Method: Proposes Reasoning Activation Potential (RAP) with two complementary estimators: 1) Causal Discrepancy Estimator (CDE) eliminates samples overly reliant on language priors by comparing multimodal vs text-only outputs; 2) Attention Confidence Estimator (ACE) discards samples dominated by irrelevant tokens using token-level self-attention. Also includes Difficulty-aware Replacement Module (DRM) to substitute trivial instances with challenging ones.

Result: Experiments on six datasets show RAP consistently achieves superior performance using only 9.3% of training data while reducing computational costs by over 43%.

Conclusion: Smaller high-value datasets can match or outperform full corpora for multimodal reasoning in MLLMs. The RAP paradigm effectively identifies cognitive samples that trigger genuine reasoning, enabling efficient training with reduced data and computational requirements.

Abstract: While multi-modal large language models (MLLMs) have made significant progress in complex reasoning tasks via reinforcement learning, it is commonly believed that extensive training data is necessary for improving multi-modal reasoning ability, inevitably leading to data redundancy and substantial computational costs. However, can smaller high-value datasets match or outperform full corpora for multi-modal reasoning in MLLMs? In this work, we challenge this assumption through a key observation: meaningful multi-modal reasoning is triggered by only a sparse subset of training samples, termed cognitive samples, whereas the majority contribute marginally. Building on this insight, we propose a novel data selection paradigm termed Reasoning Activation Potential (RAP)}, which identifies cognitive samples by estimating each sample’s potential to stimulate genuine multi-modal reasoning by two complementary estimators: 1) Causal Discrepancy Estimator (CDE) based on the potential outcome model principle, eliminates samples that overly rely on language priors by comparing outputs between multi-modal and text-only inputs; 2) Attention Confidence Estimator (ACE), which exploits token-level self-attention to discard samples dominated by irrelevant but over-emphasized tokens in intermediate reasoning stages. Moreover, we introduce a Difficulty-aware Replacement Module (DRM) to substitute trivial instances with cognitively challenging ones, thereby ensuring complexity for robust multi-modal reasoning. Experiments on six datasets show that our RAP method consistently achieves superior performance using only 9.3% of the training data, while reducing computational costs by over 43%.

Method: Hierarchical subspace latent diffusion model that directly maps visual lip movements to continuous latent space of pre-trained neural audio codec. Uses subspace decomposition module, diffusion convolution blocks (DiCB) for intra/inter-subspace interactions, and reparameterized flow matching with SLM and semantic losses.

Result: Achieves state-of-the-art generation quality on multiple benchmark datasets, surpassing existing methods in both objective and subjective evaluations.

Conclusion: SLD-L2S demonstrates the effectiveness of latent diffusion models for lip-to-speech synthesis, avoiding information loss from intermediate representations and enabling principled integration of semantic losses for improved speech quality.

Abstract: Although lip-to-speech synthesis (L2S) has achieved significant progress in recent years, current state-of-the-art methods typically rely on intermediate representations such as mel-spectrograms or discrete self-supervised learning (SSL) tokens. The potential of latent diffusion models (LDMs) in this task remains largely unexplored. In this paper, we introduce SLD-L2S, a novel L2S framework built upon a hierarchical subspace latent diffusion model. Our method aims to directly map visual lip movements to the continuous latent space of a pre-trained neural audio codec, thereby avoiding the information loss inherent in traditional intermediate representations. The core of our method is a hierarchical architecture that processes visual representations through multiple parallel subspaces, initiated by a subspace decomposition module. To efficiently enhance interactions within and between these subspaces, we design the diffusion convolution block (DiCB) as our network backbone. Furthermore, we employ a reparameterized flow matching technique to directly generate the target latent vectors. This enables a principled inclusion of speech language model (SLM) and semantic losses during training, moving beyond conventional flow matching objectives and improving synthesized speech quality. Our experiments show that SLD-L2S achieves state-of-the-art generation quality on multiple benchmark datasets, surpassing existing methods in both objective and subjective evaluations.

Method: Proposes audio-interleaved reasoning that treats audio as an active reasoning component. Uses a two-stage training framework: 1) supervised fine-tuning to teach LALMs to localize salient audio segments, and 2) reinforcement learning to incentivize proficient re-listening. Also develops a structured data generation pipeline for high-quality training data.

Result: Echo achieves overall superiority on audio comprehension benchmarks in both challenging expert-level and general-purpose tasks. Comprehensive analysis confirms the efficiency and generalizability of audio-interleaved reasoning.

Conclusion: Audio-interleaved reasoning is a promising direction for advancing audio comprehension in LALMs, enabling more human-like dynamic engagement with audio content during reasoning.

