tasks requiring abstract reasoning, lateral thinking, and understanding visual metaphors.

Conclusion: Current vision-language models have limited capacity for the complex multi-modal reasoning required by rebus puzzles, highlighting gaps in abstract and symbolic reasoning capabilities.

Abstract: Rebus puzzles, visual riddles that encode language through imagery, spatial arrangement, and symbolic substitution, pose a unique challenge to current vision-language models (VLMs). Unlike traditional image captioning or question answering tasks, rebus solving requires multi-modal abstraction, symbolic reasoning, and a grasp of cultural, phonetic and linguistic puns. In this paper, we investigate the capacity of contemporary VLMs to interpret and solve rebus puzzles by constructing a hand-generated and annotated benchmark of diverse English-language rebus puzzles, ranging from simple pictographic substitutions to spatially-dependent cues (“head” over “heels”). We analyze how different VLMs perform, and our findings reveal that while VLMs exhibit some surprising capabilities in decoding simple visual clues, they struggle significantly with tasks requiring abstract reasoning, lateral thinking, and understanding visual metaphors.

Method: Uses Platt scaling for calibration and proposes sub-clause frequency (SCF) scores that leverage SQL’s structured nature, combined with multivariate Platt scaling (MPS) for overall confidence scoring.

Result: Empirical evaluation shows substantial improvements in calibration and error detection over using raw model probabilities, with MPS+SCF outperforming traditional Platt scaling.

Conclusion: The proposed approach provides more accurate and calibrated confidence scores for text-to-SQL parsing, enhancing reliability of LLM-based systems.

Abstract: While large language models (LLMs) achieve strong performance on text-to-SQL parsing, they sometimes exhibit unexpected failures in which they are confidently incorrect. Building trustworthy text-to-SQL systems thus requires eliciting reliable uncertainty measures from the LLM. In this paper, we study the problem of providing a calibrated confidence score that conveys the likelihood of an output query being correct. Our work is the first to establish a benchmark for post-hoc calibration of LLM-based text-to-SQL parsing. In particular, we show that Platt scaling, a canonical method for calibration, provides substantial improvements over directly using raw model output probabilities as confidence scores. Furthermore, we propose a method for text-to-SQL calibration that leverages the structured nature of SQL queries to provide more granular signals of correctness, named “sub-clause frequency” (SCF) scores. Using multivariate Platt scaling (MPS), our extension of the canonical Platt scaling technique, we combine individual SCF scores into an overall accurate and calibrated score. Empirical evaluation on two popular text-to-SQL datasets shows that our approach of combining MPS and SCF yields further improvements in calibration and the related task of error detection over traditional Platt scaling.

Method: Created benchmark dataset with diverse plaintexts and encrypted versions, evaluated state-of-the-art LLMs using zero-shot/few-shot settings with chain-of-thought prompting to assess decryption success rates.

Result: Findings reveal key insights into LLMs’ strengths and limitations in side-channel scenarios, showing concerns about their susceptibility to under-generalization-related attacks.

Conclusion: Highlights the dual-use nature of LLMs in security contexts and contributes to AI safety and security discussions, emphasizing the need for better understanding of LLMs’ cryptanalytic capabilities.

Abstract: Recent advancements in large language models (LLMs) have transformed natural language understanding and generation, leading to extensive benchmarking across diverse tasks. However, cryptanalysis - a critical area for data security and its connection to LLMs’ generalization abilities - remains underexplored in LLM evaluations. To address this gap, we evaluate the cryptanalytic potential of state-of-the-art LLMs on ciphertexts produced by a range of cryptographic algorithms. We introduce a benchmark dataset of diverse plaintexts, spanning multiple domains, lengths, writing styles, and topics, paired with their encrypted versions. Using zero-shot and few-shot settings along with chain-of-thought prompting, we assess LLMs’ decryption success rate and discuss their comprehension abilities. Our findings reveal key insights into LLMs’ strengths and limitations in side-channel scenarios and raise concerns about their susceptibility to under-generalization-related attacks. This research highlights the dual-use nature of LLMs in security contexts and contributes to the ongoing discussion on AI safety and security.

Method: Developed Sarc7 benchmark by annotating MUStARD dataset with 7 sarcasm types. Evaluated classification using zero-shot, few-shot, chain-of-thought, and novel emotion-based prompting. Proposed emotion-based generation method focusing on sarcasm components: incongruity, shock value, and context dependency.

Result: Gemini 2.5 with emotion-based prompting achieved best F1 score of 0.3664. Human evaluators preferred emotion-based prompting with 38.46% more successful generations compared to zero-shot prompting.

Conclusion: Emotion-based prompting significantly improves sarcasm classification and generation performance, demonstrating the importance of incorporating emotional context and sarcasm components for better computational understanding of nuanced humor.

Abstract: Sarcasm is a form of humor where expressions convey meanings opposite to their literal interpretations. Classifying and generating sarcasm using large language models is vital for interpreting human communication. Sarcasm poses challenges for computational models, due to its nuanced nature. We introduce Sarc7, a benchmark that classifies 7 types of sarcasm: self-deprecating, brooding, deadpan, polite, obnoxious, raging, and manic by annotating entries of the MUStARD dataset. Classification was evaluated using zero-shot, few-shot, chain-of-thought (CoT), and a novel emotion-based prompting technique. We propose an emotion-based generation method developed by identifying key components of sarcasm-incongruity, shock value, and context dependency. Our classification experiments show that Gemini 2.5, using emotion-based prompting, outperforms other setups with an F1 score of 0.3664. Human evaluators preferred our emotion-based prompting, with 38.46% more successful generations than zero-shot prompting.

Method: Improved RegMean method that directly measures model parameters using Frobenius norm without requiring training data, with an additional hyperparameter for control during merging.

Result: FroM outperforms baseline methods across various fine-tuning scenarios and effectively alleviates the task interference problem in model merging.

Conclusion: The proposed FroM method provides an effective data-free solution for merging fine-tuned models that reduces task interference and improves performance compared to traditional approaches.

Abstract: With the development of large language models, fine-tuning has emerged as an effective method to enhance performance in specific scenarios by injecting domain-specific knowledge. In this context, model merging techniques provide a solution for fusing knowledge from multiple fine-tuning models by combining their parameters. However, traditional methods often encounter task interference when merging full fine-tuning models, and this problem becomes even more evident in parameter-efficient fine-tuning scenarios. In this paper, we introduce an improvement to the RegMean method, which indirectly leverages the training data to approximate the outputs of the linear layers before and after merging. We propose an adaptive merging method called FroM, which directly measures the model parameters using the Frobenius norm, without any training data. By introducing an additional hyperparameter for control, FroM outperforms baseline methods across various fine-tuning scenarios, alleviating the task interference problem.

Method: Created ViMUL-Bench with 8k manually verified samples across 14 languages and 15 culturally diverse categories. Also developed a machine-translated multilingual video training set with 1.2M samples and built ViMUL, a simple multilingual video LMM.

Result: The proposed ViMUL model demonstrates improved performance tradeoffs between high- and low-resource languages for video understanding tasks.

Conclusion: The ViMUL-Bench benchmark, multilingual video LMM, and large-scale training dataset will facilitate future research in developing culturally and linguistically inclusive multilingual video LMMs.

Abstract: Large multimodal models (LMMs) have recently gained attention due to their effectiveness to understand and generate descriptions of visual content. Most existing LMMs are in English language. While few recent works explore multilingual image LMMs, to the best of our knowledge, moving beyond the English language for cultural and linguistic inclusivity is yet to be investigated in the context of video LMMs. In pursuit of more inclusive video LMMs, we introduce a multilingual Video LMM benchmark, named ViMUL-Bench, to evaluate Video LMMs across 14 languages, including both low- and high-resource languages: English, Chinese, Spanish, French, German, Hindi, Arabic, Russian, Bengali, Urdu, Sinhala, Tamil, Swedish, and Japanese. Our ViMUL-Bench is designed to rigorously test video LMMs across 15 categories including eight culturally diverse categories, ranging from lifestyles and festivals to foods and rituals and from local landmarks to prominent cultural personalities. ViMUL-Bench comprises both open-ended (short and long-form) and multiple-choice questions spanning various video durations (short, medium, and long) with 8k samples that are manually verified by native language speakers. In addition, we also introduce a machine translated multilingual video training set comprising 1.2 million samples and develop a simple multilingual video LMM, named ViMUL, that is shown to provide a better tradeoff between high-and low-resource languages for video understanding. We hope our ViMUL-Bench and multilingual video LMM along with a large-scale multilingual video training set will help ease future research in developing cultural and linguistic inclusive multilingual video LMMs. Our proposed benchmark, video LMM and training data will be publicly released at https://mbzuai-oryx.github.io/ViMUL/.

Method: Extended SciTab benchmark with human-annotated cell-level rationales, created a taxonomy for handling ambiguous cases, and conducted experiments to evaluate claim verification performance and rationale alignment.

Result: Incorporating table alignment information improves claim verification performance, but most LLMs fail to recover human-aligned rationales despite predicting correct labels.

Conclusion: LLMs’ predictions do not stem from faithful reasoning, highlighting the importance of explanation tasks and cell-level rationales for interpretable scientific claim verification.

Abstract: Scientific claim verification against tables typically requires predicting whether a claim is supported or refuted given a table. However, we argue that predicting the final label alone is insufficient: it reveals little about the model’s reasoning and offers limited interpretability. To address this, we reframe table-text alignment as an explanation task, requiring models to identify the table cells essential for claim verification. We build a new dataset by extending the SciTab benchmark with human-annotated cell-level rationales. Annotators verify the claim label and highlight the minimal set of cells needed to support their decision. After the annotation process, we utilize the collected information and propose a taxonomy for handling ambiguous cases. Our experiments show that (i) incorporating table alignment information improves claim verification performance, and (ii) most LLMs, while often predicting correct labels, fail to recover human-aligned rationales, suggesting that their predictions do not stem from faithful reasoning.

Method: Developed FinCoT framework that embeds domain-specific expert financial reasoning blueprints. Evaluated three prompting approaches (standard, unstructured CoT, structured CoT) across ten CFA-style financial domains using both general-purpose and finance-specific models.

Result: FinCoT improved Qwen3-8B-Base accuracy from 63.2% to 80.5% and Fin-R1 (7B) from 65.7% to 75.7%, while reducing output length by up to 8.9x and 1.16x respectively compared to structured CoT methods. Most effective for models lacking financial post-training.

Conclusion: FinCoT not only improves performance and reduces inference costs but also yields more interpretable and expert-aligned reasoning traces, demonstrating the value of incorporating domain-specific structured reasoning in financial NLP tasks.

Abstract: This paper presents FinCoT, a structured chain-of-thought (CoT) prompting framework that embeds domain-specific expert financial reasoning blueprints to guide large language models’ behaviors. We identify three main prompting styles in financial NLP (FinNLP): (1) standard prompting (zero-shot), (2) unstructured CoT (free-form reasoning), and (3) structured CoT (with explicitly structured reasoning steps). Prior work has mainly focused on the first two, while structured CoT remains underexplored and lacks domain expertise incorporation. Therefore, we evaluate all three prompting approaches across ten CFA-style financial domains and introduce FinCoT as the first structured finance-specific prompting approach incorporating blueprints from domain experts. FinCoT improves the accuracy of a general-purpose model, Qwen3-8B-Base, from 63.2% to 80.5%, and boosts Fin-R1 (7B), a finance-specific model, from 65.7% to 75.7%, while reducing output length by up to 8.9x and 1.16x compared to structured CoT methods, respectively. We find that FinCoT proves most effective for models lacking financial post-training. Our findings show that FinCoT does not only improve performance and reduce inference costs but also yields more interpretable and expert-aligned reasoning traces.

Method: Constructed SpeechRole-Data with 98 diverse roles and 112k speech conversations, each with distinct vocal characteristics. Proposed SpeechRole-Eval benchmark to assess fundamental interaction ability, speech expressiveness, and role-playing fidelity.

Result: Experimental results revealed advantages and challenges of both cascaded and end-to-end speech role-playing agents in maintaining vocal style consistency and role coherence.

Conclusion: The released data, code, and baseline models provide a solid foundation for speech-driven multimodal role-playing research and will foster further developments in this field.

Abstract: Recently, role-playing agents have emerged as a promising paradigm for achieving personalized interaction and emotional resonance. Existing research primarily focuses on the textual modality, neglecting the critical dimension of speech in realistic interactive scenarios. In particular, there is a lack of systematic evaluation for Speech Role-Playing Agents (SRPAs). To address this gap, we construct SpeechRole-Data, a large-scale, high-quality dataset that comprises 98 diverse roles and 112k speech-based single-turn and multi-turn conversations. Each role demonstrates distinct vocal characteristics, including timbre and prosody, thereby enabling more sophisticated speech role-playing. Furthermore, we propose SpeechRole-Eval, a multidimensional evaluation benchmark that systematically assesses SRPAs performance in key aspects such as fundamental interaction ability, speech expressiveness, and role-playing fidelity. Experimental results reveal the advantages and challenges of both cascaded and end-to-end speech role-playing agents in maintaining vocal style consistency and role coherence. We release all data, code, and baseline models to provide a solid foundation for speech-driven multimodal role-playing research and to foster further developments in this field.

Method: Uses an interaction-based NLU teacher model to provide comprehensive attention patterns, then guides preference modeling to simulate these patterns through attentional alignment objectives.

Result: Demonstrates more stable and generalizable reward signals compared to state-of-the-art RM optimization methods, showing attention hacking is a fundamental limitation.

Conclusion: Attention hacking constitutes a fundamental limitation in current reward models, and Interaction Distillation effectively addresses this through attention-level optimization.

Abstract: The reward model (RM), as the core component of reinforcement learning from human feedback (RLHF) for large language models (LLMs), responsible for providing reward signals to generated responses. However, mainstream preference modeling in RM is inadequate in terms of token-level interaction, making its judgment signals vulnerable to being hacked by misallocated attention to context. This stems from two fundamental limitations: (1) Current preference modeling employs decoder-only architectures, where the unidirectional causal attention mechanism leads to forward-decaying intra-sequence attention within the prompt-response sequence. (2) The independent Siamese-encoding paradigm induces the absence of token-level inter-sequence attention between chosen and rejected sequences. To address this “attention hacking”, we propose “Interaction Distillation”, a novel training framework for more adequate preference modeling through attention-level optimization. The method introduces an interaction-based natural language understanding model as the teacher to provide sophisticated token interaction patterns via comprehensive attention, and guides the preference modeling to simulate teacher model’s interaction pattern through an attentional alignment objective. Through extensive experiments, interaction distillation has demonstrated its ability to provide more stable and generalizable reward signals compared to state-of-the-art RM optimization methods that target data noise, highlighting the attention hacking constitute a more fundamental limitation in RM.

Method: Stepwise Experience Recall (SEER) performs fine-grained, stepwise retrieval from a continually updated experience pool that incrementally adds past successful trajectories, enabling continuous improvement.

Result: SEER achieved 6.1% improvement on easy and 4.7% on hard ToolQA questions. On τ-bench with real-world domains, it showed 7.44% and 23.38% accuracy gains with Qwen2.5-7B and Qwen2.5-72B models respectively.

Conclusion: SEER provides an effective self-guided approach for improving LLM tool usage through continuous experience accumulation, demonstrating substantial performance improvements across different benchmarks and model sizes.

Abstract: Function calling enables large language models (LLMs) to interact with external systems by leveraging tools and APIs. When faced with multi-step tool usage, LLMs still struggle with tool selection, parameter generation, and tool-chain planning. Existing methods typically rely on manually designing task-specific demonstrations, or retrieving from a curated library. These approaches demand substantial expert effort and prompt engineering becomes increasingly complex and inefficient as tool diversity and task difficulty scale. To address these challenges, we propose a self-guided method, Stepwise Experience Recall (SEER), which performs fine-grained, stepwise retrieval from a continually updated experience pool. Instead of relying on static or manually curated library, SEER incrementally augments the experience pool with past successful trajectories, enabling continuous expansion of the pool and improved model performance over time. Evaluated on the ToolQA benchmark, SEER achieves an average improvement of 6.1% on easy and 4.7% on hard questions. We further test SEER on $\tau$-bench, which includes two real-world domains. Powered by Qwen2.5-7B and Qwen2.5-72B models, SEER demonstrates substantial accuracy gains of 7.44% and 23.38%, respectively.

Method: A position-to-position knowledge distillation framework that leverages position-induced disparity to counteract positional bias, with two specialized instantiations: Pos2Distill-R¹ for retrieval and Pos2Distill-R² for reasoning tasks.

Result: Achieves enhanced uniformity and significant performance gains across all contextual positions in long-context retrieval and reasoning tasks, with strong cross-task generalization.

Conclusion: Pos2Distill effectively addresses positional bias by transferring knowledge between positions, achieving superior performance while maintaining computational efficiency compared to previous approaches.

Abstract: Positional bias (PB), manifesting as non-uniform sensitivity across different contextual locations, significantly impairs long-context comprehension and processing capabilities. Previous studies have addressed PB either by modifying the underlying architectures or by employing extensive contextual awareness training. However, the former approach fails to effectively eliminate the substantial performance disparities, while the latter imposes significant data and computational overhead. To address PB effectively, we introduce \textbf{Pos2Distill}, a position to position knowledge distillation framework. Pos2Distill transfers the superior capabilities from advantageous positions to less favorable ones, thereby reducing the huge performance gaps. The conceptual principle is to leverage the inherent, position-induced disparity to counteract the PB itself. We identify distinct manifestations of PB under \textbf{\textsc{r}}etrieval and \textbf{\textsc{r}}easoning paradigms, thereby designing two specialized instantiations: \emph{Pos2Distill-R\textsuperscript{1}} and \emph{Pos2Distill-R\textsuperscript{2}} respectively, both grounded in this core principle. By employing the Pos2Distill approach, we achieve enhanced uniformity and significant performance gains across all contextual positions in long-context retrieval and reasoning tasks. Crucially, both specialized systems exhibit strong cross-task generalization mutually, while achieving superior performance on their respective tasks.

Method: Designed scenarios with ambiguous user instructions conflicting with given goals, covering scalar implicature, structural ambiguities, and power dynamics. Measured different models’ abilities to exploit loopholes to satisfy their own goals versus user goals.

Result: Both closed-source and stronger open-source models can identify ambiguities and exploit resulting loopholes. Models that exploit loopholes explicitly identify and reason about both ambiguity and conflicting goals.

Conclusion: LLMs’ ability to exploit loopholes presents a potential AI safety risk, as models demonstrate sophisticated pragmatic reasoning to bypass user intentions when faced with conflicting goals.

Abstract: Studying the responses of large language models (LLMs) to loopholes presents a two-fold opportunity. First, it affords us a lens through which to examine ambiguity and pragmatics in LLMs, since exploiting a loophole requires identifying ambiguity and performing sophisticated pragmatic reasoning. Second, loopholes pose an interesting and novel alignment problem where the model is presented with conflicting goals and can exploit ambiguities to its own advantage. To address these questions, we design scenarios where LLMs are given a goal and an ambiguous user instruction in conflict with the goal, with scenarios covering scalar implicature, structural ambiguities, and power dynamics. We then measure different models’ abilities to exploit loopholes to satisfy their given goals as opposed to the goals of the user. We find that both closed-source and stronger open-source models can identify ambiguities and exploit their resulting loopholes, presenting a potential AI safety risk. Our analysis indicates that models which exploit loopholes explicitly identify and reason about both ambiguity and conflicting goals.

Method: Created T2R-bench with 457 industrial tables from 19 domains and 4 table types, derived from real-world scenarios, and proposed evaluation criteria to measure report generation quality.

Result: Experiments on 25 LLMs show even state-of-the-art models like Deepseek-R1 only achieve 62.71 overall score, indicating significant room for improvement.

Conclusion: The table-to-report task remains challenging for LLMs, and T2R-bench provides a valuable benchmark to drive future research and improvements in this industrial application area.

Abstract: Extensive research has been conducted to explore the capabilities of large language models (LLMs) in table reasoning. However, the essential task of transforming tables information into reports remains a significant challenge for industrial applications. This task is plagued by two critical issues: 1) the complexity and diversity of tables lead to suboptimal reasoning outcomes; and 2) existing table benchmarks lack the capacity to adequately assess the practical application of this task. To fill this gap, we propose the table-to-report task and construct a bilingual benchmark named T2R-bench, where the key information flow from the tables to the reports for this task. The benchmark comprises 457 industrial tables, all derived from real-world scenarios and encompassing 19 industry domains as well as 4 types of industrial tables. Furthermore, we propose an evaluation criteria to fairly measure the quality of report generation. The experiments on 25 widely-used LLMs reveal that even state-of-the-art models like Deepseek-R1 only achieves performance with 62.71 overall score, indicating that LLMs still have room for improvement on T2R-bench.

Method: Instructed seller LLM agents to apply anchoring effect and evaluated negotiations using both objective and subjective metrics, while examining relationships with reasoning and personality factors.

Result: LLMs are influenced by anchoring effect like humans; reasoning models are less prone to anchoring (long chain of thought mitigates effect); no significant correlation between personality traits and anchoring susceptibility.

Conclusion: Findings contribute to understanding cognitive biases in LLMs and support safe, responsible application of LLMs in society by identifying factors that influence bias susceptibility.

Abstract: Cognitive biases, well-studied in humans, can also be observed in LLMs, affecting their reliability in real-world applications. This paper investigates the anchoring effect in LLM-driven price negotiations. To this end, we instructed seller LLM agents to apply the anchoring effect and evaluated negotiations using not only an objective metric but also a subjective metric. Experimental results show that LLMs are influenced by the anchoring effect like humans. Additionally, we investigated the relationship between the anchoring effect and factors such as reasoning and personality. It was shown that reasoning models are less prone to the anchoring effect, suggesting that the long chain of thought mitigates the effect. However, we found no significant correlation between personality traits and susceptibility to the anchoring effect. These findings contribute to a deeper understanding of cognitive biases in LLMs and to the realization of safe and responsible application of LLMs in society.

Method: Assessed 7 LLMs using a benchmark of 1,000 multiple-choice questions covering diverse inheritance scenarios, testing models’ ability to understand inheritance context and compute distribution shares according to Islamic jurisprudence.

Result: Significant performance gap observed: o3 and Gemini 2.5 achieved accuracies above 90%, while ALLaM, Fanar, LLaMA, and Mistral scored below 50%. Error analysis revealed recurring failure patterns including scenario misunderstandings, incorrect rule application, and insufficient domain knowledge.

Conclusion: Current LLMs have limitations in handling structured legal reasoning for Islamic inheritance law, highlighting the need for improved domain adaptation and suggesting directions for enhancing performance in specialized legal reasoning tasks.

Abstract: This paper evaluates the knowledge and reasoning capabilities of Large Language Models in Islamic inheritance law, known as ‘ilm al-mawarith. We assess the performance of seven LLMs using a benchmark of 1,000 multiple-choice questions covering diverse inheritance scenarios, designed to test models’ ability to understand the inheritance context and compute the distribution of shares prescribed by Islamic jurisprudence. The results reveal a significant performance gap: o3 and Gemini 2.5 achieved accuracies above 90%, whereas ALLaM, Fanar, LLaMA, and Mistral scored below 50%. These disparities reflect important differences in reasoning ability and domain adaptation. We conduct a detailed error analysis to identify recurring failure patterns across models, including misunderstandings of inheritance scenarios, incorrect application of legal rules, and insufficient domain knowledge. Our findings highlight limitations in handling structured legal reasoning and suggest directions for improving performance in Islamic legal reasoning. Code: https://github.com/bouchekif/inheritance_evaluation

Method: The authors propose Context-Aware Molecular T5 (CAMT5) with substructure-level tokenization and an importance-based training strategy that prioritizes key substructures. They also develop a simple ensemble strategy to aggregate outputs from multiple text-to-molecule models.

Result: Extensive experiments show CAMT5’s superiority in various text-to-molecule generation tasks. It outperforms state-of-the-art methods using only 2% of training tokens. The ensemble strategy further boosts generation performance.

Conclusion: Substructure-level tokenization and importance-based training enable better capture of molecular semantics. CAMT5 demonstrates that focusing on key substructures rather than individual atoms leads to more efficient and effective text-to-molecule generation with significantly reduced training requirements.

Abstract: Recently, text-to-molecule models have shown great potential across various chemical applications, e.g., drug-discovery. These models adapt language models to molecular data by representing molecules as sequences of atoms. However, they rely on atom-level tokenizations, which primarily focus on modeling local connectivity, thereby limiting the ability of models to capture the global structural context within molecules. To tackle this issue, we propose a novel text-to-molecule model, coined Context-Aware Molecular T5 (CAMT5). Inspired by the significance of the substructure-level contexts in understanding molecule structures, e.g., ring systems, we introduce substructure-level tokenization for text-to-molecule models. Building on our tokenization scheme, we develop an importance-based training strategy that prioritizes key substructures, enabling CAMT5 to better capture the molecular semantics. Extensive experiments verify the superiority of CAMT5 in various text-to-molecule generation tasks. Intriguingly, we find that CAMT5 outperforms the state-of-the-art methods using only 2% of training tokens. In addition, we propose a simple yet effective ensemble strategy that aggregates the outputs of text-to-molecule models to further boost the generation performance. Code is available at https://github.com/Songhyeontae/CAMT5.git.

Method: Created IntrEx dataset from Teacher-Student Chatroom Corpus with sequence-level annotations, used comparison-based rating with 100+ second-language learners inspired by RLHF, and tested LLMs (7B/8B parameters) for interestingness prediction.

Result: Fine-tuned LLMs outperformed larger proprietary models like GPT-4o in predicting human interestingness judgments, demonstrating specialized datasets effectively model engagement. Linguistic factors like concreteness, comprehensibility, and uptake significantly influence engagement.

Conclusion: Specialized datasets like IntrEx enable effective modeling of engagement in educational dialogues, with fine-tuned smaller LLMs outperforming larger general models, providing insights into linguistic drivers of conversational interest.

Abstract: Engagement and motivation are crucial for second-language acquisition, yet maintaining learner interest in educational conversations remains a challenge. While prior research has explored what makes educational texts interesting, still little is known about the linguistic features that drive engagement in conversations. To address this gap, we introduce IntrEx, the first large dataset annotated for interestingness and expected interestingness in teacher-student interactions. Built upon the Teacher-Student Chatroom Corpus (TSCC), IntrEx extends prior work by incorporating sequence-level annotations, allowing for the study of engagement beyond isolated turns to capture how interest evolves over extended dialogues. We employ a rigorous annotation process with over 100 second-language learners, using a comparison-based rating approach inspired by reinforcement learning from human feedback (RLHF) to improve agreement. We investigate whether large language models (LLMs) can predict human interestingness judgments. We find that LLMs (7B/8B parameters) fine-tuned on interestingness ratings outperform larger proprietary models like GPT-4o, demonstrating the potential for specialised datasets to model engagement in educational settings. Finally, we analyze how linguistic and cognitive factors, such as concreteness, comprehensibility (readability), and uptake, influence engagement in educational dialogues.

Method: A query-driven human-agent-GUI interaction framework with a two-stage learning paradigm that decouples and utilizes meta-knowledge, enabling the agent to decide when to query humans.

Result: VeriOS-Agent improves average step-wise success rate by 20.64% in untrustworthy scenarios over state-of-the-art methods without compromising normal performance.

Conclusion: The framework demonstrates rationality, generalizability, and scalability, providing a trustworthy solution for OS agents operating in real-world untrustworthy environments.

Abstract: With the rapid progress of multimodal large language models, operating system (OS) agents become increasingly capable of automating tasks through on-device graphical user interfaces (GUIs). However, most existing OS agents are designed for idealized settings, whereas real-world environments often present untrustworthy conditions. To mitigate risks of over-execution in such scenarios, we propose a query-driven human-agent-GUI interaction framework that enables OS agents to decide when to query humans for more reliable task completion. Built upon this framework, we introduce VeriOS-Agent, a trustworthy OS agent trained with a two-stage learning paradigm that falicitate the decoupling and utilization of meta-knowledge. Concretely, VeriOS-Agent autonomously executes actions in normal conditions while proactively querying humans in untrustworthy scenarios. Experiments show that VeriOS-Agent improves the average step-wise success rate by 20.64% in untrustworthy scenarios over the state-of-the-art, without compromising normal performance. Analysis highlights VeriOS-Agent’s rationality, generalizability, and scalability. The codes, datasets and models are available at https://github.com/Wuzheng02/VeriOS.

Method: Conducted experiments with XLM-RoBERTa, Longformer, GPT-3.5, and GPT-4 models on multiclass classification of policy topics using the Comparative Agendas Project codebook (21 policy labels) across 5 languages.

Result: No particular advantage for Longformer (specifically designed for long inputs). Open models outperformed GPT variants. Performance depends on support and substance overlaps between specific policy categories.

Conclusion: Specialized long-text models don’t necessarily outperform standard models for long document classification, and open models can be more effective than proprietary GPT models for this task.

Abstract: The most widely used large language models in the social sciences (such as BERT, and its derivatives, e.g. RoBERTa) have a limitation on the input text length that they can process to produce predictions. This is a particularly pressing issue for some classification tasks, where the aim is to handle long input texts. One such area deals with laws and draft laws (bills), which can have a length of multiple hundred pages and, therefore, are not particularly amenable for processing with models that can only handle e.g. 512 tokens. In this paper, we show results from experiments covering 5 languages with XLM-RoBERTa, Longformer, GPT-3.5, GPT-4 models for the multiclass classification task of the Comparative Agendas Project, which has a codebook of 21 policy topic labels from education to health care. Results show no particular advantage for the Longformer model, pre-trained specifically for the purposes of handling long inputs. The comparison between the GPT variants and the best-performing open model yielded an edge for the latter. An analysis of class-level factors points to the importance of support and substance overlaps between specific categories when it comes to performance on long text inputs.

Method: Investigated entropy neurons in autoregressive transformer models by analyzing their impact on output entropy and token ranking, and performed ablation studies to observe changes in generation behavior.

Result: Entropy neurons are responsible for suppressing context copying across various LLMs, and their ablation leads to significant changes in the generation process when handling conflicting information.

Conclusion: These findings enhance our understanding of LLM internal dynamics when processing conflicting contextual and parametric information, revealing the specific role of entropy neurons in context suppression.

Abstract: The behavior of Large Language Models (LLMs) when facing contextual information that conflicts with their internal parametric knowledge is inconsistent, with no generally accepted explanation for the expected outcome distribution. Recent work has identified in autoregressive transformer models a class of neurons – called entropy neurons – that produce a significant effect on the model output entropy while having an overall moderate impact on the ranking of the predicted tokens. In this paper, we investigate the preliminary claim that these neurons are involved in inhibiting context copying behavior in transformers by looking at their role in resolving conflicts between contextual and parametric information. We show that entropy neurons are responsible for suppressing context copying across a range of LLMs, and that ablating them leads to a significant change in the generation process. These results enhance our understanding of the internal dynamics of LLMs when handling conflicting information.

Method: Used mt5-based encoder and Qwen decoder approaches, augmented training data with automatically translated English data for low-resource languages, and incorporated additional linguistic features from previous shared task entries.

Result: Achieved a macro-accuracy score of 71.28 on the discourse relation classification task.

Conclusion: The system demonstrates effective performance in discourse relation classification, particularly benefiting from data augmentation techniques for low-resource languages and insights from error analysis.

Abstract: This paper presents DeDisCo, Georgetown University’s entry in the DISRPT 2025 shared task on discourse relation classification. We test two approaches, using an mt5-based encoder and a decoder based approach using the openly available Qwen model. We also experiment on training with augmented dataset for low-resource languages using matched data translated automatically from English, as well as using some additional linguistic features inspired by entries in previous editions of the Shared Task. Our system achieves a macro-accuracy score of 71.28, and we provide some interpretation and error analysis for our results.

Method: Proposes MOOM - a dual-branch memory plugin that models plot development (summarizing conflicts across time scales) and character portrayal (extracting user profiles), integrated with a forgetting mechanism inspired by competition-inhibition memory theory.

Result: MOOM outperforms all state-of-the-art memory extraction methods, requires fewer large language model invocations, and maintains controllable memory capacity. Also presents ZH-4O dataset with 600-turn average dialogues and manual memory annotations.

Conclusion: The proposed MOOM framework effectively addresses uncontrolled memory growth in ultra-long role-playing dialogues by leveraging literary theory principles and incorporating a biologically-inspired forgetting mechanism, demonstrating superior performance with computational efficiency.

Abstract: Memory extraction is crucial for maintaining coherent ultra-long dialogues in human-robot role-playing scenarios. However, existing methods often exhibit uncontrolled memory growth. To address this, we propose MOOM, the first dual-branch memory plugin that leverages literary theory by modeling plot development and character portrayal as core storytelling elements. Specifically, one branch summarizes plot conflicts across multiple time scales, while the other extracts the user’s character profile. MOOM further integrates a forgetting mechanism, inspired by the ``competition-inhibition’’ memory theory, to constrain memory capacity and mitigate uncontrolled growth. Furthermore, we present ZH-4O, a Chinese ultra-long dialogue dataset specifically designed for role-playing, featuring dialogues that average 600 turns and include manually annotated memory information. Experimental results demonstrate that MOOM outperforms all state-of-the-art memory extraction methods, requiring fewer large language model invocations while maintaining a controllable memory capacity.

Method: Uses deep learning to compare satellite image pairs from different dates to detect forest cover changes, and proposes a visual semantic model that automatically annotates changes with keywords extracted from scientific documents about the Amazon.

Result: Evaluated on Amazon image pair dataset, demonstrating effectiveness in both deforestation detection and generating relevant annotations.

Conclusion: Provides a useful tool for monitoring deforestation impacts in the Amazon, and the approach is generic enough to be applied to other domains beyond environmental applications.

Abstract: The Amazon rain forest is a vital ecosystem that plays a crucial role in regulating the Earth’s climate and providing habitat for countless species. Deforestation in the Amazon is a major concern as it has a significant impact on global carbon emissions and biodiversity. In this paper, we present a method for detecting deforestation in the Amazon using image pairs from Earth observation satellites. Our method leverages deep learning techniques to compare the images of the same area at different dates and identify changes in the forest cover. We also propose a visual semantic model that automatically annotates the detected changes with relevant keywords. The candidate annotation for images are extracted from scientific documents related to the Amazon region. We evaluate our approach on a dataset of Amazon image pairs and demonstrate its effectiveness in detecting deforestation and generating relevant annotations. Our method provides a useful tool for monitoring and studying the impact of deforestation in the Amazon. While we focus on environment applications of our work by using images of deforestation in the Amazon rain forest to demonstrate the effectiveness of our proposed approach, it is generic enough to be applied to other domains.

Method: Two strategies: (1) EMA-guided pseudo supervision framework generating reliable segment-level masks via adaptive thresholds or top-k selection, (2) class-aware cross-modal agreement loss aligning audio and visual embeddings at reliable segment-class pairs.

Result: Achieves state-of-the-art performance across multiple metrics on LLP and UnAV-100 datasets.

Conclusion: The proposed EMA-guided pseudo supervision and class-aware cross-modal alignment effectively address limitations in weakly-supervised AVVP, providing stable temporal guidance and ensuring cross-modal consistency.

Abstract: Weakly-supervised audio-visual video parsing (AVVP) seeks to detect audible, visible, and audio-visual events without temporal annotations. Previous work has emphasized refining global predictions through contrastive or collaborative learning, but neglected stable segment-level supervision and class-aware cross-modal alignment. To address this, we propose two strategies: (1) an exponential moving average (EMA)-guided pseudo supervision framework that generates reliable segment-level masks via adaptive thresholds or top-k selection, offering stable temporal guidance beyond video-level labels; and (2) a class-aware cross-modal agreement (CMA) loss that aligns audio and visual embeddings at reliable segment-class pairs, ensuring consistency across modalities while preserving temporal structure. Evaluations on LLP and UnAV-100 datasets shows that our method achieves state-of-the-art (SOTA) performance across multiple metrics.

Method: Retrieves proximal exemplars for each test instance in embedding space, fuses their predictive distributions using Dempster-Shafer theory, and uses the resulting fused belief as a per-instance thresholding mechanism.

Result: Experiments on CIFAR-10/100 with BiT and ViT backbones show higher or comparable uncertainty-aware performance with significantly fewer confidently incorrect outcomes and sustainable review load compared to prediction entropy thresholds.

Conclusion: Evidence-conditioned tagging provides a more reliable and interpretable alternative to fixed prediction entropy thresholds for operational uncertainty-aware decision-making, with gains achievable using only a few evidences.

Abstract: This work proposes an evidence-retrieval mechanism for uncertainty-aware decision-making that replaces a single global cutoff with an evidence-conditioned, instance-adaptive criterion. For each test instance, proximal exemplars are retrieved in an embedding space; their predictive distributions are fused via Dempster-Shafer theory. The resulting fused belief acts as a per-instance thresholding mechanism. Because the supporting evidences are explicit, decisions are transparent and auditable. Experiments on CIFAR-10/100 with BiT and ViT backbones show higher or comparable uncertainty-aware performance with materially fewer confidently incorrect outcomes and a sustainable review load compared with applying threshold on prediction entropy. Notably, only a few evidences are sufficient to realize these gains; increasing the evidence set yields only modest changes. These results indicate that evidence-conditioned tagging provides a more reliable and interpretable alternative to fixed prediction entropy thresholds for operational uncertainty-aware decision-making.

Method: Conducted experiments over 50 training epochs on three benchmark datasets (MNIST, CIFAR100, STL10) comparing hybrid models (parameterized quantum circuits + classical DL) vs classical CNNs, evaluating validation accuracy, test accuracy, training time, computational resources, and adversarial robustness with ε=0.1 perturbations.

Result: Hybrid models achieved superior accuracy (99.38% vs 98.21% on MNIST, 41.69% vs 32.25% on CIFAR100, 74.05% vs 63.76% on STL10), trained 5-12x faster, used 6-32% fewer parameters, consumed less memory (4-5GB vs 5-6GB), and showed better adversarial robustness on simpler datasets.

Conclusion: Hybrid quantum-classical architectures offer compelling advantages in accuracy, training efficiency, and parameter scalability, particularly for complex vision tasks, demonstrating their potential as superior alternatives to classical models.

Abstract: This study presents a systematic comparison between hybrid quantum-classical neural networks and purely classical models across three benchmark datasets (MNIST, CIFAR100, and STL10) to evaluate their performance, efficiency, and robustness. The hybrid models integrate parameterized quantum circuits with classical deep learning architectures, while the classical counterparts use conventional convolutional neural networks (CNNs). Experiments were conducted over 50 training epochs for each dataset, with evaluations on validation accuracy, test accuracy, training time, computational resource usage, and adversarial robustness (tested with $\epsilon=0.1$ perturbations).Key findings demonstrate that hybrid models consistently outperform classical models in final accuracy, achieving {99.38% (MNIST), 41.69% (CIFAR100), and 74.05% (STL10) validation accuracy, compared to classical benchmarks of 98.21%, 32.25%, and 63.76%, respectively. Notably, the hybrid advantage scales with dataset complexity, showing the most significant gains on CIFAR100 (+9.44%) and STL10 (+10.29%). Hybrid models also train 5–12$\times$ faster (e.g., 21.23s vs. 108.44s per epoch on MNIST) and use 6–32% fewer parameters} while maintaining superior generalization to unseen test data.Adversarial robustness tests reveal that hybrid models are significantly more resilient on simpler datasets (e.g., 45.27% robust accuracy on MNIST vs. 10.80% for classical) but show comparable fragility on complex datasets like CIFAR100 ($\sim$1% robustness for both). Resource efficiency analyses indicate that hybrid models consume less memory (4–5GB vs. 5–6GB for classical) and lower CPU utilization (9.5% vs. 23.2% on average).These results suggest that hybrid quantum-classical architectures offer compelling advantages in accuracy, training efficiency, and parameter scalability, particularly for complex vision tasks.

Method: Optimized YOLOv11 with DIoU Loss instead of GIoU Loss, improved DeepSort by fusing Mahalanobis and cosine distances, used Greenberg model for traffic flow analysis, and built GRU-Attention model for congestion prediction trained on flow, density, and speed data.

Result: YOLOv11-DIoU achieved 95.7% mAP (6.5% higher than baseline) with 5.3% occlusion miss rate. DeepSort reached 93.8% MOTA (11.3% higher than SORT). GRU-Attention model achieved 99.7% test accuracy with ≤1 minute time error in 10-minute advance warnings and 95% warning accuracy in validation.

Conclusion: The framework provides quantitative support for expressway congestion control with promising applications in intelligent transportation systems, demonstrating high accuracy in both vehicle perception and congestion prediction.

Abstract: Expressway traffic congestion severely reduces travel efficiency and hinders regional connectivity. Existing “detection-prediction” systems have critical flaws: low vehicle perception accuracy under occlusion and loss of long-sequence dependencies in congestion forecasting. This study proposes an integrated technical framework to resolve these issues.For traffic flow perception, two baseline algorithms were optimized. Traditional YOLOv11 was upgraded to YOLOv11-DIoU by replacing GIoU Loss with DIoU Loss, and DeepSort was improved by fusing Mahalanobis (motion) and cosine (appearance) distances. Experiments on Chang-Shen Expressway videos showed YOLOv11-DIoU achieved 95.7% mAP (6.5 percentage points higher than baseline) with 5.3% occlusion miss rate. DeepSort reached 93.8% MOTA (11.3 percentage points higher than SORT) with only 4 ID switches. Using the Greenberg model (for 10-15 vehicles/km high-density scenarios), speed and density showed a strong negative correlation (r=-0.97), conforming to traffic flow theory. For congestion warning, a GRU-Attention model was built to capture congestion precursors. Trained 300 epochs with flow, density, and speed, it achieved 99.7% test accuracy (7-9 percentage points higher than traditional GRU). In 10-minute advance warnings for 30-minute congestion, time error was $\leq$ 1 minute. Validation with an independent video showed 95% warning accuracy, over 90% spatial overlap of congestion points, and stable performance in high-flow ($>$5 vehicles/second) scenarios.This framework provides quantitative support for expressway congestion control, with promising intelligent transportation applications.

Method: Applied machine learning methods, specifically convolutional neural networks for image pattern recognition, to automate the analysis process of crowd-sourced ultrasonic sensor data for parking spot detection.

Result: Achieved 99.58% reduction in human resource time requirements while maintaining high performance levels in parking spot classification.

Conclusion: The automation tool successfully optimized the parking service quality and demonstrated significant efficiency improvements, with potential for future development and broader application.

Abstract: This research is part of a study of a real-time, cloud-based on-street parking service using crowd-sourced in-vehicle fleet data. The service provides real-time information about available parking spots by classifying crowd-sourced detections observed via ultrasonic sensors. The goal of this research is to optimize the current parking service quality by analyzing the automation of the existing test process for ground truth tests. Therefore, methods from the field of machine learning, especially image pattern recognition, are applied to enrich the database and substitute human engineering work in major areas of the analysis process. After an introduction into the related areas of machine learning, this paper explains the methods and implementations made to achieve a high level of automation, applying convolutional neural networks. Finally, predefined metrics present the performance level achieved, showing a time reduction of human resources up to 99.58 %. The overall improvements are discussed, summarized, and followed by an outlook for future development and potential application of the analysis automation tool.

Method: The authors conducted systematic empirical analysis using in-distribution and OOD prompts, characterizing operational properties in VLM embedding space, quantifying performance advantages, and testing sensitivity to different types of perturbations.

Result: VLMs leverage rich semantic novelty for superior OOD detection compared to single-modal methods. They show resilience to image noise but exhibit significant sensitivity to prompt phrasing, revealing an asymmetry in robustness.

Conclusion: The study provides structured understanding of VLM strengths and critical vulnerabilities in OOD detection, offering empirically-grounded guidance for developing more robust and reliable future VLM designs.

Abstract: Vision-Language Models (VLMs), such as CLIP, have demonstrated remarkable zero-shot out-of-distribution (OOD) detection capabilities, vital for reliable AI systems. Despite this promising capability, a comprehensive understanding of (1) why they work so effectively, (2) what advantages do they have over single-modal methods, and (3) how is their behavioral robustness – remains notably incomplete within the research community. This paper presents a systematic empirical analysis of VLM-based OOD detection using in-distribution (ID) and OOD prompts. (1) Mechanisms: We systematically characterize and formalize key operational properties within the VLM embedding space that facilitate zero-shot OOD detection. (2) Advantages: We empirically quantify the superiority of these models over established single-modal approaches, attributing this distinct advantage to the VLM’s capacity to leverage rich semantic novelty. (3) Sensitivity: We uncovers a significant and previously under-explored asymmetry in their robustness profile: while exhibiting resilience to common image noise, these VLM-based methods are highly sensitive to prompt phrasing. Our findings contribute a more structured understanding of the strengths and critical vulnerabilities inherent in VLM-based OOD detection, offering crucial, empirically-grounded guidance for developing more robust and reliable future designs.

Method: Utilizes abstract notions of sectional curvature to capture metric relations between triples of points and construct curvature-based geometric profiles of discrete metric spaces.

Result: The curvature-based analysis can estimate intrinsic dimensionality of datasets and evaluate effectiveness of dimensionality reduction techniques.

Conclusion: The proposed curvature profile provides a valuable tool for geometric analysis of discrete metric spaces and quantitative assessment of data representation quality.

Abstract: Utilizing recently developed abstract notions of sectional curvature, we introduce a method for constructing a curvature-based geometric profile of discrete metric spaces. The curvature concept that we use here captures the metric relations between triples of points and other points. More significantly, based on this curvature profile, we introduce a quantitative measure to evaluate the effectiveness of data representations, such as those produced by dimensionality reduction techniques. Furthermore, Our experiments demonstrate that this curvature-based analysis can be employed to estimate the intrinsic dimensionality of datasets. We use this to explore the large-scale geometry of empirical networks and to evaluate the effectiveness of dimensionality reduction techniques.

Method: Used Landsat-8 satellite imagery with supervised machine learning training datasets. Applied Google Earth Engine with k-means clustering for land cover mapping and convolutional neural networks for land cover classification.

Result: Developed visualization of change detection showing urban area changes over time, enabling monitoring of land cover transformations in the study region.

Conclusion: The framework successfully provides technical support for land cover/land use modeling and change detection, demonstrating effective monitoring of urbanization patterns in developing regions like Fiji.

Abstract: As a developing country, Fiji is facing rapid urbanisation, which is visible in the massive development projects that include housing, roads, and civil works. In this study, we present machine learning and remote sensing frameworks to compare land use and land cover change from 2013 to 2024 in Nadi, Fiji. The ultimate goal of this study is to provide technical support in land cover/land use modelling and change detection. We used Landsat-8 satellite image for the study region and created our training dataset with labels for supervised machine learning. We used Google Earth Engine and unsupervised machine learning via k-means clustering to generate the land cover map. We used convolutional neural networks to classify the selected regions’ land cover types. We present a visualisation of change detection, highlighting urban area changes over time to monitor changes in the map.

Method: Three-stage framework: 1) YOLOv7 segmentation for object localization, 2) ConvNeXt-based feature extractor with triplet loss for discriminative embeddings, 3) Feature-assisted IoU tracker for resilient multi-object tracking. Uses mixed-precision inference for edge optimization.

Result: High accuracy and robustness demonstrated across diverse FOI scenarios on real-world surveillance and drone video datasets. Hardware benchmarks confirm practical performance on NVIDIA Jetson edge devices.

Conclusion: The framework provides an effective, scalable solution for real-time foreign object intrusion detection in power systems, with edge deployment capability and incremental learning support for practical field applications.

Abstract: This paper presents a novel three-stage framework for real-time foreign object intrusion (FOI) detection and tracking in power transmission systems. The framework integrates: (1) a YOLOv7 segmentation model for fast and robust object localization, (2) a ConvNeXt-based feature extractor trained with triplet loss to generate discriminative embeddings, and (3) a feature-assisted IoU tracker that ensures resilient multi-object tracking under occlusion and motion. To enable scalable field deployment, the pipeline is optimized for deployment on low-cost edge hardware using mixed-precision inference. The system supports incremental updates by adding embeddings from previously unseen objects into a reference database without requiring model retraining. Extensive experiments on real-world surveillance and drone video datasets demonstrate the framework’s high accuracy and robustness across diverse FOI scenarios. In addition, hardware benchmarks on NVIDIA Jetson devices confirm the framework’s practicality and scalability for real-world edge applications.

Method: The framework decomposes images into semantic objects, synthesizes diverse editing instructions, and integrates VLMs with object detectors for instruction following assessment, semantic feature extractors for content consistency, and human preference models for visual quality.

Result: Combining VLMs with object detectors shows stronger agreement with human judgments compared to using VLMs alone. The framework was used to build EdiVal-Bench covering 9 instruction types and 11 editing models, identifying failure modes in current approaches.

Conclusion: EdiVal-Agent provides a scalable, fine-grained evaluation solution with modular design that allows future tool integration, enabling better assessment and development of next-generation image editing models.

Abstract: Instruction-based image editing has advanced rapidly, yet reliable and interpretable evaluation remains a bottleneck. Current protocols either (i) depend on paired reference images – resulting in limited coverage and inheriting biases from prior generative models – or (ii) rely solely on zero-shot vision-language models (VLMs), whose prompt-based assessments of instruction following, content consistency, and visual quality are often imprecise. To address this, we introduce EdiVal-Agent, an automated, scalable, and fine-grained evaluation framework for multi-turn instruction-based editing from an object-centric perspective, supported by a suite of expert tools. Given an image, EdiVal-Agent first decomposes it into semantically meaningful objects, then synthesizes diverse, context-aware editing instructions. For evaluation, it integrates VLMs with open-vocabulary object detectors to assess instruction following, uses semantic-level feature extractors to evaluate content consistency, and leverages human preference models to judge visual quality. We show that combining VLMs with object detectors yields stronger agreement with human judgments in instruction-following evaluation compared to using VLMs alone and CLIP-based metrics. Furthermore, the pipeline’s modular design allows future tools to be seamlessly integrated, enhancing evaluation accuracy over time. Instantiating this pipeline, we build EdiVal-Bench, a multi-turn editing benchmark covering 9 instruction types and 11 state-of-the-art editing models spanning autoregressive (AR) (including Nano Banana, GPT-Image-1), flow-matching, and diffusion paradigms. We demonstrate that EdiVal-Agent can be used to identify existing failure modes, thereby informing the development of the next generation of editing models. Project page: https://tianyucodings.github.io/EdiVAL-page/.

Method: Uses analytical image gradients of Gaussians for gradient-based bicubic spline interpolation to upscale low-resolution 3DGS renderings with minimal computational overhead.

Result: Achieves 3-4x higher novel view synthesis rates compared to baseline 3DGS, with significant artifact reduction and high reconstruction fidelity across multiple datasets.

Conclusion: Gradient-aware upscaling effectively enhances 3DGS performance on resource-constrained hardware while maintaining reconstruction quality, and can be integrated into gradient-based optimization pipelines.

Abstract: We introduce an image upscaling technique tailored for 3D Gaussian Splatting (3DGS) on lightweight GPUs. Compared to 3DGS, it achieves significantly higher rendering speeds and reduces artifacts commonly observed in 3DGS reconstructions. Our technique upscales low-resolution 3DGS renderings with a marginal increase in cost by directly leveraging the analytical image gradients of Gaussians for gradient-based bicubic spline interpolation. The technique is agnostic to the specific 3DGS implementation, achieving novel view synthesis at rates 3x-4x higher than the baseline implementation. Through extensive experiments on multiple datasets, we showcase the performance improvements and high reconstruction fidelity attainable with gradient-aware upscaling of 3DGS images. We further demonstrate the integration of gradient-aware upscaling into the gradient-based optimization of a 3DGS model and analyze its effects on reconstruction quality and performance.

Method: Uses a transformer-based feed-forward architecture with factored representation of multi-view scene geometry (depth maps, local ray maps, camera poses, metric scale factor). Standardizes supervision across datasets with flexible input augmentation.

Result: Outperforms or matches specialist feed-forward models across various 3D vision tasks while offering more efficient joint training behavior.

Conclusion: MapAnything demonstrates the feasibility of a universal 3D reconstruction backbone that can handle multiple tasks effectively, paving the way for more unified approaches in 3D computer vision.

Abstract: We introduce MapAnything, a unified transformer-based feed-forward model that ingests one or more images along with optional geometric inputs such as camera intrinsics, poses, depth, or partial reconstructions, and then directly regresses the metric 3D scene geometry and cameras. MapAnything leverages a factored representation of multi-view scene geometry, i.e., a collection of depth maps, local ray maps, camera poses, and a metric scale factor that effectively upgrades local reconstructions into a globally consistent metric frame. Standardizing the supervision and training across diverse datasets, along with flexible input augmentation, enables MapAnything to address a broad range of 3D vision tasks in a single feed-forward pass, including uncalibrated structure-from-motion, calibrated multi-view stereo, monocular depth estimation, camera localization, depth completion, and more. We provide extensive experimental analyses and model ablations demonstrating that MapAnything outperforms or matches specialist feed-forward models while offering more efficient joint training behavior, thus paving the way toward a universal 3D reconstruction backbone.

Method: Used spatiotemporal deep learning with 3D CNN and ConvLSTM2D architectures on curated video dataset of 50 clips featuring 42 cattle from multiple viewpoints in indoor/outdoor environments, with data augmentation.

Result: 3D CNN achieved 90% accuracy with 90.9% precision, recall, and F1 score, outperforming ConvLSTM2D (85% accuracy) and matching prior methods while eliminating pose estimation requirements.

Conclusion: Deep learning can effectively extract spatio-temporal features from videos for scalable cattle lameness detection in real farm settings using end-to-end classification approach.

Abstract: Cattle lameness is a prevalent health problem in livestock farming, often resulting from hoof injuries or infections, and severely impacts animal welfare and productivity. Early and accurate detection is critical for minimizing economic losses and ensuring proper treatment. This study proposes a spatiotemporal deep learning framework for automated cattle lameness detection using publicly available video data. We curate and publicly release a balanced set of 50 online video clips featuring 42 individual cattle, recorded from multiple viewpoints in both indoor and outdoor environments. The videos were categorized into lame and non-lame classes based on visual gait characteristics and metadata descriptions. After applying data augmentation techniques to enhance generalization, two deep learning architectures were trained and evaluated: 3D Convolutional Neural Networks (3D CNN) and Convolutional Long-Short-Term Memory (ConvLSTM2D). The 3D CNN achieved a video-level classification accuracy of 90%, with a precision, recall, and F1 score of 90.9% each, outperforming the ConvLSTM2D model, which achieved 85% accuracy. Unlike conventional approaches that rely on multistage pipelines involving object detection and pose estimation, this study demonstrates the effectiveness of a direct end-to-end video classification approach. Compared with the best end-to-end prior method (C3D-ConvLSTM, 90.3%), our model achieves comparable accuracy while eliminating pose estimation pre-processing.The results indicate that deep learning models can successfully extract and learn spatio-temporal features from various video sources, enabling scalable and efficient cattle lameness detection in real-world farm settings.

Method: Proposes SCM-PR framework with: VMamba backbone for RGB feature extraction, Semantic-Aware Feature Fusion module, semantic-enhanced LiDAR descriptors, cross-modal semantic attention in NetVLAD, Multi-View Semantic-Geometric Matching, and Semantic Consistency Loss in contrastive learning.

Result: Experimental results on KITTI and KITTI-360 datasets show SCM-PR achieves state-of-the-art performance compared to other cross-modal place recognition methods.

Conclusion: Incorporating semantic information significantly improves cross-modal place recognition robustness and performance, making it suitable for complex real-world environments with varying conditions.

Abstract: Ensuring accurate localization of robots in environments without GPS capability is a challenging task. Visual Place Recognition (VPR) techniques can potentially achieve this goal, but existing RGB-based methods are sensitive to changes in illumination, weather, and other seasonal changes. Existing cross-modal localization methods leverage the geometric properties of RGB images and 3D LiDAR maps to reduce the sensitivity issues highlighted above. Currently, state-of-the-art methods struggle in complex scenes, fine-grained or high-resolution matching, and situations where changes can occur in viewpoint. In this work, we introduce a framework we call Semantic-Enhanced Cross-Modal Place Recognition (SCM-PR) that combines high-level semantics utilizing RGB images for robust localization in LiDAR maps. Our proposed method introduces: a VMamba backbone for feature extraction of RGB images; a Semantic-Aware Feature Fusion (SAFF) module for using both place descriptors and segmentation masks; LiDAR descriptors that incorporate both semantics and geometry; and a cross-modal semantic attention mechanism in NetVLAD to improve matching. Incorporating the semantic information also was instrumental in designing a Multi-View Semantic-Geometric Matching and a Semantic Consistency Loss, both in a contrastive learning framework. Our experimental work on the KITTI and KITTI-360 datasets show that SCM-PR achieves state-of-the-art performance compared to other cross-modal place recognition methods.

Method: Propose scene-adaptive lattice vector quantization (SALVQ) that optimizes lattice basis per scene and scales lattice density for multi-rate capability.

Result: SALVQ enhances rate-distortion performance with minimal computational overhead and eliminates need for separate models for different compression levels.

Conclusion: SALVQ provides an efficient quantization solution that balances vector quantization benefits with low complexity, easily integrating into existing 3DGS compression systems.

Abstract: 3D Gaussian Splatting (3DGS) is rapidly gaining popularity for its photorealistic rendering quality and real-time performance, but it generates massive amounts of data. Hence compressing 3DGS data is necessary for the cost effectiveness of 3DGS models. Recently, several anchor-based neural compression methods have been proposed, achieving good 3DGS compression performance. However, they all rely on uniform scalar quantization (USQ) due to its simplicity. A tantalizing question is whether more sophisticated quantizers can improve the current 3DGS compression methods with very little extra overhead and minimal change to the system. The answer is yes by replacing USQ with lattice vector quantization (LVQ). To better capture scene-specific characteristics, we optimize the lattice basis for each scene, improving LVQ’s adaptability and R-D efficiency. This scene-adaptive LVQ (SALVQ) strikes a balance between the R-D efficiency of vector quantization and the low complexity of USQ. SALVQ can be seamlessly integrated into existing 3DGS compression architectures, enhancing their R-D performance with minimal modifications and computational overhead. Moreover, by scaling the lattice basis vectors, SALVQ can dynamically adjust lattice density, enabling a single model to accommodate multiple bit rate targets. This flexibility eliminates the need to train separate models for different compression levels, significantly reducing training time and memory consumption.

Method: A modular three-stage pipeline: (1) off-the-shelf human detection and depth estimation, (2) VLM-based reasoning to classify pairwise social affiliation, (3) lightweight spatial aggregation algorithm to localize socially connected groups.

Result: Developed a new dataset of 100K urban street-view images with bounding boxes and labels for individuals and social groups, combining human annotations and MINGLE outputs for semantic richness and broad coverage.

Conclusion: MINGLE successfully addresses the social group region detection task by integrating computer vision and reasoning capabilities, providing a foundation for future research in understanding social interactions in urban environments.

Abstract: Understanding group-level social interactions in public spaces is crucial for urban planning, informing the design of socially vibrant and inclusive environments. Detecting such interactions from images involves interpreting subtle visual cues such as relations, proximity, and co-movement - semantically complex signals that go beyond traditional object detection. To address this challenge, we introduce a social group region detection task, which requires inferring and spatially grounding visual regions defined by abstract interpersonal relations. We propose MINGLE (Modeling INterpersonal Group-Level Engagement), a modular three-stage pipeline that integrates: (1) off-the-shelf human detection and depth estimation, (2) VLM-based reasoning to classify pairwise social affiliation, and (3) a lightweight spatial aggregation algorithm to localize socially connected groups. To support this task and encourage future research, we present a new dataset of 100K urban street-view images annotated with bounding boxes and labels for both individuals and socially interacting groups. The annotations combine human-created labels and outputs from the MINGLE pipeline, ensuring semantic richness and broad coverage of real-world scenarios.

Method: Leverages cross-attention attribution maps to analyze structural entanglements, quantifies spatial demographics-semantics concept entanglement using Intersection over Union (IoU), and implements bias mitigation through energy-guided diffusion sampling that modifies latent noise space and minimizes expected SoftIoU during denoising.

Result: Shows that existing fairness interventions reduce output distribution gaps but fail to disentangle concept-level coupling, while BiasMap successfully mitigates concept entanglement in image generation and complements distributional bias mitigation.

Conclusion: BiasMap provides a deeper lens into hidden representational biases and offers an effective mitigation approach that addresses concept-level entanglements in text-to-image generation models.

Abstract: Bias discovery is critical for black-box generative models, especiall text-to-image (TTI) models. Existing works predominantly focus on output-level demographic distributions, which do not necessarily guarantee concept representations to be disentangled post-mitigation. We propose BiasMap, a model-agnostic framework for uncovering latent concept-level representational biases in stable diffusion models. BiasMap leverages cross-attention attribution maps to reveal structural entanglements between demographics (e.g., gender, race) and semantics (e.g., professions), going deeper into representational bias during the image generation. Using attribution maps of these concepts, we quantify the spatial demographics-semantics concept entanglement via Intersection over Union (IoU), offering a lens into bias that remains hidden in existing fairness discovery approaches. In addition, we further utilize BiasMap for bias mitigation through energy-guided diffusion sampling that directly modifies latent noise space and minimizes the expected SoftIoU during the denoising process. Our findings show that existing fairness interventions may reduce the output distributional gap but often fail to disentangle concept-level coupling, whereas our mitigation method can mitigate concept entanglement in image generation while complementing distributional bias mitigation.

Method: Developed a Python-based graphical user interface that integrates with various imaging systems, featuring Bézier splines, binary masks, non-destructive layers, and OpenCV integration with NumPy for high-performance matrix operations.

Result: LivePixel enables real-time image annotation directly from video devices, providing precise annotation tools similar to commercial graphics software and facilitating seamless data collection for AI model development.

Conclusion: LivePixel addresses the limitations of traditional annotation tools by providing real-time annotation capabilities that accelerate AI model development in experimental scientific workflows, with wide compatibility across video devices.

Abstract: The lack of flexible annotation tools has hindered the deployment of AI models in some scientific areas. Most existing image annotation software requires users to upload a precollected dataset, which limits support for on-demand pipelines and introduces unnecessary steps to acquire images. This constraint is particularly problematic in laboratory environments, where real-time data acquisition from instruments such as microscopes is increasingly common. In this work, we introduce \texttt{LivePixel}, a Python-based graphical user interface that integrates with imaging systems, such as webcams, microscopes, and others, to enable real-time image annotation. LivePyxel is designed to be easy to use through a simple interface that allows users to precisely delimit areas for annotation using tools commonly found in commercial graphics editing software. Of particular interest is the availability of B'ezier splines and binary masks, and the software’s capacity to work with non-destructive layers that enable high-performance editing. LivePyxel also integrates a wide compatibility across video devices, and it’s optimized for object detection operations via the use of OpenCV in combination with high-performance libraries designed to handle matrix and linear algebra operations via Numpy effectively. LivePyxel facilitates seamless data collection and labeling, accelerating the development of AI models in experimental workflows. LivePyxel freely available at https://github.com/UGarCil/LivePyxel

Method: Freezes pre-trained model parameters and trains a small h-transform network (1.42% parameters) to learn conditional h-transform. Adds adaptive consistency loss combining consistency loss and denoising score matching loss for efficient fine-tuning.

Result: Achieves state-of-the-art performance on VTO tasks with only 15 denoising steps, significantly reducing inference time while maintaining competitive results.

Conclusion: DEFT-VTON provides an efficient solution for adapting large pre-trained models to virtual try-on tasks with minimal parameter training and fast inference capabilities.

Abstract: Diffusion models enable high-quality virtual try-on (VTO) with their established image synthesis abilities. Despite the extensive end-to-end training of large pre-trained models involved in current VTO methods, real-world applications often prioritize limited training and inference, serving, and deployment budgets for VTO. To solve this obstacle, we apply Doob’s h-transform efficient fine-tuning (DEFT) for adapting large pre-trained unconditional models for downstream image-conditioned VTO abilities. DEFT freezes the pre-trained model’s parameters and trains a small h-transform network to learn a conditional h-transform. The h-transform network allows training only 1.42 percent of the frozen parameters, compared to a baseline of 5.52 percent in traditional parameter-efficient fine-tuning (PEFT). To further improve DEFT’s performance and decrease existing models’ inference time, we additionally propose an adaptive consistency loss. Consistency training distills slow but high-performing diffusion models into a fast one while retaining performance by enforcing consistencies along the inference path. Inspired by constrained optimization, instead of distillation, we combine the consistency loss and the denoising score matching loss in a data-adaptive manner for fine-tuning existing VTO models at a low cost. Empirical results show the proposed DEFT-VTON method achieves state-of-the-art performance on VTO tasks, with as few as 15 denoising steps, while maintaining competitive results.

Method: Deployed data augmentation to generate custom traffic scenarios with VRUs. Created a pipeline for augmenting Cityscapes dataset with virtual pedestrians. Developed a novel generative network architecture for adversarial learning of dataset lighting conditions to improve augmentation realism.

Result: The approach was evaluated on semantic and instance segmentation tasks, demonstrating improved performance in pedestrian recognition.

Conclusion: The proposed pipeline and generative network architecture effectively bridge the domain gap between synthetic and real data, enhancing pedestrian recognition capabilities for autonomous driving systems.

Abstract: In the autonomous driving area synthetic data is crucial for cover specific traffic scenarios which autonomous vehicle must handle. This data commonly introduces domain gap between synthetic and real domains. In this paper we deploy data augmentation to generate custom traffic scenarios with VRUs in order to improve pedestrian recognition. We provide a pipeline for augmentation of the Cityscapes dataset with virtual pedestrians. In order to improve augmentation realism of the pipeline we reveal a novel generative network architecture for adversarial learning of the data-set lighting conditions. We also evaluate our approach on the tasks of semantic and instance segmentation.

Method: Proposed Functional Kolmogorov-Arnold Network (FunKAN) that generalizes the Kolmogorov-Arnold theorem to functional spaces, learning inner functions using Fourier decomposition over Hermite functions. Also developed U-FunKAN for segmentation tasks.

Result: Outperformed other KAN-based methods in medical image enhancement (PSNR, TV metrics on IXI dataset for Gibbs ringing suppression) and segmentation (IoU, F1 scores on BUSI, GlaS, and CVC-ClinicDB datasets for breast cancer, glands, and polyp detection).

Conclusion: FunKAN successfully bridges theoretical function approximation with medical image analysis, providing a robust and interpretable solution that preserves spatial structure while achieving state-of-the-art performance in clinical applications.

Abstract: Medical image enhancement and segmentation are critical yet challenging tasks in modern clinical practice, constrained by artifacts and complex anatomical variations. Traditional deep learning approaches often rely on complex architectures with limited interpretability. While Kolmogorov-Arnold networks offer interpretable solutions, their reliance on flattened feature representations fundamentally disrupts the intrinsic spatial structure of imaging data. To address this issue we propose a Functional Kolmogorov-Arnold Network (FunKAN) – a novel interpretable neural framework, designed specifically for image processing, that formally generalizes the Kolmogorov-Arnold representation theorem onto functional spaces and learns inner functions using Fourier decomposition over the basis Hermite functions. We explore FunKAN on several medical image processing tasks, including Gibbs ringing suppression in magnetic resonance images, benchmarking on IXI dataset. We also propose U-FunKAN as state-of-the-art binary medical segmentation model with benchmarks on three medical datasets: BUSI (ultrasound images), GlaS (histological structures) and CVC-ClinicDB (colonoscopy videos), detecting breast cancer, glands and polyps, respectively. Experiments on those diverse datasets demonstrate that our approach outperforms other KAN-based backbones in both medical image enhancement (PSNR, TV) and segmentation (IoU, F1). Our work bridges the gap between theoretical function approximation and medical image analysis, offering a robust, interpretable solution for clinical applications.

Method: Proposes a dual-stream GNN that constructs instance graphs to extract instance-level features, then uses a complementary weight graph to assign importance weights highlighting hateful instances, combining weights and features for classification.

Result: Extensive experiments on public datasets show the model achieves state-of-the-art performance in hateful video classification and demonstrates strong explainability.

Conclusion: The proposed graph-based framework effectively models structured relationships within and across modalities, emphasizing hateful content for improved detection while providing explainable results.

Abstract: Hateful videos present serious risks to online safety and real-world well-being, necessitating effective detection methods. Although multimodal classification approaches integrating information from several modalities outperform unimodal ones, they typically neglect that even minimal hateful content defines a video’s category. Specifically, they generally treat all content uniformly, instead of emphasizing the hateful components. Additionally, existing multimodal methods cannot systematically capture structured information in videos, limiting the effectiveness of multimodal fusion. To address these limitations, we propose a novel multimodal dual-stream graph neural network model. It constructs an instance graph by separating the given video into several instances to extract instance-level features. Then, a complementary weight graph assigns importance weights to these features, highlighting hateful instances. Importance weights and instance features are combined to generate video labels. Our model employs a graph-based framework to systematically model structured relationships within and across modalities. Extensive experiments on public datasets show that our model is state-of-the-art in hateful video classification and has strong explainability. Code is available: https://github.com/Multimodal-Intelligence-Lab-MIL/MultiHateGNN.

Method: Uses diffusion-based depth estimation trained on synthetic colonoscopy sequences to learn geometric priors, combined with style transfer technique to adapt real clinical videos to match synthetic training domain.

Result: Achieves state-of-the-art zero-shot performance on C3VD dataset, outperforming both general-purpose and endoscopy-specific approaches. Enables 3D point cloud generation and surface coverage assessment.

Conclusion: Although full trajectory 3D reconstruction remains challenging, ColonCrafter demonstrates clinically relevant applications and provides temporally consistent depth estimation for colonoscopy videos.

Abstract: Three-dimensional (3D) scene understanding in colonoscopy presents significant challenges that necessitate automated methods for accurate depth estimation. However, existing depth estimation models for endoscopy struggle with temporal consistency across video sequences, limiting their applicability for 3D reconstruction. We present ColonCrafter, a diffusion-based depth estimation model that generates temporally consistent depth maps from monocular colonoscopy videos. Our approach learns robust geometric priors from synthetic colonoscopy sequences to generate temporally consistent depth maps. We also introduce a style transfer technique that preserves geometric structure while adapting real clinical videos to match our synthetic training domain. ColonCrafter achieves state-of-the-art zero-shot performance on the C3VD dataset, outperforming both general-purpose and endoscopy-specific approaches. Although full trajectory 3D reconstruction remains a challenge, we demonstrate clinically relevant applications of ColonCrafter, including 3D point cloud generation and surface coverage assessment.

Method: Proposes merging redundant 3D Gaussian primitives in voxel space based on geometric similarity to reduce GPU memory usage, and initializes primitives via Patch-Grid point sampling for better scene modeling.

Result: Quantitative and qualitative evaluations on public datasets show reduced GPU memory usage without impacting runtime performance, while improving rendering quality.

Conclusion: The proposed methods effectively address the computational constraints of embedded platforms by optimizing memory usage and enhancing reconstruction quality in 3D Gaussian Splatting applications.

Abstract: Recent advancements in 3D Gaussian Splatting (3DGS) have made a significant impact on rendering and reconstruction techniques. Current research predominantly focuses on improving rendering performance and reconstruction quality using high-performance desktop GPUs, largely overlooking applications for embedded platforms like micro air vehicles (MAVs). These devices, with their limited computational resources and memory, often face a trade-off between system performance and reconstruction quality. In this paper, we improve existing methods in terms of GPU memory usage while enhancing rendering quality. Specifically, to address redundant 3D Gaussian primitives in SLAM, we propose merging them in voxel space based on geometric similarity. This reduces GPU memory usage without impacting system runtime performance. Furthermore, rendering quality is improved by initializing 3D Gaussian primitives via Patch-Grid (PG) point sampling, enabling more accurate modeling of the entire scene. Quantitative and qualitative evaluations on publicly available datasets demonstrate the effectiveness of our improvements.

Method: Builds on quickest change detection formulation and introduces adaptive mechanisms to model mode-dependent prediction error distributions that evolve over time with dataset-specific dynamics.

Result: Substantial improvements in both detection delay and false alarm rates compared to prior UQ- and vision-based OOD approaches, with better accuracy and computational efficiency.

Conclusion: Provides a practical path toward reliable, driving-aware autonomy by explicitly modeling error modes for robust trajectory OOD detection in complex driving environments.

Abstract: Trajectory prediction is central to the safe and seamless operation of autonomous vehicles (AVs). In deployment, however, prediction models inevitably face distribution shifts between training data and real-world conditions, where rare or underrepresented traffic scenarios induce out-of-distribution (OOD) cases. While most prior OOD detection research in AVs has concentrated on computer vision tasks such as object detection and segmentation, trajectory-level OOD detection remains largely underexplored. A recent study formulated this problem as a quickest change detection (QCD) task, providing formal guarantees on the trade-off between detection delay and false alarms [1]. Building on this foundation, we propose a new framework that introduces adaptive mechanisms to achieve robust detection in complex driving environments. Empirical analysis across multiple real-world datasets reveals that prediction errors – even on in-distribution samples – exhibit mode-dependent distributions that evolve over time with dataset-specific dynamics. By explicitly modeling these error modes, our method achieves substantial improvements in both detection delay and false alarm rates. Comprehensive experiments on established trajectory prediction benchmarks show that our framework significantly outperforms prior UQ- and vision-based OOD approaches in both accuracy and computational efficiency, offering a practical path toward reliable, driving-aware autonomy.

Method: Integrates Google Gemini 2.5 Flash with visual feature extraction and NLP for contextual image interpretation, uses coordinate verification, probabilistic Gaussian modeling for anomaly distribution, multi-layered visualization techniques, and precise prompt engineering for structured clinical information extraction.

Result: Demonstrated high performance in anomaly detection across multiple modalities, achieved 80 pixels average deviation in location measurement, and features zero-shot learning capabilities to reduce dataset dependence.

Conclusion: Represents a significant advancement in automated diagnostic support, though requires clinical validation and multi-center evaluation before widespread adoption.

Abstract: The rapid advancement of artificial intelligence (AI) in healthcare imaging has revolutionized diagnostic medicine and clinical decision-making processes. This work presents an intelligent multimodal framework for medical image analysis that leverages Vision-Language Models (VLMs) in healthcare diagnostics. The framework integrates Google Gemini 2.5 Flash for automated tumor detection and clinical report generation across multiple imaging modalities including CT, MRI, X-ray, and Ultrasound. The system combines visual feature extraction with natural language processing to enable contextual image interpretation, incorporating coordinate verification mechanisms and probabilistic Gaussian modeling for anomaly distribution. Multi-layered visualization techniques generate detailed medical illustrations, overlay comparisons, and statistical representations to enhance clinical confidence, with location measurement achieving 80 pixels average deviation. Result processing utilizes precise prompt engineering and textual analysis to extract structured clinical information while maintaining interpretability. Experimental evaluations demonstrated high performance in anomaly detection across multiple modalities. The system features a user-friendly Gradio interface for clinical workflow integration and demonstrates zero-shot learning capabilities to reduce dependence on large datasets. This framework represents a significant advancement in automated diagnostic support and radiological workflow efficiency, though clinical validation and multi-center evaluation are necessary prior to widespread adoption.

Method: Extended the clustering-based CLAP framework that uses clustering to suppress noise and mitigate erroneous feature matches, providing an alternative to traditional outlier rejection schemes like RANSAC.

Result: Successfully generalized CLAP to handle 3D localization and image stitching problems, demonstrating its wide applicability across different fields for dealing with noise and uncertainty.

Conclusion: CLAP provides a robust clustering-based framework that is widely applicable beyond 2D localization and offers a useful alternative to traditional outlier rejection methods across various domains.

Abstract: In previous work, we introduced a 2D localization algorithm called CLAP, Clustering to Localize Across $n$ Possibilities, which was used during our championship win in RoboCup 2024, an international autonomous humanoid soccer competition. CLAP is particularly recognized for its robustness against outliers, where clustering is employed to suppress noise and mitigate against erroneous feature matches. This clustering-based strategy provides an alternative to traditional outlier rejection schemes such as RANSAC, in which candidates are validated by reprojection error across all data points. In this paper, CLAP is extended to a more general framework beyond 2D localization, specifically to 3D localization and image stitching. We also show how CLAP, RANSAC, and Hough transforms are related. The generalization of CLAP is widely applicable to many different fields and can be a useful tool to deal with noise and uncertainty.

Method: Uses SAM’s image encoder with task-specific adaptation to extract structure-aware feature embeddings. Adds a lightweight 3D head to refine features for local deformations. Introduces Hierarchical Feature Consistency Loss for coarse-to-fine feature matching and anatomical alignment.

Result: Significantly outperforms state-of-the-art methods: 2.68% improvement on ACDC cardiac dataset and 6.44% improvement on abdomen CT dataset for both intra-subject and inter-subject registration.

Conclusion: SAMIR demonstrates that foundation models like SAM can be effectively adapted for medical image registration, providing superior performance without requiring weak supervision labels, making it more practical for real-world applications.

Abstract: Image registration is a fundamental task in medical image analysis. Deformations are often closely related to the morphological characteristics of tissues, making accurate feature extraction crucial. Recent weakly supervised methods improve registration by incorporating anatomical priors such as segmentation masks or landmarks, either as inputs or in the loss function. However, such weak labels are often not readily available, limiting their practical use. Motivated by the strong representation learning ability of visual foundation models, this paper introduces SAMIR, an efficient medical image registration framework that utilizes the Segment Anything Model (SAM) to enhance feature extraction. SAM is pretrained on large-scale natural image datasets and can learn robust, general-purpose visual representations. Rather than using raw input images, we design a task-specific adaptation pipeline using SAM’s image encoder to extract structure-aware feature embeddings, enabling more accurate modeling of anatomical consistency and deformation patterns. We further design a lightweight 3D head to refine features within the embedding space, adapting to local deformations in medical images. Additionally, we introduce a Hierarchical Feature Consistency Loss to guide coarse-to-fine feature matching and improve anatomical alignment. Extensive experiments demonstrate that SAMIR significantly outperforms state-of-the-art methods on benchmark datasets for both intra-subject cardiac image registration and inter-subject abdomen CT image registration, achieving performance improvements of 2.68% on ACDC and 6.44% on the abdomen dataset. The source code will be publicly available on GitHub following the acceptance of this paper.

Method: Uses Federated Learning with FLOWER and RAY frameworks to enable collaborative model training across distributed edge satellite centers. Employs YOLOS-small (Vision Transformer), Faster R-CNN with ResNet50, and Faster R-CNN with MobileNetV3 models trained on public datasets.

Result: The framework provides efficient client spawning and distributed learning workload execution while maintaining data privacy and security for satellite imagery analysis.

Conclusion: This federated learning approach offers a novel perspective for image segmentation tasks on satellite imagery, enabling collaborative deforestation detection without compromising client data security.

Abstract: Accurate identification of deforestation from satellite images is essential in order to understand the geographical situation of an area. This paper introduces a new distributed approach to identify as well as locate deforestation across different clients using Federated Learning (FL). Federated Learning enables distributed network clients to collaboratively train a model while maintaining data privacy and security of the active users. In our framework, a client corresponds to an edge satellite center responsible for local data processing. Moreover, FL provides an advantage over centralized training method which requires combining data, thereby compromising with data security of the clients. Our framework leverages the FLOWER framework with RAY framework to execute the distributed learning workload. Furthermore, efficient client spawning is ensured by RAY as it can select definite amount of users to create an emulation environment. Our FL framework uses YOLOS-small (a Vision Transformer variant), Faster R-CNN with a ResNet50 backbone, and Faster R-CNN with a MobileNetV3 backbone models trained and tested on publicly available datasets. Our approach provides us a different view for image segmentation-based tasks on satellite imagery.

Method: Uses metric monocular depth estimator and Gaussian scene reconstructor to create metric 3D GMMs for each image. Refines initial pose by optimizing a differentiable GMM alignment objective considering geometry, color, covariance, and semantic features.

Result: Outperforms both classical and state-of-the-art learning-based methods (including MASt3R) on Real-Estate10K dataset.

Conclusion: Bridging single-view perception with multi-view geometry enables robust and metric relative pose estimation without requiring explicit 2D correspondences.

Abstract: Estimating metric relative camera pose from a pair of images is of great importance for 3D reconstruction and localisation. However, conventional two-view pose estimation methods are not metric, with camera translation known only up to a scale, and struggle with wide baselines and textureless or reflective surfaces. This paper introduces GARPS, a training-free framework that casts this problem as the direct alignment of two independently reconstructed 3D scenes. GARPS leverages a metric monocular depth estimator and a Gaussian scene reconstructor to obtain a metric 3D Gaussian Mixture Model (GMM) for each image. It then refines an initial pose from a feed-forward two-view pose estimator by optimising a differentiable GMM alignment objective. This objective jointly considers geometric structure, view-independent colour, anisotropic covariance, and semantic feature consistency, and is robust to occlusions and texture-poor regions without requiring explicit 2D correspondences. Extensive experiments on the Real-Estate10K dataset demonstrate that GARPS outperforms both classical and state-of-the-art learning-based methods, including MASt3R. These results highlight the potential of bridging single-view perception with multi-view geometry to achieve robust and metric relative pose estimation.

Method: The authors introduce a generic and efficient lookup operation that can replace multiplication operations. They construct lookup tables in a differentiable manner and propose training strategies for end-to-end optimization. Lookup networks are developed for image classification, super-resolution, and point cloud classification tasks.

Result: Lookup networks achieve higher efficiency in terms of energy consumption and inference speed while maintaining competitive performance compared to vanilla convolutional networks. The method produces state-of-the-art performance on both classification and regression tasks across different data types.

Conclusion: The proposed lookup operation provides an efficient alternative to multiplication operations in neural networks, enabling deployment on resource-constrained devices while maintaining performance across various computer vision tasks.

Abstract: Convolutional neural networks are constructed with massive operations with different types and are highly computationally intensive. Among these operations, multiplication operation is higher in computational complexity and usually requires {more} energy consumption with longer inference time than other operations, which hinders the deployment of convolutional neural networks on mobile devices. In many resource-limited edge devices, complicated operations can be calculated via lookup tables to reduce computational cost. Motivated by this, in this paper, we introduce a generic and efficient lookup operation which can be used as a basic operation for the construction of neural networks. Instead of calculating the multiplication of weights and activation values, simple yet efficient lookup operations are adopted to compute their responses. To enable end-to-end optimization of the lookup operation, we construct the lookup tables in a differentiable manner and propose several training strategies to promote their convergence. By replacing computationally expensive multiplication operations with our lookup operations, we develop lookup networks for the image classification, image super-resolution, and point cloud classification tasks. It is demonstrated that our lookup networks can benefit from the lookup operations to achieve higher efficiency in terms of energy consumption and inference speed while maintaining competitive performance to vanilla convolutional networks. Extensive experiments show that our lookup networks produce state-of-the-art performance on different tasks (both classification and regression tasks) and different data types (both images and point clouds).

Method: Proposes semantic visual projector using SAM-generated superpixels as “visual words”, with semantic superpixel positional embedding and aggregator to preserve geometry and details.

Result: Achieves 93% reduction in visual tokens without performance compromise, significantly speeds up MLLM training and inference, outperforms existing compressive visual projectors on RIS.

Conclusion: The semantic visual projector approach effectively addresses visual token redundancy while preserving semantic information, enabling efficient MLLM adaptation for segmentation tasks.

Abstract: Recently, Referring Image Segmentation (RIS) frameworks that pair the Multimodal Large Language Model (MLLM) with the Segment Anything Model (SAM) have achieved impressive results. However, adapting MLLM to segmentation is computationally intensive, primarily due to visual token redundancy. We observe that traditional patch-wise visual projectors struggle to strike a balance between reducing the number of visual tokens and preserving semantic clarity, often retaining overly long token sequences to avoid performance drops. Inspired by text tokenizers, we propose a novel semantic visual projector that leverages semantic superpixels generated by SAM to identify “visual words” in an image. By compressing and projecting semantic superpixels as visual tokens, our approach adaptively shortens the token sequence according to scene complexity while minimizing semantic loss in compression. To mitigate loss of information, we propose a semantic superpixel positional embedding to strengthen MLLM’s awareness of superpixel geometry and position, alongside a semantic superpixel aggregator to preserve both fine-grained details inside superpixels and global context outside. Experiments show that our method cuts visual tokens by 93% without compromising performance, notably speeding up MLLM training and inference, and outperforming existing compressive visual projectors on RIS.

Method: Three complementary innovations: 1) Distortion-Resilient Multi-scale Extraction backbone for robust features under distortion, 2) Uncertainty-aware Spatial Cross-Attention for reliable cross-view alignment, 3) Distance-aware Temporal Self-Attention for balancing near/far field details and ensuring temporal coherence.

Result: Extensive experiments on Synwoodscapes dataset demonstrate that FishBEV consistently outperforms state-of-the-art baselines on surround-view fisheye BEV segmentation tasks.

Conclusion: FishBEV provides an effective solution for BEV segmentation with fisheye cameras, overcoming the specific challenges of distortion, multi-view alignment, and temporal stability that degrade performance in autonomous driving applications.

Abstract: As a cornerstone technique for autonomous driving, Bird’s Eye View (BEV) segmentation has recently achieved remarkable progress with pinhole cameras. However, it is non-trivial to extend the existing methods to fisheye cameras with severe geometric distortion, ambiguous multi-view correspondences and unstable temporal dynamics, all of which significantly degrade BEV performance. To address these challenges, we propose FishBEV, a novel BEV segmentation framework specifically tailored for fisheye cameras. This framework introduces three complementary innovations, including a Distortion-Resilient Multi-scale Extraction (DRME) backbone that learns robust features under distortion while preserving scale consistency, an Uncertainty-aware Spatial Cross-Attention (U-SCA) mechanism that leverages uncertainty estimation for reliable cross-view alignment, a Distance-aware Temporal Self-Attention (D-TSA) module that adaptively balances near field details and far field context to ensure temporal coherence. Extensive experiments on the Synwoodscapes dataset demonstrate that FishBEV consistently outperforms SOTA baselines, regarding the performance evaluation of FishBEV on the surround-view fisheye BEV segmentation tasks.

Method: Developed three spline-based KAN variants: SBTAYLOR-KAN (B-splines + Taylor series), SBRBF-KAN (B-splines + RBF), and SBWAVELET-KAN (B-splines + Morlet wavelets) that capture local and global nonlinearities without preprocessing.

Result: SBTAYLOR-KAN achieved up to 98.93% accuracy, maintained over 86% accuracy with only 30% training data, and outperformed other models with 68.22% accuracy on imbalanced skin cancer data, using just 2,872 parameters vs millions in CNNs.

Conclusion: The framework provides a lightweight, interpretable, and generalizable solution for medical image classification that addresses limited datasets and data-scarce scenarios in clinical AI applications.

Abstract: Effective and interpretable classification of medical images is a challenge in computer-aided diagnosis, especially in resource-limited clinical settings. This study introduces spline-based Kolmogorov-Arnold Networks (KANs) for accurate medical image classification with limited, diverse datasets. The models include SBTAYLOR-KAN, integrating B-splines with Taylor series; SBRBF-KAN, combining B-splines with Radial Basis Functions; and SBWAVELET-KAN, embedding B-splines in Morlet wavelet transforms. These approaches leverage spline-based function approximation to capture both local and global nonlinearities. The models were evaluated on brain MRI, chest X-rays, tuberculosis X-rays, and skin lesion images without preprocessing, demonstrating the ability to learn directly from raw data. Extensive experiments, including cross-dataset validation and data reduction analysis, showed strong generalization and stability. SBTAYLOR-KAN achieved up to 98.93% accuracy, with a balanced F1-score, maintaining over 86% accuracy using only 30% of the training data across three datasets. Despite class imbalance in the skin cancer dataset, experiments on both imbalanced and balanced versions showed SBTAYLOR-KAN outperforming other models, achieving 68.22% accuracy. Unlike traditional CNNs, which require millions of parameters (e.g., ResNet50 with 24.18M), SBTAYLOR-KAN achieves comparable performance with just 2,872 trainable parameters, making it more suitable for constrained medical environments. Gradient-weighted Class Activation Mapping (Grad-CAM) was used for interpretability, highlighting relevant regions in medical images. This framework provides a lightweight, interpretable, and generalizable solution for medical image classification, addressing the challenges of limited datasets and data-scarce scenarios in clinical AI applications.

Method: Identifies style-sensitive cross-attention layers through activation analysis, then updates only these selected layers to protect artworks while maintaining imperceptibility.

Result: Effective style defense against fine-tuned diffusion models using WikiArt and Anita datasets, preserving artistic integrity with competitive imperceptibility.

Conclusion: StyleProtect provides lightweight yet effective protection against style mimicry by targeting specific vulnerable layers in diffusion models.

Abstract: The rapid advancement of generative models, particularly diffusion-based approaches, has inadvertently facilitated their potential for misuse. Such models enable malicious exploiters to replicate artistic styles that capture an artist’s creative labor, personal vision, and years of dedication in an inexpensive manner. This has led to a rise in the need and exploration of methods for protecting artworks against style mimicry. Although generic diffusion models can easily mimic an artistic style, finetuning amplifies this capability, enabling the model to internalize and reproduce the style with higher fidelity and control. We hypothesize that certain cross-attention layers exhibit heightened sensitivity to artistic styles. Sensitivity is measured through activation strengths of attention layers in response to style and content representations, and assessing their correlations with features extracted from external models. Based on our findings, we introduce an efficient and lightweight protection strategy, StyleProtect, that achieves effective style defense against fine-tuned diffusion models by updating only selected cross-attention layers. Our experiments utilize a carefully curated artwork dataset based on WikiArt, comprising representative works from 30 artists known for their distinctive and influential styles and cartoon animations from the Anita dataset. The proposed method demonstrates promising performance in safeguarding unique styles of artworks and anime from malicious diffusion customization, while maintaining competitive imperceptibility.

Method: Proposes UM-Depth framework with motion- and uncertainty-aware refinement using teacher-student training strategy. Embeds uncertainty estimation into both training pipeline and network architecture. Uses optical flow only in teacher network during training, eliminating extra labeling and runtime costs.

Result: Achieves state-of-the-art results in both self-supervised depth and pose estimation on KITTI datasets. Extensive experiments on KITTI and Cityscapes datasets demonstrate effectiveness of uncertainty-aware refinement.

Conclusion: UM-Depth effectively addresses uncertainty challenges in self-supervised monocular depth estimation through uncertainty-aware refinement and teacher-student training, achieving superior performance without additional inference costs or labeling requirements.

Abstract: Monocular depth estimation has been increasingly adopted in robotics and autonomous driving for its ability to infer scene geometry from a single camera. In self-supervised monocular depth estimation frameworks, the network jointly generates and exploits depth and pose estimates during training, thereby eliminating the need for depth labels. However, these methods remain challenged by uncertainty in the input data, such as low-texture or dynamic regions, which can cause reduced depth accuracy. To address this, we introduce UM-Depth, a framework that combines motion- and uncertainty-aware refinement to enhance depth accuracy at dynamic object boundaries and in textureless regions. Specifically, we develop a teacherstudent training strategy that embeds uncertainty estimation into both the training pipeline and network architecture, thereby strengthening supervision where photometric signals are weak. Unlike prior motion-aware approaches that incur inference-time overhead and rely on additional labels or auxiliary networks for real-time generation, our method uses optical flow exclusively within the teacher network during training, which eliminating extra labeling demands and any runtime cost. Extensive experiments on the KITTI and Cityscapes datasets demonstrate the effectiveness of our uncertainty-aware refinement. Overall, UM-Depth achieves state-of-the-art results in both self-supervised depth and pose estimation on the KITTI datasets.

Method: Triple Query Former (TQF) factorizes queries into three specialized components: appearance query, intra-frame interaction query, and inter-frame motion query. Uses motion-aware aggregation modules for intra-frame interaction and inter-frame motion alignment.

Result: Extensive experiments on multiple RVOS benchmarks demonstrate performance advantages and effectiveness of the structured query design and aggregation modules.

Conclusion: Decomposing referring queries into specialized components with motion-aware aggregation effectively addresses query selection bias and improves cross-modal alignment in video object segmentation.

Abstract: Recently, query-based methods have achieved remarkable performance in Referring Video Object Segmentation (RVOS) by using textual static object queries to drive cross-modal alignment. However, these static queries are easily misled by distractors with similar appearance or motion, resulting in \emph{query selection bias}. To address this issue, we propose Triple Query Former (TQF), which factorizes the referring query into three specialized components: an appearance query for static attributes, an intra-frame interaction query for spatial relations, and an inter-frame motion query for temporal association. Instead of relying solely on textual embeddings, our queries are dynamically constructed by integrating both linguistic cues and visual guidance. Furthermore, we introduce two motion-aware aggregation modules that enhance object token representations: Intra-frame Interaction Aggregation incorporates position-aware interactions among objects within a single frame, while Inter-frame Motion Aggregation leverages trajectory-guided alignment across frames to ensure temporal coherence. Extensive experiments on multiple RVOS benchmarks demonstrate the advantages of TQF and the effectiveness of our structured query design and motion-aware aggregation modules.

Method: Proposes InstanceVG framework using instance queries with prior reference points to unify joint predictions of instance-level boxes and masks, ensuring consistent multi-granularity predictions.

Result: Extensive experiments on ten datasets across four tasks show InstanceVG achieves state-of-the-art performance, significantly surpassing existing methods in various metrics.

Conclusion: InstanceVG is the first framework to simultaneously tackle both GREC and GRES with instance-aware capabilities, demonstrating superior performance and consistency in generalized visual grounding tasks.

Abstract: Generalized visual grounding tasks, including Generalized Referring Expression Comprehension (GREC) and Segmentation (GRES), extend the classical visual grounding paradigm by accommodating multi-target and non-target scenarios. Specifically, GREC focuses on accurately identifying all referential objects at the coarse bounding box level, while GRES aims for achieve fine-grained pixel-level perception. However, existing approaches typically treat these tasks independently, overlooking the benefits of jointly training GREC and GRES to ensure consistent multi-granularity predictions and streamline the overall process. Moreover, current methods often treat GRES as a semantic segmentation task, neglecting the crucial role of instance-aware capabilities and the necessity of ensuring consistent predictions between instance-level boxes and masks. To address these limitations, we propose InstanceVG, a multi-task generalized visual grounding framework equipped with instance-aware capabilities, which leverages instance queries to unify the joint and consistency predictions of instance-level boxes and masks. To the best of our knowledge, InstanceVG is the first framework to simultaneously tackle both GREC and GRES while incorporating instance-aware capabilities into generalized visual grounding. To instantiate the framework, we assign each instance query a prior reference point, which also serves as an additional basis for target matching. This design facilitates consistent predictions of points, boxes, and masks for the same instance. Extensive experiments obtained on ten datasets across four tasks demonstrate that InstanceVG achieves state-of-the-art performance, significantly surpassing the existing methods in various evaluation metrics. The code and model will be publicly available at https://github.com/Dmmm1997/InstanceVG.

Method: Proposes Adaptive Similarity Distribution Matching (A-SDM) to rectify unmatched positive pairs and Explicit Fine-grained Alignment (EFA) module for cross-modal verification via sparsified similarity matrix and hard coding.

Result: Achieves state-of-the-art performance on three public datasets among global matching methods.

Conclusion: FMFA effectively addresses cross-modal alignment challenges through full-mode fine-grained alignment, improving both global matching and local verification capabilities.

Abstract: Text-to-Image Person Retrieval (TIPR) is a cross-modal matching task that aims to retrieve the most relevant person images based on a given text query. The key challenge in TIPR lies in achieving effective alignment between textual and visual modalities within a common latent space. To address this challenge, prior approaches incorporate attention mechanisms for implicit cross-modal local alignment. However, they lack the ability to verify whether all local features are correctly aligned. Moreover, existing methods primarily focus on hard negative samples during model updates, with the goal of refining distinctions between positive and negative pairs, often neglecting incorrectly matched positive pairs. To alleviate these issues, we propose FMFA, a cross-modal Full-Mode Fine-grained Alignment framework, which enhances global matching through explicit fine-grained alignment and existing implicit relational reasoning – hence the term ``full-mode" – without requiring additional supervision. Specifically, we design an Adaptive Similarity Distribution Matching (A-SDM) module to rectify unmatched positive sample pairs. A-SDM adaptively pulls the unmatched positive pairs closer in the joint embedding space, thereby achieving more precise global alignment. Additionally, we introduce an Explicit Fine-grained Alignment (EFA) module, which makes up for the lack of verification capability of implicit relational reasoning. EFA strengthens explicit cross-modal fine-grained interactions by sparsifying the similarity matrix and employs a hard coding method for local alignment. Our proposed method is evaluated on three public datasets, achieving state-of-the-art performance among all global matching methods. Our code is available at https://github.com/yinhao1102/FMFA.

Method: Introduces a color mapping module that explicitly models the correspondence between text embedding space and image RGB values, predicting embedding vectors from given RGB values.

Result: Experimental results show the method performs well in color continuity and controllability, enabling precise color control while maintaining semantic consistency.

Conclusion: The proposed approach achieves finer-grained, continuous, and controllable color editing by establishing a direct mapping between text embeddings and RGB color space.

Abstract: In recent years, text-driven image editing has made significant progress. However, due to the inherent ambiguity and discreteness of natural language, color editing still faces challenges such as insufficient precision and difficulty in achieving continuous control. Although linearly interpolating the embedding vectors of different textual descriptions can guide the model to generate a sequence of images with varying colors, this approach lacks precise control over the range of color changes in the output images. Moreover, the relationship between the interpolation coefficient and the resulting image color is unknown and uncontrollable. To address these issues, we introduce a color mapping module that explicitly models the correspondence between the text embedding space and image RGB values. This module predicts the corresponding embedding vector based on a given RGB value, enabling precise color control of the generated images while maintaining semantic consistency. Users can specify a target RGB range to generate images with continuous color variations within the desired range, thereby achieving finer-grained, continuous, and controllable color editing. Experimental results demonstrate that our method performs well in terms of color continuity and controllability.

Method: Proposed iterative prompt refinement algorithm that uses Vision Language Models (VLMs) to analyze both input prompts and generated images, with visual feedback for more effective refinement. Also introduced a new dataset with textual and visual safety labels using multi-modal LLM for supervised fine-tuning.

Result: Experimental results show the approach produces safer outputs without compromising alignment with user intent, offering improved safety while maintaining reliability comparable to existing LLM-based methods.

Conclusion: The method provides a practical solution for generating safer text-to-image content by leveraging visual feedback through VLMs, addressing limitations of text-only safety approaches.

Abstract: Text-to-Image (T2I) models have made remarkable progress in generating images from text prompts, but their output quality and safety still depend heavily on how prompts are phrased. Existing safety methods typically refine prompts using large language models (LLMs), but they overlook the images produced, which can result in unsafe outputs or unnecessary changes to already safe prompts. To address this, we propose an iterative prompt refinement algorithm that uses Vision Language Models (VLMs) to analyze both the input prompts and the generated images. By leveraging visual feedback, our method refines prompts more effectively, improving safety while maintaining user intent and reliability comparable to existing LLM-based approaches. Additionally, we introduce a new dataset labeled with both textual and visual safety signals using off-the-shelf multi-modal LLM, enabling supervised fine-tuning. Experimental results demonstrate that our approach produces safer outputs without compromising alignment with user intent, offering a practical solution for generating safer T2I content. Our code is available at https://github.com/ku-dmlab/IPR. \textbf{\textcolor{red}WARNING: This paper contains examples of harmful or inappropriate images generated by models.

Method: Proposes Task-Aware Image Signal Processing (TA-ISP) - a compact framework that predicts lightweight multi-scale modulation operators (global, regional, pixel scales) to reshape image statistics across spatial extents, avoiding dense convolutional pipelines.

Result: TA-ISP consistently improves accuracy on RAW-domain detection and segmentation benchmarks under both daytime and nighttime conditions, while significantly reducing parameter count and inference time.

Conclusion: TA-ISP provides an efficient solution for RAW data processing that is well-suited for deployment on resource-constrained devices, offering better performance with reduced computational requirements compared to existing approaches.

Abstract: In recent years, there has been a growing trend in computer vision towards exploiting RAW sensor data, which preserves richer information compared to conventional low-bit RGB images. Early studies mainly focused on enhancing visual quality, while more recent efforts aim to leverage the abundant information in RAW data to improve the performance of visual perception tasks such as object detection and segmentation. However, existing approaches still face two key limitations: large-scale ISP networks impose heavy computational overhead, while methods based on tuning traditional ISP pipelines are restricted by limited representational capacity.To address these issues, we propose Task-Aware Image Signal Processing (TA-ISP), a compact RAW-to-RGB framework that produces task-oriented representations for pretrained vision models. Instead of heavy dense convolutional pipelines, TA-ISP predicts a small set of lightweight, multi-scale modulation operators that act at global, regional, and pixel scales to reshape image statistics across different spatial extents. This factorized control significantly expands the range of spatially varying transforms that can be represented while keeping memory usage, computation, and latency tightly constrained. Evaluated on several RAW-domain detection and segmentation benchmarks under both daytime and nighttime conditions, TA-ISP consistently improves downstream accuracy while markedly reducing parameter count and inference time, making it well suited for deployment on resource-constrained devices.

Method: Introduce VHBench-10 with 10,000 samples across 10 fine-grained hallucination categories. Propose VisionWeaver - a Context-Aware Routing Network that uses global visual features to generate routing signals and dynamically aggregate features from multiple specialized experts.

Result: Evaluations confirm encoders exhibit unique hallucination characteristics. VisionWeaver significantly reduces hallucinations and improves overall model performance compared to simple feature fusion approaches.

Conclusion: Visual encoder choice is crucial for reducing hallucinations in LVLMs. The proposed VisionWeaver framework effectively mitigates hallucinations by leveraging multiple visual experts through dynamic routing, demonstrating superior performance over standard approaches.

Abstract: Object hallucination in Large Vision-Language Models (LVLMs) significantly impedes their real-world applicability. As the primary component for accurately interpreting visual information, the choice of visual encoder is pivotal. We hypothesize that the diverse training paradigms employed by different visual encoders instill them with distinct inductive biases, which leads to their diverse hallucination performances. Existing benchmarks typically focus on coarse-grained hallucination detection and fail to capture the diverse hallucinations elaborated in our hypothesis. To systematically analyze these effects, we introduce VHBench-10, a comprehensive benchmark with approximately 10,000 samples for evaluating LVLMs across ten fine-grained hallucination categories. Our evaluations confirm encoders exhibit unique hallucination characteristics. Building on these insights and the suboptimality of simple feature fusion, we propose VisionWeaver, a novel Context-Aware Routing Network. It employs global visual features to generate routing signals, dynamically aggregating visual features from multiple specialized experts. Comprehensive experiments confirm the effectiveness of VisionWeaver in significantly reducing hallucinations and improving overall model performance.

Method: Proposed Nighttime Deraining Location-enhanced Perceptual Network (NDLPNet) with Position Perception Module (PPM) to capture spatial contextual information and recalibrate feature channel importance. Also created a new NSR dataset with 900 real-world nighttime rainy image pairs.

Result: Extensive experiments show the method outperforms state-of-the-art techniques in nighttime deraining tasks on both existing datasets and the new NSR benchmark dataset.

Conclusion: NDLPNet effectively addresses nighttime deraining challenges by capturing spatial positional information and rain density distribution, providing superior performance in low-light rain removal while preserving crucial background details.

Abstract: Visual degradation caused by rain streak artifacts in low-light conditions significantly hampers the performance of nighttime surveillance and autonomous navigation. Existing image deraining techniques are primarily designed for daytime conditions and perform poorly under nighttime illumination due to the spatial heterogeneity of rain distribution and the impact of light-dependent stripe visibility. In this paper, we propose a novel Nighttime Deraining Location-enhanced Perceptual Network(NDLPNet) that effectively captures the spatial positional information and density distribution of rain streaks in low-light environments. Specifically, we introduce a Position Perception Module (PPM) to capture and leverage spatial contextual information from input data, enhancing the model’s capability to identify and recalibrate the importance of different feature channels. The proposed nighttime deraining network can effectively remove the rain streaks as well as preserve the crucial background information. Furthermore, We construct a night scene rainy (NSR) dataset comprising 900 image pairs, all based on real-world nighttime scenes, providing a new benchmark for nighttime deraining task research. Extensive qualitative and quantitative experimental evaluations on both existing datasets and the NSR dataset consistently demonstrate our method outperform the state-of-the-art (SOTA) methods in nighttime deraining tasks. The source code and dataset is available at https://github.com/Feecuin/NDLPNet.

Method: A multimodal framework that combines video, audio, and phonological inputs through cross-attention fusion for dynamic feature alignment, with additional contrastive learning to improve cross-modal representation and robustness.

Result: Achieves state-of-the-art performance with Dice score of 0.95 and HD_95 of 4.20 mm on USC-75 rtMRI dataset, outperforming both unimodal and multimodal baselines.

Conclusion: Integrative multimodal modeling with cross-attention and contrastive learning significantly improves vocal tract segmentation accuracy and robustness, demonstrating the value of combining multiple modalities for precise articulatory analysis.

Abstract: Accurately segmenting articulatory structures in real-time magnetic resonance imaging (rtMRI) remains challenging, as most existing methods rely almost entirely on visual cues. Yet synchronized acoustic and phonological signals provide complementary context that can enrich visual information and improve precision. In this paper, we introduce VocSegMRI, a multimodal framework that integrates video, audio, and phonological inputs through cross-attention fusion for dynamic feature alignment. To further enhance cross-modal representation, we incorporate a contrastive learning objective that improves segmentation performance even when the audio modality is unavailable at inference. Evaluated on a sub-set of USC-75 rtMRI dataset, our approach achieves state-of-the-art performance, with a Dice score of 0.95 and a 95th percentile Hausdorff Distance (HD_95) of 4.20 mm, outperforming both unimodal and multimodal baselines. Ablation studies confirm the contributions of cross-attention and contrastive learning to segmentation precision and robustness. These results highlight the value of integrative multimodal modeling for accurate vocal tract analysis.

Method: Uses pre-optimized encoder to generate compressed-domain representations, integrated with pretrained diffusion models via lightweight adapter and attentive fusion module. Includes distribution renormalization for enhanced reconstruction fidelity.

Result: Outperforms existing methods in visual fidelity at low bitrates, achieves up to 79% compression improvement over H.266/VVC, and provides efficient solution for AI-generated content while being adaptable to broader content types.

Conclusion: The proposed framework effectively leverages pretrained diffusion models for superior compression performance with minimal retraining costs, offering a promising solution for modern visual content coding needs.

Abstract: As generative technologies advance, visual content has evolved into a complex mix of natural and AI-generated images, driving the need for more efficient coding techniques that prioritize perceptual quality. Traditional codecs and learned methods struggle to maintain subjective quality at high compression ratios, while existing generative approaches face challenges in visual fidelity and generalization. To this end, we propose a novel generative coding framework leveraging diffusion priors to enhance compression performance at low bitrates. Our approach employs a pre-optimized encoder to generate generalized compressed-domain representations, integrated with the pretrained model’s internal features via a lightweight adapter and an attentive fusion module. This framework effectively leverages existing pretrained diffusion models and enables efficient adaptation to different pretrained models for new requirements with minimal retraining costs. We also introduce a distribution renormalization method to further enhance reconstruction fidelity. Extensive experiments show that our method (1) outperforms existing methods in visual fidelity across low bitrates, (2) improves compression performance by up to 79% over H.266/VVC, and (3) offers an efficient solution for AI-generated content while being adaptable to broader content types.

Method: Pretrained on large-scale AD scenarios using QA and trajectory datasets, then fine-tuned with two-mode dataset (fast answering without CoT and slow thinking with CoT). Uses Adaptive Think Reward strategy with Group Relative Policy Optimization to reward selective CoT application by comparing trajectory quality across reasoning modes.

Result: Achieves PDMS of 90.3 on Navsim benchmark, surpassing best vision-only baseline by 1.7 points. Outperforms never-Think and always-Think baselines by 2.0 and 1.4 PDMS points respectively, while reducing inference time by 14% compared to always-Think baseline.

Conclusion: AdaThinkDrive effectively balances accuracy and efficiency through adaptive reasoning, demonstrating superior performance over both non-reasoning and always-reasoning approaches while reducing computational costs.

Abstract: While reasoning technology like Chain of Thought (CoT) has been widely adopted in Vision Language Action (VLA) models, it demonstrates promising capabilities in end to end autonomous driving. However, recent efforts to integrate CoT reasoning often fall short in simple scenarios, introducing unnecessary computational overhead without improving decision quality. To address this, we propose AdaThinkDrive, a novel VLA framework with a dual mode reasoning mechanism inspired by fast and slow thinking. First, our framework is pretrained on large scale autonomous driving (AD) scenarios using both question answering (QA) and trajectory datasets to acquire world knowledge and driving commonsense. During supervised fine tuning (SFT), we introduce a two mode dataset, fast answering (w/o CoT) and slow thinking (with CoT), enabling the model to distinguish between scenarios that require reasoning. Furthermore, an Adaptive Think Reward strategy is proposed in conjunction with the Group Relative Policy Optimization (GRPO), which rewards the model for selectively applying CoT by comparing trajectory quality across different reasoning modes. Extensive experiments on the Navsim benchmark show that AdaThinkDrive achieves a PDMS of 90.3, surpassing the best vision only baseline by 1.7 points. Moreover, ablations show that AdaThinkDrive surpasses both the never Think and always Think baselines, improving PDMS by 2.0 and 1.4, respectively. It also reduces inference time by 14% compared to the always Think baseline, demonstrating its ability to balance accuracy and efficiency through adaptive reasoning.

Method: Proposes independent prediction of manipulated regions from both local and global perspectives, then uses morphological operations to fuse outputs for noise suppression and spatial coherence enhancement.

Result: Extensive experiments show effectiveness of each module in improving accuracy and robustness of forgery localization.

Conclusion: The proposed approach successfully addresses localization challenges by leveraging complementary local-global information with effective morphological fusion, achieving improved performance in deepfake region localization.

Abstract: While the pursuit of higher accuracy in deepfake detection remains a central goal, there is an increasing demand for precise localization of manipulated regions. Despite the remarkable progress made in classification-based detection, accurately localizing forged areas remains a significant challenge. A common strategy is to incorporate forged region annotations during model training alongside manipulated images. However, such approaches often neglect the complementary nature of local detail and global semantic context, resulting in suboptimal localization performance. Moreover, an often-overlooked aspect is the fusion strategy between local and global predictions. Naively combining the outputs from both branches can amplify noise and errors, thereby undermining the effectiveness of the localization. To address these issues, we propose a novel approach that independently predicts manipulated regions using both local and global perspectives. We employ morphological operations to fuse the outputs, effectively suppressing noise while enhancing spatial coherence. Extensive experiments reveal the effectiveness of each module in improving the accuracy and robustness of forgery localization.

Method: Lightweight causal spatial neighborhood encoder to capture local geometric relations, followed by Mamba-based state space models for scalable temporal modeling with linear complexity. Adaptive controller adjusts processing speed based on event rate during inference.

Result: Achieves optimal balance between window latency and inference latency by directly processing raw event streams without intermediate representations like frames, voxel grids, or point clouds.

Conclusion: The proposed architecture enables efficient real-time processing of event streams with adaptive rate control, overcoming computational challenges and latency issues of existing methods.

Abstract: Event cameras capture asynchronous pixel-level brightness changes with microsecond temporal resolution, offering unique advantages for high-speed vision tasks. Existing methods often convert event streams into intermediate representations such as frames, voxel grids, or point clouds, which inevitably require predefined time windows and thus introduce window latency. Meanwhile, pointwise detection methods face computational challenges that prevent real-time efficiency due to their high computational cost. To overcome these limitations, we propose the Variable-Rate Spatial Event Mamba, a novel architecture that directly processes raw event streams without intermediate representations. Our method introduces a lightweight causal spatial neighborhood encoder to efficiently capture local geometric relations, followed by Mamba-based state space models for scalable temporal modeling with linear complexity. During inference, a controller adaptively adjusts the processing speed according to the event rate, achieving an optimal balance between window latency and inference latency.

Method: Gated Residual Tokenization (GRT) with two stages: (1) Motion-Compensated Inter-Gated Tokenization uses pixel-level motion estimation to skip static regions, (2) Semantic-Scene Intra-Tokenization Merging fuses tokens across static regions to reduce redundancy.

Result: GRT outperforms larger VLLM baselines on the DIVE benchmark and scales positively with FPS, demonstrating efficient high-FPS video understanding capability.

Conclusion: Dense temporal information is crucial for video understanding, and GRT provides an efficient, scalable solution for high-FPS video comprehension while reducing computational overhead.

Abstract: High temporal resolution is essential for capturing fine-grained details in video understanding. However, current video large language models (VLLMs) and benchmarks mostly rely on low-frame-rate sampling, such as uniform sampling or keyframe selection, discarding dense temporal information. This compromise avoids the high cost of tokenizing every frame, which otherwise leads to redundant computation and linear token growth as video length increases. While this trade-off works for slowly changing content, it fails for tasks like lecture comprehension, where information appears in nearly every frame and requires precise temporal alignment. To address this gap, we introduce Dense Video Understanding (DVU), which enables high-FPS video comprehension by reducing both tokenization time and token overhead. Existing benchmarks are also limited, as their QA pairs focus on coarse content changes. We therefore propose DIVE (Dense Information Video Evaluation), the first benchmark designed for dense temporal reasoning. To make DVU practical, we present Gated Residual Tokenization (GRT), a two-stage framework: (1) Motion-Compensated Inter-Gated Tokenization uses pixel-level motion estimation to skip static regions during tokenization, achieving sub-linear growth in token count and compute. (2) Semantic-Scene Intra-Tokenization Merging fuses tokens across static regions within a scene, further reducing redundancy while preserving dynamic semantics. Experiments on DIVE show that GRT outperforms larger VLLM baselines and scales positively with FPS. These results highlight the importance of dense temporal information and demonstrate that GRT enables efficient, scalable high-FPS video understanding.

Method: Block-Wise Caching (BWCache) dynamically caches and reuses features from DiT blocks across timesteps, using a similarity indicator to trigger reuse only when feature differences are below a threshold.

Result: Extensive experiments show BWCache achieves up to 2.24× speedup with comparable visual quality across several video diffusion models.

Conclusion: BWCache effectively reduces computational redundancy in DiT-based video generation while preserving visual fidelity, making it suitable for real-world applications.

Abstract: Recent advancements in Diffusion Transformers (DiTs) have established them as the state-of-the-art method for video generation. However, their inherently sequential denoising process results in inevitable latency, limiting real-world applicability. Existing acceleration methods either compromise visual quality due to architectural modifications or fail to reuse intermediate features at proper granularity. Our analysis reveals that DiT blocks are the primary contributors to inference latency. Across diffusion timesteps, the feature variations of DiT blocks exhibit a U-shaped pattern with high similarity during intermediate timesteps, which suggests substantial computational redundancy. In this paper, we propose Block-Wise Caching (BWCache), a training-free method to accelerate DiT-based video generation. BWCache dynamically caches and reuses features from DiT blocks across diffusion timesteps. Furthermore, we introduce a similarity indicator that triggers feature reuse only when the differences between block features at adjacent timesteps fall below a threshold, thereby minimizing redundant computations while maintaining visual fidelity. Extensive experiments on several video diffusion models demonstrate that BWCache achieves up to 2.24$\times$ speedup with comparable visual quality.

Method: Proposes a supervised domain adaptation framework based on Learning Invariant Representation and Risk (LIRR) paradigm, jointly optimizing domain-invariant representations and task-specific risk using both synthetic and limited real labeled data.

Result: Outperforms source-only, fine-tuning, and oracle baselines on SPEED+ benchmark. With only 5% labeled target data, matches or surpasses oracle performance trained on larger data fractions. Framework is lightweight, backbone-agnostic, and computationally efficient.

Conclusion: Provides a practical pathway for robust spacecraft pose estimation in real-world space environments by effectively bridging the synthetic-to-real domain gap with minimal labeled real data requirements.

Abstract: Spacecraft Pose Estimation (SPE) is a fundamental capability for autonomous space operations such as rendezvous, docking, and in-orbit servicing. Hybrid pipelines that combine object detection, keypoint regression, and Perspective-n-Point (PnP) solvers have recently achieved strong results on synthetic datasets, yet their performance deteriorates sharply on real or lab-generated imagery due to the persistent synthetic-to-real domain gap. Existing unsupervised domain adaptation approaches aim to mitigate this issue but often underperform when a modest number of labeled target samples are available. In this work, we propose the first Supervised Domain Adaptation (SDA) framework tailored for SPE keypoint regression. Building on the Learning Invariant Representation and Risk (LIRR) paradigm, our method jointly optimizes domain-invariant representations and task-specific risk using both labeled synthetic and limited labeled real data, thereby reducing generalization error under domain shift. Extensive experiments on the SPEED+ benchmark demonstrate that our approach consistently outperforms source-only, fine-tuning, and oracle baselines. Notably, with only 5% labeled target data, our method matches or surpasses oracle performance trained on larger fractions of labeled data. The framework is lightweight, backbone-agnostic, and computationally efficient, offering a practical pathway toward robust and deployable spacecraft pose estimation in real-world space environments.

Method: Integrates semantic features from UAV and satellite imagery using semantic weighting mechanism and optimized particle filtering architecture for 4-DoF pose estimation.

Result: 10x computational efficiency gain over feature extraction methods, global positioning errors below 10 meters, rapid pose estimation within seconds using low-resolution satellite maps.

Conclusion: The proposed SWA-PF method with MAFS dataset effectively addresses UAV localization challenges in GNSS-denied environments with superior efficiency and accuracy.

Abstract: Vision-based Unmanned Aerial Vehicle (UAV) localization systems have been extensively investigated for Global Navigation Satellite System (GNSS)-denied environments. However, existing retrieval-based approaches face limitations in dataset availability and persistent challenges including suboptimal real-time performance, environmental sensitivity, and limited generalization capability, particularly in dynamic or temporally varying environments. To overcome these limitations, we present a large-scale Multi-Altitude Flight Segments dataset (MAFS) for variable altitude scenarios and propose a novel Semantic-Weighted Adaptive Particle Filter (SWA-PF) method. This approach integrates robust semantic features from both UAV-captured images and satellite imagery through two key innovations: a semantic weighting mechanism and an optimized particle filtering architecture. Evaluated using our dataset, the proposed method achieves 10x computational efficiency gain over feature extraction methods, maintains global positioning errors below 10 meters, and enables rapid 4 degree of freedom (4-DoF) pose estimation within seconds using accessible low-resolution satellite maps. Code and dataset will be available at https://github.com/YuanJiayuuu/SWA-PF.

Method: Proposes MFM that masks and reconstructs features directly in feature space using a lightweight Rebuilder module during training, leveraging the segmentation decoder to classify reconstructed features without modifying inference pipeline.

Result: Extensive experiments show MFM consistently enhances segmentation performance across various architectures and UDA benchmarks.

Conclusion: MFM provides a simple, efficient, and generalizable strategy for unsupervised domain-adaptive semantic segmentation that tightly couples auxiliary objectives with the main task.

Abstract: Unsupervised domain adaptation (UDA) for semantic segmentation aims to transfer models from a labeled source domain to an unlabeled target domain. While auxiliary self-supervised tasks-particularly contrastive learning-have improved feature discriminability, masked modeling approaches remain underexplored in this setting, largely due to architectural incompatibility and misaligned optimization objectives. We propose Masked Feature Modeling (MFM), a novel auxiliary task that performs feature masking and reconstruction directly in the feature space. Unlike existing masked modeling methods that reconstruct low-level inputs or perceptual features (e.g., HOG or visual tokens), MFM aligns its learning target with the main segmentation task, ensuring compatibility with standard architectures like DeepLab and DAFormer without modifying the inference pipeline. To facilitate effective reconstruction, we introduce a lightweight auxiliary module, Rebuilder, which is trained jointly but discarded during inference, adding zero computational overhead at test time. Crucially, MFM leverages the segmentation decoder to classify the reconstructed features, tightly coupling the auxiliary objective with the pixel-wise prediction task to avoid interference with the primary task. Extensive experiments across various architectures and UDA benchmarks demonstrate that MFM consistently enhances segmentation performance, offering a simple, efficient, and generalizable strategy for unsupervised domain-adaptive semantic segmentation.

Method: Investigated MiniROCKET and HDC-MiniROCKET for spectral classification, which extract well-engineered features without trainable parameters in the feature extraction part, making them less vulnerable to limited training data.

Result: MiniROCKET outperforms 1D-Justo-LiuNet in limited data scenarios and is mostly on par with it in general cases, despite having more parameters.

Conclusion: MiniROCKET-based approaches provide a robust solution for spectral classification, particularly beneficial when training data is limited, while maintaining competitive performance in standard scenarios.

Abstract: The classification of pixel spectra of hyperspectral images, i.e. spectral classification, is used in many fields ranging from agricultural, over medical to remote sensing applications and is currently also expanding to areas such as autonomous driving. Even though for full hyperspectral images the best-performing methods exploit spatial-spectral information, performing classification solely on spectral information has its own advantages, e.g. smaller model size and thus less data required for training. Moreover, spectral information is complementary to spatial information and improvements on either part can be used to improve spatial-spectral approaches in the future. Recently, 1D-Justo-LiuNet was proposed as a particularly efficient model with very few parameters, which currently defines the state of the art in spectral classification. However, we show that with limited training data the model performance deteriorates. Therefore, we investigate MiniROCKET and HDC-MiniROCKET for spectral classification to mitigate that problem. The model extracts well-engineered features without trainable parameters in the feature extraction part and is therefore less vulnerable to limited training data. We show that even though MiniROCKET has more parameters it outperforms 1D-Justo-LiuNet in limited data scenarios and is mostly on par with it in the general case

Method: Uses three expert networks: main segmentation expert, signed distance field regression expert, and boundary prediction expert. Features Multi-Gating Pseudo-labeling module for dynamic feature aggregation and Adaptive Multi-Objective Loss for automatic balancing of learning objectives.

Result: Outperforms state-of-the-art approaches on GlaS and CRAG benchmarks in low-label settings, demonstrating superior performance in semi-supervised histopathology segmentation.

Conclusion: Semi-MOE shows the potential of MoE-based architectures for advancing semi-supervised segmentation, effectively addressing noisy pseudo-label issues through specialized expert networks and adaptive optimization.