Abstract: The maturation of Large Audio Language Models (LALMs) has raised growing expectations for them to comprehend complex audio much like humans. Current efforts primarily replicate text-based reasoning by contextualizing audio content through a one-time encoding, which introduces a critical information bottleneck. Drawing inspiration from human cognition, we propose audio-interleaved reasoning to break through this bottleneck. It treats audio as an active reasoning component, enabling sustained audio engagement and perception-grounded analysis. To instantiate it, we introduce a two-stage training framework, first teaching LALMs to localize salient audio segments through supervised fine-tuning, and then incentivizing proficient re-listening via reinforcement learning. In parallel, a structured data generation pipeline is developed to produce high-quality training data. Consequently, we present Echo, a LALM capable of dynamically re-listening to audio in demand during reasoning. On audio comprehension benchmarks, Echo achieves overall superiority in both challenging expert-level and general-purpose tasks. Comprehensive analysis further confirms the efficiency and generalizability of audio-interleaved reasoning, establishing it as a promising direction for advancing audio comprehension. Project page: https://github.com/wdqqdw/Echo.

Method: 1) Ingests unstructured PDFs via OCR and layout analysis, converting them to hierarchical text chunks; 2) Pre-generates QA knowledge base from chunks using LLM; 3) At query time, matches user questions against QA bank first, only falling back to on-the-fly generation when no suitable match is found.

Result: Experiments on OHRBench show HybridRAG provides higher answer quality and lower latency compared to standard RAG baselines.

Conclusion: HybridRAG is a practical solution for real-world chatbot applications handling large volumes of unstructured documents and high user loads under limited computational resources.

Abstract: Retrieval-Augmented Generation (RAG) has emerged as a powerful approach for grounding Large Language Model (LLM)-based chatbot responses on external knowledge. However, existing RAG studies typically assume well-structured textual sources (e.g. Wikipedia or curated datasets) and perform retrieval and generation at query time, which can limit their applicability in real-world chatbot scenarios. In this paper, we present HybridRAG, a novel and practical RAG framework towards more accurate and faster chatbot responses. First, HybridRAG ingests raw, unstructured PDF documents containing complex layouts (text, tables, figures) via Optical Character Recognition (OCR) and layout analysis, and convert them into hierarchical text chunks. Then, it pre-generates a plausible question-answer (QA) knowledge base from the organized chunks using an LLM. At query time, user questions are matched against this QA bank to retrieve immediate answers when possible, and only if no suitable QA match is found does our framework fall back to an on-the-fly response generation. Experiments on OHRBench demonstrate that our HybridRAG provides higher answer quality and lower latency compared to a standard RAG baseline. We believe that HybridRAG could be a practical solution for real-world chatbot applications that must handle large volumes of unstructured documents and lots of users under limited computational resources.

Method: Used knowledge distillation with LoRA to transfer refusal behaviors from OpenAI o1-mini to three open-source models (Meta-Llama-3-8B-Instruct, Gemma-2-2B-IT, Qwen3-8B). Employed ~28,000 multilingual jailbreak prompts from XSafety via black-box response-based PEFT, then evaluated on MultiJail benchmark.

Result: Counterintuitively, fine-tuning on “safe” refusal data increased Jailbreak Success Rate by up to 16.6 percentage points across all student models. Removing nuanced “boundary” refusals mitigated safety declines but reduced reasoning performance on GSM8K. Results showed divergent generalization to unseen languages depending on base model.

Conclusion: Knowledge distillation for multilingual safety alignment is challenging but potentially useful. The study reveals unexpected safety degradation from standard fine-tuning approaches and highlights the need for more sophisticated distillation techniques that preserve both safety and reasoning capabilities across languages.

Abstract: Large language models (LLMs) are increasingly deployed worldwide, yet their safety alignment remains predominantly English-centric. This allows for vulnerabilities in non-English contexts, especially with low-resource languages. We introduce a novel application of knowledge distillation (KD) in the context of multilingual jailbreak prevention, examining its efficacy. We distill the refusal behaviors of a proprietary teacher model (OpenAI o1-mini) with Low-Rank Adaptation (LoRA) into three open-source student models: Meta-Llama-3-8B-Instruct, Gemma-2-2B-IT, and Qwen3-8B, using ~28,000 multilingual jailbreak prompts from XSafety via black-box response-based, parameter-efficient fine-tuning (PEFT). Evaluation on the MultiJail benchmark reveals a counterintuitive behavior: standard fine-tuning on the teacher’s ``safe’’ refusal data inadvertently increases Jailbreak Success Rate (JSR) for all student models, up to 16.6 percentage points. Our experiments reveal a divergent generalization to unseen languages during distillation, with varying outcomes depending on the base model. By removing a primary source of safety degradation, nuanced `boundary’ refusals, we mitigate or even reverse safety declines in student models, although reductions in reasoning performance (GSM8K) persist. Overall, our exploratory study highlights the challenges and potential of KD as a technique for multilingual safety alignment, offering a foundation for future research in this direction.