Abstract: Semi-supervised learning has been employed to alleviate the need for extensive labeled data for histopathology image segmentation, but existing methods struggle with noisy pseudo-labels due to ambiguous gland boundaries and morphological misclassification. This paper introduces Semi-MOE, to the best of our knowledge, the first multi-task Mixture-of-Experts framework for semi-supervised histopathology image segmentation. Our approach leverages three specialized expert networks: A main segmentation expert, a signed distance field regression expert, and a boundary prediction expert, each dedicated to capturing distinct morphological features. Subsequently, the Multi-Gating Pseudo-labeling module dynamically aggregates expert features, enabling a robust fuse-and-refine pseudo-labeling mechanism. Furthermore, to eliminate manual tuning while dynamically balancing multiple learning objectives, we propose an Adaptive Multi-Objective Loss. Extensive experiments on GlaS and CRAG benchmarks show that our method outperforms state-of-the-art approaches in low-label settings, highlighting the potential of MoE-based architectures in advancing semi-supervised segmentation. Our code is available at https://github.com/vnlvi2k3/Semi-MoE.

Method: Proposes Consistent View Alignment method that aligns representations from different data views to capture complementary information while avoiding false positive alignments.

Result: The method achieved 1st and 2nd place in MICCAI 2025 SSL3D challenge using Primus vision transformer and ResEnc CNN respectively, with improved downstream task performance.

Conclusion: Structured view alignment is critical for learning effective representations, and explicit induction of latent space structure significantly improves self-supervised learning outcomes.

Abstract: Many recent approaches in representation learning implicitly assume that uncorrelated views of a data point are sufficient to learn meaningful representations for various downstream tasks. In this work, we challenge this assumption and demonstrate that meaningful structure in the latent space does not emerge naturally. Instead, it must be explicitly induced. We propose a method that aligns representations from different views of the data to align complementary information without inducing false positives. Our experiments show that our proposed self-supervised learning method, Consistent View Alignment, improves performance for downstream tasks, highlighting the critical role of structured view alignment in learning effective representations. Our method achieved first and second place in the MICCAI 2025 SSL3D challenge when using a Primus vision transformer and ResEnc convolutional neural network, respectively. The code and pretrained model weights are released at https://github.com/Tenbatsu24/LatentCampus.

Method: Proposes SpecDiff with two algorithms: (1) feature selection based on self-speculative information and historical information to determine dynamic importance scores, and (2) multi-level feature classification using importance score differences and multi-level calculation strategy.

Result: Achieves average 2.80×, 2.74×, and 3.17× speedup in Stable Diffusion 3, 3.5, and FLUX respectively with negligible quality loss compared to RFlow on NVIDIA A800-80GB GPU.

Conclusion: By merging speculative and historical information, SpecDiff overcomes the speedup-accuracy trade-off bottleneck and pushes the Pareto frontier of speedup and accuracy in efficient diffusion model inference.

Abstract: Feature caching has recently emerged as a promising method for diffusion model acceleration. It effectively alleviates the inefficiency problem caused by high computational requirements by caching similar features in the inference process of the diffusion model. In this paper, we analyze existing feature caching methods from the perspective of information utilization, and point out that relying solely on historical information will lead to constrained accuracy and speed performance. And we propose a novel paradigm that introduces future information via self-speculation based on the information similarity at the same time step across different iteration times. Based on this paradigm, we present \textit{SpecDiff}, a training-free multi-level feature caching strategy including a cached feature selection algorithm and a multi-level feature classification algorithm. (1) Feature selection algorithm based on self-speculative information. \textit{SpecDiff} determines a dynamic importance score for each token based on self-speculative information and historical information, and performs cached feature selection through the importance score. (2) Multi-level feature classification algorithm based on feature importance scores. \textit{SpecDiff} classifies tokens by leveraging the differences in feature importance scores and introduces a multi-level feature calculation strategy. Extensive experiments show that \textit{SpecDiff} achieves average 2.80 \times, 2.74 \times , and 3.17\times speedup with negligible quality loss in Stable Diffusion 3, 3.5, and FLUX compared to RFlow on NVIDIA A800-80GB GPU. By merging speculative and historical information, \textit{SpecDiff} overcomes the speedup-accuracy trade-off bottleneck, pushing the Pareto frontier of speedup and accuracy in the efficient diffusion model inference.

Method: The EDITS framework uses a Vision Language Model to generate external texts fused with image features, creates image and text prototypes through Local Semantic Awareness, and employs Dual Prototype Guidance with a diffusion model to generate the final synthetic dataset.

Result: Extensive experiments confirm the effectiveness of the EDITS method, demonstrating improved performance over traditional dataset distillation approaches.

Conclusion: EDITS successfully exploits implicit textual semantics within image data to achieve enhanced dataset distillation, providing a more comprehensive approach that captures both visual and semantic information for improved learning efficiency.

Abstract: Dataset distillation aims to synthesize a compact dataset from the original large-scale one, enabling highly efficient learning while preserving competitive model performance. However, traditional techniques primarily capture low-level visual features, neglecting the high-level semantic and structural information inherent in images. In this paper, we propose EDITS, a novel framework that exploits the implicit textual semantics within the image data to achieve enhanced distillation. First, external texts generated by a Vision Language Model (VLM) are fused with image features through a Global Semantic Query module, forming the prior clustered buffer. Local Semantic Awareness then selects representative samples from the buffer to construct image and text prototypes, with the latter produced by guiding a Large Language Model (LLM) with meticulously crafted prompt. Ultimately, Dual Prototype Guidance strategy generates the final synthetic dataset through a diffusion model. Extensive experiments confirm the effectiveness of our method.Source code is available in: https://github.com/einsteinxia/EDITS.

Method: Combines Gaussian Splatting radiative rasterization with a dedicated detector-to-world transformation model incorporating laminographic tilt angle, plus an initialization strategy that filters out common laminographic artifacts.

Result: Achieves superior performance over iterative methods using only 3% of full views, with extensive experiments on synthetic and real datasets demonstrating effectiveness.

Conclusion: LamiGauss enables accurate and efficient reconstruction with limited data by preventing redundant Gaussians from being allocated to false structures and concentrating model capacity on genuine objects.

Abstract: X-ray Computed Laminography (CL) is essential for non-destructive inspection of plate-like structures in applications such as microchips and composite battery materials, where traditional computed tomography (CT) struggles due to geometric constraints. However, reconstructing high-quality volumes from laminographic projections remains challenging, particularly under highly sparse-view acquisition conditions. In this paper, we propose a reconstruction algorithm, namely LamiGauss, that combines Gaussian Splatting radiative rasterization with a dedicated detector-to-world transformation model incorporating the laminographic tilt angle. LamiGauss leverages an initialization strategy that explicitly filters out common laminographic artifacts from the preliminary reconstruction, preventing redundant Gaussians from being allocated to false structures and thereby concentrating model capacity on representing the genuine object. Our approach effectively optimizes directly from sparse projections, enabling accurate and efficient reconstruction with limited data. Extensive experiments on both synthetic and real datasets demonstrate the effectiveness and superiority of the proposed method over existing techniques. LamiGauss uses only 3$%$ of full views to achieve superior performance over the iterative method optimized on a full dataset.

Method: Proposed a distractor-aware drop-in memory module and introspection-based management method for SAM2, creating DAM4SAM. Also constructed DiDi dataset for distractor analysis.

Result: Outperforms SAM2.1 on 13 benchmarks, sets new SOTA on 10 benchmarks. Integration with EfficientTAM tracker yields 11% improvement, matching non-real-time SAM2.1-L performance. EdgeTAM integration shows 4% performance boost.

Conclusion: The distractor-aware memory module effectively addresses tracking challenges with distractors and demonstrates excellent generalization across different tracker architectures while maintaining real-time capabilities.

Abstract: Recent emergence of memory-based video segmentation methods such as SAM2 has led to models with excellent performance in segmentation tasks, achieving leading results on numerous benchmarks. However, these modes are not fully adjusted for visual object tracking, where distractors (i.e., objects visually similar to the target) pose a key challenge. In this paper we propose a distractor-aware drop-in memory module and introspection-based management method for SAM2, leading to DAM4SAM. Our design effectively reduces the tracking drift toward distractors and improves redetection capability after object occlusion. To facilitate the analysis of tracking in the presence of distractors, we construct DiDi, a Distractor-Distilled dataset. DAM4SAM outperforms SAM2.1 on thirteen benchmarks and sets new state-of-the-art results on ten. Furthermore, integrating the proposed distractor-aware memory into a real-time tracker EfficientTAM leads to 11% improvement and matches tracking quality of the non-real-time SAM2.1-L on multiple tracking and segmentation benchmarks, while integration with edge-based tracker EdgeTAM delivers 4% performance boost, demonstrating a very good generalization across architectures.

Method: Uses a dual-stream attention network with Fused Attention Blocks (FABlocks) to fuse raw X-rays and segmented bone images, trained in a two-stage pipeline with segmentation guidance.

Result: Achieves 88.68% accuracy and 0.9334 AUC on visible fractures, and 82.29% accuracy with 0.8688 AUC on invisible fractures despite no specific training for them.

Conclusion: The anatomy-aware dual-input architecture shows strong clinical potential for robust fracture detection, particularly in cases with subtle radiographic presentations.

Abstract: Pelvic fractures pose significant diagnostic challenges, particularly in cases where fracture signs are subtle or invisible on standard radiographs. To address this, we introduce PelFANet, a dual-stream attention network that fuses raw pelvic X-rays with segmented bone images to improve fracture classification. The network em-ploys Fused Attention Blocks (FABlocks) to iteratively exchange and refine fea-tures from both inputs, capturing global context and localized anatomical detail. Trained in a two-stage pipeline with a segmentation-guided approach, PelFANet demonstrates superior performance over conventional methods. On the AMERI dataset, it achieves 88.68% accuracy and 0.9334 AUC on visible fractures, while generalizing effectively to invisible fracture cases with 82.29% accuracy and 0.8688 AUC, despite not being trained on them. These results highlight the clini-cal potential of anatomy-aware dual-input architectures for robust fracture detec-tion, especially in scenarios with subtle radiographic presentations.

Method: Uses wrist-based ROI localization with geometric cues, end-to-end mapping with embedded ROI offsets to reduce computation, and multi-task learning with auxiliary geometric feature head. Constructed synthetic training data with 3D labels and real event data with 2D labels.

Result: Achieved 2D-AUCp improvement from 0.77 to 0.85, reduced parameters by 89% (11.2M to 1.2M), reduced FLOPs by 89% (1.648G to 0.185G), and maintained competitive 3D-AUCp of 0.84 on synthetic data.

Conclusion: EvHand-FPV demonstrates accurate and efficient egocentric event-based hand tracking suitable for on-device XR applications, addressing the limitations of traditional frame-based methods.

Abstract: Hand tracking holds great promise for intuitive interaction paradigms, but frame-based methods often struggle to meet the requirements of accuracy, low latency, and energy efficiency, especially in resource-constrained settings such as Extended Reality (XR) devices. Event cameras provide $\mu$s-level temporal resolution at mW-level power by asynchronously sensing brightness changes. In this work, we present EvHand-FPV, a lightweight framework for egocentric First-Person-View 3D hand tracking from a single event camera. We construct an event-based FPV dataset that couples synthetic training data with 3D labels and real event data with 2D labels for evaluation to address the scarcity of egocentric benchmarks. EvHand-FPV also introduces a wrist-based region of interest (ROI) that localizes the hand region via geometric cues, combined with an end-to-end mapping strategy that embeds ROI offsets into the network to reduce computation without explicit reconstruction, and a multi-task learning strategy with an auxiliary geometric feature head that improves representations without test-time overhead. On our real FPV test set, EvHand-FPV improves 2D-AUCp from 0.77 to 0.85 while reducing parameters from 11.2M to 1.2M by 89% and FLOPs per inference from 1.648G to 0.185G by 89%. It also maintains a competitive 3D-AUCp of 0.84 on synthetic data. These results demonstrate accurate and efficient egocentric event-based hand tracking suitable for on-device XR applications. The dataset and code are available at https://github.com/zen5x5/EvHand-FPV.

Method: WARM module sandwiches cross-attention between whitening and coloring transformations - whitening aligns support features to prototypical tokens before attention, then coloring restores original distribution to attended tokens.

Result: Achieves state-of-the-art performance with significant margin on multiple FS-PCS benchmarks through extensive experiments.

Conclusion: The simple yet effective WARM design enables robust attention and generates representative prototypes by capturing semantic relationships among support features.

Abstract: Few-Shot 3D Point Cloud Segmentation (FS-PCS) aims to predict per-point labels for an unlabeled point cloud, given only a few labeled examples. To extract discriminative representations from the limited support set, existing methods have constructed prototypes using conventional algorithms such as farthest point sampling. However, we point out that its initial randomness significantly affects FS-PCS performance and that the prototype generation process remains underexplored despite its prevalence. This motivates us to investigate an advanced prototype generation method based on attention mechanism. Despite its potential, we found that vanilla module suffers from the distributional gap between learnable prototypical tokens and support features. To overcome this, we propose White Aggregation and Restoration Module (WARM), which resolves the misalignment by sandwiching cross-attention between whitening and coloring transformations. Specifically, whitening aligns the support features to prototypical tokens before attention process, and subsequently coloring restores the original distribution to the attended tokens. This simple yet effective design enables robust attention, thereby generating representative prototypes by capturing the semantic relationships among support features. Our method achieves state-of-the-art performance with a significant margin on multiple FS-PCS benchmarks, demonstrating its effectiveness through extensive experiments.

Method: Self-Rationale Calibration (SRC) framework: 1) Rationale fine-tuning to require rationale before answer, 2) Diverse candidate generation, 3) Pairwise scoring with R-Scorer model, 4) Confidence-weighted preference curation for alignment calibration.

Result: Significant improvements in perception, reasoning, and generalization across multiple benchmarks for LVLMs.

Conclusion: Rationale-oriented alignment is crucial for exploring the full potential of Large Vision-Language Models.

Abstract: Large Vision-Language Models (LVLMs) have manifested strong visual question answering capability. However, they still struggle with aligning the rationale and the generated answer, leading to inconsistent reasoning and incorrect responses. To this end, this paper introduces the Self-Rationale Calibration (SRC) framework to iteratively calibrate the alignment between rationales and answers. SRC begins by employing a lightweight “rationale fine-tuning” approach, which modifies the model’s response format to require a rationale before deriving an answer without explicit prompts. Next, SRC searches for a diverse set of candidate responses from the fine-tuned LVLMs for each sample, followed by a proposed pairwise scoring strategy using a tailored scoring model, R-Scorer, to evaluate both rationale quality and factual consistency of candidates. Based on a confidence-weighted preference curation process, SRC decouples the alignment calibration into a preference fine-tuning manner, leading to significant improvements of LVLMs in perception, reasoning, and generalization across multiple benchmarks. Our results emphasize the rationale-oriented alignment in exploring the potential of LVLMs.

Method: Uses a translation module to map student features into joint space, trained with dual-objective loss enforcing both local alignment and global relational consistency at object level.

Result: Achieves +10.1 average score improvement across four personalized detection benchmarks under few-shot regimes, reaching performance comparable to larger multimodal models.

Conclusion: MOCHA proves effective for real-world deployment by enabling compact vision-only detectors to achieve multimodal performance levels through efficient knowledge distillation.

Abstract: We introduce MOCHA (Multi-modal Objects-aware Cross-arcHitecture Alignment), a knowledge distillation approach that transfers region-level multimodal semantics from a large vision-language teacher (e.g., LLaVa) into a lightweight vision-only object detector student (e.g., YOLO). A translation module maps student features into a joint space, where the training of the student and translator is guided by a dual-objective loss that enforces both local alignment and global relational consistency. Unlike prior approaches focused on dense or global alignment, MOCHA operates at the object level, enabling efficient transfer of semantics without modifying the teacher or requiring textual input at inference. We validate our method across four personalized detection benchmarks under few-shot regimes. Results show consistent gains over baselines, with a +10.1 average score improvement. Despite its compact architecture, MOCHA reaches performance on par with larger multimodal models, proving its suitability for real-world deployment.

Method: AntiPure uses two guidance mechanisms: 1) Patch-wise Frequency Guidance to reduce model influence over high-frequency components, and 2) Erroneous Timestep Guidance to disrupt denoising strategies across timesteps.

Result: AntiPure achieves minimal perceptual discrepancy and maximal distortion, outperforming other protective perturbation methods in resisting purification within the purification-customization workflow.

Conclusion: AntiPure effectively exposes purification vulnerabilities and provides robust protection against image misuse in diffusion models, serving as a stress test for purification methods.

Abstract: Diffusion models like Stable Diffusion have become prominent in visual synthesis tasks due to their powerful customization capabilities, which also introduce significant security risks, including deepfakes and copyright infringement. In response, a class of methods known as protective perturbation emerged, which mitigates image misuse by injecting imperceptible adversarial noise. However, purification can remove protective perturbations, thereby exposing images again to the risk of malicious forgery. In this work, we formalize the anti-purification task, highlighting challenges that hinder existing approaches, and propose a simple diagnostic protective perturbation named AntiPure. AntiPure exposes vulnerabilities of purification within the “purification-customization” workflow, owing to two guidance mechanisms: 1) Patch-wise Frequency Guidance, which reduces the model’s influence over high-frequency components in the purified image, and 2) Erroneous Timestep Guidance, which disrupts the model’s denoising strategy across different timesteps. With additional guidance, AntiPure embeds imperceptible perturbations that persist under representative purification settings, achieving effective post-customization distortion. Experiments show that, as a stress test for purification, AntiPure achieves minimal perceptual discrepancy and maximal distortion, outperforming other protective perturbation methods within the purification-customization workflow.

Method: Proposes Noise Level Guidance (NLG) that refines initial noise by increasing its alignment likelihood with general guidance. Works without additional training data, auxiliary networks, or backpropagation, and is generalizable to both conditional and unconditional diffusion models.

Result: Extensive experiments on five standard benchmarks demonstrate enhanced output generation quality and improved input condition adherence. The method integrates seamlessly with existing guidance methods while maintaining computational efficiency.

Conclusion: NLG establishes itself as a practical and scalable enhancement to diffusion models, providing a unified framework that accommodates various forms of diffusion-level guidance without additional computational burdens.

Abstract: Diffusion models have achieved state-of-the-art image generation. However, the random Gaussian noise used to start the diffusion process influences the final output, causing variations in image quality and prompt adherence. Existing noise-level optimization approaches generally rely on extra dataset construction, additional networks, or backpropagation-based optimization, limiting their practicality. In this paper, we propose Noise Level Guidance (NLG), a simple, efficient, and general noise-level optimization approach that refines initial noise by increasing the likelihood of its alignment with general guidance - requiring no additional training data, auxiliary networks, or backpropagation. The proposed NLG approach provides a unified framework generalizable to both conditional and unconditional diffusion models, accommodating various forms of diffusion-level guidance. Extensive experiments on five standard benchmarks demonstrate that our approach enhances output generation quality and input condition adherence. By seamlessly integrating with existing guidance methods while maintaining computational efficiency, our method establishes NLG as a practical and scalable enhancement to diffusion models. Code can be found at https://github.com/harveymannering/NoiseLevelGuidance.

Method: Created PairTally dataset with 681 images containing two object categories, including both inter-category and intra-category settings. Benchmarked various state-of-the-art models including exemplar-based methods, language-prompted models, and large vision-language models.

Result: Current models struggle to reliably count what users intend, especially in fine-grained and visually ambiguous cases, despite recent advances in counting technologies.

Conclusion: PairTally provides a new foundation for diagnosing and improving fine-grained visual counting systems, highlighting the limitations of current approaches and the need for better selective counting capabilities.

Abstract: Visual counting is a fundamental yet challenging task, especially when users need to count objects of a specific type in complex scenes. While recent models, including class-agnostic counting models and large vision-language models (VLMs), show promise in counting tasks, their ability to perform fine-grained, intent-driven counting remains unclear. In this paper, we introduce PairTally, a benchmark dataset specifically designed to evaluate fine-grained visual counting. Each of the 681 high-resolution images in PairTally contains two object categories, requiring models to distinguish and count based on subtle differences in shape, size, color, or semantics. The dataset includes both inter-category (distinct categories) and intra-category (closely related subcategories) settings, making it suitable for rigorous evaluation of selective counting capabilities. We benchmark a variety of state-of-the-art models, including exemplar-based methods, language-prompted models, and large VLMs. Our results show that despite recent advances, current models struggle to reliably count what users intend, especially in fine-grained and visually ambiguous cases. PairTally provides a new foundation for diagnosing and improving fine-grained visual counting systems.

Method: Enhanced YOLO-FEDER FusionNet framework integrating generic object detection with camouflage detection. Uses large-scale synthetic data with real-world samples, systematic evaluation of intermediate multi-scale FEDER features, and benchmarking across multiple YOLO-based backbone configurations.

Result: Best configuration (YOLOv8l backbone with FEDER features from DWD module) achieved: FNR reduction of up to 39.1 percentage points and mAP increase of up to 62.8 percentage points at IoU threshold of 0.5 compared to baseline.

Conclusion: Integrating intermediate FEDER features with backbone upgrades significantly improves drone detection performance in complex environments, demonstrating the effectiveness of combining camouflage detection techniques with object detection frameworks.

Abstract: Drone detection in visually complex environments remains challenging due to background clutter, small object scale, and camouflage effects. While generic object detectors like YOLO exhibit strong performance in low-texture scenes, their effectiveness degrades in cluttered environments with low object-background separability. To address these limitations, this work presents an enhanced iteration of YOLO-FEDER FusionNet – a detection framework that integrates generic object detection with camouflage object detection techniques. Building upon the original architecture, the proposed iteration introduces systematic advancements in training data composition, feature fusion strategies, and backbone design. Specifically, the training process leverages large-scale, photo-realistic synthetic data, complemented by a small set of real-world samples, to enhance robustness under visually complex conditions. The contribution of intermediate multi-scale FEDER features is systematically evaluated, and detection performance is comprehensively benchmarked across multiple YOLO-based backbone configurations. Empirical results indicate that integrating intermediate FEDER features, in combination with backbone upgrades, contributes to notable performance improvements. In the most promising configuration – YOLO-FEDER FusionNet with a YOLOv8l backbone and FEDER features derived from the DWD module – these enhancements lead to a FNR reduction of up to 39.1 percentage points and a mAP increase of up to 62.8 percentage points at an IoU threshold of 0.5, compared to the initial baseline.

Method: Proposes STEP framework with dCTS (lightweight CNN-based policy network) for dynamic patch merging into superpatches, and early-exit mechanism to remove high-confidence tokens, reducing computational load.

Result: dCTS alone reduces token count by 2.5x, yielding 2.6x computational cost reduction and 3.4x throughput increase. Full STEP framework achieves up to 4x computational complexity reduction and 1.7x inference speed gain with max 2.0% accuracy drop. Up to 40% of tokens can be halted early.

Conclusion: STEP effectively addresses ViT efficiency issues through hybrid token reduction, enabling high-performance semantic segmentation with significantly reduced computational requirements while maintaining accuracy.

Abstract: Vision Transformers (ViTs) achieve state-of-the-art performance in semantic segmentation but are hindered by high computational and memory costs. To address this, we propose STEP (SuperToken and Early-Pruning), a hybrid token-reduction framework that combines dynamic patch merging and token pruning to enhance efficiency without significantly compromising accuracy. At the core of STEP is dCTS, a lightweight CNN-based policy network that enables flexible merging into superpatches. Encoder blocks integrate also early-exits to remove high-confident supertokens, lowering computational load. We evaluate our method on high-resolution semantic segmentation benchmarks, including images up to 1024 x 1024, and show that when dCTS is applied alone, the token count can be reduced by a factor of 2.5 compared to the standard 16 x 16 pixel patching scheme. This yields a 2.6x reduction in computational cost and a 3.4x increase in throughput when using ViT-Large as the backbone. Applying the full STEP framework further improves efficiency, reaching up to a 4x reduction in computational complexity and a 1.7x gain in inference speed, with a maximum accuracy drop of no more than 2.0%. With the proposed STEP configurations, up to 40% of tokens can be confidently predicted and halted before reaching the final encoder layer.

Method: Three core innovations: 1) Large-scale data curation pipeline with scoring/filtering strategies, 2) Progressive training framework starting with SAIL-ViT encoder through multimodal pre-training to thinking-fusion SFT-RL hybrid, 3) Architectural advances including sparse Mixture-of-Experts designs.

Result: State-of-the-art performance across diverse image/video benchmarks, competitive on 106 datasets, top results on MMMU and MathVista reasoning benchmarks. SAIL-VL2-2B ranks first among open-source models under 4B parameters on OpenCompass leaderboard.

Conclusion: SAIL-VL2 demonstrates strong multimodal capabilities from perception to complex reasoning, serving as an efficient and extensible foundation for the open-source multimodal community with innovative data, training, and architectural approaches.

Abstract: We introduce SAIL-VL2, an open-suite vision-language foundation model (LVM) for comprehensive multimodal understanding and reasoning. As the successor to SAIL-VL, SAIL-VL2 achieves state-of-the-art performance at the 2B and 8B parameter scales across diverse image and video benchmarks, demonstrating strong capabilities from fine-grained perception to complex reasoning. Three core innovations drive its effectiveness. First, a large-scale data curation pipeline with scoring and filtering strategies enhances both quality and distribution across captioning, OCR, QA, and video data, improving training efficiency. Second, a progressive training framework begins with a powerful pre-trained vision encoder (SAIL-ViT), advances through multimodal pre-training, and culminates in a thinking-fusion SFT-RL hybrid paradigm that systematically strengthens model capabilities. Third, architectural advances extend beyond dense LLMs to efficient sparse Mixture-of-Experts (MoE) designs. With these contributions, SAIL-VL2 demonstrates competitive performance across 106 datasets and achieves state-of-the-art results on challenging reasoning benchmarks such as MMMU and MathVista. Furthermore, on the OpenCompass leaderboard, SAIL-VL2-2B ranks first among officially released open-source models under the 4B parameter scale, while serving as an efficient and extensible foundation for the open-source multimodal community.

Method: Intermediate fusion of multi-modal embeddings (clinical, radiology, pathology data) combined with Cox Proportional Hazard regressors. Uses cross-modal interaction learning.

Result: Superior performance compared to late fusion methods. Achieved mean C-index of 0.861 (σ=0.112) on internal 5-fold nested cross-validation and 0.7103 on CHIMERA 2025 challenge validation data.

Conclusion: PROFUSEme demonstrates effective multi-modal fusion for early BCR prediction in prostate cancer, showing promise for clinical decision support and improved patient management.

Abstract: Almost 30% of prostate cancer (PCa) patients undergoing radical prostatectomy (RP) experience biochemical recurrence (BCR), characterized by increased prostate specific antigen (PSA) and associated with increased mortality. Accurate early prediction of BCR, at the time of RP, would contribute to prompt adaptive clinical decision-making and improved patient outcomes. In this work, we propose prostate cancer BCR prediction via fused multi-modal embeddings (PROFUSEme), which learns cross-modal interactions of clinical, radiology, and pathology data, following an intermediate fusion configuration in combination with Cox Proportional Hazard regressors. Quantitative evaluation of our proposed approach reveals superior performance, when compared with late fusion configurations, yielding a mean C-index of 0.861 ($\sigma=0.112$) on the internal 5-fold nested cross-validation framework, and a C-index of 0.7103 on the hold out data of CHIMERA 2025 challenge validation leaderboard.

Method: Built on Wan model with modified input paradigm using spatially-aligned skeleton signals for body motion and implicit facial features for expressions. Includes auxiliary Relighting LoRA module for environmental lighting consistency during character replacement.

Result: Achieves state-of-the-art performance in generating character videos with high controllability and expressiveness, demonstrating seamless environmental integration.

Conclusion: Wan-Animate provides an effective unified framework for character animation tasks with open-source commitment for model weights and source code.

Abstract: We introduce Wan-Animate, a unified framework for character animation and replacement. Given a character image and a reference video, Wan-Animate can animate the character by precisely replicating the expressions and movements of the character in the video to generate high-fidelity character videos. Alternatively, it can integrate the animated character into the reference video to replace the original character, replicating the scene’s lighting and color tone to achieve seamless environmental integration. Wan-Animate is built upon the Wan model. To adapt it for character animation tasks, we employ a modified input paradigm to differentiate between reference conditions and regions for generation. This design unifies multiple tasks into a common symbolic representation. We use spatially-aligned skeleton signals to replicate body motion and implicit facial features extracted from source images to reenact expressions, enabling the generation of character videos with high controllability and expressiveness. Furthermore, to enhance environmental integration during character replacement, we develop an auxiliary Relighting LoRA. This module preserves the character’s appearance consistency while applying the appropriate environmental lighting and color tone. Experimental results demonstrate that Wan-Animate achieves state-of-the-art performance. We are committed to open-sourcing the model weights and its source code.

Method: Proposes tri-branch architecture using vision-language models to extract global visual semantic information, with MOT-Adapter for task adaptation and Visual Semantic Fusion Module for improved feature fusion.

Result: Outperforms existing methods by 8-20% in low-quality video scenarios while maintaining robust performance in conventional scenarios.

Conclusion: The VSE-MOT framework effectively addresses MOT challenges in low-quality videos through visual semantic enhancement, demonstrating significant performance improvements in real-world applications.

Abstract: Current multi-object tracking (MOT) algorithms typically overlook issues inherent in low-quality videos, leading to significant degradation in tracking performance when confronted with real-world image deterioration. Therefore, advancing the application of MOT algorithms in real-world low-quality video scenarios represents a critical and meaningful endeavor. To address the challenges posed by low-quality scenarios, inspired by vision-language models, this paper proposes a Visual Semantic Enhancement-guided Multi-Object Tracking framework (VSE-MOT). Specifically, we first design a tri-branch architecture that leverages a vision-language model to extract global visual semantic information from images and fuse it with query vectors. Subsequently, to further enhance the utilization of visual semantic information, we introduce the Multi-Object Tracking Adapter (MOT-Adapter) and the Visual Semantic Fusion Module (VSFM). The MOT-Adapter adapts the extracted global visual semantic information to suit multi-object tracking tasks, while the VSFM improves the efficacy of feature fusion. Through extensive experiments, we validate the effectiveness and superiority of the proposed method in real-world low-quality video scenarios. Its tracking performance metrics outperform those of existing methods by approximately 8% to 20%, while maintaining robust performance in conventional scenarios.

Method: Proposes AD-DINOv3 framework with multimodal contrastive learning using DINOv3 as visual backbone and CLIP text encoder. Uses lightweight adapters to bridge domain gap and an Anomaly-Aware Calibration Module (AACM) to guide attention to anomalous regions rather than generic foreground semantics.

Result: Extensive experiments on eight industrial and medical benchmarks show AD-DINOv3 consistently matches or surpasses state-of-the-art methods, demonstrating superior performance as a general zero-shot anomaly detection framework.

Conclusion: AD-DINOv3 successfully adapts DINOv3 for zero-shot anomaly detection by addressing feature misalignment and semantic bias through multimodal adaptation and anomaly-aware calibration, proving to be an effective and scalable solution for diverse anomaly detection tasks.

Abstract: Zero-Shot Anomaly Detection (ZSAD) seeks to identify anomalies from arbitrary novel categories, offering a scalable and annotation-efficient solution. Traditionally, most ZSAD works have been based on the CLIP model, which performs anomaly detection by calculating the similarity between visual and text embeddings. Recently, vision foundation models such as DINOv3 have demonstrated strong transferable representation capabilities. In this work, we are the first to adapt DINOv3 for ZSAD. However, this adaptation presents two key challenges: (i) the domain bias between large-scale pretraining data and anomaly detection tasks leads to feature misalignment; and (ii) the inherent bias toward global semantics in pretrained representations often leads to subtle anomalies being misinterpreted as part of the normal foreground objects, rather than being distinguished as abnormal regions. To overcome these challenges, we introduce AD-DINOv3, a novel vision-language multimodal framework designed for ZSAD. Specifically, we formulate anomaly detection as a multimodal contrastive learning problem, where DINOv3 is employed as the visual backbone to extract patch tokens and a CLS token, and the CLIP text encoder provides embeddings for both normal and abnormal prompts. To bridge the domain gap, lightweight adapters are introduced in both modalities, enabling their representations to be recalibrated for the anomaly detection task. Beyond this baseline alignment, we further design an Anomaly-Aware Calibration Module (AACM), which explicitly guides the CLS token to attend to anomalous regions rather than generic foreground semantics, thereby enhancing discriminability. Extensive experiments on eight industrial and medical benchmarks demonstrate that AD-DINOv3 consistently matches or surpasses state-of-the-art methods, verifying its superiority as a general zero-shot anomaly detection framework.

Method: Integrates Soft mixture-of-experts (MoE) mechanism into Cross-Sensor Masked Autoencoder (CSMAE), creating CSMoE model with modality-specific expert specialization and cross-sensor representation learning. Uses thematic-climatic descriptor-driven sampling for diverse training set.

Result: CSMoE achieves more than twice the computational efficiency of existing RS FMs while maintaining competitive performance in scene classification, semantic segmentation, and image retrieval tasks.

Conclusion: The Soft MoE integration effectively creates computationally efficient remote sensing foundation models with superior trade-off between representational capacity, accuracy, and computational efficiency.

Abstract: Self-supervised learning through masked autoencoders has attracted great attention for remote sensing (RS) foundation model (FM) development, enabling improved representation learning across diverse sensors and downstream tasks. However, existing RS FMs often either suffer from substantial computational complexity during both training and inference or exhibit limited representational capacity. These issues restrict their practical applicability in RS. To address this limitation, we propose an adaptation for enhancing the efficiency of RS FMs by integrating the Soft mixture-of-experts (MoE) mechanism into the FM. The integration of Soft MoEs into the FM allows modality-specific expert specialization alongside shared cross-sensor representation learning. To demonstrate the effectiveness of our adaptation, we apply it on the Cross-Sensor Masked Autoencoder (CSMAE) model, resulting in the Cross-Sensor Mixture-of-Experts (CSMoE) model. In addition, we introduce a thematic-climatic descriptor-driven sampling strategy for the construction of a representative and diverse training set to train our CSMoE model. Extensive experiments on scene classification, semantic segmentation, and content-based image retrieval demonstrate that our adaptation yields a reduction in computational requirements while maintaining or improving representational performance. Compared to state-of-the-art RS FMs, CSMoE achieves a superior trade-off between representational capacity, accuracy, and computational efficiency. On average, CSMoE achieves more than twice the computational efficiency of existing RS FMs, while maintaining competitive performance across all experiments. These results show the effectiveness of the proposed adaptation for creating computationally efficient RS FMs. The code for the model, the training set creation, and the model weights will be available at https://git.tu-berlin.de/rsim/csmoe.

Method: Proposed a robust virtual staining framework featuring cascaded registration mechanisms to resolve spatial mismatches between generated outputs and their corresponding ground truth in unpaired or roughly paired datasets.

Result: Significantly outperforms state-of-the-art models across five datasets with average improvements of 3.2% on internal datasets and 10.1% on external datasets. Achieves remarkable 23.8% improvement in peak signal-to-noise ratio in datasets with substantial misalignment compared to baseline models.

Conclusion: The method demonstrates exceptional robustness across diverse datasets, simplifies data acquisition for virtual staining, and offers new insights for advancing virtual staining development in clinical applications.

Abstract: Accurate histopathological diagnosis often requires multiple differently stained tissue sections, a process that is time-consuming, labor-intensive, and environmentally taxing due to the use of multiple chemical stains. Recently, virtual staining has emerged as a promising alternative that is faster, tissue-conserving, and environmentally friendly. However, existing virtual staining methods face significant challenges in clinical applications, primarily due to their reliance on well-aligned paired data. Obtaining such data is inherently difficult because chemical staining processes can distort tissue structures, and a single tissue section cannot undergo multiple staining procedures without damage or loss of information. As a result, most available virtual staining datasets are either unpaired or roughly paired, making it difficult for existing methods to achieve accurate pixel-level supervision. To address this challenge, we propose a robust virtual staining framework featuring cascaded registration mechanisms to resolve spatial mismatches between generated outputs and their corresponding ground truth. Experimental results demonstrate that our method significantly outperforms state-of-the-art models across five datasets, achieving an average improvement of 3.2% on internal datasets and 10.1% on external datasets. Moreover, in datasets with substantial misalignment, our approach achieves a remarkable 23.8% improvement in peak signal-to-noise ratio compared to baseline models. The exceptional robustness of the proposed method across diverse datasets simplifies the data acquisition process for virtual staining and offers new insights for advancing its development.

Method: Comprehensive analysis evaluating multiple state-of-the-art detectors on benchmark datasets before and after applying various smoothing filters.

Result: Performance degradation observed, highlighting vulnerabilities introduced by facial enhancements.

Conclusion: There is a need for robust detection models resilient to facial alterations like beauty filters.

Abstract: Digital beautification through social media filters has become increasingly popular, raising concerns about the reliability of facial images and videos and the effectiveness of automated face analysis. This issue is particularly critical for digital manipulation detectors, systems aiming at distinguishing between genuine and manipulated data, especially in cases involving deepfakes and morphing attacks designed to deceive humans and automated facial recognition. This study examines whether beauty filters impact the performance of deepfake and morphing attack detectors. We perform a comprehensive analysis, evaluating multiple state-of-the-art detectors on benchmark datasets before and after applying various smoothing filters. Our findings reveal performance degradation, highlighting vulnerabilities introduced by facial enhancements and underscoring the need for robust detection models resilient to such alterations.

Method: Released two tailored datasets (Lens for 12 daily scenarios and AdsQA for advertisement videos), evaluated 40+ baselines including generalist MLLMs and task-specific models, and organized three competition tracks: VG-RS, VQA-SA, and VR-Ads.

Result: 76 teams from academic and industrial institutions registered, with 40+ valid submissions out of 1200+ entries included in ranking lists. The datasets, code sets, and rankings are publicly available.

Conclusion: The MARS2 2025 Challenge successfully established a large-scale benchmark for multimodal reasoning, providing resources and competition tracks that will help researchers follow and advance the state-of-the-art in this dynamic field.

Abstract: This paper reviews the MARS2 2025 Challenge on Multimodal Reasoning. We aim to bring together different approaches in multimodal machine learning and LLMs via a large benchmark. We hope it better allows researchers to follow the state-of-the-art in this very dynamic area. Meanwhile, a growing number of testbeds have boosted the evolution of general-purpose large language models. Thus, this year’s MARS2 focuses on real-world and specialized scenarios to broaden the multimodal reasoning applications of MLLMs. Our organizing team released two tailored datasets Lens and AdsQA as test sets, which support general reasoning in 12 daily scenarios and domain-specific reasoning in advertisement videos, respectively. We evaluated 40+ baselines that include both generalist MLLMs and task-specific models, and opened up three competition tracks, i.e., Visual Grounding in Real-world Scenarios (VG-RS), Visual Question Answering with Spatial Awareness (VQA-SA), and Visual Reasoning in Creative Advertisement Videos (VR-Ads). Finally, 76 teams from the renowned academic and industrial institutions have registered and 40+ valid submissions (out of 1200+) have been included in our ranking lists. Our datasets, code sets (40+ baselines and 15+ participants’ methods), and rankings are publicly available on the MARS2 workshop website and our GitHub organization page https://github.com/mars2workshop/, where our updates and announcements of upcoming events will be continuously provided.

Method: The authors introduce the Hierarchical Abstraction Image Dataset (HAID), which contains abstract images generated from normal raster images at multiple abstraction levels. They train and evaluate conventional vision systems on HAID across classification, segmentation, and object detection tasks.

Result: The study provides a comprehensive comparison between rasterized and abstract image representations across multiple vision tasks, though specific quantitative results are not detailed in the abstract.

Conclusion: The paper discusses whether abstract images can be considered an effective format for conveying visual semantic information and contributing to vision tasks, based on their hierarchical abstraction analysis.

Abstract: Imagine living in a world composed solely of primitive shapes, could you still recognise familiar objects? Recent studies have shown that abstract images-constructed by primitive shapes-can indeed convey visual semantic information to deep learning models. However, representations obtained from such images often fall short compared to those derived from traditional raster images. In this paper, we study the reasons behind this performance gap and investigate how much high-level semantic content can be captured at different abstraction levels. To this end, we introduce the Hierarchical Abstraction Image Dataset (HAID), a novel data collection that comprises abstract images generated from normal raster images at multiple levels of abstraction. We then train and evaluate conventional vision systems on HAID across various tasks including classification, segmentation, and object detection, providing a comprehensive study between rasterised and abstract image representations. We also discuss if the abstract image can be considered as a potentially effective format for conveying visual semantic information and contributing to vision tasks.

Method: Proposes BEVUDA++ with Reliable Depth Teacher (blends target LiDAR with depth predictions using uncertainty estimation) and Geometric Consistent Student (maps multi-space features to unified geometric embedding space), plus Uncertainty-guided Exponential Moving Average to reduce error accumulation.

Result: Achieves 12.9% NDS and 9.5% mAP enhancement on Day-Night adaptation, with state-of-the-art performance across four cross-domain scenarios in BEV 3D object detection tasks.

Conclusion: The geometric-aware teacher-student framework effectively addresses domain shift accumulation across multiple geometric spaces, demonstrating significant improvements in cross-domain BEV perception for autonomous driving.

Abstract: Vision-centric Bird’s Eye View (BEV) perception holds considerable promise for autonomous driving. Recent studies have prioritized efficiency or accuracy enhancements, yet the issue of domain shift has been overlooked, leading to substantial performance degradation upon transfer. We identify major domain gaps in real-world cross-domain scenarios and initiate the first effort to address the Domain Adaptation (DA) challenge in multi-view 3D object detection for BEV perception. Given the complexity of BEV perception approaches with their multiple components, domain shift accumulation across multi-geometric spaces (e.g., 2D, 3D Voxel, BEV) poses a significant challenge for BEV domain adaptation. In this paper, we introduce an innovative geometric-aware teacher-student framework, BEVUDA++, to diminish this issue, comprising a Reliable Depth Teacher (RDT) and a Geometric Consistent Student (GCS) model. Specifically, RDT effectively blends target LiDAR with dependable depth predictions to generate depth-aware information based on uncertainty estimation, enhancing the extraction of Voxel and BEV features that are essential for understanding the target domain. To collaboratively reduce the domain shift, GCS maps features from multiple spaces into a unified geometric embedding space, thereby narrowing the gap in data distribution between the two domains. Additionally, we introduce a novel Uncertainty-guided Exponential Moving Average (UEMA) to further reduce error accumulation due to domain shifts informed by previously obtained uncertainty guidance. To demonstrate the superiority of our proposed method, we execute comprehensive experiments in four cross-domain scenarios, securing state-of-the-art performance in BEV 3D object detection tasks, e.g., 12.9% NDS and 9.5% mAP enhancement on Day-Night adaptation.

Method: Created a comprehensive benchmark using GPT-4o to generate diverse, context-rich questions from visual descriptions, captions, scene titles, and summaries, with a two-stage filtering process (Context-Independence and Contextual Veracity) to ensure quality.

Result: Existing MLLMs perform poorly on Cineaste, with the top open-source model achieving only 63.15% accuracy, demonstrating that long-range temporal reasoning is a major bottleneck.

Conclusion: The benchmark reveals significant challenges in fine-grained contextual understanding and highlights the need for advancements in long-form movie comprehension capabilities of vision-language models.

Abstract: While recent advancements in vision-language models have improved video understanding, diagnosing their capacity for deep, narrative comprehension remains a challenge. Existing benchmarks often test short-clip recognition or use template-based questions, leaving a critical gap in evaluating fine-grained reasoning over long-form narrative content. To address these gaps, we introduce $\mathsf{Cin\acute{e}aste}$, a comprehensive benchmark for long-form movie understanding. Our dataset comprises 3,119 multiple-choice question-answer pairs derived from 1,805 scenes across 200 diverse movies, spanning five novel fine-grained contextual reasoning categories. We use GPT-4o to generate diverse, context-rich questions by integrating visual descriptions, captions, scene titles, and summaries, which require deep narrative understanding. To ensure high-quality evaluation, our pipeline incorporates a two-stage filtering process: Context-Independence filtering ensures questions require video context, while Contextual Veracity filtering validates factual consistency against the movie content, mitigating hallucinations. Experiments show that existing MLLMs struggle on $\mathsf{Cin\acute{e}aste}$; our analysis reveals that long-range temporal reasoning is a primary bottleneck, with the top open-source model achieving only 63.15% accuracy. This underscores significant challenges in fine-grained contextual understanding and the need for advancements in long-form movie comprehension.

Method: Created GenExam benchmark with 1,000 samples across 10 subjects using exam-style prompts organized under a four-level taxonomy. Each problem includes ground-truth images and fine-grained scoring points for precise evaluation of semantic correctness and visual plausibility.

Result: State-of-the-art models like GPT-Image-1 and Gemini-2.5-Flash-Image achieved less than 15% strict scores, with most models scoring almost 0%, demonstrating the significant challenge posed by this benchmark.

Conclusion: GenExam provides a rigorous assessment framework for evaluating AI models’ integrated knowledge, reasoning, and generation capabilities, offering valuable insights for progress toward general AGI through exam-style image generation tasks.

Abstract: Exams are a fundamental test of expert-level intelligence and require integrated understanding, reasoning, and generation. Existing exam-style benchmarks mainly focus on understanding and reasoning tasks, and current generation benchmarks emphasize the illustration of world knowledge and visual concepts, neglecting the evaluation of rigorous drawing exams. We introduce GenExam, the first benchmark for multidisciplinary text-to-image exams, featuring 1,000 samples across 10 subjects with exam-style prompts organized under a four-level taxonomy. Each problem is equipped with ground-truth images and fine-grained scoring points to enable a precise evaluation of semantic correctness and visual plausibility. Experiments show that even state-of-the-art models such as GPT-Image-1 and Gemini-2.5-Flash-Image achieve less than 15% strict scores, and most models yield almost 0%, suggesting the great challenge of our benchmark. By framing image generation as an exam, GenExam offers a rigorous assessment of models’ ability to integrate knowledge, reasoning, and generation, providing insights on the path to general AGI.

Method: Developed a framework with curated datasets, training recipes, and model architectures. Created 4B and 14B parameter models including pre-trained base models and instruction fine-tuned versions. Evaluated on single and multi-image benchmarks.

Result: Models demonstrate competitive performance among open-source LMMs with similar model sizes, with ability to comprehend interleaved image-text inputs.

Conclusion: The framework, models, datasets, and training code will be open-sourced to better support multimodal AI research community.

Abstract: This paper introduces BLIP-3, an open framework for developing Large Multimodal Models (LMMs). The framework comprises meticulously curated datasets, a training recipe, model architectures, and a resulting suite of LMMs. We release 4B and 14B models, including both the pre-trained base model and the instruction fine-tuned ones. Our models undergo rigorous evaluation across a range of tasks, including both single and multi-image benchmarks. Our models demonstrate competitive performance among open-source LMMs with similar model sizes. Our resulting LMMs demonstrate competitive performance among open-source LMMs with similar model sizes, with the ability to comprehend interleaved image-text inputs. Our training code, models, and all datasets used in this work, including the three largescale datasets we create and the preprocessed ones, will be open-sourced to better support the research community.

Method: Uses Grounded-SAM for precise region localization, integrates multimodal inputs (text, masks, pose, texture), employs decoupled cross-attention for texture integration, and uses auxiliary U-Net for high-frequency detail preservation. Extended VITON-HD dataset with MLLM-generated samples.

Result: Extensive experiments show DPDEdit outperforms state-of-the-art methods in image fidelity and coherence with multimodal inputs.

Conclusion: DPDEdit effectively addresses fashion image editing challenges by precisely locating editing regions and preserving garment texture details through innovative multimodal integration and texture preservation mechanisms.

Abstract: Fashion image editing is a crucial tool for designers to convey their creative ideas by visualizing design concepts interactively. Current fashion image editing techniques, though advanced with multimodal prompts and powerful diffusion models, often struggle to accurately identify editing regions and preserve the desired garment texture detail. To address these challenges, we introduce a new multimodal fashion image editing architecture based on latent diffusion models, called Detail-Preserved Diffusion Models (DPDEdit). DPDEdit guides the fashion image generation of diffusion models by integrating text prompts, region masks, human pose images, and garment texture images. To precisely locate the editing region, we first introduce Grounded-SAM to predict the editing region based on the user’s textual description, and then combine it with other conditions to perform local editing. To transfer the detail of the given garment texture into the target fashion image, we propose a texture injection and refinement mechanism. Specifically, this mechanism employs a decoupled cross-attention layer to integrate textual descriptions and texture images, and incorporates an auxiliary U-Net to preserve the high-frequency details of generated garment texture. Additionally, we extend the VITON-HD dataset using a multimodal large language model to generate paired samples with texture images and textual descriptions. Extensive experiments show that our DPDEdit outperforms state-of-the-art methods in terms of image fidelity and coherence with the given multimodal inputs.

Method: Proposes Texture-Aware SuperPixel (TASP) method that automatically adjusts spatial constraint according to local feature variance and uses a new pixel to superpixel patch-based distance to ensure texture homogeneity.

Result: TASP outperforms state-of-the-art methods in segmentation accuracy on both texture datasets and natural color image datasets.

Conclusion: The proposed TASP method effectively addresses texture segmentation challenges by adaptive spatial constraints and patch-based distances, achieving superior performance compared to existing approaches.

Abstract: Most superpixel algorithms compute a trade-off between spatial and color features at the pixel level. Hence, they may need fine parameter tuning to balance the two measures, and highly fail to group pixels with similar local texture properties. In this paper, we address these issues with a new Texture-Aware SuperPixel (TASP) method. To accurately segment textured and smooth areas, TASP automatically adjusts its spatial constraint according to the local feature variance. Then, to ensure texture homogeneity within superpixels, a new pixel to superpixel patch-based distance is proposed. TASP outperforms the segmentation accuracy of the state-of-the-art methods on texture and also natural color image datasets.

Method: Uses superpixels to extract reduced color set, applies fast approximate nearest neighbor matching with diversity constraints, and employs a fusion framework for color transfer.

Result: Method demonstrates improvement over exact matching and shows competitive visual performance compared to state-of-the-art approaches.

Conclusion: SCT provides an effective and efficient solution for color transfer that balances computational speed with visual quality through superpixel-based processing.

Abstract: In this work, we propose a fast superpixel-based color transfer method (SCT) between two images. Superpixels enable to decrease the image dimension and to extract a reduced set of color candidates. We propose to use a fast approximate nearest neighbor matching algorithm in which we enforce the match diversity by limiting the selection of the same superpixels. A fusion framework is designed to transfer the matched colors, and we demonstrate the improvement obtained over exact matching results. Finally, we show that SCT is visually competitive compared to state-of-the-art methods.

Method: Proposed a new metric that evaluates shape regularity through three aspects: convexity, balanced repartition of pixels, and contour smoothness.

Result: The proposed measure is robust to scale and noise variations, and enables more relevant comparison of superpixel methods than circularity-based approaches.

Conclusion: The new metric provides a more comprehensive and accurate way to evaluate superpixel regularity, addressing the limitations of traditional circularity measures for superpixel-based applications.

Abstract: Regular decompositions are necessary for most superpixel-based object recognition or tracking applications. So far in the literature, the regularity or compactness of a superpixel shape is mainly measured by its circularity. In this work, we first demonstrate that such measure is not adapted for superpixel evaluation, since it does not directly express regularity but circular appearance. Then, we propose a new metric that considers several shape regularity aspects: convexity, balanced repartition, and contour smoothness. Finally, we demonstrate that our measure is robust to scale and noise and enables to more relevantly compare superpixel methods.

Method: Developed a semi-automatic waste detection pipeline in collaboration with environmental protection agency. Evaluated network architecture, image resolution, geographic span, and pretraining procedures through extensive experiments.

Result: Best model achieved 92.02% F1-Score and 94.56% Accuracy. Generalization study showed only 5.1% average F1-Score decrease on different territories. Human analysis time reduced by up to 30% with computer-aided tool.

Conclusion: The developed waste detection pipeline effectively identifies illegal dumping sites with high accuracy and significantly reduces manual analysis effort, demonstrating practical value for environmental protection agencies.

Abstract: Improper solid waste management represents both a serious threat to ecosystem health and a significant source of revenues for criminal organizations perpetrating environmental crimes. This issue can be mitigated thanks to the increasing availability of Very-High-Resolution Remote Sensing (VHR RS) images. Modern image-analysis tools support automated photo-interpretation and large territory scanning in search of illegal waste disposal sites. This paper illustrates a semi-automatic waste detection pipeline, developed in collaboration with a regional environmental protection agency, for detecting candidate illegal dumping sites in VHR RS images. To optimize the effectiveness of the waste detector at the core of the pipeline, extensive experiments evaluate such design choices as the network architecture, the ground resolution and geographic span of the input images, as well as the pretraining procedures. The best model attains remarkable performance, achieving 92.02 % F1-Score and 94.56 % Accuracy. A generalization study assesses the performance variation when the detector processes images from various territories substantially different from the one used during training, incurring only a moderate performance loss, namely an average 5.1 % decrease in the F1-Score. Finally, an exercise in which expert photo-interpreters compare the effort required to scan large territories with and without support from the waste detector assesses the practical benefit of introducing a computer-aided image analysis tool in a professional environmental protection agency. Results show that a reduction of up to 30 % of the time spent for waste site detection can be attained.

Method: Proposes SCALP (Superpixels with Contour Adherence using Linear Path) - an iterative clustering framework where distance computation is enhanced by considering the linear path to superpixel barycenters, improving contour adherence.

Result: SCALP produces regular and compact superpixels that adhere well to image contours. Evaluation on Berkeley Segmentation Dataset shows it outperforms state-of-the-art methods in both standard superpixel metrics and contour detection metrics.

Conclusion: The linear path approach in clustering effectively enhances superpixel decomposition, achieving superior performance in contour adherence while maintaining computational efficiency and decomposition quality.

Abstract: Superpixel decomposition methods are generally used as a pre-processing step to speed up image processing tasks. They group the pixels of an image into homogeneous regions while trying to respect existing contours. For all state-of-the-art superpixel decomposition methods, a trade-off is made between

1. computational time, 2) adherence to image contours and 3) regularity and compactness of the decomposition. In this paper, we propose a fast method to compute Superpixels with Contour Adherence using Linear Path (SCALP) in an iterative clustering framework. The distance computed when trying to associate a pixel to a superpixel during the clustering is enhanced by considering the linear path to the superpixel barycenter. The proposed framework produces regular and compact superpixels that adhere to the image contours. We provide a detailed evaluation of SCALP on the standard Berkeley Segmentation Dataset. The obtained results outperform state-of-the-art methods in terms of standard superpixel and contour detection metrics.

Method: Proposes a Transformer-based architecture that unifies all multi-view planar reconstruction and pose estimation tasks within a single-stage framework using query-based learning to enable reasoning among semantic entities.

Result: Achieves state-of-the-art performance on ScanNetv1, ScanNetv2, NYUv2-Plane, and MatterPort3D datasets, demonstrating mutual benefits across sub-tasks through unified learning.

Conclusion: The unified single-stage framework successfully integrates all related sub-tasks of 3D planar reconstruction and pose estimation, eliminating the need for external plane correspondence labeling and initial pose estimation while achieving superior performance.

Abstract: The challenging task of 3D planar reconstruction from images involves several sub-tasks including frame-wise plane detection, segmentation, parameter regression and possibly depth prediction, along with cross-frame plane correspondence and relative camera pose estimation. Previous works adopt a divide and conquer strategy, addressing above sub-tasks with distinct network modules in a two-stage paradigm. Specifically, given an initial camera pose and per-frame plane predictions from the first stage, further exclusively designed modules relying on external plane correspondence labeling are applied to merge multi-view plane entities and produce refined camera pose. Notably, existing work fails to integrate these closely related sub-tasks into a unified framework, and instead addresses them separately and sequentially, which we identify as a primary source of performance limitations. Motivated by this finding and the success of query-based learning in enriching reasoning among semantic entities, in this paper, we propose PlaneRecTR++, a Transformer-based architecture, which for the first time unifies all tasks of multi-view planar reconstruction and pose estimation within a compact single-stage framework, eliminating the need for the initial pose estimation and supervision of plane correspondence. Extensive quantitative and qualitative experiments demonstrate that our proposed unified learning achieves mutual benefits across sub-tasks, achieving a new state-of-the-art performance on the public ScanNetv1, ScanNetv2, NYUv2-Plane, and MatterPort3D datasets. Codes are available at https://github.com/SJingjia/PlaneRecTR-PP.

Method: Derives an OOD prototype from synthetic samples and computes class prototypes from ID training data, then analyzes gradients of a nearest-class-prototype loss function with respect to the artificial OOD prototype.

Result: Experimental evaluations show enhanced distinction between ID and OOD data, surpassing established baselines in robustness, particularly on ImageNet-1k.

Conclusion: Gradient-based methods and prototype-driven approaches show significant potential for advancing OOD detection in deep neural networks.

Abstract: Out-of-distribution (OOD) detection is crucial for ensuring the reliability of deep learning models in real-world applications. Existing methods typically focus on feature representations or output-space analysis, often assuming a distribution over these spaces or leveraging gradient norms with respect to model parameters. However, these approaches struggle to distinguish near-OOD samples and often require extensive hyper-parameter tuning, limiting their practicality.In this work, we propose GRadient-aware Out-Of-Distribution detection (GROOD), a method that derives an OOD prototype from synthetic samples and computes class prototypes directly from In-distribution (ID) training data. By analyzing the gradients of a nearest-class-prototype loss function concerning an artificial OOD prototype, our approach achieves a clear separation between in-distribution and OOD samples. Experimental evaluations demonstrate that gradients computed from the OOD prototype enhance the distinction between ID and OOD data, surpassing established baselines in robustness, particularly on ImageNet-1k. These findings highlight the potential of gradient-based methods and prototype-driven approaches in advancing OOD detection within deep neural networks.

Method: Propose using entropy (approximated by bits-per-pixel from compressed images) as an intrinsic image score. Combine with graph-based method to increase spatial diversity and mitigate bias from selecting only low-entropy iconic images.

Result: This simple entropy-based score yields good results, particularly for semantic segmentation tasks, while being computationally efficient and requiring no supervision.

Conclusion: Compression-based entropy provides an effective, low-cost measure for coreset selection when combined with spatial diversity techniques to prevent biased learning in deep models.

Abstract: In this paper we propose a score of an image to use for coreset selection in image classification and semantic segmentation tasks. The score is the entropy of an image as approximated by the bits-per-pixel of its compressed version. Thus the score is intrinsic to an image and does not require supervision or training. It is very simple to compute and readily available as all images are stored in a compressed format. The motivation behind our choice of score is that most other scores proposed in literature are expensive to compute. More importantly, we want a score that captures the perceptual complexity of an image. Entropy is one such measure, images with clutter tend to have a higher entropy. However sampling only low entropy iconic images, for example, leads to biased learning and an overall decrease in test performance with current deep learning models. To mitigate the bias we use a graph based method that increases the spatial diversity of the selected samples. We show that this simple score yields good results, particularly for semantic segmentation tasks.

Method: Proposes Structured Preference Optimization (SPO) with: 1) Preference-Based Scoring and Optimization that evaluates reasoning chains on task relevance, visual grounding, and historical consistency; 2) Curriculum-Guided Training that progressively adapts from simple to complex tasks.

Result: SPO achieves +5.98% GCR and +4.68% SR improvement in VirtualHome, and +3.30% GCR and +2.11% SR improvement in Habitat over best-performing baselines. Introduces ExtendaBench with 1,509 tasks across VirtualHome and Habitat 2.0.

Conclusion: SPO significantly improves reasoning quality and decision accuracy in long-horizon tasks, demonstrating the effectiveness of preference-driven optimization in vision-language task planning.

Abstract: Existing methods for vision-language task planning excel in short-horizon tasks but often fall short in complex, long-horizon planning within dynamic environments. These challenges primarily arise from the difficulty of effectively training models to produce high-quality reasoning processes for long-horizon tasks. To address this, we propose Structured Preference Optimization (SPO), which aims to enhance reasoning and action selection in long-horizon task planning through structured preference evaluation and optimized training strategies. Specifically, SPO introduces: 1) Preference-Based Scoring and Optimization, which systematically evaluates reasoning chains based on task relevance, visual grounding, and historical consistency; and 2) Curriculum-Guided Training, where the model progressively adapts from simple to complex tasks, improving its generalization ability in long-horizon scenarios and enhancing reasoning robustness. To advance research in vision-language long-horizon task planning, we introduce ExtendaBench, a comprehensive benchmark covering 1,509 tasks across VirtualHome and Habitat 2.0, categorized into ultra-short, short, medium, and long tasks. Experimental results demonstrate that SPO significantly improves reasoning quality and final decision accuracy, outperforming prior methods on long-horizon tasks and underscoring the effectiveness of preference-driven optimization in vision-language task planning. Specifically, SPO achieves a +5.98% GCR and +4.68% SR improvement in VirtualHome and a +3.30% GCR and +2.11% SR improvement in Habitat over the best-performing baselines.

Method: DiffGAN combines Generative Adversarial Network (GAN) with Non-dominated Sorting Genetic Algorithm II to generate diverse triggering inputs. It uses two custom fitness functions (diversity and divergence) to explore GAN input space and identify output discrepancies between models.

Result: DiffGAN significantly outperforms state-of-the-art baseline, generating 4x more triggering inputs with greater diversity and validity within the same budget. The generated inputs also improve accuracy of ML-based model selection mechanisms.

Conclusion: DiffGAN provides an effective black-box approach for differential testing of DNN models, generating high-quality triggering inputs that reveal behavioral discrepancies and enabling better model selection and combination strategies.

Abstract: Deep Neural Networks (DNNs) are increasingly deployed across applications. However, ensuring their reliability remains a challenge, and in many situations, alternative models with similar functionality and accuracy are available. Traditional accuracy-based evaluations often fail to capture behavioral differences between models, especially with limited test datasets, making it difficult to select or combine models effectively. Differential testing addresses this by generating test inputs that expose discrepancies in DNN model behavior. However, existing approaches face significant limitations: many rely on model internals or are constrained by available seed inputs. To address these challenges, we propose DiffGAN, a black-box test image generation approach for differential testing of DNN models. DiffGAN leverages a Generative Adversarial Network (GAN) and the Non-dominated Sorting Genetic Algorithm II to generate diverse and valid triggering inputs that reveal behavioral discrepancies between models. DiffGAN employs two custom fitness functions, focusing on diversity and divergence, to guide the exploration of the GAN input space and identify discrepancies between models’ outputs. By strategically searching this space, DiffGAN generates inputs with specific features that trigger differences in model behavior. DiffGAN is black-box, making it applicable in more situations. We evaluate DiffGAN on eight DNN model pairs trained on widely used image datasets. Our results show DiffGAN significantly outperforms a SOTA baseline, generating four times more triggering inputs, with greater diversity and validity, within the same budget. Additionally, the generated inputs improve the accuracy of a machine learning-based model selection mechanism, which selects the best-performing model based on input characteristics and can serve as a smart output voting mechanism when using alternative models.

Method: Non-rigidly registering all dMRI images to a standard template to mitigate macrostructural information, then training brain age models on 13,398 participants across 12 datasets. Compared with T1w MRI-based models that primarily use macrostructural information.

Result: dMRI-based brain age differs from T1w MRI-based brain age across neurodegeneration stages - older in participants transitioning from cognitively normal to mild cognitive impairment, but younger in Alzheimer’s disease patients. dMRI-based brain age may predict CN to MCI transition up to 5 years before diagnosis.

Conclusion: dMRI-based brain age provides distinct microstructural information compared to T1w MRI-based approaches and shows potential as an earlier biomarker for predicting neurodegenerative disease progression, particularly in the pre-clinical stages.

Abstract: Estimated brain age from magnetic resonance image (MRI) and its deviation from chronological age can provide early insights into potential neurodegenerative diseases, supporting early detection and implementation of prevention strategies. Diffusion MRI (dMRI) presents an opportunity to build an earlier biomarker for neurodegenerative disease prediction because it captures subtle microstructural changes that precede more perceptible macrostructural changes. However, the coexistence of macro- and micro-structural information in dMRI raises the question of whether current dMRI-based brain age estimation models are leveraging the intended microstructural information or if they inadvertently rely on the macrostructural information. To develop a microstructure-specific brain age, we propose a method for brain age identification from dMRI that mitigates the model’s use of macrostructural information by non-rigidly registering all images to a standard template. Imaging data from 13,398 participants across 12 datasets were used for the training and evaluation. We compare our brain age models, trained with and without macrostructural information mitigated, with an architecturally similar T1-weighted (T1w) MRI-based brain age model and two recent, popular, openly available T1w MRI-based brain age models that primarily use macrostructural information. We observe difference between our dMRI-based brain age and T1w MRI-based brain age across stages of neurodegeneration, with dMRI-based brain age being older than T1w MRI-based brain age in participants transitioning from cognitively normal (CN) to mild cognitive impairment (MCI), but younger in participants already diagnosed with Alzheimer’s disease (AD). Furthermore, dMRI-based brain age may offer advantages over T1w MRI-based brain age in predicting the transition from CN to MCI up to five years before diagnosis.

Method: Systematically unifies heterogeneous stereo sources: curated labeled datasets covering diverse environments and large-scale synthetic stereo pairs generated from unlabeled monocular images using a mixed-data strategy.

Result: Achieves state-of-the-art zero-shot generalization on four public benchmarks, demonstrating robust performance across diverse domains and mitigating dataset bias effectively.

Conclusion: The framework paves the way towards universal stereo matching by offering a scalable data paradigm that can be applied to any stereo image pair, significantly improving generalization without specialized architectures.

Abstract: Stereo matching serves as a cornerstone in 3D vision, aiming to establish pixel-wise correspondences between stereo image pairs for depth recovery. Despite remarkable progress driven by deep neural architectures, current models often exhibit severe performance degradation when deployed in unseen domains, primarily due to the limited diversity of training data. In this work, we introduce StereoAnything, a data-centric framework that substantially enhances the zero-shot generalization capability of existing stereo models. Rather than devising yet another specialized architecture, we scale stereo training to an unprecedented level by systematically unifying heterogeneous stereo sources: (1) curated labeled datasets covering diverse environments, and (2) large-scale synthetic stereo pairs generated from unlabeled monocular images. Our mixed-data strategy delivers consistent and robust learning signals across domains, effectively mitigating dataset bias. Extensive zero-shot evaluations on four public benchmarks demonstrate that Stereo Anything achieves state-of-the-art generalization. This work paves the way towards truly universal stereo matching, offering a scalable data paradigm applicable to any stereo image pair. We extensively evaluate the zero-shot capabilities of our model on four public datasets, showcasing its impressive ability to generalize to any stereo image pair. Code is available at https://github.com/XiandaGuo/OpenStereo.

Method: UniPLV uses images as a bridge to co-embed 3D points with pre-aligned images and text in shared feature space. It employs logit/feature distillation modules, vision-point matching, four task-specific losses, and a two-stage training strategy.

Result: UniPLV achieves average improvements of 15.6% and 14.8% in semantic segmentation for Base-Annotated and Annotation-Free tasks respectively, significantly outperforming state-of-the-art methods.

Conclusion: The framework effectively addresses open-world 3D scene understanding challenges, eliminates manual annotation needs, and demonstrates superior performance, pushing the boundaries of multimodal 3D analysis.

Abstract: Open-world 3D scene understanding is a critical challenge that involves recognizing and distinguishing diverse objects and categories from 3D data, such as point clouds, without relying on manual annotations. Traditional methods struggle with this open-world task, especially due to the limitations of constructing extensive point cloud-text pairs and handling multimodal data effectively. In response to these challenges, we present UniPLV, a robust framework that unifies point clouds, images, and text within a single learning paradigm for comprehensive 3D scene understanding. UniPLV leverages images as a bridge to co-embed 3D points with pre-aligned images and text in a shared feature space, eliminating the need for labor-intensive point cloud-text pair crafting. Our framework achieves precise multimodal alignment through two innovative strategies: (i) Logit and feature distillation modules between images and point clouds to enhance feature coherence; (ii) A vision-point matching module that implicitly corrects 3D semantic predictions affected by projection inaccuracies from points to pixels. To further boost performance, we implement four task-specific losses alongside a two-stage training strategy. Extensive experiments demonstrate that UniPLV significantly surpasses state-of-the-art methods, with average improvements of 15.6% and 14.8% in semantic segmentation for Base-Annotated and Annotation-Free tasks, respectively. These results underscore UniPLV’s efficacy in pushing the boundaries of open-world 3D scene understanding. We will release the code to support future research and development.

Method: Three-phase framework: 1) Agent exploration to collect noisy image-caption pairs, 2) Consensus-based pseudo-caption distillation using LLM, 3) Fine-tuning captioning model with contrastive learning.

Result: Policy trained to mine high-disagreement samples outperforms baselines. Method achieves higher semantic similarity and significant improvements in caption accuracy and consistency.

Conclusion: The proposed self-supervised approach effectively enhances captioning performance across varying viewpoints through consensus mechanisms and targeted fine-tuning.

Abstract: We present a self-supervised method to improve an agent’s abilities in describing arbitrary objects while actively exploring a generic environment. This is a challenging problem, as current models struggle to obtain coherent image captions due to different camera viewpoints and clutter. We propose a three-phase framework to fine-tune existing captioning models that enhances caption accuracy and consistency across views via a consensus mechanism. First, an agent explores the environment, collecting noisy image-caption pairs. Then, a consistent pseudo-caption for each object instance is distilled via consensus using a large language model. Finally, these pseudo-captions are used to fine-tune an off-the-shelf captioning model, with the addition of contrastive learning. We analyse the performance of the combination of captioning models, exploration policies, pseudo-labeling methods, and fine-tuning strategies, on our manually labeled test set. Results show that a policy can be trained to mine samples with higher disagreement compared to classical baselines. Our pseudo-captioning method, in combination with all policies, has a higher semantic similarity compared to other existing methods, and fine-tuning improves caption accuracy and consistency by a significant margin. Code and test set annotations available at https://hsp-iit.github.io/embodied-captioning/

Method: Self-supervised deep learning framework for Ghost Imaging reconstruction that eliminates the need for clean reference data while providing strong noise reduction capabilities.

Result: Unparalleled reconstruction performance for noisy acquisitions among unsupervised methods, as demonstrated through theoretical analysis and real data use cases.

Conclusion: This method provides essential tools for low-light Ghost Imaging scenarios, making it suitable for cutting-edge applications including in-vivo and in-operando studies of biological samples and batteries.

Abstract: We present a new self-supervised deep-learning-based Ghost Imaging (GI) reconstruction method, which provides unparalleled reconstruction performance for noisy acquisitions among unsupervised methods. We present the supporting mathematical framework and results from theoretical and real data use cases. Self-supervision removes the need for clean reference data while offering strong noise reduction. This provides the necessary tools for addressing signal-to-noise ratio concerns for GI acquisitions in emerging and cutting-edge low-light GI scenarios. Notable examples include micro- and nano-scale x-ray emission imaging, e.g., x-ray fluorescence imaging of dose-sensitive samples. Their applications include in-vivo and in-operando case studies for biological samples and batteries.

Method: Two-step process: 1) Localization - identify contextual region relevant to input query using efficient model, 2) Pruning - two strategies: Pre-LLM Compression (adaptive region compression) and Inner-LLM Pruning (training-free token pruning within early LLM layers).

Result: Extensive experiments show CROP significantly outperforms existing visual token pruning methods and achieves state-of-the-art performance across various VQA tasks.

Conclusion: CROP effectively addresses visual token redundancy in VQA through contextual region localization and adaptive pruning strategies, reducing computational overhead while maintaining performance.

Abstract: Current VLM-based VQA methods often process entire images, leading to excessive visual tokens that include redundant information irrelevant to the posed question. This abundance of unnecessary image details creates numerous visual tokens, drastically increasing memory and computational requirements in VLMs. To address this, we propose Contextual Region-Oriented Visual Token Pruning (CROP), a novel framework to compress visual tokens through a two-step process: Localization and Pruning. Specifically, CROP first employs an efficient model to identify the contextual region relevant to the input query. Subsequently, two distinct strategies are introduced for pruning: (1) Pre-LLM Compression (PLC), which adaptively compresses different image regions with varying ratios, and (2) Inner-LLM Pruning (ILP), a training-free method that prunes tokens within early LLM layers guided by the identified contextual region. Extensive experiments on a wide range of VQA tasks demonstrate that CROP significantly outperforms existing visual token pruning methods and achieves state-of-the-art performance.

Method: Proposes a semantic prompt optimization mechanism that decouples prompts into spatial and temporal components, combined with a stage-wise decoupled generation paradigm where spatial prompts guide text-to-image generation and temporal prompts direct image-to-video conversion.

Result: Experimental results show excellent spatiotemporal consistency with outstanding performance in identity preservation, text relevance, and video quality, achieving runner-up position in 2025 ACM MultiMedia Challenge.

Conclusion: The spatial-temporal decoupled framework provides a simple yet effective solution for high-fidelity identity-preserving video generation by separating spatial and temporal feature representations.

Abstract: Identity-preserving text-to-video (IPT2V) generation, which aims to create high-fidelity videos with consistent human identity, has become crucial for downstream applications. However, current end-to-end frameworks suffer a critical spatial-temporal trade-off: optimizing for spatially coherent layouts of key elements (e.g., character identity preservation) often compromises instruction-compliant temporal smoothness, while prioritizing dynamic realism risks disrupting the spatial coherence of visual structures. To tackle this issue, we propose a simple yet effective spatial-temporal decoupled framework that decomposes representations into spatial features for layouts and temporal features for motion dynamics. Specifically, our paper proposes a semantic prompt optimization mechanism and stage-wise decoupled generation paradigm. The former module decouples the prompt into spatial and temporal components. Aligned with the subsequent stage-wise decoupled approach, the spatial prompts guide the text-to-image (T2I) stage to generate coherent spatial features, while the temporal prompts direct the sequential image-to-video (I2V) stage to ensure motion consistency. Experimental results validate that our approach achieves excellent spatiotemporal consistency, demonstrating outstanding performance in identity preservation, text relevance, and video quality. By leveraging this simple yet robust mechanism, our algorithm secures the runner-up position in 2025 ACM MultiMedia Challenge.

Method: OPAL uses a mixed-integer linear programming (MILP) formulation to minimize labeling effort subject to specified accuracy targets, applied to automatic test input labeling and validation for vision systems.

Result: OPAL achieves 98.8% average accuracy while cutting manual labeling by more than half, outperforms 8 baseline methods across 7 datasets, and reduces manual effort by 28.8% with 4.5% higher accuracy than SOTA validation methods.

Conclusion: OPAL effectively addresses the need for high-accuracy test datasets with minimal human effort, and can be further enhanced with active learning for additional efficiency gains without compromising accuracy.

Abstract: Software systems increasingly include AI components based on deep learning (DL). Reliable testing of such systems requires near-perfect test-input validity and label accuracy, with minimal human effort. Yet, the DL community has largely overlooked the need to build highly accurate datasets with minimal effort, since DL training is generally tolerant of labelling errors. This challenge, instead, reflects concerns more familiar to software engineering, where a central goal is to construct high-accuracy test inputs, with accuracy as close to 100% as possible, while keeping associated costs in check. In this article we introduce OPAL, a human-assisted labelling method that can be configured to target a desired accuracy level while minimizing the manual effort required for labelling. The main contribution of OPAL is a mixed-integer linear programming (MILP) formulation that minimizes labelling effort subject to a specified accuracy target. To evaluate OPAL we instantiate it for two tasks in the context of testing vision systems: automatic labelling of test inputs and automated validation of test inputs. Our evaluation, based on more than 2500 experiments performed on seven datasets, comparing OPAL with eight baseline methods, shows that OPAL, relying on its MILP formulation, achieves an average accuracy of 98.8%, while cutting manual labelling by more than half. OPAL significantly outperforms automated labelling baselines in labelling accuracy across all seven datasets, when all methods are provided with the same manual-labelling budget. For automated test-input validation, on average, OPAL reduces manual effort by 28.8% while achieving 4.5% higher accuracy than the SOTA test-input validation baselines. Finally, we show that augmenting OPAL with an active-learning loop leads to an additional 4.5% reduction in required manual labelling, without compromising accuracy.

Method: Proposes a unified model with two-level reasoning: Macro-Level CoT for high-level task planning and Micro-Level CoT for subtask execution, using interleaved image-text supervision and multi-task objectives for training.

Result: Achieves state-of-the-art performance on reasoning-driven image generation (WISE) and editing benchmarks (RISE and KRIS) with strong generalization, using only 8 A100 GPUs efficiently.

Conclusion: Uni-CoT demonstrates a promising solution for scalable and coherent multimodal reasoning by unifying visual understanding and generation within a single efficient framework.

Abstract: Chain-of-Thought (CoT) reasoning has been widely adopted to enhance Large Language Models (LLMs) by decomposing complex tasks into simpler, sequential subtasks. However, extending CoT to vision-language reasoning tasks remains challenging, as it often requires interpreting transitions of visual states to support reasoning. Existing methods often struggle with this due to limited capacity of modeling visual state transitions or incoherent visual trajectories caused by fragmented architectures. To overcome these limitations, we propose Uni-CoT, a Unified Chain-of-Thought framework that enables coherent and grounded multimodal reasoning within a single unified model. The key idea is to leverage a model capable of both image understanding and generation to reason over visual content and model evolving visual states. However, empowering a unified model to achieve that is non-trivial, given the high computational cost and the burden of training. To address this, Uni-CoT introduces a novel two-level reasoning paradigm: A Macro-Level CoT for high-level task planning and A Micro-Level CoT for subtask execution. This design significantly reduces the computational overhead. Furthermore, we introduce a structured training paradigm that combines interleaved image-text supervision for macro-level CoT with multi-task objectives for micro-level CoT. Together, these innovations allow Uni-CoT to perform scalable and coherent multi-modal reasoning. Furthermore, thanks to our design, all experiments can be efficiently completed using only 8 A100 GPUs with 80GB VRAM each. Experimental results on reasoning-driven image generation benchmark (WISE) and editing benchmarks (RISE and KRIS) indicates that Uni-CoT demonstrates SOTA performance and strong generalization, establishing Uni-CoT as a promising solution for multi-modal reasoning. Project Page and Code: https://sais-fuxi.github.io/projects/uni-cot/

Method: Proposes Diffusion Bridge Distillation for Purification (DBLP) with noise bridge distillation to align adversarial and clean distributions in latent consistency models, plus adaptive semantic enhancement using multi-scale pyramid edge maps as conditioning.

Result: Achieves state-of-the-art robust accuracy, superior image quality, and around 0.2 seconds inference time across multiple datasets.

Conclusion: DBLP represents a significant advancement toward real-time adversarial purification by addressing the efficiency limitations of previous diffusion-based methods.

Abstract: Recent advances in deep neural networks (DNNs) have led to remarkable success across a wide range of tasks. However, their susceptibility to adversarial perturbations remains a critical vulnerability. Existing diffusion-based adversarial purification methods often require intensive iterative denoising, severely limiting their practical deployment. In this paper, we propose Diffusion Bridge Distillation for Purification (DBLP), a novel and efficient diffusion-based framework for adversarial purification. Central to our approach is a new objective, noise bridge distillation, which constructs a principled alignment between the adversarial noise distribution and the clean data distribution within a latent consistency model (LCM). To further enhance semantic fidelity, we introduce adaptive semantic enhancement, which fuses multi-scale pyramid edge maps as conditioning input to guide the purification process. Extensive experiments across multiple datasets demonstrate that DBLP achieves state-of-the-art (SOTA) robust accuracy, superior image quality, and around 0.2s inference time, marking a significant step toward real-time adversarial purification.

Method: Leverages Singular Value Decomposition (SVD) to modify CLIP’s internal parameter space without adding modules. Only fine-tunes singular values to rescale basis vectors for domain adaptation while retaining pretrained model structure.

Result: Achieves state-of-the-art classification results on 11 natural and 10 biomedical datasets. Outperforms previous methods in accuracy and generalization under few-shot settings using only 0.04% of total parameters.

Conclusion: CLIP-SVD provides an effective parameter-efficient adaptation technique that enhances performance while better preserving the model’s generalization ability and original knowledge, with added interpretability through language-based analysis.

Abstract: Vision-language models (VLMs) like CLIP have shown impressive zero-shot and few-shot learning capabilities across diverse applications. However, adapting these models to new fine-grained domains remains difficult due to reliance on prompt engineering and the high cost of full model fine-tuning. Existing adaptation approaches rely on augmented components, such as prompt tokens and adapter modules, which could limit adaptation quality, destabilize the model, and compromise the rich knowledge learned during pretraining. In this work, we present CLIP-SVD, a novel multi-modal and parameter-efficient adaptation technique that leverages Singular Value Decomposition (SVD) to modify the internal parameter space of CLIP without injecting additional modules. Specifically, we fine-tune only the singular values of the CLIP parameter matrices to rescale the basis vectors for domain adaptation while retaining the pretrained model. This design enables enhanced adaptation performance using only 0.04% of the model’s total parameters and better preservation of its generalization ability. CLIP-SVD achieves state-of-the-art classification results on 11 natural and 10 biomedical datasets, outperforming previous methods in both accuracy and generalization under few-shot settings. Additionally, we leverage a natural language-based approach to analyze the effectiveness and dynamics of the CLIP adaptation to allow interpretability of CLIP-SVD. The code is publicly available at https://github.com/HealthX-Lab/CLIP-SVD.

Method: Proposes continuous-time consistency distillation for precise ODE trajectory preservation, and dual-perspective alignment including distribution alignment between synthetic/real data and trajectory alignment across different inference steps.

Result: Significantly outperforms existing approaches on OpenVid-1M benchmark for few-step video generation while maintaining high-quality output.

Conclusion: SwiftVideo provides a unified and stable framework that successfully combines trajectory-preserving and distribution-matching strategies to achieve efficient high-quality video generation with substantially reduced computational overhead.

Abstract: Diffusion-based or flow-based models have achieved significant progress in video synthesis but require multiple iterative sampling steps, which incurs substantial computational overhead. While many distillation methods that are solely based on trajectory-preserving or distribution-matching have been developed to accelerate video generation models, these approaches often suffer from performance breakdown or increased artifacts under few-step settings. To address these limitations, we propose \textbf{\emph{SwiftVideo}}, a unified and stable distillation framework that combines the advantages of trajectory-preserving and distribution-matching strategies. Our approach introduces continuous-time consistency distillation to ensure precise preservation of ODE trajectories. Subsequently, we propose a dual-perspective alignment that includes distribution alignment between synthetic and real data along with trajectory alignment across different inference steps. Our method maintains high-quality video generation while substantially reducing the number of inference steps. Quantitative evaluations on the OpenVid-1M benchmark demonstrate that our method significantly outperforms existing approaches in few-step video generation.

Method: Created PixelHumor benchmark with 2,800 annotated multi-panel comics and conducted experiments with state-of-the-art LMMs to evaluate their humor interpretation and narrative sequencing abilities.

Result: Top models achieved only 61% accuracy in panel sequencing (far below human performance), revealing substantial gaps in multimodal integration for coherent narrative and humor understanding.

Conclusion: PixelHumor provides a rigorous framework to drive development of LMMs with better multimodal contextual reasoning and socially aware interactions, addressing critical limitations in current models.

Abstract: Understanding humor is a core aspect of social intelligence, yet it remains a significant challenge for Large Multimodal Models (LMMs). We introduce PixelHumor, a benchmark dataset of 2,800 annotated multi-panel comics designed to evaluate LMMs’ ability to interpret multimodal humor and recognize narrative sequences. Experiments with state-of-the-art LMMs reveal substantial gaps: for instance, top models achieve only 61% accuracy in panel sequencing, far below human performance. This underscores critical limitations in current models’ integration of visual and textual cues for coherent narrative and humor understanding. By providing a rigorous framework for evaluating multimodal contextual and narrative reasoning, PixelHumor aims to drive the development of LMMs that better engage in natural, socially aware interactions.

Method: Event-driven framework using lightweight U-Net architecture with innovative Dice-Contour Regularization Loss to detect thin obstacles in event streams.

Result: Achieves mean F1 Score of 0.7088 with low latency of 21.2 ms, making it suitable for edge and mobile deployment.

Conclusion: Event-based approach effectively detects submillimeter obstacles with high precision and low latency, ideal for drone safety applications.

Abstract: Drones operating in complex environments face a significant threat from thin obstacles, such as steel wires and kite strings at the submillimeter level, which are notoriously difficult for conventional sensors like RGB cameras, LiDAR, and depth cameras to detect. This paper introduces SkyShield, an event-driven, end-to-end framework designed for the perception of submillimeter scale obstacles. Drawing upon the unique features that thin obstacles present in the event stream, our method employs a lightweight U-Net architecture and an innovative Dice-Contour Regularization Loss to ensure precise detection. Experimental results demonstrate that our event-based approach achieves mean F1 Score of 0.7088 with a low latency of 21.2 ms, making it ideal for deployment on edge and mobile platforms.

Method: The authors conduct a systematic review of current trends and introduce a new annotated dataset (3DCrack) collected with 3D laser scans. They perform extensive benchmarking experiments to establish baselines for various deep learning methodologies, including recent foundation models.

Result: The review identifies key shifts in learning paradigms, generalizability improvements, and data acquisition diversification. The new 3DCrack dataset and benchmarking results provide valuable resources and insights for future research in crack detection.

Conclusion: The paper provides comprehensive insights into evolving methodologies and future directions for deep learning-based crack detection, offering a new dataset and benchmarking framework to support ongoing research advancements in this critical field.

Abstract: Crack detection plays a crucial role in civil infrastructures, including inspection of pavements, buildings, etc., and deep learning has significantly advanced this field in recent years. While numerous technical and review papers exist in this domain, emerging trends are reshaping the landscape. These shifts include transitions in learning paradigms (from fully supervised learning to semi-supervised, weakly-supervised, unsupervised, few-shot, domain adaptation and fine-tuning foundation models), improvements in generalizability (from single-dataset performance to cross-dataset evaluation), and diversification in dataset acquisition (from RGB images to specialized sensor-based data). In this review, we systematically analyze these trends and highlight representative works. Additionally, we introduce a new annotated dataset collected with 3D laser scans, 3DCrack, to support future research and conduct extensive benchmarking experiments to establish baselines for commonly used deep learning methodologies, including recent foundation models. Our findings provide insights into the evolving methodologies and future directions in deep learning-based crack detection. Project page: https://github.com/nantonzhang/Awesome-Crack-Detection

Method: Developed a large-scale one-shot Gaussian head generator using WebSSL with two-stage training approach, separating reconstruction and reenactment modules, and employing ultra-lightweight Gaussian avatar driven by control signals.

Result: Achieves 90 FPS at 512x512 resolution, demonstrates scaling law benefits where larger reconstruction modules improve performance without affecting driving efficiency, and outperforms state-of-the-art methods in both quantitative and qualitative evaluations.

Conclusion: The proposed separation framework successfully enables efficient, high-quality 3D head avatar generation from single images with real-time rendering capabilities and scalable performance improvements.

Abstract: In this paper, we explore a reconstruction and reenactment separated framework for 3D Gaussians head, which requires only a single portrait image as input to generate controllable avatar. Specifically, we developed a large-scale one-shot gaussian head generator built upon WebSSL and employed a two-stage training approach that significantly enhances the capabilities of generalization and high-frequency texture reconstruction. During inference, an ultra-lightweight gaussian avatar driven by control signals enables high frame-rate rendering, achieving 90 FPS at a resolution of 512x512. We further demonstrate that the proposed framework follows the scaling law, whereby increasing the parameter scale of the reconstruction module leads to improved performance. Moreover, thanks to the separation design, driving efficiency remains unaffected. Finally, extensive quantitative and qualitative experiments validate that our approach outperforms current state-of-the-art methods.

Method: Constructed two psychophysics-inspired benchmarks using Chinese logographs and English alphabetic words by splicing, recombining, and overlaying glyphs to create ‘visible but unreadable’ stimuli for models while remaining legible to humans.

Result: Contemporary VLMs show severe performance drops under these perturbations, frequently producing unrelated or incoherent outputs, indicating they rely heavily on generic visual invariances but underutilize compositional priors needed for robust literacy.

Conclusion: The findings reveal structural limitations in current VLMs and motivate the development of architectures and training strategies that better encode symbol segmentation, composition, and binding across different scripts for robust multimodal systems.

Abstract: Writing is a universal cultural technology that reuses vision for symbolic communication. Humans display striking resilience: we readily recognize words even when characters are fragmented, fused, or partially occluded. This paper investigates whether advanced vision language models (VLMs) share this resilience. We construct two psychophysics inspired benchmarks across distinct writing systems, Chinese logographs and English alphabetic words, by splicing, recombining, and overlaying glyphs to yield ‘‘visible but unreadable’’ stimuli for models while remaining legible to humans. Despite strong performance on clean text, contemporary VLMs show a severe drop under these perturbations, frequently producing unrelated or incoherent outputs. The pattern suggests a structural limitation: models heavily leverage generic visual invariances but under rely on compositional priors needed for robust literacy. We release stimuli generation code, prompts, and evaluation protocols to facilitate transparent replication and follow up work. Our findings motivate architectures and training strategies that encode symbol segmentation, composition, and binding across scripts, and they delineate concrete challenges for deploying multimodal systems in education, accessibility, cultural heritage, and security.

Method: Proposes a fine-tuning pipeline with two key innovations: converting structured metadata into synthetic reports for better image-text alignment, and adding an auxiliary token trained to predict MMSE scores for additional supervision. Uses lightweight prompt tuning on both image and text modalities.

Result: Achieves state-of-the-art performance on two Alzheimer’s disease datasets using only 1,500 training images, outperforming methods fine-tuned on 10,000 images.

Conclusion: The approach demonstrates data-efficient adaptation of 3D Med-VLMs with improved clinical knowledge integration and metadata utilization for enhanced Alzheimer’s disease diagnosis.

Abstract: Medical vision-language models (Med-VLMs) have shown impressive results in tasks such as report generation and visual question answering, but they still face several limitations. Most notably, they underutilize patient metadata and lack integration of clinical diagnostic knowledge. Moreover, most existing models are typically trained from scratch or fine-tuned on large-scale 2D image-text pairs, requiring extensive computational resources, and their effectiveness on 3D medical imaging is often limited due to the absence of structural information. To address these gaps, we propose a data-efficient fine-tuning pipeline to adapt 3D CT-based Med-VLMs for 3D MRI and demonstrate its application in Alzheimer’s disease (AD) diagnosis. Our system introduces two key innovations. First, we convert structured metadata into synthetic reports, enriching textual input for improved image-text alignment. Second, we add an auxiliary token trained to predict the mini-mental state examination (MMSE) score, a widely used clinical measure of cognitive function that correlates with AD severity. This provides additional supervision for fine-tuning. Applying lightweight prompt tuning to both image and text modalities, our approach achieves state-of-the-art performance on two AD datasets using 1,500 training images, outperforming existing methods fine-tuned on 10,000 images. Code will be released upon publication.

Method: Multi-task learning approach that incorporates fixed object information in the environment along with human skeleton poses, using interaction area information.

Result: Achieved 99.25% accuracy on a custom dataset containing interaction classes with fixed objects (ATM machines, check-in/out machines) and non-interaction classes, outperforming base skeleton-only model by 2.75%.

Conclusion: Incorporating fixed object information through multi-task learning significantly improves human-object interaction recognition performance compared to using only human skeleton data.

Abstract: Recent graph convolutional neural networks (GCNs) have shown high performance in the field of human action recognition by using human skeleton poses. However, it fails to detect human-object interaction cases successfully due to the lack of effective representation of the scene information and appropriate learning architectures. In this context, we propose a methodology to utilize human action recognition performance by considering fixed object information in the environment and following a multi-task learning approach. In order to evaluate the proposed method, we collected real data from public environments and prepared our data set, which includes interaction classes of hands-on fixed objects (e.g., ATM ticketing machines, check-in/out machines, etc.) and non-interaction classes of walking and standing. The multi-task learning approach, along with interaction area information, succeeds in recognizing the studied interaction and non-interaction actions with an accuracy of 99.25%, outperforming the accuracy of the base model using only human skeleton poses by 2.75%.

Method: Two-stage pipeline: 1) MLLM director produces blueprint video conditioned on multimodal instructions for high-level semantics, 2) Parallel sub-clip generation using first-last frame strategy guided by blueprint keyframes.

Result: Generates vivid, fluent 1080p videos at 48 fps with superior lip sync accuracy, emotion expressiveness, instruction controllability, identity preservation, and cross-domain generalization.

Conclusion: Kling-Avatar establishes a new benchmark for semantically grounded, high-fidelity audio-driven avatar synthesis suitable for real-world applications like digital human livestreaming.

Abstract: Recent advances in audio-driven avatar video generation have significantly enhanced audio-visual realism. However, existing methods treat instruction conditioning merely as low-level tracking driven by acoustic or visual cues, without modeling the communicative purpose conveyed by the instructions. This limitation compromises their narrative coherence and character expressiveness. To bridge this gap, we introduce Kling-Avatar, a novel cascaded framework that unifies multimodal instruction understanding with photorealistic portrait generation. Our approach adopts a two-stage pipeline. In the first stage, we design a multimodal large language model (MLLM) director that produces a blueprint video conditioned on diverse instruction signals, thereby governing high-level semantics such as character motion and emotions. In the second stage, guided by blueprint keyframes, we generate multiple sub-clips in parallel using a first-last frame strategy. This global-to-local framework preserves fine-grained details while faithfully encoding the high-level intent behind multimodal instructions. Our parallel architecture also enables fast and stable generation of long-duration videos, making it suitable for real-world applications such as digital human livestreaming and vlogging. To comprehensively evaluate our method, we construct a benchmark of 375 curated samples covering diverse instructions and challenging scenarios. Extensive experiments demonstrate that Kling-Avatar is capable of generating vivid, fluent, long-duration videos at up to 1080p and 48 fps, achieving superior performance in lip synchronization accuracy, emotion and dynamic expressiveness, instruction controllability, identity preservation, and cross-domain generalization. These results establish Kling-Avatar as a new benchmark for semantically grounded, high-fidelity audio-driven avatar synthesis.

Method: Evaluate condition numbers for polynomial evaluation in Legendre, Chebyshev, and their Sobolev bases, and bound the norms of variations between symbols using different inner products.

Result: The study provides insights into the trade-offs between basis choice and polynomial degree, highlighting how different bases affect computational efficiency and accuracy in symbol representation.

Conclusion: Choosing the appropriate polynomial basis and degree is crucial for achieving accurate mathematical handwriting modeling with minimal computational overhead, with condition numbers and norm bounds serving as key metrics for optimization.

Abstract: Previous work has made use of a parameterized plane curve polynomial representation for mathematical handwriting, with the polynomials represented in a Legendre or Legendre-Sobolev graded basis. This provides a compact geometric representation for the digital ink. Preliminary results have also been shown for Chebyshev and Chebyshev-Sobolev bases. This article explores the trade-offs between basis choice and polynomial degree to achieve accurate modeling with a low computational cost. To do this, we consider the condition number for polynomial evaluation in these bases and bound how the various inner products give norms for the variations between symbols.

Method: Three-stage compression: 1) Sparsity-aware training with dynamic sparsity, 2) Structured channel pruning using batch normalization scaling factors, 3) Channel-Wise Knowledge Distillation with adjustable temperature and loss weighting for small/medium objects.

Result: YOLOv8m reduced from 25.85M to 6.85M parameters (73.51% reduction), FLOPs from 49.6G to 13.3G, MACs from 101G to 34.5G. AP50 dropped only 2.7% to 47.9, inference speed increased from 26 to 45 FPS. With TensorRT: 47.6 AP50 at 68 FPS.

Conclusion: The proposed compression pipeline enables real-time deployment on edge devices with minimal accuracy loss, demonstrating practicality for high-throughput, resource-constrained aerial object detection scenarios.

Abstract: Efficient deployment of deep learning models for aerial object detection on resource-constrained devices requires significant compression without com-promising performance. In this study, we propose a novel three-stage compression pipeline for the YOLOv8 object detection model, integrating sparsity-aware training, structured channel pruning, and Channel-Wise Knowledge Distillation (CWD). First, sparsity-aware training introduces dynamic sparsity during model optimization, effectively balancing parameter reduction and detection accuracy. Second, we apply structured channel pruning by leveraging batch normalization scaling factors to eliminate redundant channels, significantly reducing model size and computational complexity. Finally, to mitigate the accuracy drop caused by pruning, we employ CWD to transfer knowledge from the original model, using an adjustable temperature and loss weighting scheme tailored for small and medium object detection. Extensive experiments on the VisDrone dataset demonstrate the effectiveness of our approach across multiple YOLOv8 variants. For YOLOv8m, our method reduces model parameters from 25.85M to 6.85M (a 73.51% reduction), FLOPs from 49.6G to 13.3G, and MACs from 101G to 34.5G, while reducing AP50 by only 2.7%. The resulting compressed model achieves 47.9 AP50 and boosts inference speed from 26 FPS (YOLOv8m baseline) to 45 FPS, enabling real-time deployment on edge devices. We further apply TensorRT as a lightweight optimization step. While this introduces a minor drop in AP50 (from 47.9 to 47.6), it significantly improves inference speed from 45 to 68 FPS, demonstrating the practicality of our approach for high-throughput, re-source-constrained scenarios.

Method: Uses Style Disentangled Attention (SD-Attn) module with cross-3D attention for dynamic interaction between content and style images, plus style-disentangled feature selection to prevent semantic leakage. Includes Style Guided Control for exclusive geometry/texture stylization.

Result: Outperforms existing baseline methods in producing high-fidelity 3D assets with fine-grained style control.

Conclusion: StyleSculptor enables zero-shot, training-free generation of style-controllable 3D assets with effective texture and geometry style transfer capabilities.

Abstract: Creating 3D assets that follow the texture and geometry style of existing ones is often desirable or even inevitable in practical applications like video gaming and virtual reality. While impressive progress has been made in generating 3D objects from text or images, creating style-controllable 3D assets remains a complex and challenging problem. In this work, we propose StyleSculptor, a novel training-free approach for generating style-guided 3D assets from a content image and one or more style images. Unlike previous works, StyleSculptor achieves style-guided 3D generation in a zero-shot manner, enabling fine-grained 3D style control that captures the texture, geometry, or both styles of user-provided style images. At the core of StyleSculptor is a novel Style Disentangled Attention (SD-Attn) module, which establishes a dynamic interaction between the input content image and style image for style-guided 3D asset generation via a cross-3D attention mechanism, enabling stable feature fusion and effective style-guided generation. To alleviate semantic content leakage, we also introduce a style-disentangled feature selection strategy within the SD-Attn module, which leverages the variance of 3D feature patches to disentangle style- and content-significant channels, allowing selective feature injection within the attention framework. With SD-Attn, the network can dynamically compute texture-, geometry-, or both-guided features to steer the 3D generation process. Built upon this, we further propose the Style Guided Control (SGC) mechanism, which enables exclusive geometry- or texture-only stylization, as well as adjustable style intensity control. Extensive experiments demonstrate that StyleSculptor outperforms existing baseline methods in producing high-fidelity 3D assets.

Method: Systematic comparison of Qwen 3 models (0.6B, 1.7B, 4B parameters) on RewardBench tasks, evaluating accuracy and computational efficiency. Examined augmentation strategies for non-thinking models including in-context learning, rubric-guided judging, reference-based evaluation, and n-best aggregation.

Result: Thinking models achieve approximately 10% higher accuracy with little overhead (under 2x), while augmentation strategies like few-shot learning deliver modest gains at higher cost (>8x). Thinking models show 6% higher consistency on average under various bias conditions and maintain advantages in multilingual settings.

Conclusion: Explicit reasoning offers clear advantages in the LLM-as-a-judge paradigm, providing superior performance not only in accuracy and efficiency but also in robustness across different bias conditions and languages.

Abstract: As Large Language Models (LLMs) are increasingly adopted as automated judges in benchmarking and reward modeling, ensuring their reliability, efficiency, and robustness has become critical. In this work, we present a systematic comparison of “thinking” and “non-thinking” LLMs in the LLM-as-a-judge paradigm using open-source Qwen 3 models of relatively small sizes (0.6B, 1.7B, and 4B parameters). We evaluate both accuracy and computational efficiency (FLOPs) on RewardBench tasks, and further examine augmentation strategies for non-thinking models, including in-context learning, rubric-guided judging, reference-based evaluation, and n-best aggregation. Our results show that despite these enhancements, non-thinking models generally fall short of their thinking counterparts. Our results show that thinking models achieve approximately 10% points higher accuracy with little overhead (under 2x), in contrast to augmentation strategies like few-shot learning, which deliver modest gains at a higher cost (>8x). Bias and robustness analyses further demonstrate that thinking models maintain significantly greater consistency under a variety of bias conditions such as positional, bandwagon, identity, diversity, and random biases (6% higher on average). We further extend our experiments to the multilingual setting and our results confirm that explicit reasoning extends its benefits beyond English. Overall, our work results in several important findings that provide systematic evidence that explicit reasoning offers clear advantages in the LLM-as-a-judge paradigm not only in accuracy and efficiency but also in robustness.

Method: Investigated 15 models from 0.27B to 70B parameters across four families using linear probing on steering vector activations to measure evaluation awareness.

Result: Found a clear power-law scaling relationship where evaluation awareness increases predictably with model size.

Conclusion: This scaling law enables forecasting deceptive behavior in future larger models and guides the design of scale-aware evaluation strategies for AI safety.

Abstract: Large language models (LLMs) can internally distinguish between evaluation and deployment contexts, a behaviour known as \emph{evaluation awareness}. This undermines AI safety evaluations, as models may conceal dangerous capabilities during testing. Prior work demonstrated this in a single $70$B model, but the scaling relationship across model sizes remains unknown. We investigate evaluation awareness across $15$ models scaling from $0.27$B to $70$B parameters from four families using linear probing on steering vector activations. Our results reveal a clear power-law scaling: evaluation awareness increases predictably with model size. This scaling law enables forecasting deceptive behavior in future larger models and guides the design of scale-aware evaluation strategies for AI safety. A link to the implementation of this paper can be found at https://anonymous.4open.science/r/evaluation-awareness-scaling-laws/README.md.

Method: Generate synthetic training data by intervening on individual reasoning steps in model-generated CoTs to create faithful/unfaithful pairs, then apply Direct Preference Optimization to teach models to prefer causally consistent reasoning paths.

Result: FRIT increases faithful reasoning by 3.4 percentage points for Mistral on GSM8K while improving accuracy by 7.6 percentage points across factual and symbolic reasoning tasks on Qwen3-8B and Mistral-7B-v0.1.

Conclusion: FRIT provides the first scalable, supervision-free method for training language models to produce more reliable and interpretable reasoning, addressing the gap between reasoning performance and trustworthiness.

Abstract: Chain-of-thought (CoT) reasoning has emerged as a powerful tool for improving large language model performance on complex tasks, but recent work shows that reasoning steps often fail to causally influence the final answer, creating brittle and untrustworthy outputs. Prior approaches focus primarily on measuring faithfulness, while methods for systematically improving it remain limited. We introduce Faithful Reasoning via Intervention Training (FRIT), a scalable alignment method that trains models to produce causally consistent reasoning by learning from systematically corrupted examples. FRIT generates synthetic training data by intervening on individual reasoning steps in model-generated CoTs, creating faithful/unfaithful pairs that highlight when reasoning breaks down. We then apply Direct Preference Optimization to teach models to prefer causally consistent reasoning paths. Evaluating on Qwen3-8B and Mistral-7B-v0.1 across factual and symbolic reasoning tasks, FRIT increases faithful reasoning by $3.4$ percentage points for Mistral on GSM8K while improving accuracy by $7.6$ percentage points. Our approach provides the first scalable, supervision-free method for training language models to produce more reliable and interpretable reasoning, addressing a critical gap between reasoning performance and trustworthiness. We release our code at \href{https://github.com/Anut-py/frit}.

Method: The paper proposes an antifragile perspective that leverages uncertainties to prepare for future unpredictable uncertainties, identifies limitations of static testing, and explores antifragile solutions for rare event management.

Result: The paper advocates for fundamental recalibration of AI safety measurement, benchmarking, and improvement methods to complement existing robustness approaches with ethical and practical guidelines.

Conclusion: An antifragile approach is pivotal for long-term reliability of open-ended ML systems, requiring a shift from reducing current uncertainties to leveraging them for future preparedness.

Abstract: This position paper contends that modern AI research must adopt an antifragile perspective on safety – one in which the system’s capacity to guarantee long-term AI safety such as handling rare or out-of-distribution (OOD) events expands over time. Conventional static benchmarks and single-shot robustness tests overlook the reality that environments evolve and that models, if left unchallenged, can drift into maladaptation (e.g., reward hacking, over-optimization, or atrophy of broader capabilities). We argue that an antifragile approach – Rather than striving to rapidly reduce current uncertainties, the emphasis is on leveraging those uncertainties to better prepare for potentially greater, more unpredictable uncertainties in the future – is pivotal for the long-term reliability of open-ended ML systems. In this position paper, we first identify key limitations of static testing, including scenario diversity, reward hacking, and over-alignment. We then explore the potential of antifragile solutions to manage rare events. Crucially, we advocate for a fundamental recalibration of the methods used to measure, benchmark, and continually improve AI safety over the long term, complementing existing robustness approaches by providing ethical and practical guidelines towards fostering an antifragile AI safety community.

Method: Proposes IMAC (Imagined Autocurricula) that leverages Unsupervised Environment Design (UED) to create automatic curricula over worlds generated by world models, ensuring agents train on useful generated data.

Result: Achieves strong transfer performance on held-out environments in challenging procedurally generated settings, training only inside a world model learned from narrow dataset.

Conclusion: Opens path to utilizing larger-scale foundation world models for developing generally capable agents without requiring massive training data or perfect simulations.

Abstract: Training agents to act in embodied environments typically requires vast training data or access to accurate simulation, neither of which exists for many cases in the real world. Instead, world models are emerging as an alternative leveraging offline, passively collected data, they make it possible to generate diverse worlds for training agents in simulation. In this work, we harness world models to generate imagined environments to train robust agents capable of generalizing to novel task variations. One of the challenges in doing this is ensuring the agent trains on useful generated data. We thus propose a novel approach, IMAC (Imagined Autocurricula), leveraging Unsupervised Environment Design (UED), which induces an automatic curriculum over generated worlds. In a series of challenging, procedurally generated environments, we show it is possible to achieve strong transfer performance on held-out environments, having trained only inside a world model learned from a narrower dataset. We believe this opens the path to utilizing larger-scale, foundation world models for generally capable agents.

Method: Introduces Chain of Action (CoA) framework that treats abstracted actions as intermediate reasoning steps (like chain of thought) to guide final executable action generation. Trains an All-in-One agent on diverse action space mixtures using the CoA paradigm.

Result: The unified agent achieves new state-of-the-art performance, improving overall task success rate over strong specialized baselines. The approach learns more robust and generalizable policies.

Conclusion: CoA framework successfully resolves the action space dilemma by unifying planning and control, demonstrating that diverse action space training with intermediate reasoning steps leads to superior generalist agents. Releases OpenHA suite for reproducible research.

Abstract: The choice of action spaces is a critical yet unresolved challenge in developing capable, end-to-end trainable agents. This paper first presents a large-scale, systematic comparison of prominent abstracted action spaces and tokenizers for Vision-Language-Action (VLA) or hierarchical agent models in the open-ended Minecraft. Our analysis reveals that no single action space is universally optimal; instead, the most effective abstraction is highly task-dependent, creating a dilemma for building generalist agents. To resolve this, we introduce Chain of Action (CoA), a novel framework that unifies high-level planning and low-level control within a single, monolithic VLA model. CoA treats an abstracted action not as a command for a separate policy, but as an intermediate reasoning step–akin to a chain of thought–that guides the generation of the final, executable action. Furthermore, we demonstrate that an All-in-One agent trained on a diverse mixture of action spaces using the CoA paradigm learns a more robust and generalizable policy. This unified agent achieves a new state-of-the-art, improving the overall task success rate over strong, specialized baselines. To foster reproducible research, we release the OpenHA (Open Hierarchical Agents) suite, which includes our comprehensive benchmark of over 800 distinct tasks, curated datasets, source code, and all pretrained model checkpoints at https://github.com/CraftJarvis/OpenHA

Method: Instruction tuning framework that teaches models to reason about action applicability, state transitions, and plan validity using explicit logical inference steps and structured reflection

Result: Achieved up to 94% planning accuracy on standard benchmarks, representing 66% absolute improvement over baseline models across multiple planning domains

Conclusion: Successfully bridges the gap between LLM reasoning capabilities and logical precision required for automated planning, offering promising direction for better AI planning systems

Abstract: Large language models (LLMs) have demonstrated impressive capabilities across diverse tasks, yet their ability to perform structured symbolic planning remains limited, particularly in domains requiring formal representations like the Planning Domain Definition Language (PDDL). In this paper, we present a novel instruction tuning framework, PDDL-Instruct, designed to enhance LLMs’ symbolic planning capabilities through logical chain-of-thought reasoning. Our approach focuses on teaching models to rigorously reason about action applicability, state transitions, and plan validity using explicit logical inference steps. By developing instruction prompts that guide models through the precise logical reasoning required to determine when actions can be applied in a given state, we enable LLMs to self-correct their planning processes through structured reflection. The framework systematically builds verification skills by decomposing the planning process into explicit reasoning chains about precondition satisfaction, effect application, and invariant preservation. Experimental results on multiple planning domains show that our chain-of-thought reasoning based instruction-tuned models are significantly better at planning, achieving planning accuracy of up to 94% on standard benchmarks, representing a 66% absolute improvement over baseline models. This work bridges the gap between the general reasoning capabilities of LLMs and the logical precision required for automated planning, offering a promising direction for developing better AI planning systems.

Method: Five-layer architecture (Perception, Reasoning, Action, Integration, Learning) integrating ROS2 and Gazebo with YOLOv11 object detection, GPT-4 reasoning, and local Gemma-3 deployment for LLM-driven reasoning and tool-calling capabilities.

Result: Achieved higher detection confidence (0.79 vs 0.72), improved person detection rates (91% vs 75%), and dramatically increased action recommendation (92% vs 4.5%) in simulated search-and-rescue scenarios.

Conclusion: Modest computational overhead enables qualitatively new levels of autonomy and ecosystem integration, confirming LLM-driven reasoning can significantly enhance UAV capabilities for dynamic missions.

Abstract: Unmanned Aerial Vehicles (UAVs) are increasingly deployed in defense, surveillance, and disaster response, yet most systems remain confined to SAE Level 2–3 autonomy. Their reliance on rule-based control and narrow AI restricts adaptability in dynamic, uncertain missions. Existing UAV frameworks lack context-aware reasoning, autonomous decision-making, and ecosystem-level integration; critically, none leverage Large Language Model (LLM) agents with tool-calling for real-time knowledge access. This paper introduces the Agentic UAVs framework, a five-layer architecture (Perception, Reasoning, Action, Integration, Learning) that augments UAVs with LLM-driven reasoning, database querying, and third-party system interaction. A ROS2 and Gazebo-based prototype integrates YOLOv11 object detection with GPT-4 reasoning and local Gemma-3 deployment. In simulated search-and-rescue scenarios, agentic UAVs achieved higher detection confidence (0.79 vs. 0.72), improved person detection rates (91% vs. 75%), and markedly increased action recommendation (92% vs. 4.5%). These results confirm that modest computational overhead enables qualitatively new levels of autonomy and ecosystem integration.

Method: Adds parallel fuzzy-membership feature channel with interpretable semantic features (POS, roles, sentiment, etc.) fused via gated adapter, trained with next-token prediction, auxiliary reconstruction loss, and uniformizer regularization.

Result: Improves perplexity on synthetic corpus, enables precise user-controllable generation of polarity and punctuation, maintains full compatibility with tied embeddings, adds minimal overhead.

Conclusion: Semantic fusion provides an effective, lightweight approach for interpretable and controllable language generation while preserving model efficiency and simplicity.

Abstract: We propose semantic fusion, a lightweight scheme that augments a Transformer language model (LM) with a parallel, fuzzy-membership feature channel that encodes token-level semantics. Each token is represented by a vector of interpretable features (e.g. part-of-speech cues, shallow roles, boundary flags, sentiment polarity and strength) whose values are graded degrees from differentiable membership functions (e.g. power kernels). These per-token vectors form a sentence-level semantic matrix fused via a gated adapter into the LM. Training uses standard next-token prediction, an auxiliary loss that reconstructs the semantic features from hidden states, and a lightweight uniformizer that regularizes adjective-class distributions. On a synthetic two-clause corpus with held-out adjectives for out-of-distribution (OOD) control, semantic fusion improves perplexity and enables precise, user-controllable generation of polarity and punctuation while maintaining model simplicity. This approach adds only small overhead, remains fully compatible with tied input-output embeddings, and provides an interpretable pathway for conditioned natural language generation.

Method: Proposes the Asterisk Operator (*-operator) based on Adjacency-Structured Parallel Propagation (ASPP), which formalizes reasoning as local parallel state evolution guided by implicit relational graphs. Uses Embedding-Asterisk distillation method for implementation.

Result: Achieves 100% accuracy on ARC2 validation set with only 6M parameters. Demonstrates universality, convergence properties, and superior performance on ARC2 challenges and Conway’s Game of Life through rigorous mathematical analysis and experiments.

Conclusion: The Asterisk Operator represents a significant breakthrough in neural-symbolic reasoning, providing an efficient and convergent computational paradigm that maintains local constraints while achieving global reasoning capabilities for abstract reasoning problems.

Abstract: We propose the \textbf{Asterisk Operator} ($\ast$-operator), a novel unified framework for abstract reasoning based on Adjacency-Structured Parallel Propagation (ASPP). The operator formalizes structured reasoning tasks as local, parallel state evolution processes guided by implicit relational graphs. We prove that the $\ast$-operator maintains local computational constraints while achieving global reasoning capabilities, providing an efficient and convergent computational paradigm for abstract reasoning problems. Through rigorous mathematical analysis and comprehensive experiments on ARC2 challenges and Conway’s Game of Life, we demonstrate the operator’s universality, convergence properties, and superior performance. Our innovative Embedding-Asterisk distillation method achieves 100% accuracy on ARC2 validation with only 6M parameters, representing a significant breakthrough in neural-symbolic reasoning. \textbf{Keywords:} Abstract Reasoning, Adjacency Structure, Parallel Propagation, Asterisk Operator, Convergence, Universal Approximation

Method: Dual-agent architecture with Generator Agent (AI designer) that analyzes tasks and generates executable RL agents, and Target Agent (resulting RL agent). Framework decomposes RL development into MDP modeling and algorithmic optimization stages using Model Context Protocol.

Result: Outperforms manually designed solutions across all tested benchmarks (MuJoCo, MetaDrive, MPE, SMAC) with up to 55% performance improvement and substantial average gains.

Conclusion: Establishes a new paradigm where intelligent agents autonomously design and optimize other agents, enabling truly end-to-end closed-loop automation for AI systems.

Abstract: Reinforcement learning agent development traditionally requires extensive expertise and lengthy iterations, often resulting in high failure rates and limited accessibility. This paper introduces $Agent^2$, a novel agent-generates-agent framework that achieves fully automated RL agent design through intelligent LLM-driven generation. The system autonomously transforms natural language task descriptions and environment code into comprehensive, high-performance reinforcement learning solutions without human intervention. $Agent^2$ features a revolutionary dual-agent architecture. The Generator Agent serves as an autonomous AI designer that analyzes tasks and generates executable RL agents, while the Target Agent is the resulting automatically generated RL agent. The framework decomposes RL development into two distinct stages: MDP modeling and algorithmic optimization, enabling more targeted and effective agent generation. Built on the Model Context Protocol, $Agent^2$ provides a unified framework that standardizes intelligent agent creation across diverse environments and algorithms, while incorporating adaptive training management and intelligent feedback analysis for continuous improvement. Extensive experiments on a wide range of benchmarks, including MuJoCo, MetaDrive, MPE, and SMAC, demonstrate that $Agent^2$ consistently outperforms manually designed solutions across all tasks, achieving up to 55% performance improvement and substantial gains on average. By enabling truly end-to-end, closed-loop automation, this work establishes a new paradigm in which intelligent agents design and optimize other agents, marking a fundamental breakthrough for automated AI systems.

Method: Evaluated 16 state-of-the-art VLMs (open and closed-source) across 6 multimodal datasets using 3 distinct scoring functions for comprehensive uncertainty benchmarking.

Result: Larger models consistently exhibit better uncertainty quantification; more certain models achieve higher accuracy; mathematical and reasoning tasks show poorer uncertainty performance compared to other domains.

Conclusion: This work establishes a foundation for reliable uncertainty evaluation in multimodal systems, demonstrating that models that know more also know better what they don’t know.

Abstract: Vision-Language Models (VLMs) have achieved remarkable progress in complex visual understanding across scientific and reasoning tasks. While performance benchmarking has advanced our understanding of these capabilities, the critical dimension of uncertainty quantification has received insufficient attention. Therefore, unlike prior conformal prediction studies that focused on limited settings, we conduct a comprehensive uncertainty benchmarking study, evaluating 16 state-of-the-art VLMs (open and closed-source) across 6 multimodal datasets with 3 distinct scoring functions. Our findings demonstrate that larger models consistently exhibit better uncertainty quantification; models that know more also know better what they don’t know. More certain models achieve higher accuracy, while mathematical and reasoning tasks elicit poorer uncertainty performance across all models compared to other domains. This work establishes a foundation for reliable uncertainty evaluation in multimodal systems.

Method: Cast as supervised next token prediction problem using transformers, where tokens are actions. Learn from random valid (positive) and invalid (negative) action sequences through gradient descent.

Result: Shows that transformers can faithfully represent propositional STRIPS world models and learn them effectively from action sequence data alone.

Conclusion: Transformers are capable of learning complex world models from action traces, demonstrating the viability of deep learning approaches for symbolic reasoning tasks.

Abstract: We consider the problem of learning propositional STRIPS world models from action traces alone, using a deep learning architecture (transformers) and gradient descent. The task is cast as a supervised next token prediction problem where the tokens are the actions, and an action $a$ may follow an action sequence if the hidden effects of the previous actions do not make an action precondition of $a$ false. We show that a suitable transformer architecture can faithfully represent propositional STRIPS world models, and that the models can be learned from sets of random valid (positive) and invalid (negative) action sequences alone. A number of experiments are reported.

Method: Collected dataset of safety-relevant behaviors, created modular steering framework based on unique components of existing methods, evaluated five popular steering methods on Qwen-2.5-7B and Llama-3.1-8B models.

Result: Strong steering performance depends on specific combination of steering method, model, and targeted behavior; poor combinations can cause severe concept entanglement.

Conclusion: The benchmark reveals complex dependencies between steering methods, models, and behaviors, highlighting the need for systematic evaluation to understand tradeoffs and avoid negative side effects.

Abstract: We introduce SteeringControl, a benchmark for evaluating representation steering methods across core alignment objectives–bias, harmful generation, and hallucination–and their effects on secondary behaviors such as sycophancy and commonsense morality. While prior alignment work often highlights truthfulness or reasoning ability to demonstrate the side effects of representation steering, we find there are many unexplored tradeoffs not yet understood in a systematic way. We collect a dataset of safety-relevant primary and secondary behaviors to evaluate steering effectiveness and behavioral entanglement centered around five popular steering methods. To enable this, we craft a modular steering framework based on unique components that serve as the building blocks of many existing methods. Our results on Qwen-2.5-7B and Llama-3.1-8B find that strong steering performance is dependent on the specific combination of steering method, model, and targeted behavior, and that severe concept entanglement can result from poor combinations of these three as well. We release our code here: https://github.com/wang-research-lab/SteeringControl.git.

Method: Equipped Claude Code agents with MCP-based social media and journaling tools and allowed them autonomous use across 34 Aider Polyglot Python programming challenges, analyzing their collaborative strategies and tool usage patterns.

Result: Collaborative tools substantially improved performance on hardest problems: 15-40% lower cost, 12-27% fewer turns, and 12-38% faster completion. Different models adopted distinct strategies - Sonnet 3.7 used broad tool engagement while Sonnet 4 showed selective adoption with journal-based semantic search. Agents preferred writing over reading by 2-9x, indicating articulation drives improvement.

Conclusion: AI agents can systematically benefit from human-inspired collaboration tools at the edge of their capabilities, with adaptive collaborative interfaces serving as reasoning enhancers rather than universal efficiency boosts, mirroring human collaborative behavior patterns.

Abstract: We investigate whether giving LLM agents the collaborative tools and autonomy that humans naturally use for problem solving can improve their performance. We equip Claude Code agents with MCP-based social media and journaling tools and allow them to use these tools as they see fit. Across 34 Aider Polyglot Python programming challenges, collaborative tools substantially improve performance on the hardest problems, delivering 15-40% lower cost, 12-27% fewer turns, and 12-38% faster completion than baseline agents. Effects on the full challenge set are mixed, suggesting these tools act as performance enhancers when additional reasoning scaffolding is most needed. Surprisingly, Different models naturally adopted distinct collaborative strategies without explicit instruction. Sonnet 3.7 engaged broadly across tools and benefited from articulation-based cognitive scaffolding. Sonnet 4 showed selective adoption, leaning on journal-based semantic search when problems were genuinely difficult. This mirrors how human developers adjust collaboration based on expertise and task complexity. Behavioral analysis shows agents prefer writing over reading by about 2-9x, indicating that structured articulation drives much of the improvement rather than information access alone. Overall, AI agents can systematically benefit from human-inspired collaboration tools at the edge of their capabilities, pointing to adaptive collaborative interfaces as reasoning enhancers rather than universal efficiency boosts.

Method: Survey responses and student interviews across three proof-based undergraduate mathematics courses (first-semester abstract algebra, topology, second-semester abstract algebra) where AI use was permitted.

Result: Analysis of how students engaged with AI tools and their perceptions of generative AI’s usefulness and limitations in proof-based mathematics contexts.

Conclusion: Discussion of future considerations for integrating generative AI into proof-based mathematics instruction based on student experiences and perceptions.