Method: Extensive analysis of retrieval heads in LLMs, examining their dynamic behavior across timesteps during autoregressive generation. Validation on Needle-in-a-Haystack and multi-hop QA tasks, with quantification of differences between dynamic and static retrieval heads in a Dynamic Retrieval-Augmented Generation framework.

Result: Three core findings: (1) Retrieval heads vary dynamically across timesteps; (2) Dynamic retrieval heads are specific at each timestep and cannot be effectively replaced by static retrieval heads; (3) Hidden states encode predictive signals for future retrieval head patterns, indicating internal planning mechanisms.

Conclusion: The study reveals dynamic temporal patterns in LLM retrieval heads that were overlooked by static analyses, providing new insights into the internal mechanisms of LLMs and their planning capabilities during generation.

Abstract: Recent studies have identified “retrieval heads” in Large Language Models (LLMs) responsible for extracting information from input contexts. However, prior works largely rely on static statistics aggregated across datasets, identifying heads that perform retrieval on average. This perspective overlooks the fine-grained temporal dynamics of autoregressive generation. In this paper, we investigate retrieval heads from a dynamic perspective. Through extensive analysis, we establish three core claims: (1) Dynamism: Retrieval heads vary dynamically across timesteps; (2) Irreplaceability: Dynamic retrieval heads are specific at each timestep and cannot be effectively replaced by static retrieval heads; and (3) Correlation: The model’s hidden state encodes a predictive signal for future retrieval head patterns, indicating an internal planning mechanism. We validate these findings on the Needle-in-a-Haystack task and a multi-hop QA task, and quantify the differences on the utility of dynamic and static retrieval heads in a Dynamic Retrieval-Augmented Generation framework. Our study provides new insights into the internal mechanisms of LLMs.

Method: Created ALME benchmark with 57,602 controlled audio-text conflict stimuli across 8 languages. Tested Gemini 2.0 Flash and other audio-LLMs, comparing audio-text vs text-text conflicts. Conducted interventions: forced transcription, framing text as corrupted, and fine-tuning ablations (audio projection layer vs LoRA on LLM).

Result: Models follow text 10x more often in audio-text conflicts (16.6%) vs text-text conflicts (1.6%). Audio-only accuracy (97.2%) exceeds cascade accuracy (93.9%). Forced transcription increases text dominance (19% to 33%). Framing text as corrupted reduces text dominance by 80%. Fine-tuning audio projection layer increases dominance (+26.5%), while LoRA on LLM halves it (-23.9%).

Conclusion: Text dominance reflects accessibility asymmetry in arbitration, not information content differences. The issue resides in LLM’s reasoning rather than audio encoder. Modality arbitration is a distinct reliability dimension not captured by standard benchmarks.

Abstract: When audio and text conflict, speech-enabled language models follow the text 10 times more often than when arbitrating between two text sources, even when explicitly instructed to trust the audio. Using ALME, a benchmark of 57,602 controlled audio-text conflict stimuli across 8 languages, we find that Gemini 2.0 Flash exhibits 16.6% text dominance under audio-text conflict versus 1.6% under text-text conflict with identical reliability cues. This gap is not explained by audio quality: audio-only accuracy (97.2%) exceeds cascade accuracy (93.9%), indicating audio embeddings preserve more information than text transcripts. We propose that text dominance reflects an asymmetry not in information content but in arbitration accessibility: how easily the model can reason over competing representations. This framework explains otherwise puzzling findings. Forcing transcription before answering increases text dominance (19% to 33%), sacrificing audio’s information advantage without improving accessibility. Framing text as ``deliberately corrupted’’ reduces text dominance by 80%. A fine-tuning ablation provides interventional evidence: training only the audio projection layer increases text dominance (+26.5%), while LoRA on the language model halves it ($-$23.9%), localizing text dominance to the LLM’s reasoning rather than the audio encoder. Experiments across four state-of-the-art audio-LLMs and 8 languages show consistent trends with substantial cross-linguistic and cross-model variation, establishing modality arbitration as a distinct reliability dimension not captured by standard speech benchmarks.