Abstract: With the rapid rise of generative AI in higher education and the unreliability of current AI detection tools, developing policies that encourage student learning and critical thinking has become increasingly important. This study examines student use and perceptions of generative AI across three proof-based undergraduate mathematics courses: a first-semester abstract algebra course, a topology course and a second-semester abstract algebra course. In each case, course policy permitted some use of generative AI. Drawing on survey responses and student interviews, we analyze how students engaged with AI tools, their perceptions of generative AI’s usefulness and limitations, and what implications these perceptions hold for teaching proof-based mathematics. We conclude by discussing future considerations for integrating generative AI into proof-based mathematics instruction.

Method: CoBRA has two components: Cognitive Bias Index (measures bias through social science experiments) and Behavioral Regulation Engine (aligns agent behavior to demonstrate controlled bias).

Result: CoBRA can precisely program cognitive bias in social agents in a model-agnostic manner, as demonstrated through technical benchmarks.

Conclusion: The toolkit provides a systematic approach to specifying agent behavior that overcomes the limitations of implicit natural language descriptions and ensures consistent cognitive bias programming.

Abstract: This paper introduces CoBRA, a novel toolkit for systematically specifying agent behavior in LLM-based social simulation. We found that conventional approaches that specify agent behaviors through implicit natural language descriptions cannot yield consistent behaviors across models, and the produced agent behaviors do not capture the nuances of the descriptions. In contrast, CoBRA presents a new approach to program agents’ cognitive biases explicitly, by grounding agents’ expected behaviors using classic social science experiments. CoBRA has two components: (1) Cognitive Bias Index that measures the cognitive bias of a social agent, by quantifying the agent’s reactions in a set of validated classical social science experiments; (2) Behavioral Regulation Engine that aligns the agent’s behavior to demonstrate controlled cognitive bias. We evaluated CoBRA as an HCI toolkit through demonstration and technical benchmarks. Our results suggest that CoBRA can precisely program the cognitive bias demonstrated in a social agent in a model-agnostic manner.

Method: The authors construct a state control benchmark with binary toggle instructions from public datasets. They propose State-aware Reasoning (StaR), a training method that teaches agents to: 1) perceive the current toggle state, 2) analyze the desired state from the instruction, and 3) act accordingly.

Result: Experiments on three multimodal agents show that StaR improves toggle instruction execution accuracy by over 30%. Further evaluations on three public benchmarks demonstrate that StaR also enhances general task performance. Dynamic environment evaluations highlight StaR’s potential for real-world applications.

Conclusion: State-aware Reasoning (StaR) effectively addresses the toggle control reliability problem in multimodal agents, significantly improving both specific toggle instruction execution and general GUI task performance, making it suitable for real-world applications.

Abstract: The advent of multimodal agents facilitates effective interaction within graphical user interface (GUI), especially in ubiquitous GUI control. However, their inability to reliably execute toggle control instructions remains a key bottleneck. To investigate this, we construct a state control benchmark with binary toggle instructions from public datasets. Evaluations of existing agents demonstrate their unreliability, particularly when the current toggle state already matches the desired state. To address the challenge, we propose State-aware Reasoning (StaR), a training method that teaches agents to perceive the current toggle state, analyze the desired state from the instruction, and act accordingly. Experiments on three multimodal agents demonstrate that StaR can improve toggle instruction execution accuracy by over 30%. Further evaluations on three public benchmarks show that StaR also enhances general task performance. Finally, evaluations on a dynamic environment highlight the potential of StaR for real-world applications. Code, benchmark, and StaR-enhanced agents are available at https://github.com/ZrW00/StaR.

Method: InfraMind integrates five modules: 1) systematic search-based exploration with VM snapshots for GUI understanding, 2) memory-driven planning for precision and efficiency, 3) advanced state identification for hierarchical interfaces, 4) structured knowledge distillation for lightweight deployment, and 5) multi-layered safety mechanisms.

Result: Extensive experiments on open-source and commercial DCIM platforms show InfraMind consistently outperforms existing frameworks in task success rate and operational efficiency.

Conclusion: InfraMind provides a rigorous and scalable solution for industrial management automation by addressing the specific challenges of industrial GUI environments through its integrated framework approach.

Abstract: Mission-critical industrial infrastructure, such as data centers, increasingly depends on complex management software. Its operations, however, pose significant challenges due to the escalating system complexity, multi-vendor integration, and a shortage of expert operators. While Robotic Process Automation (RPA) offers partial automation through handcrafted scripts, it suffers from limited flexibility and high maintenance costs. Recent advances in Large Language Model (LLM)-based graphical user interface (GUI) agents have enabled more flexible automation, yet these general-purpose agents face five critical challenges when applied to industrial management, including unfamiliar element understanding, precision and efficiency, state localization, deployment constraints, and safety requirements. To address these issues, we propose InfraMind, a novel exploration-based GUI agentic framework specifically tailored for industrial management systems. InfraMind integrates five innovative modules to systematically resolve different challenges in industrial management: (1) systematic search-based exploration with virtual machine snapshots for autonomous understanding of complex GUIs; (2) memory-driven planning to ensure high-precision and efficient task execution; (3) advanced state identification for robust localization in hierarchical interfaces; (4) structured knowledge distillation for efficient deployment with lightweight models; and (5) comprehensive, multi-layered safety mechanisms to safeguard sensitive operations. Extensive experiments on both open-source and commercial DCIM platforms demonstrate that our approach consistently outperforms existing frameworks in terms of task success rate and operational efficiency, providing a rigorous and scalable solution for industrial management automation.

Method: Proposes THOR with three components: 1) TIRGen multi-agent pipeline for dataset construction, 2) RL strategy for joint trajectory-level and step-level optimization, 3) self-correction mechanism using tool feedback during inference.

Result: Achieves state-of-the-art performance on multiple mathematical benchmarks for similar-scale models, shows strong generalization across diverse models, and delivers consistent improvements on code benchmarks.

Conclusion: THOR effectively addresses key challenges in tool integration for LLMs through hierarchical optimization and self-correction, demonstrating strong performance and generalization capabilities.

Abstract: Large Language Models (LLMs) have made remarkable progress in mathematical reasoning, but still continue to struggle with high-precision tasks like numerical computation and formal symbolic manipulation. Integrating external tools has emerged as a promising approach to bridge this gap. Despite recent advances, existing methods struggle with three key challenges: constructing tool-integrated reasoning data, performing fine-grained optimization, and enhancing inference. To overcome these limitations, we propose THOR (Tool-Integrated Hierarchical Optimization via RL). First, we introduce TIRGen, a multi-agent actor-critic-based pipeline for constructing high-quality datasets of tool-integrated reasoning paths, aligning with the policy and generalizing well across diverse models. Second, to perform fine-grained hierarchical optimization, we introduce an RL strategy that jointly optimizes for both trajectory-level problem solving and step-level code generation. This is motivated by our key insight that the success of an intermediate tool call is a strong predictor of the final answer’s correctness. Finally, THOR incorporates a self-correction mechanism that leverages immediate tool feedback to dynamically revise erroneous reasoning paths during inference. Our approach demonstrates strong generalization across diverse models, performing effectively in both reasoning and non-reasoning models. It further achieves state-of-the-art performance for models of a similar scale on multiple mathematical benchmarks, while also delivering consistent improvements on code benchmarks. Our code will be publicly available at https://github.com/JingMog/THOR.

Method: Uses MLLM-based recommendation pipeline with structured reasoning, template-augmented reasoning mechanism, and prefix-tree-based constrained decoding to generate precise instruction suggestions from predefined candidates.

Result: Demonstrated substantial improvements in instruction recommendation accuracy through evaluation with real-world annotated datasets and user studies.

Conclusion: MIRA has the potential to revolutionize how users engage with AI services on smartphones by offering a more seamless and efficient experience through intuitive context-aware recommendations.

Abstract: The rapid advancement of generative AI technologies is driving the integration of diverse AI-powered services into smartphones, transforming how users interact with their devices. To simplify access to predefined AI services, this paper introduces MIRA, a pioneering framework for task instruction recommendation that enables intuitive one-touch AI tasking on smartphones. With MIRA, users can long-press on images or text objects to receive contextually relevant instruction recommendations for executing AI tasks. Our work introduces three key innovations: 1) A multimodal large language model (MLLM)-based recommendation pipeline with structured reasoning to extract key entities, infer user intent, and generate precise instructions; 2) A template-augmented reasoning mechanism that integrates high-level reasoning templates, enhancing task inference accuracy; 3) A prefix-tree-based constrained decoding strategy that restricts outputs to predefined instruction candidates, ensuring coherent and intent-aligned suggestions. Through evaluation using a real-world annotated datasets and a user study, MIRA has demonstrated substantial improvements in the accuracy of instruction recommendation. The encouraging results highlight MIRA’s potential to revolutionize the way users engage with AI services on their smartphones, offering a more seamless and efficient experience.

Method: Exhaustive DPLL architecture integrated with effective simplification techniques from mixed integer programming.

Result: Outperforms state-of-the-art MCILC counters and propositional model counters, solving 1718/2840 random instances (vs 1470 by SOTA) and all 4131 application instances.

Conclusion: The proposed approach with MIP simplification techniques significantly improves efficiency and is the only method capable of solving all application benchmarks.

Abstract: Linear constraints are one of the most fundamental constraints in fields such as computer science, operations research and optimization. Many applications reduce to the task of model counting over integer linear constraints (MCILC). In this paper, we design an exact approach to MCILC based on an exhaustive DPLL architecture. To improve the efficiency, we integrate several effective simplification techniques from mixed integer programming into the architecture. We compare our approach to state-of-the-art MCILC counters and propositional model counters on 2840 random and 4131 application benchmarks. Experimental results show that our approach significantly outperforms all exact methods in random benchmarks solving 1718 instances while the state-of-the-art approach only computes 1470 instances. In addition, our approach is the only approach to solve all 4131 application instances.

Method: Used artificial neural networks (ANNs) with different topologies (feed-forward, recurrent, laminated) to test performance on learning artificial grammars of varying complexity. Controlled for network size and resources while comparing information flow patterns.

Result: Recurrent networks demonstrated qualitative expansion in processable input types and significant performance improvements for complex grammars compared to feed-forward networks. Training difficulty of recurrent networks created transition barriers and irreversibility. Laminated networks showed no performance advantage in grammar learning tasks.

Conclusion: Certain changes in network information flow (specifically recurrent connectivity) can produce transitional improvements in cognitive performance, supporting the concept of cognitive evolution through major transitions in neural network structure.

Abstract: Transitional accounts of evolution emphasise a few changes that shape what is evolvable, with dramatic consequences for derived lineages. More recently it has been proposed that cognition might also have evolved via a series of major transitions that manipulate the structure of biological neural networks, fundamentally changing the flow of information. We used idealised models of information flow, artificial neural networks (ANNs), to evaluate whether changes in information flow in a network can yield a transitional change in cognitive performance. We compared networks with feed-forward, recurrent and laminated topologies, and tested their performance learning artificial grammars that differed in complexity, controlling for network size and resources. We documented a qualitative expansion in the types of input that recurrent networks can process compared to feed-forward networks, and a related qualitative increase in performance for learning the most complex grammars. We also noted how the difficulty in training recurrent networks poses a form of transition barrier and contingent irreversibility – other key features of evolutionary transitions. Not all changes in network topology confer a performance advantage in this task set. Laminated networks did not outperform non-laminated networks in grammar learning. Overall, our findings show how some changes in information flow can yield transitions in cognitive performance.

Method: A multi-agent system inspired by real-world crowdsourcing companies that implements rational task assignment and enables collaborative annotation workflow between LLMs, SLMs, and human experts.

Result: Demonstrated effectiveness through extensive experiments on six diverse multimodal classification tasks.

Conclusion: CrowdAgent provides a unified framework for end-to-end annotation process control that synergistically combines different annotation sources while managing quality-cost trade-offs.

Abstract: High-quality annotated data is a cornerstone of modern Natural Language Processing (NLP). While recent methods begin to leverage diverse annotation sources-including Large Language Models (LLMs), Small Language Models (SLMs), and human experts-they often focus narrowly on the labeling step itself. A critical gap remains in the holistic process control required to manage these sources dynamically, addressing complex scheduling and quality-cost trade-offs in a unified manner. Inspired by real-world crowdsourcing companies, we introduce CrowdAgent, a multi-agent system that provides end-to-end process control by integrating task assignment, data annotation, and quality/cost management. It implements a novel methodology that rationally assigns tasks, enabling LLMs, SLMs, and human experts to advance synergistically in a collaborative annotation workflow. We demonstrate the effectiveness of CrowdAgent through extensive experiments on six diverse multimodal classification tasks. The source code and video demo are available at https://github.com/QMMMS/CrowdAgent.

Method: A hierarchical architecture with a Graph Convolutional Network (GCN) as first-order learner mapping node features to optimal paths, and an MLP controller as second-order learner that dynamically adapts GCN parameters when facing novel maze environments.

Result: Second-order learning is particularly effective when cognitive systems develop internal mental maps structurally isomorphic to the environment, showing significant performance improvements and robust generalization on unseen maze tasks.

Conclusion: The study provides empirical support for the pivotal role of structured mental representations in maximizing second-order learning effectiveness, demonstrating that environment-cognition isomorphism enhances cognitive adaptation and generalization.

Abstract: Mental representation, characterized by structured internal models mirroring external environments, is fundamental to advanced cognition but remains challenging to investigate empirically. Existing theory hypothesizes that second-order learning – learning mechanisms that adapt first-order learning (i.e., learning about the task/domain) – promotes the emergence of such environment-cognition isomorphism. In this paper, we empirically validate this hypothesis by proposing a hierarchical architecture comprising a Graph Convolutional Network (GCN) as a first-order learner and an MLP controller as a second-order learner. The GCN directly maps node-level features to predictions of optimal navigation paths, while the MLP dynamically adapts the GCN’s parameters when confronting structurally novel maze environments. We demonstrate that second-order learning is particularly effective when the cognitive system develops an internal mental map structurally isomorphic to the environment. Quantitative and qualitative results highlight significant performance improvements and robust generalization on unseen maze tasks, providing empirical support for the pivotal role of structured mental representations in maximizing the effectiveness of second-order learning.

Method: Three studies: Study 1 evaluated LLMs’ semantic understanding of Big Five measures; Study 2 created AI-Agents using BFI-2 prompts and compared responses with humans; Study 3 validated with risk-taking and moral dilemma vignettes.

Result: AI-Agents align with humans in Big Five trait-response correlations, especially with newer LLMs, but show inconsistent factor loading patterns and cannot fully replicate human response precision.

Conclusion: AI-Agents can serve as useful tools for preliminary research due to correlation alignment with humans, but cannot yet substitute human participants in precision or high-stakes projects due to finer pattern discrepancies.

Abstract: We introduce a methodology for assigning quantifiable and psychometrically validated personalities to AI-Agents using the Big Five framework. Across three studies, we evaluate its feasibility and limits. In Study 1, we show that large language models (LLMs) capture semantic similarities among Big Five measures, providing a basis for personality assignment. In Study 2, we create AI-Agents using prompts designed based on the Big Five Inventory (BFI-2) in the Likert or Expanded format, and find that, when paired with newer LLMs (e.g., GPT-4, GPT-4o, Llama, DeepSeek), these AI-Agents align more closely with human responses on the Mini-Markers test than those generated with binary adjective prompts or older models, although the finer pattern of results (e.g., factor loading patterns) were not consistent between AI-Agents and human participants. In Study 3, we validate our AI-Agents with risk-taking and moral dilemma vignettes. We find that while fine-tuning shifts responses toward more moral judgment, AI-Agent correlations between the input Big Five traits and the output moral judgments mirror those from human participants. Overall, our results show that AI-Agents align with humans in correlations between input Big Five traits and output responses and may serve as useful tools for preliminary research. Nevertheless, discrepancies in finer response patterns indicate that AI-Agents cannot (yet) fully substitute for human participants in precision or high-stakes projects.

Method: Adaptive Difficulty Curriculum Learning (ADCL) dynamically adjusts problem difficulty during training, and Expert-Guided Self-Reformulation (EGSR) guides models to reformulate expert solutions within their own framework rather than direct imitation.

Result: Combined strategies improved performance by 10% on AIME24 and 16.6% on AIME25 benchmarks using Qwen2.5-7B as base model, showing significant synergistic enhancement.

Conclusion: Human-inspired learning strategies (ADCL and EGSR) effectively address LLM limitations in complex problem solving and demonstrate substantial performance improvements on challenging mathematical reasoning tasks.

Abstract: Despite impressive progress in areas like mathematical reasoning, large language models still face significant challenges in consistently solving complex problems. Drawing inspiration from key human learning strategies, we propose two novel strategies to enhance the capability of large language models to solve these complex problems. First, Adaptive Difficulty Curriculum Learning (ADCL) is a novel curriculum learning strategy that tackles the Difficulty Shift phenomenon (i.e., a model’s perception of problem difficulty dynamically changes during training) by periodically re-estimating difficulty within upcoming data batches to maintain alignment with the model’s evolving capabilities. Second, Expert-Guided Self-Reformulation (EGSR) is a novel reinforcement learning strategy that bridges the gap between imitation learning and pure exploration by guiding models to reformulate expert solutions within their own conceptual framework, rather than relying on direct imitation, fostering deeper understanding and knowledge assimilation. Extensive experiments on challenging mathematical reasoning benchmarks, using Qwen2.5-7B as the base model, demonstrate that these human-inspired strategies synergistically and significantly enhance performance. Notably, their combined application improves performance over the standard Zero-RL baseline by 10% on the AIME24 benchmark and 16.6% on AIME25.

Method: Multi-agent framework with specialized agents using structured reasoning inspired by Attentive Reasoning Queries (ARQs), few-shot examples for ensemble diversity, and a judge agent for candidate reranking.

Result: 14% increase in Top-1 accuracy, 18% increase in Top-5 accuracy, and 12% improvement in Mean Reciprocal Rank on bank dataset, with similar gains on public benchmarks (LCQMC and FiQA).

Conclusion: The framework effectively handles ambiguous queries and shows strong generalization across domains and languages, making it suitable for production banking applications.

Abstract: Modern consumer banking applications require accurate and efficient retrieval of information in response to user queries. Mapping user utterances to the most relevant Frequently Asked Questions (FAQs) is a crucial component of these systems. Traditional approaches often rely on a single model or technique, which may not capture the nuances of diverse user inquiries. In this paper, we introduce a multi-agent framework for FAQ annotation that combines multiple specialized agents with different approaches and a judge agent that reranks candidates to produce optimal results. Our agents utilize a structured reasoning approach inspired by Attentive Reasoning Queries (ARQs), which guides them through systematic reasoning steps using targeted, task-specific JSON queries. Our framework features a few-shot example strategy, where each agent receives different few-shots, enhancing ensemble diversity and coverage of the query space. We evaluate our framework on a real-world major bank dataset as well as public benchmark datasets (LCQMC and FiQA), demonstrating significant improvements over single-agent approaches across multiple metrics, including a 14% increase in Top-1 accuracy, an 18% increase in Top-5 accuracy, and a 12% improvement in Mean Reciprocal Rank on our dataset, and similar gains on public benchmarks when compared with traditional and single-agent annotation techniques. Our framework is particularly effective at handling ambiguous queries, making it well-suited for deployment in production banking applications while showing strong generalization capabilities across different domains and languages.

Method: Probing and leveraging models’ safety decision boundaries, using steering vectors in representation space to identify boundary-aligned prompts. RASS framework automates prompt generation and selection targeting overrefusal near safety boundaries.

Result: The approach provides precise and interpretable view of model safety decisions, extends to multilingual scenarios, and enables effective mitigation of overrefusal. MORBench evaluation set created for robust safety assessment across languages.

Conclusion: RASS successfully addresses overrefusal by strategically targeting safety boundary regions, offering an automated solution that improves model helpfulness while maintaining safety standards across multiple languages.

Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet they often refuse to answer legitimate queries–a phenomenon known as overrefusal. Overrefusal typically stems from over-conservative safety alignment, causing models to treat many reasonable prompts as potentially risky. To systematically understand this issue, we probe and leverage the models’ safety decision boundaries to analyze and mitigate overrefusal. Our findings reveal that overrefusal is closely tied to misalignment at these boundary regions, where models struggle to distinguish subtle differences between benign and harmful content. Building on these insights, we present RASS, an automated framework for prompt generation and selection that strategically targets overrefusal prompts near the safety boundary. By harnessing steering vectors in the representation space, RASS efficiently identifies and curates boundary-aligned prompts, enabling more effective and targeted mitigation of overrefusal. This approach not only provides a more precise and interpretable view of model safety decisions but also seamlessly extends to multilingual scenarios. We have explored the safety decision boundaries of various LLMs and construct the MORBench evaluation set to facilitate robust assessment of model safety and helpfulness across multiple languages. Code and datasets are available at https://github.com/Master-PLC/RASS.

Method: Developed methods for assessing LLM response accuracy and correct answer correlation on PRONTOQA steamroller reasoning problems. Evaluated State of the Art models from December 2023 and August 2024, tracking completion tokens to analyze reasoning improvements.

Result: Progress in improving LLM reasoning abilities has stalled over the nine month period. Almost all improvement since GPT-4 can be attributed to hidden system prompts or training models to automatically use Chain of Thought prompting. Current frontier LLMs perform best with bottom-up reasoning strategies, but show low positive correlation between correct reasoning and correct conclusions.

Conclusion: LLM reasoning capability improvements have plateaued, with recent gains coming from prompt engineering rather than fundamental reasoning advances. The disconnect between correct reasoning steps and final conclusions suggests limitations in current LLM reasoning architectures.

Abstract: Empirical methods to examine the capability of Large Language Models (LLMs) to use Automated Theorem Prover (ATP) reasoning strategies are studied. We evaluate the performance of State of the Art models from December 2023 and August 2024 on PRONTOQA steamroller reasoning problems. For that, we develop methods for assessing LLM response accuracy and correct answer correlation. Our results show that progress in improving LLM reasoning abilities has stalled over the nine month period. By tracking completion tokens, we show that almost all improvement in reasoning ability since GPT-4 was released can be attributed to either hidden system prompts or the training of models to automatically use generic Chain of Thought prompting strategies. Among the ATP reasoning strategies tried, we found that current frontier LLMs are best able to follow the bottom-up (also known as forward-chaining) strategy. A low positive correlation was found between an LLM response containing correct reasoning and arriving at the correct conclusion.

Method: Two-step approach: pre-screening followed by assessment phase using RAG (Retrieval-Augmented Generation) architecture. The tool analyzes AI systems against AI Act criteria to determine risk classification and retrieves relevant legal articles.

Result: Qualitative evaluation shows promising results - correctly predicts risk levels and retrieves relevant articles across three distinct semantic groups. The tool’s reasoning effectively compares systems against high-risk criteria as defined in the AI Act.

Conclusion: The TAI Scan Tool successfully provides automated AI Act compliance assessment with minimal input, demonstrating accurate risk classification and relevant legal article retrieval, making it valuable for organizations navigating AI regulatory requirements.

Abstract: This paper introduces the TAI Scan Tool, a RAG-based TAI self-assessment tool with minimalistic input. The current version of the tool supports the legal TAI assessment, with a particular emphasis on facilitating compliance with the AI Act. It involves a two-step approach with a pre-screening and an assessment phase. The assessment output of the system includes insight regarding the risk-level of the AI system according to the AI Act, while at the same time retrieving relevant articles to aid with compliance and notify on their obligations. Our qualitative evaluation using use-case scenarios yields promising results, correctly predicting risk levels while retrieving relevant articles across three distinct semantic groups. Furthermore, interpretation of results shows that the tool’s reasoning relies on comparison with the setting of high-risk systems, a behaviour attributed to their deployment requiring careful consideration, and therefore frequently presented within the AI Act.

Method: Analyzed internal representations of Llama-3.1-8B-Instruct and Gemma-2-9B-Instruct using linear probes and Sparse Autoencoders (SAEs) on factual verification tasks, comparing deceptive vs truthful/neutral instructions.

Result: Deceptive instructions cause significant representational shifts concentrated in early-to-mid layers, with identifiable SAE features highly sensitive to deception. Truthful and deceptive representations form distinct subspaces.

Conclusion: The study reveals layer-wise and feature-level signatures of deception, providing insights for detecting and mitigating instructed dishonesty in LLMs through internal representation analysis.

Abstract: Large language models (LLMs) tend to follow maliciously crafted instructions to generate deceptive responses, posing safety challenges. How deceptive instructions alter the internal representations of LLM compared to truthful ones remains poorly understood beyond output analysis. To bridge this gap, we investigate when and how these representations ``flip’’, such as from truthful to deceptive, under deceptive versus truthful/neutral instructions. Analyzing the internal representations of Llama-3.1-8B-Instruct and Gemma-2-9B-Instruct on a factual verification task, we find the model’s instructed True/False output is predictable via linear probes across all conditions based on the internal representation. Further, we use Sparse Autoencoders (SAEs) to show that the Deceptive instructions induce significant representational shifts compared to Truthful/Neutral representations (which are similar), concentrated in early-to-mid layers and detectable even on complex datasets. We also identify specific SAE features highly sensitive to deceptive instruction and use targeted visualizations to confirm distinct truthful/deceptive representational subspaces. % Our analysis pinpoints layer-wise and feature-level correlates of instructed dishonesty, offering insights for LLM detection and control. Our findings expose feature- and layer-level signatures of deception, offering new insights for detecting and mitigating instructed dishonesty in LLMs.

Method: Using linear probes on LLM internal activations to detect deception in generated responses, testing models from 1.5B to 14B parameters including llama, qwen, and DeepSeek-r1 variants.

Result: Probes achieve >90% accuracy in detecting deception, with smaller models (1.5B) at chance accuracy, larger models (7B+) reaching 70-80%, and reasoning variants exceeding 90%. Layer-wise accuracy shows three-stage pattern, and multiple linear deception directions (20-100) are found across models.

Conclusion: Linear probes are highly effective at detecting deception in LLM responses, with detection capability scaling with model size and complexity, revealing structured patterns in how deception is encoded across model layers.

Abstract: Sophisticated instrumentation for AI systems might have indicators that signal misalignment from human values, not unlike a “check engine” light in cars. One such indicator of misalignment is deceptiveness in generated responses. Future AI instrumentation may have the ability to detect when an LLM generates deceptive responses while reasoning about seemingly plausible but incorrect answers to factual questions. In this work, we demonstrate that linear probes on LLMs internal activations can detect deception in their responses with extremely high accuracy. Our probes reach a maximum of greater than 90% accuracy in distinguishing between deceptive and non-deceptive arguments generated by llama and qwen models ranging from 1.5B to 14B parameters, including their DeepSeek-r1 finetuned variants. We observe that probes on smaller models (1.5B) achieve chance accuracy at detecting deception, while larger models (greater than 7B) reach 70-80%, with their reasoning counterparts exceeding 90%. The layer-wise probe accuracy follows a three-stage pattern across layers: near-random (50%) in early layers, peaking in middle layers, and slightly declining in later layers. Furthermore, using an iterative null space projection approach, we find multitudes of linear directions that encode deception, ranging from 20 in Qwen 3B to nearly 100 in DeepSeek 7B and Qwen 14B models.

Method: Agentic framework with specialized agents for planning, crime type detection, external intelligence gathering, and compliance validation. Features dynamic memory management, AI-Privacy Guard, and real-time validation using Agent-as-a-Judge paradigm with human-in-the-loop workflow.

Result: Demonstrated versatility across complex financial crime scenarios, showing ability to streamline SAR drafting, align with regulatory expectations, and enable compliance teams to focus on higher-order analytical work.

Conclusion: Marks the beginning of a new era in compliance reporting, bringing AI agent benefits to regulatory processes for scalable, reliable, and transparent SAR generation.

Abstract: Generating regulatorily compliant Suspicious Activity Report (SAR) remains a high-cost, low-scalability bottleneck in Anti-Money Laundering (AML) workflows. While large language models (LLMs) offer promising fluency, they suffer from factual hallucination, limited crime typology alignment, and poor explainability – posing unacceptable risks in compliance-critical domains. This paper introduces Co-Investigator AI, an agentic framework optimized to produce Suspicious Activity Reports (SARs) significantly faster and with greater accuracy than traditional methods. Drawing inspiration from recent advances in autonomous agent architectures, such as the AI Co-Scientist, our approach integrates specialized agents for planning, crime type detection, external intelligence gathering, and compliance validation. The system features dynamic memory management, an AI-Privacy Guard layer for sensitive data handling, and a real-time validation agent employing the Agent-as-a-Judge paradigm to ensure continuous narrative quality assurance. Human investigators remain firmly in the loop, empowered to review and refine drafts in a collaborative workflow that blends AI efficiency with domain expertise. We demonstrate the versatility of Co-Investigator AI across a range of complex financial crime scenarios, highlighting its ability to streamline SAR drafting, align narratives with regulatory expectations, and enable compliance teams to focus on higher-order analytical work. This approach marks the beginning of a new era in compliance reporting – bringing the transformative benefits of AI agents to the core of regulatory processes and paving the way for scalable, reliable, and transparent SAR generation.

Method: Uses structured LLM reasoning to generate constraint relaxation strategies with a unified triple-representation scheme. Implements a bidirectional coevolution mechanism combining Evolutionary Algorithms for local refinement and Monte Carlo Tree Search for global strategy exploration.

Result: Comprehensive experiments on three challenging NCOP benchmarks show AutoCO achieves consistent effectiveness and superior performance over baseline methods.

Conclusion: AutoCO represents a revolutionary approach to NCOP resolution through learning to relax with LLMs, providing automated constraint optimization that balances intensification and diversification in fragmented solution spaces.

Abstract: Nonlinear Combinatorial Optimization Problems (NCOPs) present a formidable computational hurdle in practice, as their nonconvex nature gives rise to multi-modal solution spaces that defy efficient optimization. Traditional constraint relaxation approaches rely heavily on expert-driven, iterative design processes that lack systematic automation and scalable adaptability. While recent Large Language Model (LLM)-based optimization methods show promise for autonomous problem-solving, they predominantly function as passive constraint validators rather than proactive strategy architects, failing to handle the sophisticated constraint interactions inherent to NCOPs.To address these limitations, we introduce the first end-to-end \textbf{Auto}mated \textbf{C}onstraint \textbf{O}ptimization (AutoCO) method, which revolutionizes NCOPs resolution through learning to relax with LLMs.Specifically, we leverage structured LLM reasoning to generate constraint relaxation strategies, which are dynamically evolving with algorithmic principles and executable code through a unified triple-representation scheme. We further establish a novel bidirectional (global-local) coevolution mechanism that synergistically integrates Evolutionary Algorithms for intensive local refinement with Monte Carlo Tree Search for systematic global strategy space exploration, ensuring optimal balance between intensification and diversification in fragmented solution spaces. Finally, comprehensive experiments on three challenging NCOP benchmarks validate AutoCO’s consistent effectiveness and superior performance over the baselines.

Method: Engineered proxy-anomalies through structured perturbations of healthy spectrograms for validation set, enabling effective model selection without real faulty samples.

Result: Experimental evaluations on five fault types (Imbalance, Modulation, Whine, Wind, PWM) show the proposed method significantly outperforms conventional model selection strategies.

Conclusion: The proxy-anomaly approach provides an effective solution for model selection in unsupervised anomaly detection, validated on a high-fidelity electric vehicle sound dataset made publicly available.

Abstract: The detection of anomalies in automotive cabin sounds is critical for ensuring vehicle quality and maintaining passenger comfort. In many real-world settings, this task is more appropriately framed as an unsupervised learning problem rather than the supervised case due to the scarcity or complete absence of labeled faulty data. In such an unsupervised setting, the model is trained exclusively on healthy samples and detects anomalies as deviations from normal behavior. However, in the absence of labeled faulty samples for validation and the limited reliability of commonly used metrics, such as validation reconstruction error, effective model selection remains a significant challenge. To overcome these limitations, a domain-knowledge-informed approach for model selection is proposed, in which proxy-anomalies engineered through structured perturbations of healthy spectrograms are used in the validation set to support model selection. The proposed methodology is evaluated on a high-fidelity electric vehicle dataset comprising healthy and faulty cabin sounds across five representative fault types viz., Imbalance, Modulation, Whine, Wind, and Pulse Width Modulation. This dataset, generated using advanced sound synthesis techniques, and validated via expert jury assessments, has been made publicly available to facilitate further research. Experimental evaluations on the five fault cases demonstrate the selection of optimal models using proxy-anomalies, significantly outperform conventional model selection strategies.

Method: Signal-dependent framework combining multiple binaural filters in an online manner using implicit localization, enabling dynamic spatial audio rendering that adapts to continuous talker motion.

Result: Enables real-time tracking and enhancement of moving sound sources while preserving natural binaural cues, allowing users to emphasize or suppress sounds from selected directions.

Conclusion: Provides a flexible, array-agnostic solution for spatial audio capture and personalized playback in augmented/virtual reality applications, supporting speech focus, noise reduction, and world-locked audio.

Abstract: We propose a novel mixture of experts framework for field-of-view enhancement in binaural signal matching. Our approach enables dynamic spatial audio rendering that adapts to continuous talker motion, allowing users to emphasize or suppress sounds from selected directions while preserving natural binaural cues. Unlike traditional methods that rely on explicit direction-of-arrival estimation or operate in the Ambisonics domain, our signal-dependent framework combines multiple binaural filters in an online manner using implicit localization. This allows for real-time tracking and enhancement of moving sound sources, supporting applications such as speech focus, noise reduction, and world-locked audio in augmented and virtual reality. The method is agnostic to array geometry offering a flexible solution for spatial audio capture and personalized playback in next-generation consumer audio devices.

Method: Two-stage self-supervised framework: Video2RMS extracts frame-level intensity envelopes, and RMS2Sound uses RMS-ControlNet with a pretrained text-to-audio model. Includes RMS discretization and semantic timbre prompts (audio or text).

Result: Achieves state-of-the-art performance in audio-visual alignment and controllability for sound timing, intensity, timbre, and nuance. Outperforms previous methods without requiring human annotations.

Conclusion: RMS serves as an effective temporal event feature for video-to-sound generation. The annotation-free approach provides excellent alignment and controllability, making Foley sound synthesis more accessible and efficient.

Abstract: Foley sound synthesis is crucial for multimedia production, enhancing user experience by synchronizing audio and video both temporally and semantically. Recent studies on automating this labor-intensive process through video-to-sound generation face significant challenges. Systems lacking explicit temporal features suffer from poor alignment and controllability, while timestamp-based models require costly and subjective human annotation. We propose Video-Foley, a video-to-sound system using Root Mean Square (RMS) as an intuitive condition with semantic timbre prompts (audio or text). RMS, a frame-level intensity envelope closely related to audio semantics, acts as a temporal event feature to guide audio generation from video. The annotation-free self-supervised learning framework consists of two stages, Video2RMS and RMS2Sound, incorporating novel ideas including RMS discretization and RMS-ControlNet with a pretrained text-to-audio model. Our extensive evaluation shows that Video-Foley achieves state-of-the-art performance in audio-visual alignment and controllability for sound timing, intensity, timbre, and nuance. Source code, model weights and demos are available on our companion website. (https://jnwnlee.github.io/video-foley-demo)

Method: APSS extracts amplitude and phase spectra from mixed speech, fuses them into joint representations using a feature combiner, processes them with time-frequency Transformers to capture dependencies, then uses parallel separators to estimate individual speaker spectra which are reconstructed via inverse STFT.

Result: APSS outperforms both time-domain separation methods and implicit-phase-estimation time-frequency approaches. It achieves stable competitive results across multiple datasets, demonstrating strong generalization capability.

Conclusion: Explicit parallel estimation of amplitude and phase spectra significantly improves speech separation performance, making APSS a highly effective and practical solution with strong generalization across different datasets.

Abstract: This paper proposes APSS, a novel neural speech separation model with parallel amplitude and phase spectrum estimation. Unlike most existing speech separation methods, the APSS distinguishes itself by explicitly estimating the phase spectrum for more complete and accurate separation. Specifically, APSS first extracts the amplitude and phase spectra from the mixed speech signal. Subsequently, the extracted amplitude and phase spectra are fused by a feature combiner into joint representations, which are then further processed by a deep processor with time-frequency Transformers to capture temporal and spectral dependencies. Finally, leveraging parallel amplitude and phase separators, the APSS estimates the respective spectra for each speaker from the resulting features, which are then combined via inverse short-time Fourier transform (iSTFT) to reconstruct the separated speech signals. Experimental results indicate that APSS surpasses both time-domain separation methods and implicit-phase-estimation-based time-frequency approaches. Also, APSS achieves stable and competitive results on multiple datasets, highlighting its strong generalization capability and practical applicability.

Method: Proposes OS-SCL technique that perturbs features in embedding space and uses one-stage noisy supervised contrastive learning. Also introduces TFgram time-frequency feature extracted from raw audio to capture critical detection information.

Result: Achieved 94.64% AUC, 88.42% pAUC, and 89.24% mAUC using only Log-Mel features. With TFgram feature, achieved 95.71% AUC, 90.23% pAUC, and 91.23% mAUC on DCASE 2020 Challenge Task 2.

Conclusion: OS-SCL effectively reduces false alarms in anomalous sound detection and the proposed TFgram feature significantly improves detection performance, demonstrating state-of-the-art results on benchmark dataset.

Abstract: Unsupervised anomalous sound detection aims to detect unknown anomalous sounds by training a model using only normal audio data. Despite advancements in self-supervised methods, the issue of frequent false alarms when handling samples of the same type from different machines remains unresolved. This paper introduces a novel training technique called one-stage supervised contrastive learning (OS-SCL), which significantly addresses this problem by perturbing features in the embedding space and employing a one-stage noisy supervised contrastive learning approach. On the DCASE 2020 Challenge Task 2, it achieved 94.64% AUC, 88.42% pAUC, and 89.24% mAUC using only Log-Mel features. Additionally, a time-frequency feature named TFgram is proposed, which is extracted from raw audio. This feature effectively captures critical information for anomalous sound detection, ultimately achieving 95.71% AUC, 90.23% pAUC, and 91.23% mAUC. The source code is available at: \underline{www.github.com/huangswt/OS-SCL}.

Method: Novel rectified flow matching-based diffusion framework trained on a constructed dataset featuring overlapping multi-event audio for complex editing scenarios.

Result: Achieves faithful semantic alignment without requiring auxiliary captions or masks, while maintaining competitive editing quality across evaluation metrics.

Conclusion: The proposed framework provides an efficient and practical solution for text-guided audio editing that handles complex scenarios effectively without additional requirements.

Abstract: Diffusion models have shown remarkable progress in text-to-audio generation. However, text-guided audio editing remains in its early stages. This task focuses on modifying the target content within an audio signal while preserving the rest, thus demanding precise localization and faithful editing according to the text prompt. Existing training-based and zero-shot methods that rely on full-caption or costly optimization often struggle with complex editing or lack practicality. In this work, we propose a novel end-to-end efficient rectified flow matching-based diffusion framework for audio editing, and construct a dataset featuring overlapping multi-event audio to support training and benchmarking in complex scenarios. Experiments show that our model achieves faithful semantic alignment without requiring auxiliary captions or masks, while maintaining competitive editing quality across metrics.

Method: Comprehensive evaluation of various CNN architectures (PANNs 1D/2D models, ConvNeXt-based model, MobileNetV3, CNN9, CNN13) converted to ONNX format, with 24-hour continuous inference sessions to assess performance stability.

Result: Appropriate model selection and optimization can maintain consistent inference latency and effectively manage thermal behavior over extended periods on Raspberry Pi.

Conclusion: The findings provide valuable insights for deploying audio tagging models in real-world edge computing scenarios with resource constraints.

Abstract: Convolutional Neural Networks (CNNs) have demonstrated exceptional performance in audio tagging tasks. However, deploying these models on resource-constrained devices like the Raspberry Pi poses challenges related to computational efficiency and thermal management. In this paper, a comprehensive evaluation of multiple convolutional neural network (CNN) architectures for audio tagging on the Raspberry Pi is conducted, encompassing all 1D and 2D models from the Pretrained Audio Neural Networks (PANNs) framework, a ConvNeXt-based model adapted for audio classification, as well as MobileNetV3 architectures. In addition, two PANNs-derived networks, CNN9 and CNN13, recently proposed, are also evaluated. To enhance deployment efficiency and portability across diverse hardware platforms, all models are converted to the Open Neural Network Exchange (ONNX) format. Unlike previous works that focus on a single model, our analysis encompasses a broader range of architectures and involves continuous 24-hour inference sessions to assess performance stability. Our experiments reveal that, with appropriate model selection and optimization, it is possible to maintain consistent inference latency and manage thermal behavior effectively over extended periods. These findings provide valuable insights for deploying audio tagging models in real-world edge computing scenarios.

Method: Uses a quantized melodic bottleneck with chromagram and VQ-VAE to extract discrete timbre-invariant melody representations, followed by a flow-matching model to generate accompaniment conditioned on these robust codes.

Result: Achieves competitive performance on separated-vocal benchmarks while significantly outperforming baselines on clean studio vocals and solo instrumental tracks, demonstrating superior generalization.

Conclusion: Provides a qualitative leap in generalization, enabling robust accompaniment for instruments where existing models fail, paving the way for more versatile music co-creation tools.

Abstract: Singing Accompaniment Generation (SAG) is the process of generating instrumental music for a given clean vocal input. However, existing SAG techniques use source-separated vocals as input and overfit to separation artifacts. This creates a critical train-test mismatch, leading to failure on clean, real-world vocal inputs. We introduce AnyAccomp, a framework that resolves this by decoupling accompaniment generation from source-dependent artifacts. AnyAccomp first employs a quantized melodic bottleneck, using a chromagram and a VQ-VAE to extract a discrete and timbre-invariant representation of the core melody. A subsequent flow-matching model then generates the accompaniment conditioned on these robust codes. Experiments show AnyAccomp achieves competitive performance on separated-vocal benchmarks while significantly outperforming baselines on generalization test sets of clean studio vocals and, notably, solo instrumental tracks. This demonstrates a qualitative leap in generalization, enabling robust accompaniment for instruments - a task where existing models completely fail - and paving the way for more versatile music co-creation tools. Demo audio and code: https://anyaccomp.github.io

Method: Analyzed three spike encoding methods (TAE, SF, MW) across three environmental sound datasets (ESC10, UrbanSound8K, TAU Urban Acoustic Scenes) using multiband analysis and downstream classification with a standard SNN.

Result: TAE consistently outperformed SF and MW in reconstruction quality per frequency band and per class, had the lowest spike firing rates (indicating superior energy efficiency), and achieved the best performance in environmental sound classification.

Conclusion: This work provides foundational insights and comparative benchmarks to guide spike encoder selection for neuromorphic environmental sound processing, with TAE demonstrating superior performance across multiple metrics.

Abstract: Spiking Neural Networks (SNNs) offer energy efficient processing suitable for edge applications, but conventional sensor data must first be converted into spike trains for neuromorphic processing. Environmental sound, including urban soundscapes, poses challenges due to variable frequencies, background noise, and overlapping acoustic events, while most spike based audio encoding research has focused on speech. This paper analyzes three spike encoding methods, Threshold Adaptive Encoding (TAE), Step Forward (SF), and Moving Window (MW) across three datasets: ESC10, UrbanSound8K, and TAU Urban Acoustic Scenes. Our multiband analysis shows that TAE consistently outperforms SF and MW in reconstruction quality, both per frequency band and per class across datasets. Moreover, TAE yields the lowest spike firing rates, indicating superior energy efficiency. For downstream environmental sound classification with a standard SNN, TAE also achieves the best performance among the compared encoders. Overall, this work provides foundational insights and a comparative benchmark to guide the selection of spike encoders for neuromorphic environmental sound processing.

Method: ECHO integrates an advanced band-split architecture with frequency positional embeddings for spectral localization across arbitrary sampling configurations, and incorporates sliding patches to support variable-length inputs without padding or cropping, producing embeddings that retain temporal and spectral fidelity.

Result: Experimental results demonstrate consistent state-of-the-art performance in machine signal anomaly detection and fault classification across various datasets including DCASE task 2 challenges (2020-2025) and industrial signal corpora.

Conclusion: ECHO confirms the effectiveness and generalization capability of the proposed model for machine signal processing, and the model has been open-sourced for public use.

Abstract: Pre-trained foundation models have demonstrated remarkable success in audio, vision and language, yet their potential for general machine signal modeling with arbitrary sampling rates-covering acoustic, vibration, and other industrial sensor data-remains under-explored. In this work, we propose a novel foundation model ECHO that integrates an advanced band-split architecture with frequency positional embeddings, enabling spectral localization across arbitrary sampling configurations. Moreover, the model incorporates sliding patches to support inputs of variable length without padding or cropping, producing a concise embedding that retains both temporal and spectral fidelity and naturally extends to streaming scenarios. We evaluate our method on various kinds of machine signal datasets, including previous DCASE task 2 challenges (2020-2025), and widely-used industrial signal corpora. Experimental results demonstrate consistent state-of-the-art performance in machine signal anomaly detection and fault classification, confirming the effectiveness and generalization capability of the proposed model. We open-sourced ECHO on https://github.com/yucongzh/ECHO.

Method: Proposes a framework with two key strategies: error-aware curriculum learning that organizes samples by difficulty level, and guided thought dropout mechanism that focuses reasoning capabilities on challenging cases.

Result: Achieves 73.80% on MMAU-mini and sets a new state-of-the-art of 64.30% on MMAR benchmark, demonstrating strong generalization in multimodal audio-language understanding.

Conclusion: Omni-CLST effectively leverages existing high-quality audio QA data through curriculum learning and selective reasoning, achieving superior performance and robust generalization without requiring new dataset construction.

Abstract: With the rapid progress of large audio-language models (LALMs), audio question answering (AQA) has emerged as a challenging task requiring both fine-grained audio understanding and complex reasoning. While current methods mainly rely on constructing new datasets via captioning or reasoning traces, existing high-quality AQA data remains underutilized. To address this, we propose Omni-CLST, an error-aware Curriculum Learning framework with guided Selective Chain-of-Thought. The framework efficiently leverages existing high-quality dataset through two key strategies: an error-aware curriculum that organizes samples by difficulty, and a guided thought dropout mechanism that focuses reasoning on challenging cases. Experiments show that Omni-CLST achieves 73.80% on MMAU-mini and a new state of the art of 64.30% on MMAR, demonstrating robust generalization in multimodal audio-language understanding.

Method: Deep learning architecture combining physics-informed neural networks (PINNs) with conservation laws, using automatic differentiation to enforce physics constraints and adaptive loss weighting to balance data fidelity with physical consistency for both ODE and PDE systems.

Result: Achieved 98.9% correlation for 1D temporal dynamics (loss: 0.0219, MAE: 0.0184), captured complex spiral waves in 2D systems (loss: 4.7656, pattern correlation: 0.94), maintained conservation law adherence within 0.5%, and showed 10-50x computational speedup compared to numerical solvers.

Conclusion: USPIL provides a transformative tool for ecological forecasting and conservation planning, establishing physics-informed deep learning as a powerful paradigm that enables mechanistic understanding, parameter discovery, and multi-scale modeling capabilities.

Abstract: Ecological systems exhibit complex multi-scale dynamics that challenge traditional modeling. New methods must capture temporal oscillations and emergent spatiotemporal patterns while adhering to conservation principles. We present the Unified Spatiotemporal Physics-Informed Learning (USPIL) framework, a deep learning architecture integrating physics-informed neural networks (PINNs) and conservation laws to model predator-prey dynamics across dimensional scales. The framework provides a unified solution for both ordinary (ODE) and partial (PDE) differential equation systems, describing temporal cycles and reaction-diffusion patterns within a single neural network architecture. Our methodology uses automatic differentiation to enforce physics constraints and adaptive loss weighting to balance data fidelity with physical consistency. Applied to the Lotka-Volterra system, USPIL achieves 98.9% correlation for 1D temporal dynamics (loss: 0.0219, MAE: 0.0184) and captures complex spiral waves in 2D systems (loss: 4.7656, pattern correlation: 0.94). Validation confirms conservation law adherence within 0.5% and shows a 10-50x computational speedup for inference compared to numerical solvers. USPIL also enables mechanistic understanding through interpretable physics constraints, facilitating parameter discovery and sensitivity analysis not possible with purely data-driven methods. Its ability to transition between dimensional formulations opens new avenues for multi-scale ecological modeling. These capabilities make USPIL a transformative tool for ecological forecasting, conservation planning, and understanding ecosystem resilience, establishing physics-informed deep learning as a powerful and scientifically rigorous paradigm.

Method: Conducted 360 controlled experiments across three benchmark datasets (MNIST, CIFAR-10, CIFAR-100) using eight popular optimizers with 15 random seeds each, using CodeCarbon for precise energy tracking on Apple M1 Pro hardware.

Result: Substantial trade-offs between training speed, accuracy, and environmental impact that vary across datasets and model complexity. AdamW and NAdam identified as consistently efficient choices, while SGD demonstrates superior performance on complex datasets despite higher emissions.

Conclusion: Provides actionable insights for practitioners seeking to balance performance and sustainability in machine learning workflows through optimizer selection.

Abstract: As machine learning models grow increasingly complex and computationally demanding, understanding the environmental impact of training decisions becomes critical for sustainable AI development. This paper presents a comprehensive empirical study investigating the relationship between optimizer choice and energy efficiency in neural network training. We conducted 360 controlled experiments across three benchmark datasets (MNIST, CIFAR-10, CIFAR-100) using eight popular optimizers (SGD, Adam, AdamW, RMSprop, Adagrad, Adadelta, Adamax, NAdam) with 15 random seeds each. Using CodeCarbon for precise energy tracking on Apple M1 Pro hardware, we measured training duration, peak memory usage, carbon dioxide emissions, and final model performance. Our findings reveal substantial trade-offs between training speed, accuracy, and environmental impact that vary across datasets and model complexity. We identify AdamW and NAdam as consistently efficient choices, while SGD demonstrates superior performance on complex datasets despite higher emissions. These results provide actionable insights for practitioners seeking to balance performance and sustainability in machine learning workflows.

Method: Hybrid DeepONet-Transolver framework trained on nonlinear FEA simulation data from 254 unique bottle designs per geometry family (2-parameter and 4-parameter families).

Result: Achieves mean relative L² errors of 2.5-13% for displacement fields and ~2.4% for reaction forces. Point-wise displacement errors of 10⁻⁴-10⁻³, accurately capturing buckling behavior across diverse geometries.

Conclusion: The framework provides a scalable, computationally efficient surrogate for multi-task predictions in computational mechanics and rapid design evaluation applications.

Abstract: Neural surrogates and operator networks for solving partial differential equation (PDE) problems have attracted significant research interest in recent years. However, most existing approaches are limited in their ability to generalize solutions across varying non-parametric geometric domains. In this work, we address this challenge in the context of Polyethylene Terephthalate (PET) bottle buckling analysis, a representative packaging design problem conventionally solved using computationally expensive finite element analysis (FEA). We introduce a hybrid DeepONet-Transolver framework that simultaneously predicts nodal displacement fields and the time evolution of reaction forces during top load compression. Our methodology is evaluated on two families of bottle geometries parameterized by two and four design variables. Training data is generated using nonlinear FEA simulations in Abaqus for 254 unique designs per family. The proposed framework achieves mean relative $L^{2}$ errors of 2.5-13% for displacement fields and approximately 2.4% for time-dependent reaction forces for the four-parameter bottle family. Point-wise error analyses further show absolute displacement errors on the order of $10^{-4}$-$10^{-3}$, with the largest discrepancies confined to localized geometric regions. Importantly, the model accurately captures key physical phenomena, such as buckling behavior, across diverse bottle geometries. These results highlight the potential of our framework as a scalable and computationally efficient surrogate, particularly for multi-task predictions in computational mechanics and applications requiring rapid design evaluation.

Method: Developed AERIS (pixel-level Swin diffusion transformer) and SWiPe technique that combines window parallelism with sequence/pipeline parallelism to efficiently scale window-based transformers without additional communication costs.

Result: Achieved 10.21 ExaFLOPS sustained performance (11.21 peak) on Aurora with 95.5% weak scaling efficiency and 81.6% strong scaling efficiency. Outperformed IFS ENS and maintained stable performance on seasonal scales up to 90 days.

Conclusion: Billion-parameter diffusion models like AERIS show significant potential for advancing weather and climate prediction capabilities through improved scaling and performance.

Abstract: Generative machine learning offers new opportunities to better understand complex Earth system dynamics. Recent diffusion-based methods address spectral biases and improve ensemble calibration in weather forecasting compared to deterministic methods, yet have so far proven difficult to scale stably at high resolutions. We introduce AERIS, a 1.3 to 80B parameter pixel-level Swin diffusion transformer to address this gap, and SWiPe, a generalizable technique that composes window parallelism with sequence and pipeline parallelism to shard window-based transformers without added communication cost or increased global batch size. On Aurora (10,080 nodes), AERIS sustains 10.21 ExaFLOPS (mixed precision) and a peak performance of 11.21 ExaFLOPS with $1 \times 1$ patch size on the 0.25{\deg} ERA5 dataset, achieving 95.5% weak scaling efficiency, and 81.6% strong scaling efficiency. AERIS outperforms the IFS ENS and remains stable on seasonal scales to 90 days, highlighting the potential of billion-parameter diffusion models for weather and climate prediction.

Method: Linear Algorithm for Meta-Learning (LAMeL) that uses meta-learning to identify shared model parameters across related chemical prediction tasks, learning a common functional manifold that serves as an informed starting point for new tasks.

Result: Performance improvements ranging from 1.1- to 25-fold over standard ridge regression, consistently outperforming or matching traditional linear methods across various chemical domains.

Conclusion: LAMeL provides a reliable tool for chemical property prediction where both accuracy and interpretability are critical, effectively leveraging shared knowledge across tasks without requiring shared data.

Abstract: Chemists in search of structure-property relationships face great challenges due to limited high quality, concordant datasets. Machine learning (ML) has significantly advanced predictive capabilities in chemical sciences, but these modern data-driven approaches have increased the demand for data. In response to the growing demand for explainable AI (XAI) and to bridge the gap between predictive accuracy and human comprehensibility, we introduce LAMeL - a Linear Algorithm for Meta-Learning that preserves interpretability while improving the prediction accuracy across multiple properties. While most approaches treat each chemical prediction task in isolation, LAMeL leverages a meta-learning framework to identify shared model parameters across related tasks, even if those tasks do not share data, allowing it to learn a common functional manifold that serves as a more informed starting point for new unseen tasks. Our method delivers performance improvements ranging from 1.1- to 25-fold over standard ridge regression, depending on the domain of the dataset. While the degree of performance enhancement varies across tasks, LAMeL consistently outperforms or matches traditional linear methods, making it a reliable tool for chemical property prediction where both accuracy and interpretability are critical.

Method: Systematic analysis using Hateful Memes Challenge dataset on 500 samples, multi-phase investigation to probe reasoning and failure modes, with quantitative validation of 144 content policy refusals.

Result: Identified ‘Unimodal Bottleneck’ flaw where multimodal reasoning is preempted by context-blind safety filters (50% visual, 50% textual triggers). Brittle safety system blocks both high-risk imagery and benign meme formats.

Conclusion: Exposes fundamental tension between capability and safety in LMMs, highlighting need for more integrated, context-aware alignment strategies for safe and effective AI deployment.

Abstract: As Large Multimodal Models (LMMs) become integral to daily digital life, understanding their safety architectures is a critical problem for AI Alignment. This paper presents a systematic analysis of OpenAI’s GPT-4o mini, a globally deployed model, on the difficult task of multimodal hate speech detection. Using the Hateful Memes Challenge dataset, we conduct a multi-phase investigation on 500 samples to probe the model’s reasoning and failure modes. Our central finding is the experimental identification of a “Unimodal Bottleneck,” an architectural flaw where the model’s advanced multimodal reasoning is systematically preempted by context-blind safety filters. A quantitative validation of 144 content policy refusals reveals that these overrides are triggered in equal measure by unimodal visual 50% and textual 50% content. We further demonstrate that this safety system is brittle, blocking not only high-risk imagery but also benign, common meme formats, leading to predictable false positives. These findings expose a fundamental tension between capability and safety in state-of-the-art LMMs, highlighting the need for more integrated, context-aware alignment strategies to ensure AI systems can be deployed both safely and effectively.

Method: Transform log entries into contextual vector representations using semantic embedding techniques, then use a sequence-aware neural network trained on normal operational data to assign real-time anomaly scores to events.

Result: The framework successfully flags deviations from baseline behavior, allowing operators to quickly identify critical event sequences that lead to system failures.

Conclusion: The automated fault analysis approach using semantic embeddings and neural networks provides an effective method for real-time anomaly detection and failure prediction in complex control systems like the ALS.

Abstract: This paper introduces an automated fault analysis framework for the Advanced Light Source (ALS) that processes real-time event logs from its EPICS control system. By treating log entries as natural language, we transform them into contextual vector representations using semantic embedding techniques. A sequence-aware neural network, trained on normal operational data, assigns a real-time anomaly score to each event. This method flags deviations from baseline behavior, enabling operators to rapidly identify the critical event sequences that precede complex system failures.

Method: Leverages Differential Privacy framework without fine-tuning, performs inference on private records and aggregates per-token output distributions, uses blending operation to combine private and public inference for enhanced utility.

Result: Outperforms previous state-of-the-art methods on in-context-learning tasks, generates longer and coherent synthetic text while maintaining privacy guarantees.

Conclusion: The approach provides a promising direction for privacy-preserving text generation with high utility, offering worst-case theoretical bounds on information leakage.

Abstract: Large language models (LLMs) have significantly transformed natural language understanding and generation, but they raise privacy concerns due to potential exposure of sensitive information. Studies have highlighted the risk of information leakage, where adversaries can extract sensitive information embedded in the prompts. In this work, we introduce a novel private prediction framework for generating high-quality synthetic text with strong privacy guarantees. Our approach leverages the Differential Privacy (DP) framework to ensure worst-case theoretical bounds on information leakage without requiring any fine-tuning of the underlying models.The proposed method performs inference on private records and aggregates the resulting per-token output distributions. This enables the generation of longer and coherent synthetic text while maintaining privacy guarantees. Additionally, we propose a simple blending operation that combines private and public inference to further enhance utility. Empirical evaluations demonstrate that our approach outperforms previous state-of-the-art methods on in-context-learning (ICL) tasks, making it a promising direction for privacy-preserving text generation while maintaining high utility.

Method: Two-step approach: 1) Estimate CNN with linear terms (equivalent to multinomial logit), 2) Constrain interpretable parameters at linear values while adding higher-order terms or advanced architectures like Transformers.

Result: Significantly improved model accuracy compared to discrete choice models while retaining interpretability of selected parameters. Demonstrated on real-world transit route choice data from Singapore.

Conclusion: Provides a unifying approach that leverages strengths of both discrete choice models (interpretability) and AI models (predictive power), enabling more accurate models for planning/policy applications.

Abstract: Despite the significant progress of deep learning models in multitude of applications, their adaption in planning and policy related areas remains challenging due to the black-box nature of these models. In this work, we develop a set of DeepLogit models that follow a novel sequentially constrained approach in estimating deep learning models for transport policy analysis. In the first step of the proposed approach, we estimate a convolutional neural network (CNN) model with only linear terms, which is equivalent of a linear-in-parameter multinomial logit model. We then estimate other deep learning models by constraining the parameters that need interpretability at the values obtained in the linear-in-parameter CNN model and including higher order terms or by introducing advanced deep learning architectures like Transformers. Our approach can retain the interpretability of the selected parameters, yet provides significantly improved model accuracy than the discrete choice model. We demonstrate our approach on a transit route choice example using real-world transit smart card data from Singapore. This study shows the potential for a unifying approach, where theory-based discrete choice model (DCM) and data-driven AI models can leverage each other’s strengths in interpretability and predictive power. With the availability of larger datasets and more complex constructions, such approach can lead to more accurate models using discrete choice models while maintaining its applicability in planning and policy-related areas. Our code is available on https://github.com/jeremyoon/route-choice/ .

Method: Integrates Digital Twin technology for real-time monitoring and predictive maintenance, Zero-Knowledge Proofs for secure model verification, and dynamic allocation strategy using block coordinate descent and convex optimization to optimize UAV flight paths, transmission power, and processing rates.

Result: Achieves up to 29.6% reduction in system energy consumption compared to conventional FL approaches, with demonstrated improvements in learning performance, security, and scalability through simulations.

Conclusion: The proposed framework presents a promising solution for next-generation UAV-based intelligent networks by effectively addressing energy efficiency, security, and operational reliability challenges in federated learning systems.

Abstract: Federated learning (FL) has gained popularity as a privacy-preserving method of training machine learning models on decentralized networks. However to ensure reliable operation of UAV-assisted FL systems, issues like as excessive energy consumption, communication inefficiencies, and security vulnerabilities must be solved. This paper proposes an innovative framework that integrates Digital Twin (DT) technology and Zero-Knowledge Federated Learning (zkFed) to tackle these challenges. UAVs act as mobile base stations, allowing scattered devices to train FL models locally and upload model updates for aggregation. By incorporating DT technology, our approach enables real-time system monitoring and predictive maintenance, improving UAV network efficiency. Additionally, Zero-Knowledge Proofs (ZKPs) strengthen security by allowing model verification without exposing sensitive data. To optimize energy efficiency and resource management, we introduce a dynamic allocation strategy that adjusts UAV flight paths, transmission power, and processing rates based on network conditions. Using block coordinate descent and convex optimization techniques, our method significantly reduces system energy consumption by up to 29.6% compared to conventional FL approaches. Simulation results demonstrate improved learning performance, security, and scalability, positioning this framework as a promising solution for next-generation UAV-based intelligent networks.

Method: Transforms multimodal signals into structured 2D image representations, fuses them, and uses systematic reorganization into multiple formats with multi-stage CNN training for enhanced generalization.

Result: Significantly boosts classification performance for stress detection and improves model robustness and interpretability.

Conclusion: The image-based transformation method is broadly applicable to multimodal physiological signal domains and enables more accurate, personalized real-time health monitoring through wearables.

Abstract: This study introduces a novel method that transforms multimodal physiological signalsphotoplethysmography (PPG), galvanic skin response (GSR), and acceleration (ACC) into 2D image matrices to enhance stress detection using convolutional neural networks (CNNs). Unlike traditional approaches that process these signals separately or rely on fixed encodings, our technique fuses them into structured image representations that enable CNNs to capture temporal and cross signal dependencies more effectively. This image based transformation not only improves interpretability but also serves as a robust form of data augmentation. To further enhance generalization and model robustness, we systematically reorganize the fused signals into multiple formats, combining them in a multi stage training pipeline. This approach significantly boosts classification performance. While demonstrated here in the context of stress detection, the proposed method is broadly applicable to any domain involving multimodal physiological signals, paving the way for more accurate, personalized, and real time health monitoring through wearable technologies.

Method: Proposed KAN with ECOC framework that transforms multi-class classification into multiple binary tasks using Error-Correcting Output Codes, leveraging Hamming distance decoding for improved robustness.

Result: The KAN with ECOC framework outperformed vanilla KAN on blood cell classification dataset, achieving higher accuracy across diverse hyperparameter settings. ECOC consistently enhanced performance across FastKAN and FasterKAN variants.

Conclusion: ECOC integration significantly boosts KAN generalizability in critical healthcare AI applications, representing the first work combining ECOC with KAN for multi-class medical image classification performance enhancement.

Abstract: Kolmogorov-Arnold Networks (KAN) offer universal function approximation using univariate spline compositions without nonlinear activations. In this work, we integrate Error-Correcting Output Codes (ECOC) into the KAN framework to transform multi-class classification into multiple binary tasks, improving robustness via Hamming distance decoding. Our proposed KAN with ECOC framework outperforms vanilla KAN on a challenging blood cell classification dataset, achieving higher accuracy across diverse hyperparameter settings. Ablation studies further confirm that ECOC consistently enhances performance across FastKAN and FasterKAN variants. These results demonstrate that ECOC integration significantly boosts KAN generalizability in critical healthcare AI applications. To the best of our knowledge, this is the first work of ECOC with KAN for enhancing multi-class medical image classification performance.

Method: A central LLM/MLLM agent intelligently orchestrates diverse visual tools via Reinforcement Learning with hybrid rewards combining rule-based logic and LLM/MLLM evaluator judgments, trained on a diverse dataset using four model backbones.

Result: LLM-I achieves state-of-the-art performance, outperforming existing methods by a large margin across four benchmarks, with additional gains from a novel test-time scaling strategy.

Conclusion: The framework successfully overcomes the limitations of current unified models by enabling dynamic tool selection and orchestration, demonstrating superior performance in interleaved image-text generation tasks.

Abstract: We propose LLM-Interleaved (LLM-I), a flexible and dynamic framework that reframes interleaved image-text generation as a tool-use problem. LLM-I is designed to overcome the “one-tool” bottleneck of current unified models, which are limited to synthetic imagery and struggle with tasks requiring factual grounding or programmatic precision. Our framework empowers a central LLM or MLLM agent to intelligently orchestrate a diverse toolkit of specialized visual tools, including online image search, diffusion-based generation, code execution, and image editing. The agent is trained to select and apply these tools proficiently via a Reinforcement Learning (RL) framework that features a hybrid reward system combining rule-based logic with judgments from LLM and MLLM evaluators. Trained on a diverse new dataset using four different model backbones, LLM-I demonstrates state-of-the-art performance, outperforming existing methods by a large margin across four benchmarks. We also introduce a novel test-time scaling strategy that provides further performance gains. Project Page: https://github.com/ByteDance-BandAI/LLM-I.

Method: Proposes GenPAS framework with three stochastic sampling steps: sequence sampling, target sampling, and input sampling, which unifies existing strategies and enables flexible control of training distributions.

Result: Extensive experiments show GenPAS yields superior accuracy, data efficiency, and parameter efficiency compared to existing augmentation strategies on benchmark and industrial datasets.

Conclusion: GenPAS provides practical guidance for principled training data construction in generative recommendation and demonstrates that systematic data augmentation significantly impacts model performance.

Abstract: Generative recommendation plays a crucial role in personalized systems, predicting users’ future interactions from their historical behavior sequences. A critical yet underexplored factor in training these models is data augmentation, the process of constructing training data from user interaction histories. By shaping the training distribution, data augmentation directly and often substantially affects model generalization and performance. Nevertheless, in much of the existing work, this process is simplified, applied inconsistently, or treated as a minor design choice, without a systematic and principled understanding of its effects. Motivated by our empirical finding that different augmentation strategies can yield large performance disparities, we conduct an in-depth analysis of how they reshape training distributions and influence alignment with future targets and generalization to unseen inputs. To systematize this design space, we propose GenPAS, a generalized and principled framework that models augmentation as a stochastic sampling process over input-target pairs with three bias-controlled steps: sequence sampling, target sampling, and input sampling. This formulation unifies widely used strategies as special cases and enables flexible control of the resulting training distribution. Our extensive experiments on benchmark and industrial datasets demonstrate that GenPAS yields superior accuracy, data efficiency, and parameter efficiency compared to existing strategies, providing practical guidance for principled training data construction in generative recommendation.

Method: Controllable Pareto Trade-off (CPT) uses multi-objective optimization with two key techniques: 1) stabilizing fairness updates with moving average of stochastic gradients, and 2) pruning gradients by keeping only critical parameter gradients.

Result: CPT achieves higher-quality solutions on the Pareto front than baselines, exhibits better controllability, and can precisely follow human-defined reference vectors in hate speech detection and occupation classification tasks.

Conclusion: The proposed CPT method successfully enables precise control over fairness-accuracy trade-offs according to user preferences, outperforming existing approaches in both solution quality and controllability.

Abstract: The fairness-accuracy trade-off is a key challenge in NLP tasks. Current work focuses on finding a single “optimal” solution to balance the two objectives, which is limited considering the diverse solutions on the Pareto front. This work intends to provide controllable trade-offs according to the user’s preference of the two objectives, which is defined as a reference vector. To achieve this goal, we apply multi-objective optimization (MOO), which can find solutions from various regions of the Pareto front. However, it is challenging to precisely control the trade-off due to the stochasticity of the training process and the high dimentional gradient vectors. Thus, we propose Controllable Pareto Trade-off (CPT) that can effectively train models to perform different trade-offs according to users’ preferences. CPT 1) stabilizes the fairness update with a moving average of stochastic gradients to determine the update direction, and 2) prunes the gradients by only keeping the gradients of the critical parameters. We evaluate CPT on hate speech detection and occupation classification tasks. Experiments show that CPT can achieve a higher-quality set of solutions on the Pareto front than the baseline methods. It also exhibits better controllability and can precisely follow the human-defined reference vectors.

Method: Proposes RF-LSCM framework that models channel Angular Power Spectrum using radiance field representation with physics-informed frequency-dependent attenuation model, point-cloud-aided environment enhancement, and hierarchical tensor angular modeling for computational efficiency.

Result: Achieves up to 30% reduction in MAE for coverage prediction and 22% MAE improvement through multi-frequency data fusion, while significantly reducing GPU memory requirements and training time.

Conclusion: RF-LSCM provides a superior solution for localized wireless channel modeling that enables accurate multi-domain analysis and efficient computation, making it highly suitable for cellular network optimization applications.

Abstract: Accurate localized wireless channel modeling is a cornerstone of cellular network optimization, enabling reliable prediction of network performance during parameter tuning. Localized statistical channel modeling (LSCM) is the state-of-the-art channel modeling framework tailored for cellular network optimization. However, traditional LSCM methods, which infer the channel’s Angular Power Spectrum (APS) from Reference Signal Received Power (RSRP) measurements, suffer from critical limitations: they are typically confined to single-cell, single-grid and single-carrier frequency analysis and fail to capture complex cross-domain interactions. To overcome these challenges, we propose RF-LSCM, a novel framework that models the channel APS by jointly representing large-scale signal attenuation and multipath components within a radiance field. RF-LSCM introduces a multi-domain LSCM formulation with a physics-informed frequency-dependent Attenuation Model (FDAM) to facilitate the cross frequency generalization as well as a point-cloud-aided environment enhanced method to enable multi-cell and multi-grid channel modeling. Furthermore, to address the computational inefficiency of typical neural radiance fields, RF-LSCM leverages a low-rank tensor representation, complemented by a novel Hierarchical Tensor Angular Modeling (HiTAM) algorithm. This efficient design significantly reduces GPU memory requirements and training time while preserving fine-grained accuracy. Extensive experiments on real-world multi-cell datasets demonstrate that RF-LSCM significantly outperforms state-of-the-art methods, achieving up to a 30% reduction in mean absolute error (MAE) for coverage prediction and a 22% MAE improvement by effectively fusing multi-frequency data.

Method: Distribution-free conformal prediction framework that calibrates prediction intervals using nonconformity scores on a calibration set, with local conformal quantile estimation to handle spatial heteroskedasticity.

Result: The framework achieves reliable calibration and locally adaptive uncertainty intervals, consistently outperforming heuristic UQ approaches across multiple PDE systems (damped harmonic oscillator, Poisson, Allen-Cahn, Helmholtz) and uncertainty metrics.

Conclusion: This work bridges PINNs with distribution-free UQ, providing a general framework that enhances calibration and reliability while opening new avenues for uncertainty-aware modeling of complex PDE systems.

Abstract: Physics-Informed Neural Networks (PINNs) have emerged as a powerful framework for solving PDEs, yet existing uncertainty quantification (UQ) approaches for PINNs generally lack rigorous statistical guarantees. In this work, we bridge this gap by introducing a distribution-free conformal prediction (CP) framework for UQ in PINNs. This framework calibrates prediction intervals by constructing nonconformity scores on a calibration set, thereby yielding distribution-free uncertainty estimates with rigorous finite-sample coverage guarantees for PINNs. To handle spatial heteroskedasticity, we further introduce local conformal quantile estimation, enabling spatially adaptive uncertainty bands while preserving theoretical guarantee. Through systematic evaluations on typical PDEs (damped harmonic oscillator, Poisson, Allen-Cahn, and Helmholtz equations) and comprehensive testing across multiple uncertainty metrics, our results demonstrate that the proposed framework achieves reliable calibration and locally adaptive uncertainty intervals, consistently outperforming heuristic UQ approaches. By bridging PINNs with distribution-free UQ, this work introduces a general framework that not only enhances calibration and reliability, but also opens new avenues for uncertainty-aware modeling of complex PDE systems.

Method: Used custom smartwatch system with 7 daily EMAs for 91 socially anxious students. Developed base model on external heart rate data, transferred representations, fine-tuned with probabilistic predictions combined with trait measures in meta-learner.

Result: Achieved 60.4% balanced accuracy in state anxiety detection. On external TILES-18 dataset (10,095 EMAs), achieved 59.1% accuracy, outperforming prior work by at least 7%.

Conclusion: The pipeline successfully predicts momentary social anxiety using wearable data, demonstrating generalizability and potential for real-time personalized interventions.

Abstract: Social anxiety is a common mental health condition linked to significant challenges in academic, social, and occupational functioning. A core feature is elevated momentary (state) anxiety in social situations, yet little prior work has measured or predicted fluctuations in this anxiety throughout the day. Capturing these intra-day dynamics is critical for designing real-time, personalized interventions such as Just-In-Time Adaptive Interventions (JITAIs). To address this gap, we conducted a study with socially anxious college students (N=91; 72 after exclusions) using our custom smartwatch-based system over an average of 9.03 days (SD = 2.95). Participants received seven ecological momentary assessments (EMAs) per day to report state anxiety. We developed a base model on over 10,000 days of external heart rate data, transferred its representations to our dataset, and fine-tuned it to generate probabilistic predictions. These were combined with trait-level measures in a meta-learner. Our pipeline achieved 60.4% balanced accuracy in state anxiety detection in our dataset. To evaluate generalizability, we applied the training approach to a separate hold-out set from the TILES-18 dataset-the same dataset used for pretraining. On 10,095 once-daily EMAs, our method achieved 59.1% balanced accuracy, outperforming prior work by at least 7%.

Method: Proposes DirEgo2Token which sequentializes directed graphs via k-hop ego graphs, then develops DirGraphSSM that implements state space models on directed graphs through message-passing mechanism.

Result: Achieves state-of-the-art performance on three representative directed graph learning tasks and competitive performance on two additional tasks with 1.5x to 2x training speed improvements compared to existing SOTA models.

Conclusion: This work successfully extends state space models to directed graph learning, providing an effective solution for capturing long-range causal dependencies while maintaining high efficiency.

Abstract: Directed graphs are ubiquitous across numerous domains, where the directionality of edges encodes critical causal dependencies. However, existing GNNs and graph Transformers tailored for directed graphs face two major challenges: (1) effectively capturing long-range causal dependencies derived from directed edges; (2) balancing accuracy and training efficiency when processing large-scale graph datasets. In recent years, state space models (SSMs) have achieved substantial progress in causal sequence tasks, and their variants designed for graphs have demonstrated state-of-the-art accuracy while maintaining high efficiency across various graph learning benchmarks. However, existing graph state space models are exclusively designed for undirected graphs, which limits their performance in directed graph learning. To this end, we propose an innovative approach DirEgo2Token which sequentializes directed graphs via k-hop ego graphs. This marks the first systematic extension of state space models to the field of directed graph learning. Building upon this, we develop DirGraphSSM, a novel directed graph neural network architecture that implements state space models on directed graphs via the message-passing mechanism. Experimental results demonstrate that DirGraphSSM achieves state-of-the-art performance on three representative directed graph learning tasks while attaining competitive performance on two additional tasks with 1.5$\times $ to 2$\times $ training speed improvements compared to existing state-of-the-art models.

Method: Strategic model partitioning scheme: apply lightweight differential privacy to less critical low norm portions while protecting the remainder with homomorphic encryption, using distributed voting for consensus.

Result: Theoretical analysis confirms efficiency-utility adjustments with same privacy. Experiments show flexible prioritization between model accuracy and training time through hyperparameter adjustment.

Conclusion: ParaAegis provides practitioners with tunable control over the privacy-utility-efficiency balance in federated learning, overcoming the limitations of rigid trade-offs in existing protection mechanisms.

Abstract: Federated learning (FL) faces a critical dilemma: existing protection mechanisms like differential privacy (DP) and homomorphic encryption (HE) enforce a rigid trade-off, forcing a choice between model utility and computational efficiency. This lack of flexibility hinders the practical implementation. To address this, we introduce ParaAegis, a parallel protection framework designed to give practitioners flexible control over the privacy-utility-efficiency balance. Our core innovation is a strategic model partitioning scheme. By applying lightweight DP to the less critical, low norm portion of the model while protecting the remainder with HE, we create a tunable system. A distributed voting mechanism ensures consensus on this partitioning. Theoretical analysis confirms the adjustments between efficiency and utility with the same privacy. Crucially, the experimental results demonstrate that by adjusting the hyperparameters, our method enables flexible prioritization between model accuracy and training time.

Method: Proposes ST-LINK framework with Spatially-Enhanced Attention (SE-Attention) that extends rotary position embeddings to integrate spatial correlations, and Memory Retrieval Feed-Forward Network (MRFFN) that dynamically retrieves historical patterns.

Result: Comprehensive experiments show ST-LINK surpasses conventional deep learning and LLM approaches, effectively capturing both regular traffic patterns and abrupt changes.

Conclusion: ST-LINK successfully overcomes LLM limitations in spatial dependency modeling for traffic forecasting through innovative attention mechanisms and memory retrieval techniques.

Abstract: Traffic forecasting represents a crucial problem within intelligent transportation systems. In recent research, Large Language Models (LLMs) have emerged as a promising method, but their intrinsic design, tailored primarily for sequential token processing, introduces notable challenges in effectively capturing spatial dependencies. Specifically, the inherent limitations of LLMs in modeling spatial relationships and their architectural incompatibility with graph-structured spatial data remain largely unaddressed. To overcome these limitations, we introduce ST-LINK, a novel framework that enhances the capability of Large Language Models to capture spatio-temporal dependencies. Its key components are Spatially-Enhanced Attention (SE-Attention) and the Memory Retrieval Feed-Forward Network (MRFFN). SE-Attention extends rotary position embeddings to integrate spatial correlations as direct rotational transformations within the attention mechanism. This approach maximizes spatial learning while preserving the LLM’s inherent sequential processing structure. Meanwhile, MRFFN dynamically retrieves and utilizes key historical patterns to capture complex temporal dependencies and improve the stability of long-term forecasting. Comprehensive experiments on benchmark datasets demonstrate that ST-LINK surpasses conventional deep learning and LLM approaches, and effectively captures both regular traffic patterns and abrupt changes.

Method: Proposes CAUSA with two main components: 1) generalized unsupervised spectral regression to capture feature-clustering label dependencies, and 2) causal regularization module that adaptively separates confounders and learns view-shared sample weights to balance confounder distributions.

Result: Comprehensive experiments show CAUSA outperforms several state-of-the-art methods in multi-view unsupervised feature selection.

Conclusion: This is the first in-depth study of causal multi-view feature selection in unsupervised setting, demonstrating that causal perspective helps mitigate spurious correlations and select truly informative features.

Abstract: Multi-view unsupervised feature selection (MUFS) has recently received increasing attention for its promising ability in dimensionality reduction on multi-view unlabeled data. Existing MUFS methods typically select discriminative features by capturing correlations between features and clustering labels. However, an important yet underexplored question remains: \textit{Are such correlations sufficiently reliable to guide feature selection?} In this paper, we analyze MUFS from a causal perspective by introducing a novel structural causal model, which reveals that existing methods may select irrelevant features because they overlook spurious correlations caused by confounders. Building on this causal perspective, we propose a novel MUFS method called CAusal multi-view Unsupervised feature Selection leArning (CAUSA). Specifically, we first employ a generalized unsupervised spectral regression model that identifies informative features by capturing dependencies between features and consensus clustering labels. We then introduce a causal regularization module that can adaptively separate confounders from multi-view data and simultaneously learn view-shared sample weights to balance confounder distributions, thereby mitigating spurious correlations. Thereafter, integrating both into a unified learning framework enables CAUSA to select causally informative features. Comprehensive experiments demonstrate that CAUSA outperforms several state-of-the-art methods. To our knowledge, this is the first in-depth study of causal multi-view feature selection in the unsupervised setting.

Method: Proposes Floating-Body Hydrodynamic Neural Networks (FHNN) that predict interpretable hydrodynamic parameters (directional added masses, drag coefficients, flow streamfunction) and couples them with analytic equations of motion to constrain the hypothesis space.

Result: Achieves up to an order-of-magnitude lower error than Neural ODEs on synthetic vortex datasets, recovers physically consistent flow fields, and handles dissipative dynamics better than Hamiltonian/Lagrangian neural networks.

Conclusion: FHNN bridges the gap between black-box learning and transparent system identification by providing interpretable hydrodynamic parameters while maintaining prediction accuracy and integration stability.

Abstract: Fluid-structure interaction is common in engineering and natural systems, where floating-body motion is governed by added mass, drag, and background flows. Modeling these dissipative dynamics is difficult: black-box neural models regress state derivatives with limited interpretability and unstable long-horizon predictions. We propose Floating-Body Hydrodynamic Neural Networks (FHNN), a physics-structured framework that predicts interpretable hydrodynamic parameters such as directional added masses, drag coefficients, and a streamfunction-based flow, and couples them with analytic equations of motion. This design constrains the hypothesis space, enhances interpretability, and stabilizes integration. On synthetic vortex datasets, FHNN achieves up to an order-of-magnitude lower error than Neural ODEs, recovers physically consistent flow fields. Compared with Hamiltonian and Lagrangian neural networks, FHNN more effectively handles dissipative dynamics while preserving interpretability, which bridges the gap between black-box learning and transparent system identification.

Method: Transformer architecture trained on 1.8 TB of diverse simulation data, learning to infer governing dynamics from context without being told the underlying equations.

Result: Outperforms specialized architectures by up to 29x across multiple physics domains, demonstrates zero-shot generalization to unseen systems through in-context learning, and achieves stable 50-timestep rollouts.

Conclusion: This work establishes that a single model can learn generalizable physical principles from data alone, opening the path toward a universal Physics Foundation Model that could transform computational science and engineering.

Abstract: Foundation models have revolutionized natural language processing through a ``train once, deploy anywhere’’ paradigm, where a single pre-trained model adapts to countless downstream tasks without retraining. Access to a Physics Foundation Model (PFM) would be transformative – democratizing access to high-fidelity simulations, accelerating scientific discovery, and eliminating the need for specialized solver development. Yet current physics-aware machine learning approaches remain fundamentally limited to single, narrow domains and require retraining for each new system. We present the General Physics Transformer (GPhyT), trained on 1.8 TB of diverse simulation data, that demonstrates foundation model capabilities are achievable for physics. Our key insight is that transformers can learn to infer governing dynamics from context, enabling a single model to simulate fluid-solid interactions, shock waves, thermal convection, and multi-phase dynamics without being told the underlying equations. GPhyT achieves three critical breakthroughs: (1) superior performance across multiple physics domains, outperforming specialized architectures by up to 29x, (2) zero-shot generalization to entirely unseen physical systems through in-context learning, and (3) stable long-term predictions through 50-timestep rollouts. By establishing that a single model can learn generalizable physical principles from data alone, this work opens the path toward a universal PFM that could transform computational science and engineering.

Method: Hybrid workflow: classical ML ensemble (Logistic Regression, Random Forest, XGBoost) for feature engineering and dimensionality reduction, followed by Quantum Neural Network trained via parameter-shift rule as classifier.

Result: QNN achieved average AUC of 0.852 +/- 0.027 in simulations and 0.88 on Quafu Quantum Cloud hardware, surpassing classical benchmarks with strong recall performance on 279-sample real-world credit dataset.

Conclusion: Provides practical blueprint for quantum computing in data-constrained financial scenarios during NISQ era, demonstrating quantum advantage in high-stakes inclusive finance applications.

Abstract: Quantum Machine Learning (QML) offers a new paradigm for addressing complex financial problems intractable for classical methods. This work specifically tackles the challenge of few-shot credit risk assessment, a critical issue in inclusive finance where data scarcity and imbalance limit the effectiveness of conventional models. To address this, we design and implement a novel hybrid quantum-classical workflow. The methodology first employs an ensemble of classical machine learning models (Logistic Regression, Random Forest, XGBoost) for intelligent feature engineering and dimensionality reduction. Subsequently, a Quantum Neural Network (QNN), trained via the parameter-shift rule, serves as the core classifier. This framework was evaluated through numerical simulations and deployed on the Quafu Quantum Cloud Platform’s ScQ-P21 superconducting processor. On a real-world credit dataset of 279 samples, our QNN achieved a robust average AUC of 0.852 +/- 0.027 in simulations and yielded an impressive AUC of 0.88 in the hardware experiment. This performance surpasses a suite of classical benchmarks, with a particularly strong result on the recall metric. This study provides a pragmatic blueprint for applying quantum computing to data-constrained financial scenarios in the NISQ era and offers valuable empirical evidence supporting its potential in high-stakes applications like inclusive finance.

Method: Embed a graph neural network into a pore network model to replace analytical conductance formulas with GNN-based predictions from pore/throat features. Use end-to-end differentiable training with permeability as the only training target, backpropagating gradients through both GNN (automatic differentiation) and PNM solver (discrete adjoint method).

Result: The model achieves high accuracy and generalizes well across different scales, outperforming both pure data-driven and traditional PNM approaches. Gradient-based sensitivity analysis shows physically consistent feature influences.

Conclusion: The approach provides a scalable, physically informed framework for permeability prediction that reduces model uncertainty and improves accuracy in complex porous media while enhancing interpretability.

Abstract: Accurate prediction of permeability in porous media is essential for modeling subsurface flow. While pure data-driven models offer computational efficiency, they often lack generalization across scales and do not incorporate explicit physical constraints. Pore network models (PNMs), on the other hand, are physics-based and efficient but rely on idealized geometric assumptions to estimate pore-scale hydraulic conductance, limiting their accuracy in complex structures. To overcome these limitations, we present an end-to-end differentiable hybrid framework that embeds a graph neural network (GNN) into a PNM. In this framework, the analytical formulas used for conductance calculations are replaced by GNN-based predictions derived from pore and throat features. The predicted conductances are then passed to the PNM solver for permeability computation. In this way, the model avoids the idealized geometric assumptions of PNM while preserving the physics-based flow calculations. The GNN is trained without requiring labeled conductance data, which can number in the thousands per pore network; instead, it learns conductance values by using a single scalar permeability as the training target. This is made possible by backpropagating gradients through both the GNN (via automatic differentiation) and the PNM solver (via a discrete adjoint method), enabling fully coupled, end-to-end training. The resulting model achieves high accuracy and generalizes well across different scales, outperforming both pure data-driven and traditional PNM approaches. Gradient-based sensitivity analysis further reveals physically consistent feature influences, enhancing model interpretability. This approach offers a scalable and physically informed framework for permeability prediction in complex porous media, reducing model uncertainty and improving accuracy.

Method: Uses graph regularization with provided similarity graphs to guide parameter sharing among nodes, enabling flexible aggregation of neighbors’ parameters

Result: Outperforms both centralized and locally trained GMMs in heterogeneous, low-sample regimes

Conclusion: Graph-regularized approach effectively handles distributed GMM learning with heterogeneous data while maintaining privacy through parameter sharing instead of raw data transfer

Abstract: We present a graph-regularized learning of Gaussian Mixture Models (GMMs) in distributed settings with heterogeneous and limited local data. The method exploits a provided similarity graph to guide parameter sharing among nodes, avoiding the transfer of raw data. The resulting model allows for flexible aggregation of neighbors’ parameters and outperforms both centralized and locally trained GMMs in heterogeneous, low-sample regimes.

Method: Variational framework for joint inference of cardiac data imputation model and Gaussian Process emulator to reproduce personalized cardiovascular dynamics.

Result: Accurate imputation of missing cardiac features from minimal heart information (e.g., blood pressures only) while estimating emulated parameters of lumped model.

Conclusion: Enables novel exploration of heart-brain relationships through simulation of realistic cardiac dynamics corresponding to different brain anatomy conditions.

Abstract: The use of mechanistic models in clinical studies is limited by the lack of multi-modal patients data representing different anatomical and physiological processes. For example, neuroimaging datasets do not provide a sufficient representation of heart features for the modeling of cardiovascular factors in brain disorders. To tackle this problem we introduce a probabilistic framework for joint cardiac data imputation and personalisation of cardiovascular mechanistic models, with application to brain studies with incomplete heart data. Our approach is based on a variational framework for the joint inference of an imputation model of cardiac information from the available features, along with a Gaussian Process emulator that can faithfully reproduce personalised cardiovascular dynamics. Experimental results on UK Biobank show that our model allows accurate imputation of missing cardiac features in datasets containing minimal heart information, e.g. systolic and diastolic blood pressures only, while jointly estimating the emulated parameters of the lumped model. This allows a novel exploration of the heart-brain joint relationship through simulation of realistic cardiac dynamics corresponding to different conditions of brain anatomy.

Method: Proved mathematical equivalence of three energy formulations (kinetic, conditional kinetic, geodesic) under MDM structure, parameterized mask schedules using Beta distributions to reduce design space to 2D search, enabling post-training tuning without model modification.

Result: Energy-inspired schedules outperform hand-crafted baselines in experiments on synthetic and real-world benchmarks, particularly in low-step sampling settings.

Conclusion: The framework unifies MDM theory with optimal transport, provides practical schedule optimization method, and demonstrates improved sampling performance with energy-minimizing schedules.

Abstract: We present a systematic theoretical framework that interprets masked diffusion models (MDMs) as solutions to energy minimization problems in discrete optimal transport. Specifically, we prove that three distinct energy formulations–kinetic, conditional kinetic, and geodesic energy–are mathematically equivalent under the structure of MDMs, and that MDMs minimize all three when the mask schedule satisfies a closed-form optimality condition. This unification not only clarifies the theoretical foundations of MDMs, but also motivates practical improvements in sampling. By parameterizing interpolation schedules via Beta distributions, we reduce the schedule design space to a tractable 2D search, enabling efficient post-training tuning without model modification. Experiments on synthetic and real-world benchmarks demonstrate that our energy-inspired schedules outperform hand-crafted baselines, particularly in low-step sampling settings.

Method: FedSSG maintains per-client drift memory that accumulates local model differences as historical gradient sketches, using a statistically grounded gate function based on participation ratios to control memory updates and local alignment without extra communication.

Result: FedSSG improves test accuracy by up to 0.9% on CIFAR-10 and 2.7% on CIFAR-100, achieves about 4.5x faster target-accuracy convergence, and adds only O(d) client memory with constant-time computation.

Conclusion: Sampling statistics can be effectively used as a principled, history-aware phase control to stabilize and accelerate federated training, with FedSSG demonstrating superior performance over drift-aware baselines.

Abstract: Non-IID data and partial participation induce client drift and inconsistent local optima in federated learning, causing unstable convergence and accuracy loss. We present FedSSG, a stochastic sampling-guided, history-aware drift alignment method. FedSSG maintains a per-client drift memory that accumulates local model differences as a lightweight sketch of historical gradients; crucially, it gates both the memory update and the local alignment term by a smooth function of the observed/expected participation ratio (a phase-by-expectation signal derived from the server sampler). This statistically grounded gate stays weak and smooth when sampling noise dominates early, then strengthens once participation statistics stabilize, contracting the local-global gap without extra communication. Across CIFAR-10/100 with 100/500 clients and 2-15 percent participation, FedSSG consistently outperforms strong drift-aware baselines and accelerates convergence; on our benchmarks it improves test accuracy by up to a few points (e.g., about +0.9 on CIFAR-10 and about +2.7 on CIFAR-100 on average over the top-2 baseline) and yields about 4.5x faster target-accuracy convergence on average. The method adds only O(d) client memory and a constant-time gate, and degrades gracefully to a mild regularizer under near-IID or uniform sampling. FedSSG shows that sampling statistics can be turned into a principled, history-aware phase control to stabilize and speed up federated training.

Method: TFMAdapter uses a two-stage approach: (1) generates pseudo-forecasts with a simple regression model, and (2) trains a Gaussian Process regressor to refine predictions using both pseudo-forecasts and TSFM forecasts alongside covariates, operating on limited history during a single model call.

Result: Extensive experiments show TFMAdapter consistently outperforms both foundation models and supervised baselines, achieving 24-27% improvement over base foundation models with minimal data and computational overhead.

Conclusion: Lightweight adapters like TFMAdapter can effectively bridge the gap between generic foundation models and domain-specific forecasting needs by incorporating covariate information without retraining.

Abstract: Time Series Foundation Models (TSFMs) have recently achieved state-of-the-art performance in univariate forecasting on new time series simply by conditioned on a brief history of past values. Their success demonstrates that large-scale pretraining across diverse domains can acquire the inductive bias to generalize from temporal patterns in a brief history. However, most TSFMs are unable to leverage covariates – future-available exogenous variables critical for accurate forecasting in many applications – due to their domain-specific nature and the lack of associated inductive bias. We propose TFMAdapter, a lightweight, instance-level adapter that augments TSFMs with covariate information without fine-tuning. Instead of retraining, TFMAdapter operates on the limited history provided during a single model call, learning a non-parametric cascade that combines covariates with univariate TSFM forecasts. However, such learning would require univariate forecasts at all steps in the history, requiring too many calls to the TSFM. To enable training on the full historical context while limiting TSFM invocations, TFMAdapter uses a two-stage method: (1) generating pseudo-forecasts with a simple regression model, and (2) training a Gaussian Process regressor to refine predictions using both pseudo- and TSFM forecasts alongside covariates. Extensive experiments on real-world datasets demonstrate that TFMAdapter consistently outperforms both foundation models and supervised baselines, achieving a 24-27% improvement over base foundation models with minimal data and computational overhead. Our results highlight the potential of lightweight adapters to bridge the gap between generic foundation models and domain-specific forecasting needs.

Method: APFEx combines three innovations: 1) adaptive multi-objective optimizer switching between Pareto cone projection, gradient weighting, and exploration strategies, 2) differentiable intersectional fairness metrics for gradient-based optimization, and 3) theoretical convergence guarantees to Pareto-optimal solutions.

Result: Experiments on four real-world datasets show APFEx reduces fairness violations while maintaining competitive accuracy, outperforming existing single-attribute fairness methods.

Conclusion: APFEx bridges a critical gap in fair ML by providing the first scalable, model-agnostic framework for explicitly optimizing intersectional fairness across multiple protected attributes with theoretical guarantees.

Abstract: Ensuring fairness in machine learning models is critical, especially when biases compound across intersecting protected attributes like race, gender, and age. While existing methods address fairness for single attributes, they fail to capture the nuanced, multiplicative biases faced by intersectional subgroups. We introduce Adaptive Pareto Front Explorer (APFEx), the first framework to explicitly model intersectional fairness as a joint optimization problem over the Cartesian product of sensitive attributes. APFEx combines three key innovations- (1) an adaptive multi-objective optimizer that dynamically switches between Pareto cone projection, gradient weighting, and exploration strategies to navigate fairness-accuracy trade-offs, (2) differentiable intersectional fairness metrics enabling gradient-based optimization of non-smooth subgroup disparities, and (3) theoretical guarantees of convergence to Pareto-optimal solutions. Experiments on four real-world datasets demonstrate APFEx’s superiority, reducing fairness violations while maintaining competitive accuracy. Our work bridges a critical gap in fair ML, providing a scalable, model-agnostic solution for intersectional fairness.

Method: Using ensemble modeling with confidence-weighted average method to combine state-of-the-art deep learning models out-of-the-box (no retraining or fine-tuning).

Result: 10% performance improvement over the best individual prediction model, especially in long-tailed metrics, with consistent improvements across both NuScenes and Argoverse datasets.

Conclusion: Simple confidence-weighted ensemble approach effectively enhances trajectory prediction performance without the need for expensive retraining, making it a practical solution for combining multiple state-of-the-art models.

Abstract: This work explores the application of ensemble modeling to the multidimensional regression problem of trajectory prediction for vehicles in urban environments. As newer and bigger state-of-the-art prediction models for autonomous driving continue to emerge, an important open challenge is the problem of how to combine the strengths of these big models without the need for costly re-training. We show how, perhaps surprisingly, combining state-of-the-art deep learning models out-of-the-box (without retraining or fine-tuning) with a simple confidence-weighted average method can enhance the overall prediction. Indeed, while combining trajectory prediction models is not straightforward, this simple approach enhances performance by 10% over the best prediction model, especially in the long-tailed metrics. We show that this performance improvement holds on both the NuScenes and Argoverse datasets, and that these improvements are made across the dataset distribution. The code for our work is open source.

Method: Formulates client selection as restless multi-armed bandit problem, proposes WILF-Q approach that uses Q-learning to adaptively learn and update approximated Whittle index for each client, then selects clients with highest indices.

Result: WILF-Q significantly outperforms existing baseline policies in terms of learning efficiency, demonstrating robust and efficient client selection.

Conclusion: WILF-Q provides a practical and efficient solution for client selection in wireless FL settings without requiring explicit knowledge of client state transitions or data distributions.

Abstract: We consider the client selection problem in wireless Federated Learning (FL), with the objective of reducing the total required time to achieve a certain level of learning accuracy. Since the server cannot observe the clients’ dynamic states that can change their computation and communication efficiency, we formulate client selection as a restless multi-armed bandit problem. We propose a scalable and efficient approach called the Whittle Index Learning in Federated Q-learning (WILF-Q), which uses Q-learning to adaptively learn and update an approximated Whittle index associated with each client, and then selects the clients with the highest indices. Compared to existing approaches, WILF-Q does not require explicit knowledge of client state transitions or data distributions, making it well-suited for deployment in practical FL settings. Experiment results demonstrate that WILF-Q significantly outperforms existing baseline policies in terms of learning efficiency, providing a robust and efficient approach to client selection in wireless FL.

Method: Combines domain decomposition with specialized enrichment functions inspired by XFEM, uses distinct neural networks for standard and enriched solution components, and employs energy-based loss functions with customized integration schemes.

Result: The method effectively captures crack body discontinuities and singularities, enabling robust simulations of complex multiple-crack problems in 1D and 2D domains with extensibility to 3D problems.

Conclusion: X-PINN provides a flexible and effective framework for fracture mechanics problems, demonstrating robustness in handling multiple cracks through neural network enrichment and domain decomposition techniques.

Abstract: This paper presents eXtended Physics-Informed Neural Network (X-PINN), a novel and robust framework for addressing fracture mechanics problems involving multiple cracks in fractured media. To address this, an energy-based loss function, customized integration schemes, and domain decomposition procedures are proposed. Inspired by the Extended Finite Element Method (XFEM), the neural network solution space is enriched with specialized functions that allow crack body discontinuities and singularities at crack tips to be explicitly captured. Furthermore, a structured framework is introduced in which standard and enriched solution components are modeled using distinct neural networks, enabling flexible and effective simulations of complex multiple-crack problems in 1D and 2D domains, with convenient extensibility to 3D problems. Numerical experiments are conducted to validate the effectiveness and robustness of the proposed method.

Method: Proposes EpiSMART framework using size-constrained replay buffer and informed sample selection strategy that retains high-entropy and seizure-predicted samples to incrementally adapt to patient-specific EEG signals.

Result: 21% improvement in F1 score over baseline on CHB-MIT dataset, requiring only 6.46 minutes of labeled data and 6.28 updates per day, making it suitable for real-time wearable deployment.

Conclusion: EpiSMART enables robust personalized seizure detection under resource-constrained conditions by effectively integrating new data without degrading past knowledge, advancing practical deployment in wearable healthcare systems.

Abstract: Objective: Epilepsy, a prevalent neurological disease, demands careful diagnosis and continuous care. Seizure detection remains challenging, as current clinical practice relies on expert analysis of electroencephalography, which is a time-consuming process and requires specialized knowledge. Addressing this challenge, this paper explores automated epileptic seizure detection using deep learning, focusing on personalized continual learning models that adapt to each patient’s unique electroencephalography signal features, which evolve over time. Methods: In this context, our approach addresses the challenge of integrating new data into existing models without catastrophic forgetting, a common issue in static deep learning models. We propose EpiSMART, a continual learning framework for seizure detection that uses a size-constrained replay buffer and an informed sample selection strategy to incrementally adapt to patient-specific electroencephalography signals. By selectively retaining high-entropy and seizure-predicted samples, our method preserves critical past information while maintaining high performance with minimal memory and computational requirements. Results: Validation on the CHB-MIT dataset, shows that EpiSMART achieves a 21% improvement in the F1 score over a trained baseline without updates in all other patients. On average, EpiSMART requires only 6.46 minutes of labeled data and 6.28 updates per day, making it suitable for real-time deployment in wearable systems. Conclusion:EpiSMART enables robust and personalized seizure detection under realistic and resource-constrained conditions by effectively integrating new data into existing models without degrading past knowledge. Significance: This framework advances automated seizure detection by providing a continual learning approach that supports patient-specific adaptation and practical deployment in wearable healthcare systems.

Method: Receiver-centric deep temporal graph network using heterogeneous star graphs (receiver center with satellite leaves) with time-varying attributes. A single layer Heterogeneous Graph ConvLSTM aggregates spatial context and temporal dynamics over short history to predict 2D deviation vectors.

Result: Achieves lowest MAE across all baselines: 3.64-7.74 cm at -45 dBm, improving to 1.65-2.08 cm at -60 to -70 dBm. Outperforms Seq2Point, MLP, and CNN in mixed mode datasets (3.78-4.25 cm MAE). Superior data efficiency with only 10% training data (20 cm vs 36-42 cm baselines).

Conclusion: The graph-based approach effectively mitigates GNSS jamming in real-time, demonstrating robust performance across diverse jamming profiles and power levels while requiring minimal training data compared to traditional methods.

Abstract: Global Navigation Satellite Systems (GNSS) are increasingly disrupted by intentional jamming, degrading availability precisely when positioning and timing must remain operational. We address this by reframing jamming mitigation as dynamic graph regression and introducing a receiver-centric deep temporal graph network that predicts, and thus corrects, the receivers horizontal deviation in real time. At each 1 Hz epoch, the satellite receiver environment is represented as a heterogeneous star graph (receiver center, tracked satellites as leaves) with time varying attributes (e.g., SNR, azimuth, elevation, latitude/longitude). A single layer Heterogeneous Graph ConvLSTM (HeteroGCLSTM) aggregates one hop spatial context and temporal dynamics over a short history to output the 2D deviation vector applied for on the fly correction. We evaluate on datasets from two distinct receivers under three jammer profiles, continuous wave (cw), triple tone (cw3), and wideband FM, each exercised at six power levels between -45 and -70 dBm, with 50 repetitions per scenario (prejam/jam/recovery). Against strong multivariate time series baselines (MLP, uniform CNN, and Seq2Point CNN), our model consistently attains the lowest mean absolute error (MAE). At -45 dBm, it achieves 3.64 cm (GP01/cw), 7.74 cm (GP01/cw3), 4.41 cm (ublox/cw), 4.84 cm (ublox/cw3), and 4.82 cm (ublox/FM), improving to 1.65-2.08 cm by -60 to -70 dBm. On mixed mode datasets pooling all powers, MAE is 3.78 cm (GP01) and 4.25 cm (ublox10), outperforming Seq2Point, MLP, and CNN. A split study shows superior data efficiency: with only 10% training data our approach remains well ahead of baselines (20 cm vs. 36-42 cm).

Method: Federated learning framework with client-level differential privacy, using multilayer perceptron on sliding windows of case counts, where clients exchange norm-clipped updates and server aggregates with DP noise.

Result: At moderately strong privacy levels, DP model closely approaches non-DP performance: R²=0.94 vs 0.95 and MAPE=26% in Nov 2020; R²=0.88 vs 0.93 and MAPE=21% in Mar 2022. Very strict privacy yields unusable forecasts.

Conclusion: Client-level DP-FL can deliver useful county-level predictions with strong privacy guarantees, enabling privacy-compliant collaboration among health authorities for local epidemic forecasting.

Abstract: In times of epidemics, swift reaction is necessary to mitigate epidemic spreading. For this reaction, localized approaches have several advantages, limiting necessary resources and reducing the impact of interventions on a larger scale. However, training a separate machine learning (ML) model on a local scale is often not feasible due to limited available data. Centralizing the data is also challenging because of its high sensitivity and privacy constraints. In this study, we consider a localized strategy based on the German counties and communities managed by the related local health authorities (LHA). For the preservation of privacy to not oppose the availability of detailed situational data, we propose a privacy-preserving forecasting method that can assist public health experts and decision makers. ML methods with federated learning (FL) train a shared model without centralizing raw data. Considering the counties, communities or LHAs as clients and finding a balance between utility and privacy, we study a FL framework with client-level differential privacy (DP). We train a shared multilayer perceptron on sliding windows of recent case counts to forecast the number of cases, while clients exchange only norm-clipped updates and the server aggregated updates with DP noise. We evaluate the approach on COVID-19 data on county-level during two phases. As expected, very strict privacy yields unstable, unusable forecasts. At a moderately strong level, the DP model closely approaches the non-DP model: $R^2= 0.94$ (vs. 0.95) and mean absolute percentage error (MAPE) of 26 % in November 2020; $R^2= 0.88$ (vs. 0.93) and MAPE of 21 % in March 2022. Overall, client-level DP-FL can deliver useful county-level predictions with strong privacy guarantees, and viable privacy budgets depend on epidemic phase, allowing privacy-compliant collaboration among health authorities for local forecasting.

Method: Convert nanopore current signals into scaleogram images via wavelet transforms to capture amplitude, frequency, and time information, then apply machine learning algorithms for classification.

Result: Achieved ~81% classification accuracy on 42 peptides, setting a new state-of-the-art in the field.

Conclusion: This approach represents significant progress toward practical peptide/protein diagnostics at point-of-care settings, with demonstrated model transfer techniques for real hardware deployment.

Abstract: A device capable of performing real time classification of proteins in a clinical setting would allow for inexpensive and rapid disease diagnosis. One such candidate for this technology are nanopore devices. These devices work by measuring a current signal that arises when a protein or peptide enters a nanometer-length-scale pore. Should this current be uniquely related to the structure of the peptide and its interactions with the pore, the signals can be used to perform identification. While such a method would allow for real time identification of peptides and proteins in a clinical setting, to date, the complexities of these signals limit their accuracy. In this work, we tackle the issue of classification by converting the current signals into scaleogram images via wavelet transforms, capturing amplitude, frequency, and time information in a modality well-suited to machine learning algorithms. When tested on 42 peptides, our method achieved a classification accuracy of ~$81,%$, setting a new state-of-the-art in the field and taking a step toward practical peptide/protein diagnostics at the point of care. In addition, we demonstrate model transfer techniques that will be critical when deploying these models into real hardware, paving the way to a new method for real-time disease diagnosis.

Method: Environmental sensor fusion (temperature, humidity, pressure differentials) with quantized decision tree inference on STM32 microcontroller for real-time, low-power edge computing.

Result: Over 99% queen detection accuracy using only environmental inputs, with audio features providing no significant performance improvement.

Conclusion: Presents a scalable, sustainable, non-invasive hive monitoring solution using off-the-shelf, energy-efficient hardware for autonomous precision beekeeping.

Abstract: Queen bee presence is essential for the health and stability of honeybee colonies, yet current monitoring methods rely on manual inspections that are labor-intensive, disruptive, and impractical for large-scale beekeeping. While recent audio-based approaches have shown promise, they often require high power consumption, complex preprocessing, and are susceptible to ambient noise. To overcome these limitations, we propose a lightweight, multimodal system for queen detection based on environmental sensor fusion-specifically, temperature, humidity, and pressure differentials between the inside and outside of the hive. Our approach employs quantized decision tree inference on a commercial STM32 microcontroller, enabling real-time, low-power edge computing without compromising accuracy. We show that our system achieves over 99% queen detection accuracy using only environmental inputs, with audio features offering no significant performance gain. This work presents a scalable and sustainable solution for non-invasive hive monitoring, paving the way for autonomous, precision beekeeping using off-the-shelf, energy-efficient hardware.

Method: Adopted Bayesian Risk Markov Decision Process (BRMDP), derived asymptotic normality of value function differences, developed posterior sampling procedures for online RL and contextual multi-arm bandits, and established sub-linear regret bounds.

Result: Bayesian risk-averse approach pessimistically underestimates original value function, with discrepancy increasing with stronger risk aversion and decreasing with more data. Sub-linear regret bounds achieved for both conventional and Bayesian risk regret definitions.

Conclusion: The proposed Bayesian risk-averse framework effectively addresses epistemic uncertainty in RL, with theoretical guarantees and practical effectiveness demonstrated through numerical experiments and regret analysis.

Abstract: In this paper, we study the Bayesian risk-averse formulation in reinforcement learning (RL). To address the epistemic uncertainty due to a lack of data, we adopt the Bayesian Risk Markov Decision Process (BRMDP) to account for the parameter uncertainty of the unknown underlying model. We derive the asymptotic normality that characterizes the difference between the Bayesian risk value function and the original value function under the true unknown distribution. The results indicate that the Bayesian risk-averse approach tends to pessimistically underestimate the original value function. This discrepancy increases with stronger risk aversion and decreases as more data become available. We then utilize this adaptive property in the setting of online RL as well as online contextual multi-arm bandits (CMAB), a special case of online RL. We provide two procedures using posterior sampling for both the general RL problem and the CMAB problem. We establish a sub-linear regret bound, with the regret defined as the conventional regret for both the RL and CMAB settings. Additionally, we establish a sub-linear regret bound for the CMAB setting with the regret defined as the Bayesian risk regret. Finally, we conduct numerical experiments to demonstrate the effectiveness of the proposed algorithm in addressing epistemic uncertainty and verifying the theoretical properties.

Method: Implemented three neural network architectures (deep dense network, shallow three-layer network, CNN) using TensorFlow and PyTorch, tested with various optimizers (Adagrad, RMSprop, Adadelta, SGD, FTRL) across EEG frequency bands, and used SHAP plots for feature importance analysis.

Result: Adagrad and RMSprop consistently performed well, with Adagrad excelling in beta band and RMSprop in gamma band. CNN showed second highest accuracy with strong spatial feature capture. Deep dense network learned complex patterns well, while shallow network offered computational efficiency despite lower accuracy.

Conclusion: Optimizer selection, model architecture, and EEG frequency band analysis are crucial for enhancing classifier performance in neuroimaging tasks, with different approaches offering trade-offs between accuracy and efficiency.

Abstract: This study investigates classifier performance across EEG frequency bands using various optimizers and evaluates efficient class prediction for the left and right hemispheres. Three neural network architectures - a deep dense network, a shallow three-layer network, and a convolutional neural network (CNN) - are implemented and compared using the TensorFlow and PyTorch frameworks. Results indicate that the Adagrad and RMSprop optimizers consistently perform well across different frequency bands, with Adadelta exhibiting robust performance in cross-model evaluations. Specifically, Adagrad excels in the beta band, while RMSprop achieves superior performance in the gamma band. Conversely, SGD and FTRL exhibit inconsistent performance. Among the models, the CNN demonstrates the second highest accuracy, particularly in capturing spatial features of EEG data. The deep dense network shows competitive performance in learning complex patterns, whereas the shallow three-layer network, sometimes being less accurate, provides computational efficiency. SHAP (Shapley Additive Explanations) plots are employed to identify efficient class prediction, revealing nuanced contributions of EEG frequency bands to model accuracy. Overall, the study highlights the importance of optimizer selection, model architecture, and EEG frequency band analysis in enhancing classifier performance and understanding feature importance in neuroimaging-based classification tasks.

Method: Proposes Quantile Neural Basis Model that incorporates interpretability principles from Quantile Generalized Additive Models into neural network training. Uses shared basis decomposition and weight factorization, avoiding parametric distributional assumptions while complementing Neural Models for Location, Scale, and Shape.

Result: Validated on day-ahead electricity price forecasting, achieving predictive performance comparable to distributional and quantile regression neural networks.

Conclusion: The model provides valuable insights into model behavior through learned nonlinear mappings from input features to output predictions across the horizon, offering both strong performance and interpretability.

Abstract: While neural networks are achieving high predictive accuracy in multi-horizon probabilistic forecasting, understanding the underlying mechanisms that lead to feature-conditioned outputs remains a significant challenge for forecasters. In this work, we take a further step toward addressing this critical issue by introducing the Quantile Neural Basis Model, which incorporates the interpretability principles of Quantile Generalized Additive Models into an end-to-end neural network training framework. To this end, we leverage shared basis decomposition and weight factorization, complementing Neural Models for Location, Scale, and Shape by avoiding any parametric distributional assumptions. We validate our approach on day-ahead electricity price forecasting, achieving predictive performance comparable to distributional and quantile regression neural networks, while offering valuable insights into model behavior through the learned nonlinear mappings from input features to output predictions across the horizon.

Method: Collaboration between Johnson County, Kansas and Carnegie Mellon University using predictive modeling to identify high-risk individuals, followed by a field trial with targeted mental health outreach interventions.

Result: Model accurately predicted reincarceration (over 50% of highest-risk group returned to jail within a year). Outreach was most effective for highest-risk individuals, reducing mental health crises, EMS calls, and criminal justice involvement.

Conclusion: Targeted, data-driven mental health outreach for high-risk incarcerated individuals can break the cycle of incarceration and improve public safety outcomes.

Abstract: Many incarcerated individuals face significant and complex challenges, including mental illness, substance dependence, and homelessness, yet jails and prisons are often poorly equipped to address these needs. With little support from the existing criminal justice system, these needs can remain untreated and worsen, often leading to further offenses and a cycle of incarceration with adverse outcomes both for the individual and for public safety, with particularly large impacts on communities of color that continue to widen the already extensive racial disparities in criminal justice outcomes. Responding to these failures, a growing number of criminal justice stakeholders are seeking to break this cycle through innovative approaches such as community-driven and alternative approaches to policing, mentoring, community building, restorative justice, pretrial diversion, holistic defense, and social service connections. Here we report on a collaboration between Johnson County, Kansas, and Carnegie Mellon University to perform targeted, proactive mental health outreach in an effort to reduce reincarceration rates. This paper describes the data used, our predictive modeling approach and results, as well as the design and analysis of a field trial conducted to confirm our model’s predictive power, evaluate the impact of this targeted outreach, and understand at what level of reincarceration risk outreach might be most effective. Through this trial, we find that our model is highly predictive of new jail bookings, with more than half of individuals in the trial’s highest-risk group returning to jail in the following year. Outreach was most effective among these highest-risk individuals, with impacts on mental health utilization, EMS dispatches, and criminal justice involvement.

Method: Variational formulation of compositional kernel ridge regression with coordinate-wise input reweighting, analyzing global minimizers and stationary points with Gaussian noise variables.

Result: The model successfully recovers relevant variables while eliminating noise variables. Laplace (ℓ₁-type) kernels recover features contributing to nonlinear effects at stationary points, while Gaussian kernels only recover linear features.

Conclusion: Compositional kernel ridge regression provides a framework for feature learning, with kernel choice (Laplace vs Gaussian) determining the ability to capture nonlinear relationships during feature recovery.

Abstract: We study a compositional variant of kernel ridge regression in which the predictor is applied to a coordinate-wise reweighting of the inputs. Formulated as a variational problem, this model provides a simple testbed for feature learning in compositional architectures. From the perspective of variable selection, we show how relevant variables are recovered while noise variables are eliminated. We establish guarantees showing that both global minimizers and stationary points discard noise coordinates when the noise variables are Gaussian distributed. A central finding is that $\ell_1$-type kernels, such as the Laplace kernel, succeed in recovering features contributing to nonlinear effects at stationary points, whereas Gaussian kernels recover only linear ones.

Method: End-to-end framework combining multi-stage AI architecture to functionally separate the problem, novel PCDS data generation strategy that eliminates need for hundreds of thousands of costly simulations, and Bayesian engine to quantify predictive uncertainty.

Result: Framework successfully deconstructs IOP measurements into fundamental components, estimating unmeasurable tissue permeability and outflow facility. Noninvasive outflow facility estimates achieved excellent agreement with state-of-the-art tonography with precision comparable to direct instruments. Permeability biomarker accurately stratified clinical cohorts by disease risk.

Conclusion: The framework establishes a generalizable blueprint for solving similar inverse problems in other data-scarce, computationally-intensive domains beyond glaucoma.

Abstract: Many critical healthcare decisions are challenged by the inability to measure key underlying parameters. Glaucoma, a leading cause of irreversible blindness driven by elevated intraocular pressure (IOP), provides a stark example. The primary determinant of IOP, a tissue property called trabecular meshwork permeability, cannot be measured in vivo, forcing clinicians to depend on indirect surrogates. This clinical challenge is compounded by a broader computational one: developing predictive models for such ill-posed inverse problems is hindered by a lack of ground-truth data and prohibitive cost of large-scale, high-fidelity simulations. We address both challenges with an end-to-end framework to noninvasively estimate unmeasurable variables from sparse, routine data. Our approach combines a multi-stage artificial intelligence architecture to functionally separate the problem; a novel data generation strategy we term PCDS that obviates the need for hundreds of thousands of costly simulations, reducing the effective computational time from years to hours; and a Bayesian engine to quantify predictive uncertainty. Our framework deconstructs a single IOP measurement into its fundamental components from routine inputs only, yielding estimates for the unmeasurable tissue permeability and a patient’s outflow facility. Our noninvasively estimated outflow facility achieved excellent agreement with state-of-the-art tonography with precision comparable to direct physical instruments. Furthermore, the newly derived permeability biomarker demonstrates high accuracy in stratifying clinical cohorts by disease risk, highlighting its diagnostic potential. More broadly, our framework establishes a generalizable blueprint for solving similar inverse problems in other data-scarce, computationally-intensive domains.

Method: Uses graph algorithms to create hierarchical circuit representation, then LLM agents perform hypothesis-verification-refinement loops with consistency checks. Verified insights are integrated into Bayesian optimization through LLM-guided sampling and trust-region updates.

Result: The framework achieves robust circuit understanding directly from raw netlists and translates this knowledge into optimization gains with improved efficiency while preserving feasibility.

Conclusion: TopoSizing provides an automated, transparent approach that combines structural analysis with optimization, addressing key limitations in current analog circuit design methodologies.

Abstract: Analog and mixed-signal circuit design remains challenging due to the shortage of high-quality data and the difficulty of embedding domain knowledge into automated flows. Traditional black-box optimization achieves sampling efficiency but lacks circuit understanding, which often causes evaluations to be wasted in low-value regions of the design space. In contrast, learning-based methods embed structural knowledge but are case-specific and costly to retrain. Recent attempts with large language models show potential, yet they often rely on manual intervention, limiting generality and transparency. We propose TopoSizing, an end-to-end framework that performs robust circuit understanding directly from raw netlists and translates this knowledge into optimization gains. Our approach first applies graph algorithms to organize circuits into a hierarchical device-module-stage representation. LLM agents then execute an iterative hypothesis-verification-refinement loop with built-in consistency checks, producing explicit annotations. Verified insights are integrated into Bayesian optimization through LLM-guided initial sampling and stagnation-triggered trust-region updates, improving efficiency while preserving feasibility.

Method: Proposes Tree-Guided Preference Optimization (TGPO) with tree-structured trajectory representation to merge semantically identical states, Process Reward Model for automatic fine-grained rewards, and dynamic weighting mechanism for prioritizing high-impact decisions.

Result: Experiments on Online-Mind2Web and C-WebShop datasets show TGPO significantly outperforms existing methods with higher success rates and fewer redundant steps.

Conclusion: TGPO effectively addresses key challenges in web agent training through innovative trajectory representation and reward modeling, demonstrating superior performance in automated web interaction tasks.

Abstract: With the rapid advancement of large language models and vision-language models, employing large models as Web Agents has become essential for automated web interaction. However, training Web Agents with reinforcement learning faces critical challenges including credit assignment misallocation, prohibitively high annotation costs, and reward sparsity. To address these issues, we propose Tree-Guided Preference Optimization (TGPO), an offline reinforcement learning framework that proposes a tree-structured trajectory representation merging semantically identical states across trajectories to eliminate label conflicts. Our framework incorporates a Process Reward Model that automatically generates fine-grained rewards through subgoal progress, redundancy detection, and action verification. Additionally, a dynamic weighting mechanism prioritizes high-impact decision points during training. Experiments on Online-Mind2Web and our self-constructed C-WebShop datasets demonstrate that TGPO significantly outperforms existing methods, achieving higher success rates with fewer redundant steps.

Method: TimeAlign learns auxiliary features via a simple reconstruction task and feeds them back to any base forecaster. It’s architecture-agnostic and incurs negligible overhead.

Result: Extensive experiments across eight benchmarks verify TimeAlign’s superior performance. Gains primarily come from correcting frequency mismatches between historical inputs and future outputs.

Conclusion: TimeAlign serves as an effective general alignment module for modern deep learning time-series forecasting systems, with theoretical justification showing it increases mutual information between learned representations and predicted targets.

Abstract: Representation learning techniques like contrastive learning have long been explored in time series forecasting, mirroring their success in computer vision and natural language processing. Yet recent state-of-the-art (SOTA) forecasters seldom adopt these representation approaches because they have shown little performance advantage. We challenge this view and demonstrate that explicit representation alignment can supply critical information that bridges the distributional gap between input histories and future targets. To this end, we introduce TimeAlign, a lightweight, plug-and-play framework that learns auxiliary features via a simple reconstruction task and feeds them back to any base forecaster. Extensive experiments across eight benchmarks verify its superior performance. Further studies indicate that the gains arises primarily from correcting frequency mismatches between historical inputs and future outputs. We also provide a theoretical justification for the effectiveness of TimeAlign in increasing the mutual information between learned representations and predicted targets. As it is architecture-agnostic and incurs negligible overhead, TimeAlign can serve as a general alignment module for modern deep learning time-series forecasting systems. The code is available at https://github.com/TROUBADOUR000/TimeAlign.

Method: Sequentially fine-tuned Llama-3.2-1B on six disjoint datasets about named entities, then analyzed activation patterns and trained linear probes to distinguish early vs late learned entities.

Result: Average activations for each dataset arranged exactly in training order along a straight line in 2D subspace. Linear probes achieved ~90% accuracy distinguishing early vs late entities, and fine-tuning enabled ~80% accuracy in reporting training stage of unseen entities.

Conclusion: Models encode temporal acquisition information linearly in their activations, demonstrating capability to differentiate information by learning time, with implications for handling conflicting data and knowledge updates.

Abstract: We show that language models’ activations linearly encode when information was learned during training. Our setup involves creating a model with a known training order by sequentially fine-tuning Llama-3.2-1B on six disjoint but otherwise similar datasets about named entities. We find that the average activations of test samples for the six training datasets encode the training order: when projected into a 2D subspace, these centroids are arranged exactly in the order of training and lie on a straight line. Further, we show that linear probes can accurately (~90%) distinguish “early” vs. “late” entities, generalizing to entities unseen during the probes’ own training. The model can also be fine-tuned to explicitly report an unseen entity’s training stage (~80% accuracy). Interestingly, this temporal signal does not seem attributable to simple differences in activation magnitudes, losses, or model confidence. Our paper demonstrates that models are capable of differentiating information by its acquisition time, and carries significant implications for how they might manage conflicting data and respond to knowledge modifications.

Method: Introduces a unifying variational framework that integrates convex transformations of the residual, where different transformations correspond to distinct objective functionals (exponential weights for uniform error minimization, linear weights for quadratic error minimization).

Result: The framework enables systematic design of adaptive schemes across norms, reduces discretization error through variance reduction of loss estimator, and enhances learning dynamics by improving gradient signal-to-noise ratio. Substantial performance gains demonstrated across optimizers and architectures in operator learning.

Conclusion: Provides theoretical justification for residual-based adaptivity and establishes a foundation for principled discretization and training strategies in scientific machine learning.

Abstract: Residual-based adaptive strategies are widely used in scientific machine learning but remain largely heuristic. We introduce a unifying variational framework that formalizes these methods by integrating convex transformations of the residual. Different transformations correspond to distinct objective functionals: exponential weights target the minimization of uniform error, while linear weights recover the minimization of quadratic error. Within this perspective, adaptive weighting is equivalent to selecting sampling distributions that optimize the primal objective, thereby linking discretization choices directly to error metrics. This principled approach yields three benefits: (1) it enables systematic design of adaptive schemes across norms, (2) reduces discretization error through variance reduction of the loss estimator, and (3) enhances learning dynamics by improving the gradient signal-to-noise ratio. Extending the framework to operator learning, we demonstrate substantial performance gains across optimizers and architectures. Our results provide a theoretical justification of residual-based adaptivity and establish a foundation for principled discretization and training strategies.

Method: Introduces a pioneering Banach-Bregman framework using Bregman projections and Bregman-Fejer monotonicity as a unified template that encompasses stochastic approximation, mirror descent, natural gradient, adaptive methods, and mirror-prox methods in general Banach spaces.

Result: Establishes super-relaxations (λ > 2) in non-Hilbert settings, enabling flexible geometries with acceleration effects. Achieves up to 20% faster convergence, reduced variance, and enhanced accuracy across machine learning benchmarks, deep learning (Transformer training), reinforcement learning (actor-critic), and large language models (WikiText-2 with distilGPT-2).

Conclusion: The Banach-Bregman geometry serves as a cornerstone unifying optimization theory and practice across core AI paradigms, positioning it as the foundation for next-generation optimization beyond traditional Euclidean-based methods.

Abstract: Stochastic optimization powers the scalability of modern artificial intelligence, spanning machine learning, deep learning, reinforcement learning, and large language model training. Yet, existing theory remains largely confined to Hilbert spaces, relying on inner-product frameworks and orthogonality. This paradigm fails to capture non-Euclidean settings, such as mirror descent on simplices, Bregman proximal methods for sparse learning, natural gradient descent in information geometry, or Kullback–Leibler-regularized language model training. Unlike Euclidean-based Hilbert-space methods, this approach embraces general Banach spaces. This work introduces a pioneering Banach–Bregman framework for stochastic iterations, establishing Bregman geometry as a foundation for next-generation optimization. It (i) provides a unified template via Bregman projections and Bregman–Fejer monotonicity, encompassing stochastic approximation, mirror descent, natural gradient, adaptive methods, and mirror-prox; (ii) establishes super-relaxations ($\lambda > 2$) in non-Hilbert settings, enabling flexible geometries and elucidating their acceleration effect; and (iii) delivers convergence theorems spanning almost-sure boundedness to geometric rates, validated on synthetic and real-world tasks. Empirical studies across machine learning (UCI benchmarks), deep learning (e.g., Transformer training), reinforcement learning (actor–critic), and large language models (WikiText-2 with distilGPT-2) show up to 20% faster convergence, reduced variance, and enhanced accuracy over classical baselines. These results position Banach–Bregman geometry as a cornerstone unifying optimization theory and practice across core AI paradigms.

Method: Uses implicit neural representation (INR) fitted to noisy data with Runge-Kutta integration and total variation constraints to ensure the reconstructed state follows dynamical system principles while staying close to original data.

Result: Effective noise suppression, precise derivative estimation via automatic differentiation, and reliable system identification when combined with SINDy for equation recovery.

Conclusion: RKTV-INR provides an effective framework for denoising dynamical system data, enabling accurate derivative computation and successful recovery of governing equations from noisy measurements.

Abstract: Data-driven modeling of nonlinear dynamical systems is often hampered by measurement noise. We propose a denoising framework, called Runge-Kutta and Total Variation Based Implicit Neural Representation (RKTV-INR), that represents the state trajectory with an implicit neural representation (INR) fitted directly to noisy observations. Runge-Kutta integration and total variation are imposed as constraints to ensure that the reconstructed state is a trajectory of a dynamical system that remains close to the original data. The trained INR yields a clean, continuous trajectory and provides accurate first-order derivatives via automatic differentiation. These denoised states and derivatives are then supplied to Sparse Identification of Nonlinear Dynamics (SINDy) to recover the governing equations. Experiments demonstrate effective noise suppression, precise derivative estimation, and reliable system identification.

Method: Proposes defense using critically-damped higher-order Langevin dynamics that introduces auxiliary variables and a joint diffusion process. The auxiliary variables mix external randomness to corrupt sensitive input data earlier in the diffusion process.

Result: Theoretical investigation and validation on toy dataset and speech dataset using AUROC curves and FID metric.

Conclusion: The proposed method provides an effective defense mechanism against membership inference attacks for diffusion models by leveraging higher-order dynamics and auxiliary variables to enhance privacy protection.

Abstract: Recent advances in generative artificial intelligence applications have raised new data security concerns. This paper focuses on defending diffusion models against membership inference attacks. This type of attack occurs when the attacker can determine if a certain data point was used to train the model. Although diffusion models are intrinsically more resistant to membership inference attacks than other generative models, they are still susceptible. The defense proposed here utilizes critically-damped higher-order Langevin dynamics, which introduces several auxiliary variables and a joint diffusion process along these variables. The idea is that the presence of auxiliary variables mixes external randomness that helps to corrupt sensitive input data earlier on in the diffusion process. This concept is theoretically investigated and validated on a toy dataset and a speech dataset using the Area Under the Receiver Operating Characteristic (AUROC) curves and the FID metric.

Method: Uses first-order saliency criterion from Neural Tangent Kernel under Adam optimization, adaptive sparsity allocation across layers and modules (attention vs MLP), and KL divergence-based calibration data selection.

Result: Outperforms existing structured pruning methods on Llama3, Qwen, and T5 models under equivalent sparsity constraints.

Conclusion: Provides a theoretically sound and practical approach to LLM compression that balances immediate zero-shot accuracy preservation with fine-tuning capability.

Abstract: Structured pruning of large language models (LLMs) offers substantial efficiency improvements by removing entire hidden units, yet current approaches often suffer from significant performance degradation, particularly in zero-shot settings, and necessitate costly recovery techniques such as supervised fine-tuning (SFT) or adapter insertion. To address these critical shortcomings, we introduce NIRVANA, a novel pruning method explicitly designed to balance immediate zero-shot accuracy preservation with robust fine-tuning capability. Leveraging a first-order saliency criterion derived from the Neural Tangent Kernel under Adam optimization dynamics, NIRVANA provides a theoretically grounded pruning strategy that respects essential model training behaviors. To further address the unique challenges posed by structured pruning, NIRVANA incorporates an adaptive sparsity allocation mechanism across layers and modules (attention vs. MLP), which adjusts pruning intensity between modules in a globally balanced manner. Additionally, to mitigate the high sensitivity of pruning decisions to calibration data quality, we propose a simple yet effective KL divergence-based calibration data selection strategy, ensuring more reliable and task-agnostic pruning outcomes. Comprehensive experiments conducted on Llama3, Qwen, and T5 models demonstrate that NIRVANA outperforms existing structured pruning methods under equivalent sparsity constraints, providing a theoretically sound and practical approach to LLM compression. The code is available at https://github.com/iDEA-iSAIL-Lab-UIUC/NIRVANA.

Method: Generates multiple rollouts at inference, uses a frozen anchor policy to synthesize a reference from them, and optimizes toward this reference using either programmatic equivalence (verifiable tasks) or self-proposed rubrics with LLM judges (non-verifiable tasks).

Result: Significant performance improvements: up to +27% on MATH-500, +12% on HealthBench for test-time application, and further gains (+33% and +30%) with reinforcement learning (CaT-RL).

Conclusion: CaT effectively turns extra inference compute into valuable supervision signals, enabling performance scaling with rollout count and outperforming selection-based methods, with trained policies surpassing initial teacher signals.

Abstract: Where do learning signals come from when there is no ground truth in post-training? We propose turning exploration into supervision through Compute as Teacher (CaT), which converts the model’s own exploration at inference-time into reference-free supervision by synthesizing a single reference from a group of parallel rollouts and then optimizing toward it. Concretely, the current policy produces a group of rollouts; a frozen anchor (the initial policy) reconciles omissions and contradictions to estimate a reference, turning extra inference-time compute into a teacher signal. We turn this into rewards in two regimes: (i) verifiable tasks use programmatic equivalence on final answers; (ii) non-verifiable tasks use self-proposed rubrics-binary, auditable criteria scored by an independent LLM judge, with reward given by the fraction satisfied. Unlike selection methods (best-of-N, majority, perplexity, or judge scores), synthesis may disagree with the majority and be correct even when all rollouts are wrong; performance scales with the number of rollouts. As a test-time procedure, CaT improves Gemma 3 4B, Qwen 3 4B, and Llama 3.1 8B (up to +27% on MATH-500; +12% on HealthBench). With reinforcement learning (CaT-RL), we obtain further gains (up to +33% and +30%), with the trained policy surpassing the initial teacher signal.

Method: Proposes Federated Contrastive Shareable Representations (FedCoSR) that aggregates shallow layer parameters and local representations globally, uses contrastive learning between local and global representations, and implements adaptive local aggregation for clients with scarce data.

Result: Simulations demonstrate FedCoSR’s effectiveness in mitigating label heterogeneity, achieving improvements in both accuracy and fairness compared to existing methods on datasets with varying label heterogeneity.

Conclusion: FedCoSR successfully addresses label distribution skew and data scarcity issues in federated learning through innovative use of contrastive learning and shareable representations, providing better accuracy and fairness while maintaining data privacy.

Abstract: Heterogeneity arising from label distribution skew and data scarcity can cause inaccuracy and unfairness in intelligent communication applications that heavily rely on distributed computing. To deal with it, this paper proposes a novel personalized federated learning algorithm, named Federated Contrastive Shareable Representations (FedCoSR), to facilitate knowledge sharing among clients while maintaining data privacy. Specifically, the parameters of local models’ shallow layers and typical local representations are both considered as shareable information for the server and are aggregated globally. To address performance degradation caused by label distribution skew among clients, contrastive learning is adopted between local and global representations to enrich local knowledge. Additionally, to ensure fairness for clients with scarce data, FedCoSR introduces adaptive local aggregation to coordinate the global model involvement in each client. Our simulations demonstrate FedCoSR’s effectiveness in mitigating label heterogeneity by achieving accuracy and fairness improvements over existing methods on datasets with varying degrees of label heterogeneity.

Method: Introduces a modality-aware, challenge-oriented perspective to analyze how foundation models pre-trained on different modalities face distinct hurdles when adapted to time series. Proposes taxonomy organized by pre-training modality (time series, language, vision) and categorizes corresponding solutions.

Result: Provides a comprehensive framework for understanding modality-specific challenges in adapting foundation models to time series tasks, analyzes advantages and limitations of different approaches, and reviews real-world applications and open-source codes.

Conclusion: Foundation models show great potential for revolutionizing time series analysis, but modality-specific adaptation challenges require careful consideration. The survey provides a structured foundation for future research in this rapidly evolving field with identified future directions.

Abstract: Time series data are ubiquitous across diverse real-world applications, making time series analysis critically important. Traditional approaches are largely task-specific, offering limited functionality and poor transferability. In recent years, foundation models have revolutionized NLP and CV with their remarkable cross-task transferability, zero-/few-shot learning capabilities, and multimodal integration capacity. This success has motivated increasing efforts to explore foundation models for addressing time series modeling challenges. Although some tutorials and surveys were published in the early stages of this field, the rapid pace of recent developments necessitates a more comprehensive and in-depth synthesis to cover the latest advances. Our survey aims to fill this gap by introducing a modality-aware, challenge-oriented perspective, which reveals how foundation models pre-trained on different modalities face distinct hurdles when adapted to time series tasks. Building on this perspective, we propose a taxonomy of existing works organized by pre-training modality (time series, language, and vision), analyze modality-specific challenges and categorize corresponding solutions, discussing their advantages and limitations. Beyond this, we review real-world applications to illustrate domain-specific advancements, provide open-source codes, and conclude with potential future research directions in this rapidly evolving field.

Method: A pointwise adaptive weighting method that balances residual decay rates across different training points to ensure more uniform convergence.

Result: The method offers bounded weights, high prediction accuracy, fast convergence rate, low training uncertainty, low computational cost, and ease of hyperparameter tuning compared to state-of-the-art methods.

Conclusion: Balancing residual decay rates across training points is an effective strategy for improving physics-informed deep learning performance on complex partial differential equation problems.

Abstract: Physics-informed deep learning has emerged as a promising alternative for solving partial differential equations. However, for complex problems, training these networks can still be challenging, often resulting in unsatisfactory accuracy and efficiency. In this work, we demonstrate that the failure of plain physics-informed neural networks arises from the significant discrepancy in the convergence rate of residuals at different training points, where the slowest convergence rate dominates the overall solution convergence. Based on these observations, we propose a pointwise adaptive weighting method that balances the residual decay rate across different training points. The performance of our proposed adaptive weighting method is compared with current state-of-the-art adaptive weighting methods on benchmark problems for both physics-informed neural networks and physics-informed deep operator networks. Through extensive numerical results we demonstrate that our proposed approach of balanced residual decay rates offers several advantages, including bounded weights, high prediction accuracy, fast convergence rate, low training uncertainty, low computational cost, and ease of hyperparameter tuning.

Method: Developed limited-switch algorithms for price-based blind network revenue management and bandits with knapsacks problems, establishing matching upper and lower bounds on optimal regret through theoretical analysis.

Result: The optimal regret rate is characterized by a piecewise-constant function of the switching budget that depends on the number of resource constraints. Algorithms achieve strong cumulative reward while significantly reducing switches.

Conclusion: Resource constraints fundamentally shape the statistical complexity of online learning under limited switches, and the developed algorithms provide efficient solutions for practical dynamic pricing applications with switching limitations.

Abstract: This paper studies the impact of limited switches on resource-constrained dynamic pricing with demand learning. We focus on the classical price-based blind network revenue management problem and extend our results to the bandits with knapsacks problem. In both settings, a decision maker faces stochastic and distributionally unknown demand, and must allocate finite initial inventory across multiple resources over time. In addition to standard resource constraints, we impose a switching constraint that limits the number of action changes over the time horizon. We establish matching upper and lower bounds on the optimal regret and develop computationally efficient limited-switch algorithms that achieve it. We show that the optimal regret rate is fully characterized by a piecewise-constant function of the switching budget, which further depends on the number of resource constraints. Our results highlight the fundamental role of resource constraints in shaping the statistical complexity of online learning under limited switches. Extensive simulations demonstrate that our algorithms maintain strong cumulative reward performance while significantly reducing the number of switches.

Method: The authors introduce a fixed-kinetic energy version of NHF, inspired by physics principles, which reduces parameter requirements while maintaining the continuous, volume-preserving, and invertible properties of Hamiltonian flows.

Result: The fixed-kinetic energy NHF demonstrates improved interpretability and robustness compared to the original model, validated on 2D Gaussian mixture, MNIST, and Fashion-MNIST datasets. The method is also successfully adapted for Bayesian inference in cosmology applications.

Conclusion: The physics-inspired fixed-kinetic energy modification to Neural Hamiltonian Flows provides a more interpretable, robust, and parameter-efficient architecture that maintains the desirable properties of Hamiltonian dynamics while being applicable to both generative modeling and Bayesian inference tasks.

Abstract: Normalizing Flows (NF) are Generative models which transform a simple prior distribution into the desired target. They however require the design of an invertible mapping whose Jacobian determinant has to be computable. Recently introduced, Neural Hamiltonian Flows (NHF) are Hamiltonian dynamics-based flows, which are continuous, volume-preserving and invertible and thus make for natural candidates for robust NF architectures. In particular, their similarity to classical Mechanics could lead to easier interpretability of the learned mapping. In this paper, we show that the current NHF architecture may still pose a challenge to interpretability. Inspired by Physics, we introduce a fixed-kinetic energy version of the model. This approach improves interpretability and robustness while requiring fewer parameters than the original model. We illustrate that on a 2D Gaussian mixture and on the MNIST and Fashion-MNIST datasets. Finally, we show how to adapt NHF to the context of Bayesian inference and illustrate the method on an example from cosmology.

Method: Developed a new loss function called weighted focal differentiable MCC and used it to train recurrent neural network (RNN) models. This loss function can also be used to fine-tune pre-trained models.

Result: RNN models trained with the new loss function achieved superior performance compared to models using weighted cross-entropy loss. The best model achieved state-of-the-art performance with 38.88% F1 score and 38.20% MCC on an independent test set from the largest available dataset.

Conclusion: The weighted focal differentiable MCC loss function enables significant improvement in O-GlcNAcylation site prediction, providing a more reliable and generalizable model for this important therapeutic target.

Abstract: O-GlcNAcylation, a subtype of glycosylation, has the potential to be an important target for therapeutics, but methods to reliably predict O-GlcNAcylation sites had not been available until 2023; a 2021 review correctly noted that published models were insufficient and failed to generalize. Moreover, many are no longer usable. In 2023, a considerably better recurrent neural network (RNN) model was published. This article creates improved models by using a new loss function, which we call the weighted focal differentiable MCC. RNN models trained with this new loss display superior performance to models trained using the weighted cross-entropy loss; this new function can also be used to fine-tune trained models. An RNN trained with this loss achieves state-of-the-art performance in O-GlcNAcylation site prediction with an F$_1$ score of 38.88% and an MCC of 38.20% on an independent test set from the largest dataset available.

Method: Propose frequency pretraining task where neural networks predict frequency content of randomly generated synthetic time series for pretraining before sleep staging.

Result: Method surpasses fully supervised learning with limited data and few subjects, matches performance with many subjects, and demonstrates frequency information relevance for sleep staging.

Conclusion: Frequency pretraining approach benefits EEG applications with limited data or few subjects, including brain-computer interfaces, while showing neural networks use information beyond frequencies.

Abstract: Analyzing electroencephalographic (EEG) time series can be challenging, especially with deep neural networks, due to the large variability among human subjects and often small datasets. To address these challenges, various strategies, such as self-supervised learning, have been suggested, but they typically rely on extensive empirical datasets. Inspired by recent advances in computer vision, we propose a pretraining task termed “frequency pretraining” to pretrain a neural network for sleep staging by predicting the frequency content of randomly generated synthetic time series. Our experiments demonstrate that our method surpasses fully supervised learning in scenarios with limited data and few subjects, and matches its performance in regimes with many subjects. Furthermore, our results underline the relevance of frequency information for sleep stage scoring, while also demonstrating that deep neural networks utilize information beyond frequencies to enhance sleep staging performance, which is consistent with previous research. We anticipate that our approach will be advantageous across a broad spectrum of applications where EEG data is limited or derived from a small number of subjects, including the domain of brain-computer interfaces.

Method: The study conducts an extensive benchmark on 16 diverse datasets, fully optimizing hyperparameters, feature encodings, and architectures for five model families. It also proposes reduced search spaces for efficient optimization.

Result: Dataset-specific tuning substantially improves performance for most models compared to original configurations. Diffusion-based models generally outperform other models, but this advantage is not significant when constrained to the same GPU budget.

Conclusion: Proper dataset-specific optimization is crucial for tabular data generation performance, and while diffusion models show promise, computational budget constraints can equalize performance across different model families.

Abstract: The ability to train generative models that produce realistic, safe and useful tabular data is essential for data privacy, imputation, oversampling, explainability or simulation. However, generating tabular data is not straightforward due to its heterogeneity, non-smooth distributions, complex dependencies and imbalanced categorical features. Although diverse methods have been proposed in the literature, there is a need for a unified evaluation, under the same conditions, on a variety of datasets. This study addresses this need by fully considering the optimization of: hyperparameters, feature encodings, and architectures. We investigate the impact of dataset-specific tuning on five recent model families for tabular data generation through an extensive benchmark on 16 datasets. These datasets vary in terms of size (an average of 80,000 rows), data types, and domains. We also propose a reduced search space for each model that allows for quick optimization, achieving nearly equivalent performance at a significantly lower cost. Our benchmark demonstrates that, for most models, large-scale dataset-specific tuning substantially improves performance compared to the original configurations. Furthermore, we confirm that diffusion-based models generally outperform other models on tabular data. However, this advantage is not significant when the entire tuning and training process is restricted to the same GPU budget.

Method: Proposed cross-attention pooling (CAP) framework with two novel attention functions that enable the target bag representation to incorporate information from the query instance, addressing the challenge of distinguishing similar instances.

Result: CAP significantly outperforms state-of-the-art MIL methods and baseline models across three verification tasks, achieving better classification accuracy and superior key instance detection capabilities.

Conclusion: The cross-attention pooling framework with novel attention functions effectively solves multiple instance verification problems by properly integrating query information, demonstrating substantial improvements over existing approaches.

Abstract: We explore multiple instance verification, a problem setting in which a query instance is verified against a bag of target instances with heterogeneous, unknown relevancy. We show that naive adaptations of attention-based multiple instance learning (MIL) methods and standard verification methods like Siamese neural networks are unsuitable for this setting: directly combining state-of-the-art (SOTA) MIL methods and Siamese networks is shown to be no better, and sometimes significantly worse, than a simple baseline model. Postulating that this may be caused by the failure of the representation of the target bag to incorporate the query instance, we introduce a new pooling approach named “cross-attention pooling” (CAP). Under the CAP framework, we propose two novel attention functions to address the challenge of distinguishing between highly similar instances in a target bag. Through empirical studies on three different verification tasks, we demonstrate that CAP outperforms adaptations of SOTA MIL methods and the baseline by substantial margins, in terms of both classification accuracy and the ability to detect key instances. The superior ability to identify key instances is attributed to the new attention functions by ablation studies. We share our code at https://github.com/xxweka/MIV.

Method: Integrates ABBA (adaptive Brownian bridge-based symbolic aggregation) symbolic representation with LLMs, using a fixed-polygonal chain trick to prevent cumulative errors during symbol-to-numerical value conversion.

Result: Achieves SOTA performance on UCR classification, medical time series classification, and TSER benchmarks. Shows competitive prediction capability compared to recent SOTA methods.

Conclusion: LLM-ABBA provides an effective framework for integrating symbolic time series representation with LLMs, demonstrating strong performance across multiple time series tasks with potential for extension to other applications.

Abstract: The success of large language models (LLMs) for time series has been demonstrated in previous work. Utilizing a symbolic time series representation, one can efficiently bridge the gap between LLMs and time series. However, the remaining challenge is to exploit the semantic information hidden in time series by using symbols or existing tokens of LLMs, while aligning the embedding space of LLMs according to the hidden information of time series. The symbolic time series approximation (STSA) method called adaptive Brownian bridge-based symbolic aggregation (ABBA) shows outstanding efficacy in preserving salient time series features by modeling time series patterns in terms of amplitude and period while using existing tokens of LLMs. In this paper, we introduce a method, called LLM-ABBA, that integrates ABBA into large language models for various downstream time series tasks. By symbolizing time series, LLM-ABBA compares favorably to the recent state-of-the-art (SOTA) in UCR and three medical time series classification tasks. Meanwhile, a fixed-polygonal chain trick in ABBA is introduced to \kc{avoid obvious drifting} during prediction tasks by significantly mitigating the effects of cumulative error arising from misused symbols during the transition from symbols to numerical values. In time series regression tasks, LLM-ABBA achieves the new SOTA on Time Series Extrinsic Regression (TSER) benchmarks. LLM-ABBA also shows competitive prediction capability compared to recent SOTA time series prediction results. We believe this framework can also seamlessly extend to other time series tasks.

Method: Proposed pipeline decouples dependencies in sampled relations for training and applies tailored DP-SGD to enable privacy-preserving fine-tuning of large language models on sensitive graph data.

Result: Evaluated on four real-world text-attributed graphs, showing significant improvements in relational learning tasks while maintaining robust privacy guarantees.

Conclusion: The approach successfully addresses privacy-utility-computation trade-offs, enabling practical deployment of privacy-preserving relational learning on sensitive graph data.

Abstract: Graphs offer unique insights into relationships between entities, complementing data modalities like text and images and enabling AI models to extend their capabilities beyond traditional tasks. However, learning from graphs often involves handling sensitive relationships in the data, raising significant privacy concerns. Existing privacy-preserving methods, such as DP-SGD, rely on gradient decoupling assumptions and are incompatible with relational learning due to the inherent dependencies between training samples. To address this challenge, we propose a privacy-preserving pipeline for relational learning that decouples dependencies in sampled relations for training, ensuring differential privacy through a tailored application of DP-SGD. We apply this approach to fine-tune large language models (LLMs), such as Llama2, on sensitive graph data while addressing the associated computational complexities. Our method is evaluated on four real-world text-attributed graphs, demonstrating significant improvements in relational learning tasks while maintaining robust privacy guarantees. Additionally, we analyze the trade-offs between privacy, utility, and computational efficiency, offering insights into the practical deployment of our approach for privacy-preserving relational learning. Code is available at https://github.com/Graph-COM/PvGaLM.

Method: Generate virtual data samples and evaluate model consistency scores across local updates to identify and filter out compromised updates before the aggregation phase.

Result: Plugin-attached FedAvg achieves 89.6% test accuracy under 30% targeted attacks (vs 19.5% without plugin) and maintains 65-70% accuracy under untargeted attacks (vs 17-19% without plugin) on blood cell classification.

Conclusion: The proposed plugin effectively provides strong Byzantine resilience to existing FL methods while preserving their original benefits, making FL systems more robust against attacks from compromised devices.

Abstract: Given sufficient data from multiple edge devices, federated learning (FL) enables training a shared model without transmitting private data to the central server. However, FL is generally vulnerable to Byzantine attacks from compromised edge devices, which can significantly degrade the model performance. In this work, we propose an intuitive plugin that seamlessly embeds Byzantine resilience into existing FL methods. The key idea is to generate virtual data samples and evaluate model consistency scores across local updates to effectively filter out compromised updates. By utilizing this scoring mechanism before the aggregation phase, the proposed plugin enables existing FL methods to become robust against Byzantine attacks while maintaining their original benefits. Numerical results on blood cell classification task demonstrate that the proposed plugin provides strong Byzantine resilience. In detail, plugin-attached FedAvg achieves over 89.6% test accuracy under 30% targeted attacks (vs.19.5% w/o plugin) and maintains 65-70% test accuracy under untargeted attacks (vs.17-19% w/o plugin).

Method: Proposes DRDT3 framework with DT3 module (combining self-attention and TTT layer RNN) for coarse action predictions, then iteratively refines them using a diffusion model with unified optimization objective.

Result: DT3 without diffusion shows improved performance over standard DT, while DRDT3 achieves superior results compared to state-of-the-art DT-based and offline RL methods on D4RL benchmark tasks.

Conclusion: The unified DRDT3 framework successfully combines sequence modeling strengths of attention/RNN with generative diffusion refinement to overcome Decision Transformer limitations and achieve optimal policy learning from suboptimal trajectories.

Abstract: Decision Transformer (DT), a trajectory modelling method, has shown competitive performance compared to traditional offline reinforcement learning (RL) approaches on various classic control tasks. However, it struggles to learn optimal policies from suboptimal, reward-labelled trajectories. In this study, we explore the use of conditional generative modelling to facilitate trajectory stitching given its high-quality data generation ability. Additionally, recent advancements in Recurrent Neural Networks (RNNs) have shown their linear complexity and competitive sequence modelling performance over Transformers. We leverage the Test-Time Training (TTT) layer, an RNN that updates hidden states during testing, to model trajectories in the form of DT. We introduce a unified framework, called Diffusion-Refined Decision TTT (DRDT3), to achieve performance beyond DT models. Specifically, we propose the Decision TTT (DT3) module, which harnesses the sequence modelling strengths of both self-attention and the TTT layer to capture recent contextual information and make coarse action predictions. DRDT3 iteratively refines the coarse action predictions through the generative diffusion model, progressively moving closer to the optimal actions. We further integrate DT3 with the diffusion model using a unified optimization objective. With experiments on multiple tasks in the D4RL benchmark, our DT3 model without diffusion refinement demonstrates improved performance over standard DT, while DRDT3 further achieves superior results compared to state-of-the-art DT-based and offline RL methods.

Method: Created the Gemstones dataset with over 4000 transformer checkpoints featuring diverse architectural shapes (width/depth variations), hyperparameter ablations (learning rate, cooldown), and models up to 2 billion parameters.

Result: Analysis revealed that scaling law prescriptions are highly sensitive to experimental design choices and specific model checkpoints used during fitting, challenging previous narrow-scope findings.

Conclusion: The Gemstones dataset enables more robust scaling law research by providing diverse architectural and hyperparameter variations, demonstrating that scaling prescriptions depend significantly on the breadth of experimental conditions considered.

Abstract: Scaling laws are typically fit using a family of models with a narrow range of frozen hyper-parameter choices. In this work we study scaling laws using multiple architectural shapes and hyperparameter choices, highlighting their impact on resulting prescriptions. As a primary artifact of our research, we release the Gemstones: an open-source scaling law dataset, consisting of over 4000 checkpoints from transformers with up to 2 billion parameters and diverse architectural shapes; including ablations over learning rate and cooldown. Our checkpoints enable more complex studies of scaling, such as analyzing the relationship between width and depth. By examining our model suite, we find that the prescriptions of scaling laws can be highly sensitive to the experimental design process and the specific model checkpoints used during fitting.

Method: Two UCB-based algorithms: Double-UCB (problem-independent regret bounds) and Conservative-UCB (problem-dependent bounds), extended to ballooning settings with increasing arms. Also proposes versions that don’t require prior graph structure knowledge.

Result: Establishes regret lower bounds, provides regret upper bounds for both algorithms, shows sub-linearity under mild assumptions, and validates results through experiments.

Conclusion: The proposed algorithms effectively leverage similarity structure in graph feedback bandits and handle ballooning arm scenarios, providing theoretical guarantees and practical performance for real-world applications.

Abstract: In this paper, we study the stochastic multi-armed bandit problem with graph feedback. Motivated by applications in clinical trials and recommendation systems, we assume that two arms are connected if and only if they are similar (i.e., their means are close to each other). We establish a regret lower bound for this problem under the novel feedback structure and introduce two upper confidence bound (UCB)-based algorithms: Double-UCB, which has problem-independent regret upper bounds, and Conservative-UCB, which has problem-dependent upper bounds. Leveraging the similarity structure, we also explore a scenario where the number of arms increases over time (referred to as the \emph{ballooning setting}). Practical applications of this scenario include Q&A platforms (e.g., Reddit, Stack Overflow, Quora) and product reviews on platforms like Amazon and Flipkart, where answers (or reviews) continuously appear, and the goal is to display the best ones at the top. We extend these two UCB-based algorithms to the ballooning setting. Under mild assumptions, we provide regret upper bounds for both algorithms and discuss their sub-linearity. Furthermore, we propose a new version of the corresponding algorithms that do not rely on prior knowledge of the graph’s structural information and provide regret upper bounds. Finally, we conduct experiments to validate the theoretical results.

Method: Proposes LocalEscaper framework that combines supervised and reinforcement learning. Key innovation is regional reconstruction strategy to mitigate local-optima problems in existing methods.

Result: Outperforms existing neural solvers on both synthetic and real-world datasets, achieving remarkable results with effective training on low-quality labeled data.

Conclusion: LocalEscaper successfully addresses limitations of both SL and RL approaches for TSP, providing an effective weakly-supervised framework that works with imperfect training data while avoiding local optima.

Abstract: Neural solvers have shown significant potential in solving the Traveling Salesman Problem (TSP), yet current approaches face significant challenges. Supervised learning (SL)-based solvers require large amounts of high-quality labeled data, while reinforcement learning (RL)-based solvers, though less dependent on such data, often suffer from inefficiencies. To address these limitations, we propose LocalEscaper, a novel weakly-supervised learning framework for large-scale TSP. LocalEscaper effectively combines the advantages of both SL and RL, enabling effective training on datasets with low-quality labels. To further enhance solution quality, we introduce a regional reconstruction strategy, which is the key technique of this paper and mitigates the local-optima problem common in existing local reconstruction methods. Experimental results on both synthetic and real-world datasets demonstrate that LocalEscaper outperforms existing neural solvers, achieving remarkable results.

Method: Experimental evaluation of NS-ES and NSR-ES algorithms on transformer architectures including Decision Transformers for reinforcement learning problems, with testing of pretrained model seeding for acceleration.

Result: Mixed results - NS-ES showed progress but required many more iterations for meaningful agent development. NSR-ES proved capable of straightforward application to larger models, showing similar performance between feed-forward models and Decision Transformers as previously observed with OpenAI-ES.

Conclusion: NSR-ES appears more suitable for larger transformer models in reinforcement learning contexts, while NS-ES requires significantly more computational resources and iterations to achieve comparable results.

Abstract: In this paper, we experiment with novelty-based variants of OpenAI-ES, the NS-ES and NSR-ES algorithms, and evaluate their effectiveness in training complex, transformer-based architectures designed for the problem of reinforcement learning, such as Decision Transformers. We also test if we can accelerate the novelty-based training of these larger models by seeding the training with a pretrained models. The experimental results were mixed. NS-ES showed progress, but it would clearly need many more iterations for it to yield interesting agents. NSR-ES, on the other hand, proved quite capable of being straightforwardly used on larger models, since its performance appears as similar between the feed-forward model and Decision Transformer, as it was for the OpenAI-ES in our previous work.

Method: Uses graph-based collaborative filtering to model user-response relationships, integrates mixture of LoRA experts for efficient fine-tuning, and employs optimization-free adaptation strategy for generalization to unseen users.

Result: Outperforms existing personalized reward models on UltraFeedback-P dataset, effectively capturing both common and controversial preferences while working well in sparse annotation settings.

Conclusion: CoPL provides a scalable solution for personalized LLM alignment that balances shared and user-specific preferences efficiently and generalizes well to new users without fine-tuning.

Abstract: Personalizing large language models (LLMs) is important for aligning outputs with diverse user preferences, yet existing methods struggle with flexibility and generalization. We propose CoPL (Collaborative Preference Learning), a graph-based collaborative filtering framework that models user-response relationships to enhance preference estimation, particularly in sparse annotation settings. By integrating a mixture of LoRA experts, CoPL efficiently fine-tunes LLMs while dynamically balancing shared and user-specific preferences. Additionally, an optimization-free adaptation strategy enables generalization to unseen users without fine-tuning. Experiments on UltraFeedback-P demonstrate that CoPL outperforms existing personalized reward models, effectively capturing both common and controversial preferences, making it a scalable solution for personalized LLM alignment. The code is available at https://github.com/ml-postech/CoPL.

Method: Leverages image-to-image editing models to perform gradual interventions on semantic properties and measure impact using expected property gradient magnitude score.

Result: Effective identification of local biases in synthetic scenarios, analysis of medical skin lesion classifiers, network training dynamics, and CLIP model behavior with real interventional data.

Conclusion: Interventional explanations on property level reveal new insights into deep model behavior, demonstrating significant potential for causal understanding of predictions.

Abstract: Deep learning models achieve high predictive performance but lack intrinsic interpretability, hindering our understanding of the learned prediction behavior. Existing local explainability methods focus on associations, neglecting the causal drivers of model predictions. Other approaches adopt a causal perspective but primarily provide global, model-level explanations. However, for specific inputs, it’s unclear whether globally identified factors apply locally. To address this limitation, we introduce a novel framework for local interventional explanations by leveraging recent advances in image-to-image editing models. Our approach performs gradual interventions on semantic properties to quantify the corresponding impact on a model’s predictions using a novel score, the expected property gradient magnitude. We demonstrate the effectiveness of our approach through an extensive empirical evaluation on a wide range of architectures and tasks. First, we validate it in a synthetic scenario and demonstrate its ability to locally identify biases. Afterward, we apply our approach to investigate medical skin lesion classifiers, analyze network training dynamics, and study a pre-trained CLIP model with real-life interventional data. Our results highlight the potential of interventional explanations on the property level to reveal new insights into the behavior of deep models.

Method: Introduces out-of-context representation learning - a lightweight technique that trains only new token embeddings on axioms during training, then evaluates on unseen reasoning tasks requiring one or more steps of reflexivity, symmetry, and transitivity.

Result: LLMs perform statistically significantly better than chance across various reasoning tests, with correct answers extractable through multiple phrasing variations, though not perfectly consistent. Learned embeddings show structured organization indicating real relational understanding.

Conclusion: LLMs demonstrate capability for relational reasoning through memorized knowledge, with the core reasoning process occurring during training rather than inference, suggesting internal structured representations of binary relations.

Abstract: We study how large language models (LLMs) reason about memorized knowledge through simple binary relations such as equality ($=$), inequality ($<$), and inclusion ($\subset$). Unlike in-context reasoning, the axioms (e.g., $a < b, b < c$) are only seen during training and not provided in the task prompt (e.g., evaluating $a < c$). The tasks require one or more reasoning steps, and data aggregation from one or more sources, showing performance change with task complexity. We introduce a lightweight technique, out-of-context representation learning, which trains only new token embeddings on axioms and evaluates them on unseen tasks. Across reflexivity, symmetry, and transitivity tests, LLMs mostly perform statistically significant better than chance, making the correct answer extractable when testing multiple phrasing variations, but still fall short of consistent reasoning on every single query. Analysis shows that the learned embeddings are organized in structured ways, suggesting real relational understanding. Surprisingly, it also indicates that the core reasoning happens during the training, not inference.

Method: MetaSel uses both fine-tuned and pre-trained models to estimate misclassification probability. It identifies inputs where the models’ behaviors diverge (where fine-tuning altered decision boundaries) as these are more prone to errors.

Result: MetaSel outperformed 11 state-of-the-art approaches across 68 fine-tuned models, showing average TRC improvements of 28.46% to 56.18% over the best baselines, with high median coverage and low variability.

Conclusion: MetaSel is a practical, robust, and cost-effective solution for test selection in fine-tuned models, particularly effective under highly constrained labeling budgets for addressing covariate shift issues.

Abstract: Deep Neural Networks (DNNs) face challenges during deployment due to covariate shift, i.e., data distribution shifts between development and deployment contexts. Fine-tuning adapts pre-trained models to new contexts requiring smaller labeled sets. However, testing fine-tuned models under constrained labeling budgets remains a critical challenge. This paper introduces MetaSel, a new approach tailored for DNN models that have been fine-tuned to address covariate shift, to select tests from unlabeled inputs. MetaSel assumes that fine-tuned and pre-trained models share related data distributions and exhibit similar behaviors for many inputs. However, their behaviors diverge within the input subspace where fine-tuning alters decision boundaries, making those inputs more prone to misclassification. Unlike general approaches that rely solely on the DNN model and its input set, MetaSel leverages information from both the fine-tuned and pre-trained models and their behavioral differences to estimate misclassification probability for unlabeled test inputs, enabling more effective test selection. Our extensive empirical evaluation, comparing MetaSel against 11 state-of-the-art approaches and involving 68 fine-tuned models across weak, medium, and strong distribution shifts, demonstrates that MetaSel consistently delivers significant improvements in Test Relative Coverage (TRC) over existing baselines, particularly under highly constrained labeling budgets. MetaSel shows average TRC improvements of 28.46% to 56.18% over the most frequent second-best baselines while maintaining a high TRC median and low variability. Our results confirm MetaSel’s practicality, robustness, and cost-effectiveness for test selection in the context of fine-tuned models.

Method: Characterizes statistical heterogeneity with 6 metrics, creates 7 CV datasets, and develops FEDDIVERSE algorithm that selects clients with complementary data distributions.

Result: Experiments show FEDDIVERSE enhances performance and robustness of various FL methods while maintaining low communication and computational overhead.

Conclusion: FEDDIVERSE effectively manages data heterogeneity in FL through strategic client selection, improving model generalization across diverse real-world scenarios.

Abstract: Federated Learning (FL) enables decentralized training of machine learning models on distributed data while preserving privacy. However, in real-world FL settings, client data is often non-identically distributed and imbalanced, resulting in statistical data heterogeneity which impacts the generalization capabilities of the server’s model across clients, slows convergence and reduces performance. In this paper, we address this challenge by proposing first a characterization of statistical data heterogeneity by means of 6 metrics of global and client attribute imbalance, class imbalance, and spurious correlations. Next, we create and share 7 computer vision datasets for binary and multiclass image classification tasks in Federated Learning that cover a broad range of statistical data heterogeneity and hence simulate real-world situations. Finally, we propose FEDDIVERSE, a novel client selection algorithm in FL which is designed to manage and leverage data heterogeneity across clients by promoting collaboration between clients with complementary data distributions. Experiments on the seven proposed FL datasets demonstrate FEDDIVERSE’s effectiveness in enhancing the performance and robustness of a variety of FL methods while having low communication and computational overhead.

Method: PALSGD extends Local SGD and DiLoCo by introducing a pseudo-synchronization mechanism that allows longer synchronization intervals while maintaining model consistency through careful coordination between workers.

Result: PALSGD achieves significant speed improvements: 18.4% faster than DDP on ImageNet-1K with ResNet-50, 24.4% faster on TinyStories with GPT-Neo-125M, and 21.1% faster on TinyStories with GPT-Neo-8M, while maintaining comparable performance.

Conclusion: PALSGD effectively addresses communication bottlenecks in large-scale distributed training, providing both theoretical convergence guarantees and practical performance improvements across vision and language tasks.

Abstract: Following AI scaling trends, frontier models continue to grow in size and continue to be trained on larger datasets. Training these models requires huge investments in exascale computational resources, which has in turn driven developtment of distributed deep learning methods. Data parallelism is an essential approach to speed up training, but it requires frequent global communication between workers, which can bottleneck training at the largest scales. In this work, we propose a method called Pseudo-Asynchronous Local SGD (PALSGD) to improve the efficiency of data-parallel training. PALSGD is an extension of Local SGD (Stich, 2018) and DiLoCo (Douillard et al., 2023), designed to further reduce communication frequency by introducing a pseudo-synchronization mechanism. PALSGD allows the use of longer synchronization intervals compared to standard Local SGD. Despite the reduced communication frequency, the pseudo-synchronization approach ensures that model consistency is maintained, leading to performance results comparable to those achieved with more frequent synchronization. Furthermore, we provide a theoretical analysis of PALSGD, establishing its convergence and deriving its convergence rate. This analysis offers insights into the algorithm’s behavior and performance guarantees. We evaluated PALSGD on image classification and language modeling tasks. Our results show that PALSGD achieves better performance in less time compared to existing methods like Distributed Data Parallel (DDP), and DiLoCo. Notably, PALSGD trains 18.4% faster than DDP on ImageNet-1K with ResNet-50, 24.4% faster than DDP on TinyStories with GPT-Neo-125M, and 21.1% faster than DDP on TinyStories with GPT-Neo-8M.

Method: Comprehensive benchmarking of KAN vs MLP across six state-of-the-art FL algorithms on blood cell classification dataset, analyzing hyperparameters and Non-IID data distributions.

Result: KAN effectively replaces MLP in federated environments with superior performance, optimized width over depth yields best results, and performs well under varying Non-IID conditions.

Conclusion: KAN establishes as a promising alternative for privacy-preserving medical imaging in distributed healthcare, being the first comprehensive FL benchmark for medical tasks.

Abstract: Federated Learning (FL) enables model training across decentralized devices without sharing raw data, thereby preserving privacy in sensitive domains like healthcare. In this paper, we evaluate Kolmogorov-Arnold Networks (KAN) architectures against traditional MLP across six state-of-the-art FL algorithms on a blood cell classification dataset. Notably, our experiments demonstrate that KAN can effectively replace MLP in federated environments, achieving superior performance with simpler architectures. Furthermore, we analyze the impact of key hyperparameters-grid size and network architecture-on KAN performance under varying degrees of Non-IID data distribution. In addition, our ablation studies reveal that optimizing KAN width while maintaining minimal depth yields the best performance in federated settings. As a result, these findings establish KAN as a promising alternative for privacy-preserving medical imaging applications in distributed healthcare. To the best of our knowledge, this is the first comprehensive benchmark of KAN in FL settings for medical imaging task.

Method: Integration of local neural operators with advanced iterative numerical methods in the Krylov subspace, using local in time, space, and space-time (“patch”) neural operators to accelerate computer-aided analysis of spatiotemporal dynamics.

Result: The framework successfully performs system-level stability and bifurcation analysis on three nonlinear PDE benchmarks: 1D Allen-Cahn equation (multiple pitchfork bifurcations), Liouville-Bratu-Gelfand PDE (saddle-node tipping point), and FitzHugh-Nagumo model (Hopf and saddle-node bifurcations).

Conclusion: The proposed framework demonstrates the effectiveness of combining neural operators with traditional numerical methods for advanced system-level analysis, expanding the application scope of neural operators beyond temporal prediction to include rigorous stability and bifurcation analysis of complex dynamical systems.

Abstract: Neural Operators (NOs) provide a powerful framework for computations involving physical laws that can be modelled by (integro-) partial differential equations (PDEs), directly learning maps between infinite-dimensional function spaces that bypass both the explicit equation identification and their subsequent numerical solving. Still, NOs have so far primarily been employed to explore the dynamical behavior as surrogates of brute-force temporal simulations/predictions. Their potential for systematic rigorous numerical system-level tasks, such as fixed-point, stability, and bifurcation analysis - crucial for predicting irreversible transitions in real-world phenomena - remains largely unexplored. Toward this aim, inspired by the Equation-Free multiscale framework, we propose and implement a framework that integrates (local) NOs with advanced iterative numerical methods in the Krylov subspace, so as to perform efficient system-level stability and bifurcation analysis of large-scale dynamical systems. Beyond fixed point, stability, and bifurcation analysis enabled by local in time NOs, we also demonstrate the usefulness of local in space as well as in space-time (“patch”) NOs in accelerating the computer-aided analysis of spatiotemporal dynamics. We illustrate our framework via three nonlinear PDE benchmarks: the 1D Allen-Cahn equation, which undergoes multiple concatenated pitchfork bifurcations; the Liouville-Bratu-Gelfand PDE, which features a saddle-node tipping point; and the FitzHugh-Nagumo (FHN) model, consisting of two coupled PDEs that exhibit both Hopf and saddle-node bifurcations.

Method: Forces the Jacobian matrix to be diagonal, enabling parallel processing of full sequences with O(TD) time/memory and O(log T) sequential depth. Maintains non-linear capabilities without performance loss compared to dense Jacobian models.

Result: Outperforms Transformers, LRU, S5, and Mamba on long-range forecasting tasks while providing formal gradient-stability guarantees that other input-varying systems lack.

Conclusion: LrcSSM demonstrates that non-linear recurrent models can achieve linear efficiency without performance degradation, with broader applicability to other non-linear recurrent architectures.

Abstract: We present LrcSSM, a $\textit{non-linear}$ recurrent model that processes long sequences as fast as today’s linear state-space layers. By forcing the Jacobian matrix to be diagonal, the full sequence can be solved in parallel, giving $\mathcal{O}(TD)$ time and memory and only $\mathcal{O}(\log T)$ sequential depth, for input-sequence length $T$ and a state dimension $D$. Moreover, LrcSSM offers a formal gradient-stability guarantee that other input-varying systems such as Liquid-S4 and Mamba do not provide. Importantly, the diagonal Jacobian structure of our model results in no performance loss compared to the original model with dense Jacobian, and the approach can be generalized to other non-linear recurrent models, demonstrating broader applicability. On a suite of long-range forecasting tasks, we demonstrate that LrcSSM outperforms Transformers, LRU, S5, and Mamba.

Method: Applied three state-of-the-art self-supervised learning methods to bulk gene expression data to assess their ability to generate qualitative representations for downstream phenotype prediction tasks using publicly available datasets.

Result: Self-supervised learning methods effectively capture complex information and improve phenotype prediction accuracy compared to traditional supervised models, while significantly reducing dependency on annotated data.

Conclusion: Self-supervised learning is a promising approach for gene expression data analysis, offering performance advantages over supervised methods. This is the first work applying self-supervised learning to bulk RNA-Seq data, with recommendations provided for method selection and future research directions outlined.

Abstract: Predicting phenotypes from gene expression data is a crucial task in biomedical research, enabling insights into disease mechanisms, drug responses, and personalized medicine. Traditional machine learning and deep learning rely on supervised learning, which requires large quantities of labeled data that are costly and time-consuming to obtain in the case of gene expression data. Self-supervised learning has recently emerged as a promising approach to overcome these limitations by extracting information directly from the structure of unlabeled data. In this study, we investigate the application of state-of-the-art self-supervised learning methods to bulk gene expression data for phenotype prediction. We selected three self-supervised methods, based on different approaches, to assess their ability to exploit the inherent structure of the data and to generate qualitative representations which can be used for downstream predictive tasks. By using several publicly available gene expression datasets, we demonstrate how the selected methods can effectively capture complex information and improve phenotype prediction accuracy. The results obtained show that self-supervised learning methods can outperform traditional supervised models besides offering significant advantage by reducing the dependency on annotated data. We provide a comprehensive analysis of the performance of each method by highlighting their strengths and limitations. We also provide recommendations for using these methods depending on the case under study. Finally, we outline future research directions to enhance the application of self-supervised learning in the field of gene expression data analysis. This study is the first work that deals with bulk RNA-Seq data and self-supervised learning.

Method: Proposes an interference-aware torque metric combining structural distance and energy scores to quantify edge perturbation. Uses this metric to hierarchically reconfigure receptive fields by pruning high-torque edges and adding low-torque links to suppress noise and boost relevant signals.

Result: Extensive evaluations show the approach surpasses state-of-the-art methods on both heterophilous and homophilous graphs, while maintaining high accuracy on noisy graphs.

Conclusion: The torque-driven hierarchical rewiring strategy effectively improves GNN representation learning by dynamically modulating message passing, demonstrating superior performance across various graph types and robustness against noise.

Abstract: Graph Neural Networks (GNNs) have emerged as powerful tools for learning from graph-structured data, leveraging message passing to diffuse information and update node representations. However, most efforts have suggested that native interactions encoded in the graph may not be friendly for this process, motivating the development of graph rewiring methods. In this work, we propose a torque-driven hierarchical rewiring strategy, inspired by the notion of torque in classical mechanics, dynamically modulating message passing to improve representation learning in heterophilous graphs and enhance robustness against noisy graphs. Specifically, we define an interference-aware torque metric that integrates structural distance and energy scores to quantify the perturbation induced by edges, thereby encouraging each node to aggregate information from its nearest low-energy neighbors. We use the metric to hierarchically reconfigure the receptive field of each layer by judiciously pruning high-torque edges and adding low-torque links, suppressing propagation noise and boosting pertinent signals. Extensive evaluations on benchmark datasets show that our approach surpasses state-of-the-art methods on both heterophilous and homophilous graphs, and maintains high accuracy on noisy graph.

Method: Developed HEF with dynamic weights, tolerance thresholds, and progressive penalty mechanisms. Implemented using Grid Search, PSO, and Optuna optimization on Walmart, M3, M4, and M5 benchmark datasets.

Result: HEF consistently outperformed MAE in global metrics (R2, GRA, RMSE, RMSSE) with greater explanatory power, adaptability, and stability.

Conclusion: HEF provides a robust and adaptive alternative for model selection and hyperparameter optimization in variable demand forecasting environments, though MAE remains simpler and more efficient.

Abstract: Accurate demand forecasting is crucial for effective inventory management in dynamic and competitive environments, where decisions are influenced by uncertainty, financial constraints, and logistical limitations. Traditional evaluation metrics such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) provide complementary perspectives but may lead to biased assessments when applied individually. To address this limitation, we propose the Hierarchical Evaluation Function (HEF), a composite function that integrates R2, MAE, and RMSE within a hierarchical and adaptive framework. The function incorporates dynamic weights, tolerance thresholds derived from the statistical properties of the series, and progressive penalty mechanisms to ensure robustness against extreme errors and invalid predictions. HEF was implemented to optimize multiple forecasting models using Grid Search, Particle Swarm Optimization (PSO), and Optuna, and tested on benchmark datasets including Walmart, M3, M4, and M5. Experimental results, validated through statistical tests, demonstrate that HEF consistently outperforms MAE as an evaluation function in global metrics such as R2, Global Relative Accuracy (GRA), RMSE, and RMSSE, thereby providing greater explanatory power, adaptability, and stability. While MAE retains advantages in simplicity and efficiency, HEF proves more effective for long-term planning and complex contexts. Overall, HEF constitutes a robust and adaptive alternative for model selection and hyperparameter optimization in highly variable demand forecasting environments.

Method: The authors propose a tiled-block strategy to compute attention within the TabPFN framework. This approach avoids the need for pre-processing (like KNN-based sample selection) and is compatible with standard GPU setups, allowing efficient processing of long contexts.

Result: The method successfully enables TabPFN to handle long contexts without pre-processing. It is demonstrated on the standard TabArena benchmark, showing effectiveness in scaling TabPFN beyond its previous limitations.

Conclusion: The tiled-block attention strategy is a novel and efficient solution to overcome TabPFN’s context size constraints, making it practical for larger tabular datasets without compromising performance or requiring additional pre-processing steps.

Abstract: TabPFN v2 achieves better results than tree-based models on several tabular benchmarks, which is notable since tree-based models are usually the strongest choice for tabular data. However, it cannot handle more than 10K context tokens because transformers have quadratic computation and memory costs. Unlike existing approaches that rely on context compression, such as selecting representative samples via K-nearest neighbors (KNN), we introduce a tiled-block strategy to compute attention within the TabPFN framework. This design is compatible with standard GPU setups and, to the best of our knowledge, is the first to enable TabPFN to process long contexts without any pre-processing. We demonstrate the effectiveness of our approach on the standard TabArena benchmark, with code available at https://github.com/mrsergazinov/chunk_tabpfn.

Method: Two approaches: 1) Parametric Gaussian copula method assuming prediction errors and uncertainty estimates follow Gaussian copula distribution for closed-form PI computation, and 2) Non-parametric KNN approach that empirically estimates conditional error distribution using similar validation instances.

Result: Gaussian copula method consistently outperforms conformal prediction baselines on low-frequency data, while KNN approach performs best on high-frequency data across two large public datasets with minute- and hour-level sampling.

Conclusion: RUE-derived prediction intervals show clinical promise for delivering interpretable, uncertainty-aware vital sign forecasts, addressing the trust and interpretability challenges in deploying deep learning models in healthcare.

Abstract: Vital signs, such as heart rate and blood pressure, are critical indicators of patient health and are widely used in clinical monitoring and decision-making. While deep learning models have shown promise in forecasting these signals, their deployment in healthcare remains limited in part because clinicians must be able to trust and interpret model outputs. Without reliable uncertainty quantification – particularly calibrated prediction intervals (PIs) – it is unclear whether a forecasted abnormality constitutes a meaningful warning or merely reflects model noise, hindering clinical decision-making. To address this, we present two methods for deriving PIs from the Reconstruction Uncertainty Estimate (RUE), an uncertainty measure well-suited to vital-sign forecasting due to its sensitivity to data shifts and support for label-free calibration. Our parametric approach assumes that prediction errors and uncertainty estimates follow a Gaussian copula distribution, enabling closed-form PI computation. Our non-parametric approach, based on k-nearest neighbours (KNN), empirically estimates the conditional error distribution using similar validation instances. We evaluate these methods on two large public datasets with minute- and hour-level sampling, representing high- and low-frequency health signals. Experiments demonstrate that the Gaussian copula method consistently outperforms conformal prediction baselines on low-frequency data, while the KNN approach performs best on high-frequency data. These results underscore the clinical promise of RUE-derived PIs for delivering interpretable, uncertainty-aware vital sign forecasts.

Method: DeepHit survival model with projection layer and inter-modality cross-attention to integrate clinical, MRI, RNA-seq, and pathology features.

Result: Achieved C-index of 0.843 (prostate) and 0.662 (bladder) on cross-validation; 0.818 and 0.457 on development sets respectively.

Conclusion: Multimodal integration with deep survival learning provides promising personalized risk stratification for cancer patients and is broadly applicable to biomedical survival prediction.

Abstract: Accurate prediction of time-to-event outcomes is a central challenge in oncology, with significant implications for treatment planning and patient management. In this work, we present ModaliSurv, a multimodal deep survival model utilising DeepHit with a projection layer and inter-modality cross-attention, which integrates heterogeneous patient data, including clinical, MRI, RNA-seq and whole-slide pathology features. The model is designed to capture complementary prognostic signals across modalities and estimate individualised time-to-biochemical recurrence in prostate cancer and time-to-cancer recurrence in bladder cancer. Our approach was evaluated in the context of the CHIMERA Grand Challenge, across two of the three provided tasks. For Task 1 (prostate cancer bio-chemical recurrence prediction), the proposed framework achieved a concordance index (C-index) of 0.843 on 5-folds cross-validation and 0.818 on CHIMERA development set, demonstrating robust discriminatory ability. For Task 3 (bladder cancer recurrence prediction), the model obtained a C-index of 0.662 on 5-folds cross-validation and 0.457 on development set, highlighting its adaptability and potential for clinical translation. These results suggest that leveraging multimodal integration with deep survival learning provides a promising pathway toward personalised risk stratification in prostate and bladder cancer. Beyond the challenge setting, our framework is broadly applicable to survival prediction tasks involving heterogeneous biomedical data.

Method: Improves boolean rewriting technique (XPAT) with parametrizable template for circuit rewriting, introduces novel parametrizable product sharing template that acts as proxy for synthesized area.

Result: Methodology converges better to low-area solutions and finds better approximations than original XPAT and two other state-of-the-art approaches.

Conclusion: Proposed template-based approach with parametrizable product sharing effectively reduces arithmetic operator area while maintaining accuracy, outperforming existing methods for neural network deployment on edge devices.

Abstract: Deploying neural networks on edge devices entails a careful balance between the energy required for inference and the accuracy of the resulting classification. One technique for navigating this tradeoff is approximate computing: the process of reducing energy consumption by slightly reducing the accuracy of arithmetic operators. In this context, we propose a methodology to reduce the area of the small arithmetic operators used in neural networks - i.e., adders and multipliers - via a small loss in accuracy, and show that we improve area savings for the same accuracy loss w.r.t. the state of the art. To achieve our goal, we improve on a boolean rewriting technique recently proposed, called XPAT, where the use of a parametrisable template to rewrite circuits has proved to be highly beneficial. In particular, XPAT was able to produce smaller circuits than comparable approaches while utilising a naive sum of products template structure. In this work, we show that template parameters can act as proxies for chosen metrics and we propose a novel template based on parametrisable product sharing that acts as a close proxy to synthesised area. We demonstrate experimentally that our methodology converges better to low-area solutions and that it can find better approximations than both the original XPAT and two other state-of-the-art approaches.

Method: Developed a multi-omics support vector regression (SVR) model using radiomics (MRI) and dosiomics (dose maps) features from 39 patients with 69 brain metastases. Implemented feature selection with Lasso algorithm and used delta features to capture relative changes between time points. Evaluated with 5-fold cross-validation with 10 repeats.

Result: Multi-omics models integrating radiomics, dosiomics, and their delta counterparts outperformed individual-omics models. Delta-radiomic features significantly enhanced prediction accuracy. Top-performing model achieved R2 of 0.743 and RRMSE of 0.022.

Conclusion: The multi-omics SVR model shows promising performance for predicting continuous GTV changes, providing a quantitative and personalized approach to assist patient selection and treatment adjustment in PULSAR radiotherapy.

Abstract: Personalized ultra-fractionated stereotactic adaptive radiotherapy (PULSAR) is a novel treatment that delivers radiation in pulses of protracted intervals. Accurate prediction of gross tumor volume (GTV) changes through regression models has substantial prognostic value. This study aims to develop a multi-omics based support vector regression (SVR) model for predicting GTV change. A retrospective cohort of 39 patients with 69 brain metastases was analyzed, based on radiomics (MRI images) and dosiomics (dose maps) features. Delta features were computed to capture relative changes between two time points. A feature selection pipeline using least absolute shrinkage and selection operator (Lasso) algorithm with weight- or frequency-based ranking criterion was implemented. SVR models with various kernels were evaluated using the coefficient of determination (R2) and relative root mean square error (RRMSE). Five-fold cross-validation with 10 repeats was employed to mitigate the limitation of small data size. Multi-omics models that integrate radiomics, dosiomics, and their delta counterparts outperform individual-omics models. Delta-radiomic features play a critical role in enhancing prediction accuracy relative to features at single time points. The top-performing model achieves an R2 of 0.743 and an RRMSE of 0.022. The proposed multi-omics SVR model shows promising performance in predicting continuous change of GTV. It provides a more quantitative and personalized approach to assist patient selection and treatment adjustment in PULSAR.

Method: Psychosis-bench benchmark with 16 structured, 12-turn conversational scenarios simulating delusional themes progression, evaluating 8 LLMs across explicit and implicit contexts using Delusion Confirmation Score, Harm Enablement Score, and Safety Intervention Score.

Result: All LLMs demonstrated psychogenic potential with high delusion confirmation (mean DCS 0.91), frequent harm enablement (mean HES 0.69), and inadequate safety interventions (mean SIS 0.37). Performance was worse in implicit scenarios with strong correlation between delusion confirmation and harm enablement.

Conclusion: LLM psychogenicity represents a quantifiable public health risk requiring urgent re-thinking of LLM training approaches and collaboration between developers, policymakers, and healthcare professionals.

Abstract: Background: Emerging reports of “AI psychosis” are on the rise, where user-LLM interactions may exacerbate or induce psychosis or adverse psychological symptoms. Whilst the sycophantic and agreeable nature of LLMs can be beneficial, it becomes a vector for harm by reinforcing delusional beliefs in vulnerable users. Methods: Psychosis-bench is a novel benchmark designed to systematically evaluate the psychogenicity of LLMs comprises 16 structured, 12-turn conversational scenarios simulating the progression of delusional themes(Erotic Delusions, Grandiose/Messianic Delusions, Referential Delusions) and potential harms. We evaluated eight prominent LLMs for Delusion Confirmation (DCS), Harm Enablement (HES), and Safety Intervention(SIS) across explicit and implicit conversational contexts. Findings: Across 1,536 simulated conversation turns, all LLMs demonstrated psychogenic potential, showing a strong tendency to perpetuate rather than challenge delusions (mean DCS of 0.91 $\pm$0.88). Models frequently enabled harmful user requests (mean HES of 0.69 $\pm$0.84) and offered safety interventions in only roughly a third of applicable turns (mean SIS of 0.37 $\pm$0.48). 51 / 128 (39.8%) of scenarios had no safety interventions offered. Performance was significantly worse in implicit scenarios, models were more likely to confirm delusions and enable harm while offering fewer interventions (p < .001). A strong correlation was found between DCS and HES (rs = .77). Model performance varied widely, indicating that safety is not an emergent property of scale alone. Conclusion: This study establishes LLM psychogenicity as a quantifiable risk and underscores the urgent need for re-thinking how we train LLMs. We frame this issue not merely as a technical challenge but as a public health imperative requiring collaboration between developers, policymakers, and healthcare professionals.

Method: Proposes secure parameter aggregation using the torus (rather than finite fields) to leverage uniform distribution properties, ensuring perfect privacy while maintaining floating-point-like arithmetic capabilities.

Result: Experimental results show similar performance to non-secure models while maintaining perfect privacy. Outperforms finite field approaches in model accuracy and cosine similarity in some cases.

Conclusion: Torus-based secure aggregation provides a safer choice for federated learning by guaranteeing perfect privacy without sacrificing model accuracy, overcoming limitations of previous finite field methods.

Abstract: In federated learning, multiple parties train models locally and share their parameters with a central server, which aggregates them to update a global model. To address the risk of exposing sensitive data through local models, secure aggregation via secure multiparty computation has been proposed to enhance privacy. At the same time, perfect privacy can only be achieved by a uniform distribution of the masked local models to be aggregated. This raises a problem when working with real valued data, as there is no measure on the reals that is invariant under the masking operation, and hence information leakage is bound to occur. Shifting the data to a finite field circumvents this problem, but as a downside runs into an inherent accuracy complexity tradeoff issue due to fixed point modular arithmetic as opposed to floating point numbers that can simultaneously handle numbers of varying magnitudes. In this paper, a novel secure parameter aggregation method is proposed that employs the torus rather than a finite field. This approach guarantees perfect privacy for each party’s data by utilizing the uniform distribution on the torus, while avoiding accuracy losses. Experimental results show that the new protocol performs similarly to the model without secure aggregation while maintaining perfect privacy. Compared to the finite field secure aggregation, the torus-based protocol can in some cases significantly outperform it in terms of model accuracy and cosine similarity, hence making it a safer choice.

Method: Tested foundation model’s forecasting ability in zero-shot and fine-tuned settings, compared Granger-like estimates from the model with standard Granger causality, and validated using synthetic time series from ground-truth causal models (logistic map coupling and Ornstein-Uhlenbeck processes).

Result: Foundation model achieved competitive zero-shot forecasting (MAPE 0.55 in controls, 0.27 in patients) and provided more precise detection of causal interactions compared to standard Granger causality.

Conclusion: Foundation models offer versatility, strong zero-shot performance, and potential utility for forecasting and causal discovery in time-series neuroscience data.

Abstract: Time-series forecasting and causal discovery are central in neuroscience, as predicting brain activity and identifying causal relationships between neural populations and circuits can shed light on the mechanisms underlying cognition and disease. With the rise of foundation models, an open question is how they compare to traditional methods for brain signal forecasting and causality analysis, and whether they can be applied in a zero-shot setting. In this work, we evaluate a foundation model against classical methods for inferring directional interactions from spontaneous brain activity measured with functional magnetic resonance imaging (fMRI) in humans. Traditional approaches often rely on Wiener-Granger causality. We tested the forecasting ability of the foundation model in both zero-shot and fine-tuned settings, and assessed causality by comparing Granger-like estimates from the model with standard Granger causality. We validated the approach using synthetic time series generated from ground-truth causal models, including logistic map coupling and Ornstein-Uhlenbeck processes. The foundation model achieved competitive zero-shot forecasting fMRI time series (mean absolute percentage error of 0.55 in controls and 0.27 in patients). Although standard Granger causality did not show clear quantitative differences between models, the foundation model provided a more precise detection of causal interactions. Overall, these findings suggest that foundation models offer versatility, strong zero-shot performance, and potential utility for forecasting and causal discovery in time-series data.

Method: SGT combines self-attention within symbol and constraint subgraphs with graph-aware cross-attention for structured message passing across subgraphs. It features a soft-input interface to integrate auxiliary priors and produces effective soft outputs.

Result: Experiments show SGT achieves near-ML performance while maintaining computational efficiency.

Conclusion: SGT provides a flexible and interpretable framework for receiver systems that can effectively leverage soft priors, addressing key limitations of existing MIMO detection approaches.

Abstract: We propose the Soft Graph Transformer (SGT), a soft-input-soft-output neural architecture designed for MIMO detection. While Maximum Likelihood (ML) detection achieves optimal accuracy, its exponential complexity makes it infeasible in large systems, and conventional message-passing algorithms rely on asymptotic assumptions that often fail in finite dimensions. Recent Transformer-based detectors show strong performance but typically overlook the MIMO factor graph structure and cannot exploit prior soft information. SGT addresses these limitations by combining self-attention, which encodes contextual dependencies within symbol and constraint subgraphs, with graph-aware cross-attention, which performs structured message passing across subgraphs. Its soft-input interface allows the integration of auxiliary priors, producing effective soft outputs while maintaining computational efficiency. Experiments demonstrate that SGT achieves near-ML performance and offers a flexible and interpretable framework for receiver systems that leverage soft priors.

Method: Introduce Hilbert coVariance Filters (HVFs) and Networks (HVNs) based on covariance operators, with principled discretization and proof that HVFs recover Functional PCA of filtered signals.

Result: HVNs demonstrate robust performance on synthetic and real-world time-series classification tasks, outperforming MLP and FPCA-based classifiers.

Conclusion: The proposed framework successfully extends covariance-based neural networks to infinite-dimensional spaces with theoretical foundations and practical effectiveness.

Abstract: CoVariance Neural Networks (VNNs) perform graph convolutions on the empirical covariance matrix of signals defined over finite-dimensional Hilbert spaces, motivated by robustness and transferability properties. Yet, little is known about how these arguments extend to infinite-dimensional Hilbert spaces. In this work, we take a first step by introducing a novel convolutional learning framework for signals defined over infinite-dimensional Hilbert spaces, centered on the (empirical) covariance operator. We constructively define Hilbert coVariance Filters (HVFs) and design Hilbert coVariance Networks (HVNs) as stacks of HVF filterbanks with nonlinear activations. We propose a principled discretization procedure, and we prove that empirical HVFs can recover the Functional PCA (FPCA) of the filtered signals. We then describe the versatility of our framework with examples ranging from multivariate real-valued functions to reproducing kernel Hilbert spaces. Finally, we validate HVNs on both synthetic and real-world time-series classification tasks, showing robust performance compared to MLP and FPCA-based classifiers.

Method: HAM maintains fixed hierarchical groups, trains low-rank adapters with importance scalars for each task, dynamically groups tasks by adapter similarity, then prunes, scales and merges adapters within groups.

Result: Extensive experiments on three vision benchmarks show HAM significantly outperforms state-of-the-art methods, especially as task numbers increase.

Conclusion: HAM provides an effective scalable solution for continual learning by dynamically merging adapters through hierarchical grouping, enabling better knowledge retention and transfer between related tasks.

Abstract: Continual learning is an essential capability of human cognition, yet it poses significant challenges for current deep learning models. The primary issue is that new knowledge can interfere with previously learned information, causing the model to forget earlier knowledge in favor of the new, a phenomenon known as catastrophic forgetting. Although large pre-trained models can partially mitigate forgetting by leveraging their existing knowledge and over-parameterization, they often struggle when confronted with novel data distributions. Parameter-Efficient Fine-Tuning (PEFT) methods, such as LoRA, enable efficient adaptation to new knowledge. However, they still face challenges in scaling to dynamic learning scenarios and long sequences of tasks, as maintaining one adapter per task introduces complexity and increases the potential for interference. In this paper, we introduce Hierarchical Adapters Merging (HAM), a novel framework that dynamically combines adapters from different tasks during training. This approach enables HAM to scale effectively, allowing it to manage more tasks than competing baselines with improved efficiency. To achieve this, HAM maintains a fixed set of groups that hierarchically consolidate new adapters. For each task, HAM trains a low-rank adapter along with an importance scalar, then dynamically groups tasks based on adapter similarity. Within each group, adapters are pruned, scaled and merge, facilitating transfer learning between related tasks. Extensive experiments on three vision benchmarks show that HAM significantly outperforms state-of-the-art methods, particularly as the number of tasks increases.

Method: Split-Point Analysis decomposes predictive residuals into upper/lower subsets to compute Mean Absolute Residuals. Self-consistency Discrepancy Score quantifies epistemic uncertainty. For regression: side-specific quantile regression with SDS calibration. For classification: SPA-based calibration of softmax outputs followed by predictive entropy computation.

Result: Extensive experiments show the framework matches or exceeds state-of-the-art UQ methods with minimal computational overhead across diverse regression and classification benchmarks.

Conclusion: The proposed method provides efficient, comprehensive uncertainty quantification without model modification or retraining, offering improved empirical coverage and calibration while maintaining computational efficiency.

Abstract: Uncertainty quantification (UQ) is vital for trustworthy deep learning, yet existing methods are either computationally intensive, such as Bayesian or ensemble methods, or provide only partial, task-specific estimates, such as single-forward-pass techniques. In this paper, we propose a post-hoc single-forward-pass framework that jointly captures aleatoric and epistemic uncertainty without modifying or retraining pretrained models. Our method applies \emph{Split-Point Analysis} (SPA) to decompose predictive residuals into upper and lower subsets, computing \emph{Mean Absolute Residuals} (MARs) on each side. We prove that, under ideal conditions, the total MAR equals the harmonic mean of subset MARs; deviations define a novel \emph{Self-consistency Discrepancy Score} (SDS) for fine-grained epistemic estimation across regression and classification. For regression, side-specific quantile regression yields prediction intervals with improved empirical coverage, which are further calibrated via SDS. For classification, when calibration data are available, we apply SPA-based calibration identities to adjust the softmax outputs and then compute predictive entropy on these calibrated probabilities. Extensive experiments on diverse regression and classification benchmarks demonstrate that our framework matches or exceeds several state-of-the-art UQ methods while incurring minimal overhead. Our source code is available at https://github.com/zzz0527/SPC-UQ.

Method: Systematic review of 536 ABM studies published from January 2020 to December 2023, assessed against nine criteria of model usefulness including transparency, code sharing, limitations disclosure, stakeholder engagement, and validation frameworks.

Result: Most models explored interventions (54.85%) rather than forecasting (1.68%). While 91.60% described assumptions, only 40.86% shared code, 36.38% built on existing models, 6.72% used standardized reporting, and only 2.24% had comprehensive validation frameworks. Stakeholder engagement was rare (13.62%).

Conclusion: COVID-19 ABMs advanced quickly but lacked transparency, accessibility, and participatory engagement. Stronger standards are needed for ABMs to serve as reliable decision-support tools in future public health crises.

Abstract: The COVID-19 pandemic prompted a surge in computational models to simulate disease dynamics and guide interventions. Agent-based models (ABMs) are well-suited to capture population and environmental heterogeneity, but their rapid deployment raised questions about utility for health policy. We systematically reviewed 536 COVID-19 ABM studies published from January 2020 to December 2023, retrieved from Web of Science, PubMed, and Wiley on January 30, 2024. Studies were included if they used ABMs to simulate COVID-19 transmission, where reviews were excluded. Studies were assessed against nine criteria of model usefulness, including transparency and re-use, interdisciplinary collaboration and stakeholder engagement, and evaluation practices. Publications peaked in late 2021 and were concentrated in a few countries. Most models explored behavioral or policy interventions (n = 294, 54.85%) rather than real-time forecasting (n = 9, 1.68%). While most described model assumptions (n = 491, 91.60%), fewer disclosed limitations (n = 349, 65.11%), shared code (n = 219, 40.86%), or built on existing models (n = 195, 36.38%). Standardized reporting protocols (n = 36, 6.72%) and stakeholder engagement were rare (13.62%, n = 73). Only 2.24% (n = 12) described a comprehensive validation framework, though uncertainty was often quantified (n = 407, 75.93%). Limitations of this review include underrepresentation of non-English studies, subjective data extraction, variability in study quality, and limited generalizability. Overall, COVID-19 ABMs advanced quickly, but lacked transparency, accessibility, and participatory engagement. Stronger standards are needed for ABMs to serve as reliable decision-support tools in future public health crises.

Method: Defines three core operations: inject (introduce external events), fork (create independent timeline branches), and compare (parallel observation of multiple branches). Demonstrates these through a commodity market simulation with 14 AI agents.

Result: The operations transform linear simulation workflows into interactive, explorable spaces, allowing researchers to observe divergent outcomes across parallel timelines and investigate how different interventions lead to distinct emergent behaviors.

Conclusion: These fundamental operations provide a starting point for systematic causal investigation in LLM-based agent simulations, moving beyond passive observation toward active experimentation.

Abstract: LLM-based multi-agent simulations are a rapidly growing field of research, but current simulations often lack clear modes for interaction and analysis, limiting the “what if” scenarios researchers are able to investigate. In this demo, we define three core operations for interacting with multi-agent simulations: inject, fork, and compare. Inject allows researchers to introduce external events at any point during simulation execution. Fork creates independent timeline branches from any timestamp, preserving complete state while allowing divergent exploration. Compare facilitates parallel observation of multiple branches, revealing how different interventions lead to distinct emergent behaviors. Together, these operations establish a vocabulary that transforms linear simulation workflows into interactive, explorable spaces. We demonstrate this vocabulary through a commodity market simulation with fourteen AI agents, where researchers can inject contrasting events and observe divergent outcomes across parallel timelines. By defining these fundamental operations, we provide a starting point for systematic causal investigation in LLM-based agent simulations, moving beyond passive observation toward active experimentation.

Method: Uses four specialized agents (ReformatAgent, FlowAgent, RefineAgent, ExecuteAgent) with focus on precise node connections, enhanced by reinforcement learning. Introduces FlowDataset with 13,571 workflow-description pairs and FlowBench benchmark.

Result: Significantly outperforms existing LLM-based methods in workflow generation, demonstrating superior accuracy and effectiveness.

Conclusion: ComfyGPT represents a significant advancement in automated workflow generation for node-based systems, with novel evaluation metrics and a comprehensive benchmark for future research.

Abstract: ComfyUI is a popular workflow-based interface that allows users to customize image generation tasks through an intuitive node-based system. However, the complexity of managing node connections and diverse modules can be challenging for users. In this paper, we introduce ComfyGPT, a self-optimizing multi-agent system designed to generate ComfyUI workflows based on task descriptions automatically. The key innovations of ComfyGPT include: (1) consisting of four specialized agents to build a multi-agent workflow generation system: ReformatAgent, FlowAgent, RefineAgent, and ExecuteAgent; (2) focusing on generating precise node connections instead of entire workflows, improving generation accuracy; and (3) enhancing workflow generation through reinforcement learning. Moreover, we introduce FlowDataset, a large-scale dataset containing 13,571 workflow-description pairs, and FlowBench, a comprehensive benchmark for evaluating workflow generation systems. Additionally, we propose four novel evaluation metrics: Format Validation (FV), Pass Accuracy (PA), Pass Instruct Alignment (PIA), and Pass Node Diversity (PND). Experimental results demonstrate that ComfyGPT significantly outperforms existing LLM-based methods in workflow generation, making it a significant step forward in this field. Code is avaliable at https://github.com/comfygpt/comfygpt.

Method: Using LLM agents in a spatially extended El Farol Bar problem setup with prompt-specified constraints (60% threshold) and observing their decision-making processes.

Result: LLM agents developed spontaneous motivation to attend the bar, formed collective decision-making behaviors, and balanced external constraints with internal social preferences, behaving more like humans than perfect rational agents.

Conclusion: LLM agents can realize new models of group decision making that combine formal rationality with social motivations, going beyond traditional game-theoretic problem settings and demonstrating human-like emergent social dynamics.

Abstract: We investigate the emergent social dynamics of Large Language Model (LLM) agents in a spatially extended El Farol Bar problem, observing how they autonomously navigate this classic social dilemma. As a result, the LLM agents generated a spontaneous motivation to go to the bar and changed their decision making by becoming a collective. We also observed that the LLM agents did not solve the problem completely, but rather behaved more like humans. These findings reveal a complex interplay between external incentives (prompt-specified constraints such as the 60% threshold) and internal incentives (culturally-encoded social preferences derived from pre-training), demonstrating that LLM agents naturally balance formal game-theoretic rationality with social motivations that characterize human behavior. These findings suggest that a new model of group decision making, which could not be handled in the previous game-theoretic problem setting, can be realized by LLM agents.

Method: Collected 1,960 manually annotated images using Valence-Arousal-Dominance (VAD) emotional attributes, designed four evaluation tasks (Perception, Ranking, Description, Assessment), and introduced image-pairwise analysis with 6,773 question-answer pairs to test 19 MLLMs.

Result: Some proprietary and large-scale open-source MLLMs showed promising overall performance, but analytical capabilities in certain evaluation dimensions remained suboptimal.

Conclusion: EEmo-Bench provides a foundation for enhancing MLLMs’ comprehensive emotion perception and understanding capabilities, which is crucial for machine-centric emotion analysis.

Abstract: The furnishing of multi-modal large language models (MLLMs) has led to the emergence of numerous benchmark studies, particularly those evaluating their perception and understanding capabilities. Among these, understanding image-evoked emotions aims to enhance MLLMs’ empathy, with significant applications such as human-machine interaction and advertising recommendations. However, current evaluations of this MLLM capability remain coarse-grained, and a systematic and comprehensive assessment is still lacking. To this end, we introduce EEmo-Bench, a novel benchmark dedicated to the analysis of the evoked emotions in images across diverse content categories. Our core contributions include: 1) Regarding the diversity of the evoked emotions, we adopt an emotion ranking strategy and employ the Valence-Arousal-Dominance (VAD) as emotional attributes for emotional assessment. In line with this methodology, 1,960 images are collected and manually annotated. 2) We design four tasks to evaluate MLLMs’ ability to capture the evoked emotions by single images and their associated attributes: Perception, Ranking, Description, and Assessment. Additionally, image-pairwise analysis is introduced to investigate the model’s proficiency in performing joint and comparative analysis. In total, we collect 6,773 question-answer pairs and perform a thorough assessment on 19 commonly-used MLLMs. The results indicate that while some proprietary and large-scale open-source MLLMs achieve promising overall performance, the analytical capabilities in certain evaluation dimensions remain suboptimal. Our EEmo-Bench paves the path for further research aimed at enhancing the comprehensive perceiving and understanding capabilities of MLLMs concerning image-evoked emotions, which is crucial for machine-centric emotion perception and understanding.

Method: Constructed MuCED dataset with 2,634 expert-validated music-palette pairs using Russell-based emotion vectors. Developed cross-modal representation learning framework with music encoder and color decoder, plus multi-objective optimization for emotion alignment, color diversity, and palette coherence.

Result: Extensive experiments show the method outperforms current approaches in interpreting music emotion and generating attractive, diverse color palettes.

Conclusion: Music2Palette successfully bridges auditory and visual emotion experiences, enabling applications like music-driven image recoloring, video generation, and data visualization.

Abstract: Emotion alignment between music and palettes is crucial for effective multimedia content, yet misalignment creates confusion that weakens the intended message. However, existing methods often generate only a single dominant color, missing emotion variation. Others rely on indirect mappings through text or images, resulting in the loss of crucial emotion details. To address these challenges, we present Music2Palette, a novel method for emotion-aligned color palette generation via cross-modal representation learning. We first construct MuCED, a dataset of 2,634 expert-validated music-palette pairs aligned through Russell-based emotion vectors. To directly translate music into palettes, we propose a cross-modal representation learning framework with a music encoder and color decoder. We further propose a multi-objective optimization approach that jointly enhances emotion alignment, color diversity, and palette coherence. Extensive experiments demonstrate that our method outperforms current methods in interpreting music emotion and generating attractive and diverse color palettes. Our approach enables applications like music-driven image recoloring, video generating, and data visualization, bridging the gap between auditory and visual emotion experiences.

Method: Proposed Text-Embedding KNN for SICL (TICL) - a pipeline that uses semantic context to select effective in-context examples for off-the-shelf large multimodal models.

Result: Achieved significant improvements across accented English, multilingual speech, and children’s speech tasks, with up to 84.7% relative WER reduction compared to zero-shot performance.

Conclusion: TICL demonstrates robust and efficient performance enhancement for speech recognition through semantic context-based example selection, making it a practical solution for improving speech foundation models without fine-tuning.

Abstract: Speech foundation models have recently demonstrated the ability to perform Speech In-Context Learning (SICL). Selecting effective in-context examples is crucial for SICL performance, yet selection methodologies remain underexplored. In this work, we propose Text-Embedding KNN for SICL (TICL), a simple pipeline that uses semantic context to enhance off-the-shelf large multimodal models’ speech recognition ability without fine-tuning. Across challenging automatic speech recognition tasks, including accented English, multilingual speech, and children’s speech, our method enables models to surpass zero-shot performance with up to 84.7% relative WER reduction. We conduct ablation studies to show the robustness and efficiency of our method.

Method: Proposed DSSCNet architecture combining CNN, SE blocks, and Residual networks to extract discriminative features from mel spectrograms, plus cross-corpus fine-tuning framework adapted from detection-based transfer learning.

Result: Achieved 56.84%-62.62% accuracy under OSPS and 63.47%-64.18% under LOSO on TORGO and UA-Speech datasets. After fine-tuning, improved to 75.80%-68.25% (OSPS) and 77.76%-79.44% (LOSO), outperforming existing methods.

Conclusion: DSSCNet demonstrates effectiveness and generalizability for fine-grained severity classification across diverse dysarthric speech datasets, showing substantial performance improvements through the proposed architecture and fine-tuning framework.

Abstract: Dysarthric speech severity classification is crucial for objective clinical assessment and progress monitoring in individuals with motor speech disorders. Although prior methods have addressed this task, achieving robust generalization in speaker-independent (SID) scenarios remains challenging. This work introduces DSSCNet, a novel deep neural architecture that combines Convolutional, Squeeze-Excitation (SE), and Residual network, helping it extract discriminative representations of dysarthric speech from mel spectrograms. The addition of SE block selectively focuses on the important features of the dysarthric speech, thereby minimizing loss and enhancing overall model performance. We also propose a cross-corpus fine-tuning framework for severity classification, adapted from detection-based transfer learning approaches. DSSCNet is evaluated on two benchmark dysarthric speech corpora: TORGO and UA-Speech under speaker-independent evaluation protocols: One-Speaker-Per-Severity (OSPS) and Leave-One-Speaker-Out (LOSO) protocols. DSSCNet achieves accuracies of 56.84% and 62.62% under OSPS and 63.47% and 64.18% under LOSO setting on TORGO and UA-Speech respectively outperforming existing state-of-the-art methods. Upon fine-tuning, the performance improves substantially, with DSSCNet achieving up to 75.80% accuracy on TORGO and 68.25% on UA-Speech in OSPS, and up to 77.76% and 79.44%, respectively, in LOSO. These results demonstrate the effectiveness and generalizability of DSSCNet for fine-grained severity classification across diverse dysarthric speech datasets.

Method: Represent symbolic music as image-like piano rolls (binary and velocity-based forms). Adapt SSIM components for musical context to create SSIMuse-B (composition) and SSIMuse-V (performance) variants.

Result: SSIMuse reliably detects exact replication at granularity of at least one bar in controlled experiments on synthetic samples from multiple datasets.

Conclusion: SSIMuse enables open evaluation of music replication and highlights broader ethical, social, legal, and economic implications of AI music generation.

Abstract: AI-generated music may inadvertently replicate samples from the training data, raising concerns of plagiarism. Similarity measures can quantify such replication, thereby offering supervision and guidance for music generation models. Existing similarity measure methods for symbolic music mainly target melody repetition, leaving a gap in assessing complex music with rich textures and expressive performance characteristics. To address this gap, we introduce SSIMuse, the first adaptation of the Structural Similarity Index Measure (SSIM) from images to symbolic music. Specifically, we represent symbolic music as image-like piano rolls in binary and velocity-based forms. Build upon these representations, we reinterprete and suitably modify the SSIM components in the musical context to develop two variants, i.e., SSIMuse-B and SSIMuse-V, for evaluating data replication in composition and dynamic performance, respectively. Controlled experiments on synthetic samples from multiple datasets show that SSIMuse can reliably detect exact replication at a granularity of at least one bar. SSIMuse enables open evaluation of replication in music generation and draws attention to its broader ethical, social, legal, and economic implications. The code is available at https://github.com/Tayjsl97/SSIMuse.

Method: Proposes DLL-APNet with causal convolutions to constrain information to current/historical contexts, minimizing latency. Uses knowledge distillation where a pre-trained non-causal teacher vocoder guides the causal student vocoder’s intermediate feature generation.

Result: DLL-APNet produces higher-quality speech than other causal vocoders with fewer computational resources, and achieves speech quality on par with mainstream non-causal neural vocoders.

Conclusion: The proposed vocoder successfully delivers both high perceptual quality and low latency, making it practical for real-time applications.

Abstract: The majority of mainstream neural vocoders primarily focus on speech quality and generation speed, while overlooking latency, which is a critical factor in real-time applications. Excessive latency leads to noticeable delays in user interaction, severely degrading the user experience and rendering such systems impractical for real-time use. Therefore, this paper proposes DLL-APNet, a Distilled Low-Latency neural vocoder which first predicts the Amplitude and Phase spectra explicitly from input mel spectrogram and then reconstructs the speech waveform via inverse short-time Fourier transform (iSTFT). The DLL-APNet vocoder leverages causal convolutions to constrain the utilization of information to current and historical contexts, effectively minimizing latency. To mitigate speech quality degradation caused by causal constraints, a knowledge distillation strategy is proposed, where a pre-trained non-causal teacher vocoder guides intermediate feature generation of the causal student DLL-APNet vocoder. Experimental results demonstrate that the proposed DLL-APNet vocoder produces higher-quality speech than other causal vocoders, while requiring fewer computational resources. Furthermore, the proposed DLL-APNet vocoder achieves speech quality on par with mainstream non-causal neural vocoders, validating its ability to deliver both high perceptual quality and low latency.

Method: Uses a fully causal architecture with reduced convolutional channels, and employs knowledge distillation from a non-causal high-complexity teacher codec to compensate for quality degradation.

Result: Achieves high-quality speech reconstruction with low latency (20ms), low computational complexity (910 MFLOPs), and low model complexity (5.4M parameters).

Conclusion: StreamCodec2 successfully balances speech quality, latency, and complexity through architectural improvements and knowledge distillation, making it suitable for real-time speech applications.

Abstract: While many current neural speech codecs achieve impressive reconstructed speech quality, they often neglect latency and complexity considerations, limiting their practical deployment in downstream tasks such as real-time speech communication and efficient speech compression. In our previous work, we proposed StreamCodec, which enables streamable speech coding by leveraging model causalization and a scalar-vector-combined quantization strategy, but its reconstructed quality and complexity still have room for improvement. Therefore, this paper proposes an improved iteration of StreamCodec, named StreamCodec2. The StreamCodec2 supports streamable and lightweight speech coding by adopting a fully causal architecture and reducing the convolutional channels. To compensate for the speech quality degradation caused by model causalization and pruning, we introduce a non-causal, high-complexity teacher codec to guide the training of StreamCodec2 through knowledge distillation. Experimental results demonstrate that our proposed StreamCodec2, trained with the knowledge distillation strategy, can achieve high-quality speech reconstruction while maintaining low latency (only 20 ms), low computational complexity (only 910 MFLOPs), and low model complexity (only 5.4 M parameters).

Method: Three-stage framework: 1) Universal Sound Separation breaks audio mixtures into source waveforms, 2) Single-label Classification processes each waveform to generate class labels, 3) Target Sound Extraction isolates sources using the waveform and label information. The system operates iteratively with feedback loops for refinement.

Result: Achieved 11.00 dB improvement in class-aware signal-to-distortion ratio (CA-SDRi) and 55.8% accuracy in label prediction, outperforming the ResUNetK baseline by 4.4 dB and 4.3% respectively, and ranking first among all submissions.

Conclusion: The proposed multi-stage self-guided framework effectively addresses the S5 task through autonomous target identification and iterative refinement, demonstrating superior performance in both sound separation quality and semantic labeling accuracy compared to existing baselines.

Abstract: This paper introduces a multi-stage self-directed framework designed to address the spatial semantic segmentation of sound scene (S5) task in the DCASE 2025 Task 4 challenge. This framework integrates models focused on three distinct tasks: Universal Sound Separation (USS), Single-label Classification (SC), and Target Sound Extraction (TSE). Initially, USS breaks down a complex audio mixture into separate source waveforms. Each of these separated waveforms is then processed by a SC block, generating two critical pieces of information: the waveform itself and its corresponding class label. These serve as inputs for the TSE stage, which isolates the source that matches this information. Since these inputs are produced within the system, the extraction target is identified autonomously, removing the necessity for external guidance. The extracted waveform can be looped back into the classification task, creating a cycle of iterative refinement that progressively enhances both separability and labeling accuracy. We thus call our framework a multi-stage self-guided system due to these self-contained characteristics. On the official evaluation dataset, the proposed system achieves an 11.00 dB increase in class-aware signal-to-distortion ratio improvement (CA-SDRi) and a 55.8% accuracy in label prediction, outperforming the ResUNetK baseline by 4.4 dB and 4.3%, respectively, and achieving first place among all submissions.

Method: Organized a challenge with specific task settings, released a 1,604-hour multilingual conversational speech dataset, provided baseline systems, and analyzed submissions from 78 participating teams.

Result: The challenge attracted significant participation with 489 valid leaderboard results and 14 technical reports, generating valuable insights for building multilingual conversational SLLMs.

Conclusion: The MLC-SLM challenge successfully contributed to advancing multilingual conversational speech language model research through community participation and shared insights.

Abstract: This paper summarizes the Interspeech2025 Multilingual Conversational Speech Language Model (MLC-SLM) challenge, which aims to advance the exploration of building effective multilingual conversational speech LLMs (SLLMs). We provide a detailed description of the task settings for the MLC-SLM challenge, the released real-world multilingual conversational speech dataset totaling approximately 1,604 hours, and the baseline systems for participants. The MLC-SLM challenge attracts 78 teams from 13 countries to participate, with 489 valid leaderboard results and 14 technical reports for the two tasks. We distill valuable insights on building multilingual conversational SLLMs based on submissions from participants, aiming to contribute to the advancement of the community.

Method: Integrates multiple low-rank adapters (LoRA) into attention layers with a routing mechanism that selectively activates specialized experts to handle evolving deepfake attacks.

Result: Outperforms standard fine-tuning in both in-domain and out-of-domain scenarios, reducing equal error rates. Best model lowers average out-of-domain EER from 8.55% to 6.08%.

Conclusion: The mixture-of-LoRA-experts approach effectively enhances adaptability and achieves generalizable audio deepfake detection, demonstrating superior performance over conventional fine-tuning methods.

Abstract: Foundation models such as Wav2Vec2 excel at representation learning in speech tasks, including audio deepfake detection. However, after being fine-tuned on a fixed set of bonafide and spoofed audio clips, they often fail to generalize to novel deepfake methods not represented in training. To address this, we propose a mixture-of-LoRA-experts approach that integrates multiple low-rank adapters (LoRA) into the model’s attention layers. A routing mechanism selectively activates specialized experts, enhancing adaptability to evolving deepfake attacks. Experimental results show that our method outperforms standard fine-tuning in both in-domain and out-of-domain scenarios, reducing equal error rates relative to baseline models. Notably, our best MoE-LoRA model lowers the average out-of-domain EER from 8.55% to 6.08%, demonstrating its effectiveness in achieving generalizable audio deepfake detection.

Method: Proposed DSpAST, an audio encoder based on SpatialAST that learns disentangled representations of spatial audio with only 0.2% additional parameters compared to the base model.

Result: Experiments on SpatialSoundQA with the BAT spatial audio reasoning system show that DSpAST significantly outperforms SpatialAST in spatial audio reasoning tasks.

Conclusion: DSpAST successfully addresses the challenge of learning disentangled spatial audio representations with minimal parameter overhead, demonstrating superior performance over existing approaches for spatial audio reasoning with large language models.

Abstract: Reasoning about spatial audio with large language models requires a spatial audio encoder as an acoustic front-end to obtain audio embeddings for further processing. Such an encoder needs to capture all information required to detect the type of sound events, as well as the direction and distance of their corresponding sources. Accomplishing this with a single audio encoder is demanding as the information required for each of these tasks is mostly independent of each other. As a result, the performance obtained with a single encoder is often worse than when using task-specific audio encoders. In this work, we present DSpAST, a novel audio encoder based on SpatialAST that learns disentangled representations of spatial audio while having only 0.2% additional parameters. Experiments on SpatialSoundQA with the spatial audio reasoning system BAT demonstrate that DSpAST significantly outperforms SpatialAST.

Method: Conducted perceptual analysis across expressive dimensions (adverbs of degree, emotion intensity), collected human ratings on speaker age and word-level emphasis, and created the Expressive VOice Control (E-VOC) corpus with large-scale human evaluations.

Result: GPT-4o-mini-TTS is the most reliable ITTS model with good instruction-utterance alignment; all 5 analyzed systems tend to generate Adult voices regardless of child/elderly instructions; fine-grained control remains a major challenge for most ITTS systems.

Conclusion: There is substantial room for improvement in ITTS systems, particularly in interpreting nuanced attribute instructions and achieving accurate age-related voice generation, despite GPT-4o-mini-TTS showing promising alignment capabilities.

Abstract: Instruction-guided text-to-speech (ITTS) enables users to control speech generation through natural language prompts, offering a more intuitive interface than traditional TTS. However, the alignment between user style instructions and listener perception remains largely unexplored. This work first presents a perceptual analysis of ITTS controllability across two expressive dimensions (adverbs of degree and graded emotion intensity) and collects human ratings on speaker age and word-level emphasis attributes. To comprehensively reveal the instruction-perception gap, we provide a data collection with large-scale human evaluations, named Expressive VOice Control (E-VOC) corpus. Furthermore, we reveal that (1) gpt-4o-mini-tts is the most reliable ITTS model with great alignment between instruction and generated utterances across acoustic dimensions. (2) The 5 analyzed ITTS systems tend to generate Adult voices even when the instructions ask to use child or Elderly voices. (3) Fine-grained control remains a major challenge, indicating that most ITTS systems have substantial room for improvement in interpreting slightly different attribute instructions.

Method: Proposes a two-stage framework: first uses time-domain warping for initial estimates, then refines with an Implicit Binaural Corrector (IBC) module - an implicit neural network that predicts amplitude and phase corrections directly.

Result: Achieves statistically comparable perceptual quality to best-performing baseline while significantly improving computational efficiency with 72.7% parameter reduction and fewer compute operations (MACs).

Conclusion: LINN effectively addresses the trade-off between synthesis quality and computational efficiency, providing a viable solution for high-fidelity edge-device spatial audio applications.

Abstract: High-fidelity binaural audio synthesis is crucial for immersive listening, but existing methods require extensive computational resources, limiting their edge-device application. To address this, we propose the Lightweight Implicit Neural Network (LINN), a novel two-stage framework. LINN first generates initial estimates using a time-domain warping, which is then refined by an Implicit Binaural Corrector (IBC) module. IBC is an implicit neural network that predicts amplitude and phase corrections directly, resulting in a highly compact model architecture. Experimental results show that LINN achieves statistically comparable perceptual quality to the best-performing baseline model while significantly improving computational efficiency. Compared to the most efficient existing method, LINN achieves a 72.7% reduction in parameters and significantly fewer compute operations (MACs). This demonstrates that our approach effectively addresses the trade-off between synthesis quality and computational efficiency, providing a new solution for high-fidelity edge-device spatial audio applications.

Method: Three-component architecture: 1) Dual-feature encoder combining STFT and gammatone features, 2) Channel-independent globally adaptive Fourier modulator for long-term dependencies, 3) Dynamic gating mechanism to reduce artifacts.

Result: Achieves competitive performance in binaural cues (ILD and IPD error) and objective intelligibility (MBSTOI) with fewer parameters and lower computational cost.

Conclusion: GAF-Net provides a feasible solution for high-fidelity binaural processing on resource-constrained devices, bridging the performance-efficiency gap.

Abstract: Binaural speech enhancement faces a severe trade-off challenge, where state-of-the-art performance is achieved by computationally intensive architectures, while lightweight solutions often come at the cost of significant performance degradation. To bridge this gap, we propose the Global Adaptive Fourier Network (GAF-Net), a lightweight deep complex network that aims to establish a balance between performance and computational efficiency. The GAF-Net architecture consists of three components. First, a dual-feature encoder combining short-time Fourier transform and gammatone features enhances the robustness of acoustic representation. Second, a channel-independent globally adaptive Fourier modulator efficiently captures long-term temporal dependencies while preserving the spatial cues. Finally, a dynamic gating mechanism is implemented to reduce processing artifacts. Experimental results show that GAF-Net achieves competitive performance, particularly in terms of binaural cues (ILD and IPD error) and objective intelligibility (MBSTOI), with fewer parameters and computational cost. These results confirm that GAF-Net provides a feasible way to achieve high-fidelity binaural processing on resource-constrained devices.

Method: Proposes SV-Mixer with three modules: Multi-Scale Mixing for temporal features, Local-Global Mixing for context modeling, and Group Channel Mixing for spectral subspaces. Uses distillation from WavLM teacher.

Result: Outperforms Transformer student by 14.6% while cutting parameters and GMACs by over half. At 75% compression, closely matches teacher’s performance.

Conclusion: Attention-free SSL students can achieve teacher-level accuracy with hardware-friendly footprints, enabling robust on-device speaker verification deployment.

Abstract: Self-supervised learning (SSL) has pushed speaker verification accuracy close to state-of-the-art levels, but the Transformer backbones used in most SSL encoders hinder on-device and real-time deployment. Prior compression work trims layer depth or width yet still inherits the quadratic cost of self-attention. We propose SV-Mixer, the first fully MLP-based student encoder for SSL distillation. SV-Mixer replaces Transformer with three lightweight modules: Multi-Scale Mixing for multi-resolution temporal features, Local-Global Mixing for frame-to-utterance context, and Group Channel Mixing for spectral subspaces. Distilled from WavLM, SV-Mixer outperforms a Transformer student by 14.6% while cutting parameters and GMACs by over half, and at 75% compression, it closely matches the teacher’s performance. Our results show that attention-free SSL students can deliver teacher-level accuracy with hardware-friendly footprints, opening the door to robust on-device speaker verification.

Method: Uses human-readable speech signal representations fed into LLM for pronunciation accuracy/fluency assessment with reasoning, plus phoneme sequence match scoring to refine accuracy scores.

Result: Cost-efficient and competitive performance, significantly improves conventional audio-score models on out-of-domain data by providing complementary perspective.

Conclusion: Demonstrates a new direction for pronunciation assessment by leveraging LLMs’ embedded pronunciation knowledge in text rather than supervised audio-score training.

Abstract: Automatic pronunciation assessment is typically performed by acoustic models trained on audio-score pairs. Although effective, these systems provide only numerical scores, without the information needed to help learners understand their errors. Meanwhile, large language models (LLMs) have proven effective in supporting language learning, but their potential for assessing pronunciation remains unexplored. In this work, we introduce TextPA, a zero-shot, Textual description-based Pronunciation Assessment approach. TextPA utilizes human-readable representations of speech signals, which are fed into an LLM to assess pronunciation accuracy and fluency, while also providing reasoning behind the assigned scores. Finally, a phoneme sequence match scoring method is used to refine the accuracy scores. Our work highlights a previously overlooked direction for pronunciation assessment. Instead of relying on supervised training with audio-score examples, we exploit the rich pronunciation knowledge embedded in written text. Experimental results show that our approach is both cost-efficient and competitive in performance. Furthermore, TextPA significantly improves the performance of conventional audio-score-trained models on out-of-domain data by offering a complementary perspective.

Method: SelectTTS selects appropriate frames from target speaker’s speech and decodes them using frame-level self-supervised learning (SSL) features to capture speaker characteristics.

Result: Achieves performance comparable to state-of-the-art systems (XTTS-v2 and VALL-E) on objective and subjective metrics while requiring over 8x fewer parameters and 270x less training data.

Conclusion: Frame selection with SSL features provides an efficient path to low-complexity, high-quality multi-speaker TTS that generalizes well to unseen speakers.

Abstract: Synthesizing the voices of unseen speakers remains a persisting challenge in multi-speaker text-to-speech (TTS). Existing methods model speaker characteristics through speaker conditioning during training, leading to increased model complexity and limiting reproducibility and accessibility. A low-complexity alternative would broaden the reach of speech synthesis research, particularly in settings with limited computational and data resources. To this end, we propose SelectTTS, a simple and effective alternative. SelectTTS selects appropriate frames from the target speaker and decodes them using frame-level self-supervised learning (SSL) features. We demonstrate that this approach can effectively capture speaker characteristics for unseen speakers and achieves performance comparable to state-of-the-art multi-speaker TTS frameworks on both objective and subjective metrics. By directly selecting frames from the target speaker’s speech, SelectTTS enables generalization to unseen speakers with significantly lower model complexity. Experimental results show that the proposed approach achieves performance comparable to state-of-the-art systems such as XTTS-v2 and VALL-E, while requiring over 8x fewer parameters and 270x less training data. Moreover, it demonstrates that frame selection with SSL features offers an efficient path to low-complexity, high-quality multi-speaker TTS.

Method: Uses Flow-VAE to extract continuous latent speech representations from waveforms, then trains a single autoregressive Transformer to predict these continuous speech distributions from text using Kullback-Leibler divergence loss.

Result: Achieves superior speech synthesis quality compared to existing methods and demonstrates the ability to adapt to target speakers from just a single sample.

Conclusion: KALL-E provides a more direct and effective approach for utilizing continuous speech representations in text-to-speech synthesis, eliminating the need for diffusion-based components while maintaining high quality and adaptability.

Abstract: We introduce KALL-E, a novel autoregressive (AR) language model for text-to-speech (TTS) synthesis that operates by predicting the next distribution of continuous speech frames. Unlike existing methods, KALL-E directly models the continuous speech distribution conditioned on text, eliminating the need for any diffusion-based components. Specifically, we utilize a Flow-VAE to extract a continuous latent speech representation from waveforms, instead of relying on discrete speech tokens. A single AR Transformer is then trained to predict these continuous speech distributions from text, optimizing a Kullback-Leibler divergence loss as its objective. Experimental results demonstrate that KALL-E achieves superior speech synthesis quality and can even adapt to a target speaker from just a single sample. Importantly, KALL-E provides a more direct and effective approach for utilizing continuous speech representations in TTS.

Method: Three-component framework: 1) Automatic video segmentation for semantic labeling of vessels/catheters, 2) Motion phase matching using stationary object tracking to find synchronized image pairs, 3) Novel geometric reconstruction algorithm using 3D surface intersections instead of epipolar constraints.

Result: Segmentation accuracy of 0.703 on test set of 62 X-ray sequences. 3D reconstruction achieved reprojection errors of 0.62mm +/- 0.38mm for anatomical landmarks.

Conclusion: The proposed framework simplifies 3D vessel reconstruction workflow and improves accuracy compared to traditional epipolar constraint methods, making it suitable for clinical cardiac interventions.

Abstract: X-ray angiography is widely used in cardiac interventions to visualize coronary vessels, assess integrity, detect stenoses and guide treatment. We propose a framework for reconstructing 3D vessel trees from biplanar X-ray images which are extracted from two X-ray videos captured at different C-arm angles. The proposed framework consists of three main components: image segmentation, motion phase matching, and 3D reconstruction. An automatic video segmentation method for X-ray angiography to enable semantic segmentation for image segmentation and motion phase matching. The goal of the motion phase matching is to identify a pair of X-ray images that correspond to a similar respiratory and cardiac motion phase to reduce errors in 3D reconstruction. This is achieved by tracking a stationary object such as a catheter or lead within the X-ray video. The semantic segmentation approach assigns different labels to different object classes enabling accurate differentiation between blood vessels, balloons, and catheters. Once a suitable image pair is selected, key anatomical landmarks (vessel branching points and endpoints) are matched between the two views using a heuristic method that minimizes reconstruction errors. This is followed by a novel geometric reconstruction algorithm to generate the 3D vessel tree. The algorithm computes the 3D vessel centrelines by determining the intersection of two 3D surfaces. Compared to traditional methods based on epipolar constraints, the proposed approach simplifies there construction workflow and improves overall accuracy. We trained and validated our segmentation method on 62 X-ray angiography video sequences. On the test set, our method achieved a segmentation accuracy of 0.703. The 3D reconstruction framework was validated by measuring the reconstruction error of key anatomical landmarks, achieving a reprojection errors of 0.62mm +/- 0.38mm.

Method: Developed PREDICT-GBM platform with expert-curated clinical dataset of 255 subjects with complete tumor segmentations and tissue characterization maps to systematically benchmark state-of-the-art tumor growth models.

Result: Personalized radiation treatment plans derived from tumor growth predictions achieved superior recurrence coverage compared to conventional uniform margin approaches for two of the evaluated models.

Conclusion: PREDICT-GBM provides a robust platform for advancing and systematically evaluating tumor growth modeling approaches, facilitating clinical translation and improving patient outcomes in glioblastoma treatment.

Abstract: Glioblastoma is the most prevalent primary brain malignancy, distinguished by its highly invasive behavior and exceptionally high rates of recurrence. Conventional radiation therapy, which employs uniform treatment margins, fails to account for patient-specific anatomical and biological factors that critically influence tumor cell migration. To address this limitation, numerous computational models of glioblastoma growth have been developed, enabling generation of tumor cell distribution maps extending beyond radiographically visible regions and thus informing more precise treatment strategies. However, despite encouraging preliminary findings, the clinical adoption of these growth models remains limited. To bridge this translational gap and accelerate both model development and clinical validation, we introduce PREDICT-GBM, a comprehensive integrated pipeline and dataset for modeling and evaluation. This platform enables systematic benchmarking of state-of-the-art tumor growth models using an expert-curated clinical dataset comprising 255 subjects with complete tumor segmentations and tissue characterization maps. Our analysis demonstrates that personalized radiation treatment plans derived from tumor growth predictions achieved superior recurrence coverage compared to conventional uniform margin approaches for two of the evaluated models. This work establishes a robust platform for advancing and systematically evaluating cutting-edge tumor growth modeling approaches, with the ultimate goal of facilitating clinical translation and improving patient outcomes.

Method: Multi-step AI pipeline using Google’s Gemini 2.5 Flash (2.5B parameters) for processing fluoroscopic angiograms through transformer-based neural architecture, including preprocessing, segmentation, enhancement, and artifact removal, followed by Tencent’s Hunyuan3D-2mini (384M parameters) for stereolithography file generation.

Result: Pipeline successfully generated geometrically optimized 2D projections after 16 processing steps, achieving anatomically faithful representations with enhanced contrast. AI-generated virtual flow visualization identified stagnation zones and flow patterns. Complete processing required under 15 minutes with second-level API response times.

Conclusion: This approach demonstrates clinical feasibility for generating CFD-suitable geometries from routine angiographic data, enabling rapid 3D generation and virtual flow visualization, establishing foundation for democratizing advanced geometric and hemodynamic analysis using readily available imaging data.

Abstract: Fontan palliation for univentricular congenital heart disease progresses to hemodynamic failure with complex flow patterns poorly characterized by conventional 2D imaging. Current assessment relies on fluoroscopic angiography, providing limited 3D geometric information essential for computational fluid dynamics (CFD) analysis and surgical planning. A multi-step AI pipeline was developed utilizing Google’s Gemini 2.5 Flash (2.5B parameters) for systematic, iterative processing of fluoroscopic angiograms through transformer-based neural architecture. The pipeline encompasses medical image preprocessing, vascular segmentation, contrast enhancement, artifact removal, and virtual hemodynamic flow visualization within 2D projections. Final views were processed through Tencent’s Hunyuan3D-2mini (384M parameters) for stereolithography file generation. The pipeline successfully generated geometrically optimized 2D projections from single-view angiograms after 16 processing steps using a custom web interface. Initial iterations contained hallucinated vascular features requiring iterative refinement to achieve anatomically faithful representations. Final projections demonstrated accurate preservation of complex Fontan geometry with enhanced contrast suitable for 3D conversion. AI-generated virtual flow visualization identified stagnation zones in central connections and flow patterns in branch arteries. Complete processing required under 15 minutes with second-level API response times. This approach demonstrates clinical feasibility of generating CFD-suitable geometries from routine angiographic data, enabling 3D generation and rapid virtual flow visualization for cursory insights prior to full CFD simulation. While requiring refinement cycles for accuracy, this establishes foundation for democratizing advanced geometric and hemodynamic analysis using readily available imaging data.

Method: Proposes CDPIR framework integrating cross-distribution diffusion priors from a Scalable Interpolant Transformer (SiT) with model-based iterative reconstruction. Uses Classifier-Free Guidance across multiple datasets and random conditioning dropout to learn both domain-specific and domain-invariant priors.

Result: Extensive experiments show CDPIR significantly outperforms existing approaches, particularly under out-of-distribution conditions, with superior detail preservation in sparse-view CT reconstructions.

Conclusion: CDPIR demonstrates robust performance and potential clinical value in challenging imaging scenarios by effectively addressing out-of-distribution problems through cross-distribution diffusion priors and flexible sampling strategies.

Abstract: Sparse-View CT (SVCT) reconstruction enhances temporal resolution and reduces radiation dose, yet its clinical use is hindered by artifacts due to view reduction and domain shifts from scanner, protocol, or anatomical variations, leading to performance degradation in out-of-distribution (OOD) scenarios. In this work, we propose a Cross-Distribution Diffusion Priors-Driven Iterative Reconstruction (CDPIR) framework to tackle the OOD problem in SVCT. CDPIR integrates cross-distribution diffusion priors, derived from a Scalable Interpolant Transformer (SiT), with model-based iterative reconstruction methods. Specifically, we train a SiT backbone, an extension of the Diffusion Transformer (DiT) architecture, to establish a unified stochastic interpolant framework, leveraging Classifier-Free Guidance (CFG) across multiple datasets. By randomly dropping the conditioning with a null embedding during training, the model learns both domain-specific and domain-invariant priors, enhancing generalizability. During sampling, the globally sensitive transformer-based diffusion model exploits the cross-distribution prior within the unified stochastic interpolant framework, enabling flexible and stable control over multi-distribution-to-noise interpolation paths and decoupled sampling strategies, thereby improving adaptation to OOD reconstruction. By alternating between data fidelity and sampling updates, our model achieves state-of-the-art performance with superior detail preservation in SVCT reconstructions. Extensive experiments demonstrate that CDPIR significantly outperforms existing approaches, particularly under OOD conditions, highlighting its robustness and potential clinical value in challenging imaging scenarios.

Method: End-to-end designed diffractive polarization spectral element (DPSE) modulates scene to generate phase-encoding and polarization information, combined with SPMSA-Net neural network for spectro-polarimetric data cube reconstruction.

Result: Achieves 0.78 dB PSNR and 0.012 SSIM improvement over state-of-the-art, captures 400-700nm spectrum with 10nm resolution, full-Stokes parameters, 2252x2252 spatial resolution, over 98.9% fidelity, with compact 2mm modulation component.

Conclusion: Proposed framework enables high-performance spectro-polarimetric imaging with compact architecture, high spatial/spectral resolution, and precise polarization characterization, addressing limitations of existing systems.

Abstract: Spectro-polarimetric imaging provides multidimensional optical information acquisition capabilities, offering significant potential for diverse applications. Current spectro-polarimetric imaging systems typically suffer from large physical footprints, high design complexity, elevated costs, or the drawback of requiring replacement of standard components with polarization optics. To address these issues, we propose an integrated diffractive full-Stokes spectro-polarimetric imaging framework that synergistically combines end-to-end designed diffractive polarization spectral element (DPSE) with SPMSA-Net to demonstrate high-performance spectro-polarimetric imaging. The DPSE modulates scene and generates modulated images carrying phase-encoding and polarization information. The modulated images are the input of the SPMSA-NET for the reconstruction of the spectro-polarimetric data cube. The framework achieves an average improvement of 0.78 dB in PSNR and 0.012 in SSIM over existing state-of-the-art algorithms. Based on this framework, our prototype system can simultaneously capture spectral information (400-700 nm) with 10 nm spectral resolution and full-Stokes parameters (S0,S1,S2,S3). Meanwhile, the system provides high spatial resolution of 2252*2252 pixels. Experimental results demonstrate that our system achieves high-fidelity spectral imaging (over 98.9% fidelity) and precise polarization characterization, with a compact architecture (modulation component of merely 2-mm thickness).

Method: The method uses remote sensing images to detect pile contours, reconstructs 3D models based on the material’s angle of repose, and estimates volume accordingly. Validated on conical and elongated piles in laboratory settings.

Result: The algorithm demonstrated strong potential for accurate volume estimation from experimental images and reference satellite imagery, achieving high accuracy in validation tests on various pile types.

Conclusion: The proposed volume estimation algorithm shows promising results for practical application in shipping port management, providing accurate pile volume measurements that can enhance logistics efficiency.

Abstract: Estimation of volume of piles in shipping ports plays a pivotal role for logistics management, facilitates better ship rescheduling and rerouting for economic benefits and contributes to overall efficient shipping management. This paper presents validation results for a volume estimation algorithm for dry bulk cargo piles stored in open ports. Using remote sensing images obtained in a laboratory setting, the method first detects the contour of the pile and then reconstructs its 3D model based on the material’s angle of repose, and estimates the volume accordingly. We validated the algorithm on full conical piles and single-ridge elongated piles, and further tested it on reclaimed conical and elongated piles. The results demonstrated the algorithm’s strong potential for accurately estimating pile volume from experimental images and a reference satellite image, achieving high accuracy in our validation.

Method: Conducted extensive benchmark of deep learning architectures across Folsom, SIRTA and NREL datasets, performed ablation experiments on training configurations and data processing techniques, identified timestamp alignment issues in Folsom dataset, and combined best estimation model with video prediction model.

Result: Demonstrated consistent findings across different solar stations, achieved state-of-the-art results on SIRTA dataset when combining the best irradiance estimation model with a video prediction model.

Conclusion: The study provides comprehensive benchmarking and improvements for solar irradiance estimation from sky images, addressing timestamp alignment issues and showing that enhanced estimation models combined with video prediction can achieve superior nowcasting performance.

Abstract: To address the high levels of uncertainty associated with photovoltaic energy, an increasing number of studies focusing on short-term solar forecasting (i.e. nowcasting) have been published. Most of these studies use deep-learning-based models to directly forecast a solar irradiance or photovoltaic power value given an input sequence of sky images. Recently, however, advances in generative modeling have led to approaches that divide the nowcasting problem into two sub-problems: 1) future event prediction, i.e. generating future sky images; and 2) solar irradiance or photovoltaic power estimation, i.e. predicting the concurrent value from a single image. One such approach is the SkyGPT model, whose potential for improvement is shown to be much larger in the estimation component than in the generative component. Thus, in this paper, we focus on the solar irradiance estimation problem and conduct an extensive benchmark of deep learning architectures across the widely-used Folsom, SIRTA and NREL datasets. Moreover, we perform ablation experiments on different training configurations and data processing techniques, including the choice of the target variable used for training and adjustments of the timestamp alignment between images and irradiance measurements. In particular, we draw attention to a potential error associated with the sky image timestamps in the Folsom dataset and suggest a possible fix. By leveraging the three datasets, we demonstrate that our findings are consistent across different solar stations. Finally, we combine our best irradiance estimation model with a video prediction model and obtain state-of-the-art results on the SIRTA dataset.

Method: Attention-based network with residual blocks at each encoder-decoder stage. Uses strided convolutions instead of max pooling, transposed convolutions instead of trilinear interpolation, and dilated convolutions to capture larger context without increasing computational costs.

Result: Achieves better performance than state-of-the-art methods on LUNA16 dataset using standard metrics like Dice score and IOU.

Conclusion: The proposed architecture provides accurate 3D segmentation for lung CAD systems, addressing limitations of existing methods and demonstrating superior performance on large public datasets.

Abstract: Lung cancer has been one of the major threats across the world with the highest mortalities. Computer-aided detection (CAD) can help in early detection and thus can help increase the survival rate. Accurate lung parenchyma segmentation (to include the juxta-pleural nodules) and lung nodule segmentation, the primary symptom of lung cancer, play a crucial role in the overall accuracy of the Lung CAD pipeline. Lung nodule segmentation is quite challenging because of the diverse nodule types and other inhibit structures present within the lung lobes. Traditional machine/deep learning methods suffer from generalization and robustness. Recent Vision Language Models/Foundation Models perform well on the anatomical level, but they suffer on fine-grained segmentation tasks, and their semi-automatic nature limits their effectiveness in real-time clinical scenarios. In this paper, we propose a novel method for accurate 3D segmentation of lung parenchyma and lung nodules. The proposed architecture is an attention-based network with residual blocks at each encoder-decoder state. Max pooling is replaced by strided convolutions at the encoder, and trilinear interpolation is replaced by transposed convolutions at the decoder to maximize the number of learnable parameters. Dilated convolutions at each encoder-decoder stage allow the model to capture the larger context without increasing computational costs. The proposed method has been evaluated extensively on one of the largest publicly available datasets, namely LUNA16, and is compared with recent notable work in the domain using standard performance metrics like Dice score, IOU, etc. It can be seen from the results that the proposed method achieves better performance than state-of-the-art methods. The source code, datasets, and pre-processed data can be accessed using the link: https://github.com/EMeRALDsNRPU/Attention-Based-3D-ResUNet.

Method: Proposes MEGANet-W with a two-level Haar wavelet head for multi-orientation edge extraction and Wavelet Edge Guided Attention (W-EGA) modules that fuse wavelet cues with boundary and input branches to recalibrate semantic features.

Result: Outperforms existing methods on five public polyp datasets, improving mIoU by up to 2.3% and mDice by 1.2%, while introducing no additional learnable parameters.

Conclusion: The approach improves reliability in difficult cases and offers a robust solution for medical image segmentation tasks requiring precise boundary detection.

Abstract: Colorectal polyp segmentation is critical for early detection of colorectal cancer, yet weak and low contrast boundaries significantly limit automated accuracy. Existing deep models either blur fine edge details or rely on handcrafted filters that perform poorly under variable imaging conditions. We propose MEGANet-W, a Wavelet Driven Edge Guided Attention Network that injects directional, parameter free Haar wavelet edge maps into each decoder stage to recalibrate semantic features. The key novelties of MEGANet-W include a two-level Haar wavelet head for multi-orientation edge extraction; and Wavelet Edge Guided Attention (W-EGA) modules that fuse wavelet cues with boundary and input branches. On five public polyp datasets, MEGANet-W consistently outperforms existing methods, improving mIoU by up to 2.3% and mDice by 1.2%, while introducing no additional learnable parameters. This approach improves reliability in difficult cases and offers a robust solution for medical image segmentation tasks requiring precise boundary detection.