Method: 1) Annotated a plasma physics corpus with 16 classes for nested NER; 2) Developed entity-specific model specialization using independent BERT-CRF models for individual entity types; 3) Integrated hyperparameter optimization to enhance model performance.

Result: The approach successfully extracts nested named entities from plasma physics text, providing a foundation for researchers to navigate and analyze scientific literature in this domain.

Conclusion: This work advances entity recognition in plasma physics and supports researchers in scientific literature analysis, though it focuses on text-only NER rather than multimodal understanding.

Abstract: Named Entity Recognition (NER) is an important task in natural language processing that aims to identify and extract key entities from unstructured text. We present a novel application of NER in plasma physics research articles and address the challenges of extracting specialized entities from scientific text in this domain. Research articles in plasma physics often contain highly complex and context-rich content that must be extracted to enable, e.g., advanced search. We propose a lightweight approach based on encoder-transformers and conditional random fields to extract (nested) named entities from plasma physics research articles. First, we annotate a plasma physics corpus with 16 classes specifically designed for the nested NER task. Second, we evaluate an entity-specific model specialization approach, where independent BERT-CRF models are trained to recognize individual entity types in plasma physics text. Third, we integrate an optimization process to systematically fine-tune hyperparameters and enhance model performance. Our work contributes to the advancement of entity recognition in plasma physics and also provides a foundation to support researchers in navigating and analyzing scientific literature.

Method: Uses sliding-window self-attention in an ergodic streaming-encoder ASR model to achieve bounded, low-latency inference while preserving strong local context, replacing full global attention with carefully designed local attention.

Result: Achieves state-of-the-art word error rates across standard benchmarks, attaining accuracy on-par with models 6x their size while running significantly faster on edge devices.

Conclusion: Carefully designed local attention can be competitive with full attention accuracy at a fraction of the size and latency cost, enabling new possibilities for interactive speech interfaces on edge devices.

Abstract: Latency-critical speech applications (e.g., live transcription, voice commands, and real-time translation) demand low time-to-first-token (TTFT) and high transcription accuracy, particularly on resource-constrained edge devices. Full-attention Transformer encoders remain a strong accuracy baseline for automatic speech recognition (ASR) because every frame can directly attend to every other frame, which resolves otherwise locally ambiguous acoustics using distant lexical context. However, this global dependency incurs quadratic complexity in sequence length, inducing an inherent “encode-the-whole-utterance” latency profile. For streaming use cases, this causes TTFT to grow linearly with utterance length as the encoder must process the entire prefix before any decoder token can be emitted. To better meet the needs of on-device, streaming ASR use cases we introduce Moonshine v2, an ergodic streaming-encoder ASR model that employs sliding-window self-attention to achieve bounded, low-latency inference while preserving strong local context. Our models achieve state of the art word error rates across standard benchmarks, attaining accuracy on-par with models 6x their size while running significantly faster. These results demonstrate that carefully designed local attention is competitive with the accuracy of full attention at a fraction of the size and latency cost, opening new possibilities for interactive speech interfaces on edge devices.

Method: Created RECOM benchmark with 15,000 recent Reddit questions (Sept 2025) and community answers. Evaluated 4 open-source LLMs using lexical metrics (BLEU, ROUGE), semantic similarity (BERTScore, MoverScore, cosine similarity), and logical inference (NLI).

Result: Found semantic-lexical paradox: >99% cosine similarity but <8% BLEU-1 overlap. MoverScore showed 51-53% alignment. Model scale didn’t predict performance (Mistral-7B outperformed GPT-OSS-20B). Contradiction rates <7% in NLI analysis.

Conclusion: Lexical metrics are unreliable for evaluating abstractive generation; multi-dimensional frameworks capturing semantic fidelity beyond surface-level matching are needed.

Abstract: Large Language Models (LLMs) are increasingly deployed for open-domain question answering, yet their alignment with human perspectives on temporally recent information remains underexplored. We introduce RECOM (Reddit Evaluation for Correspondence of Models), a benchmark dataset of 15,000 recent Reddit questions from September 2025 paired with community-derived reference answers. We investigate how four open-source LLMs (Llama3.1-8B, Mistral-7B, Gemma-2-9B, and GPT-OSS-20B) respond to these questions, evaluating alignment using lexical metrics (BLEU, ROUGE), semantic similarity (BERTScore, MoverScore, cosine similarity), and logical inference (NLI). Our central finding is a striking semantic-lexical paradox: all models achieve over 99% cosine similarity with references despite less than 8% BLEU-1 overlap, a 90+ percentage point gap indicating that models preserve meaning through extensive paraphrasing rather than lexical reproduction. MoverScore (51-53%) confirms this pattern, occupying an intermediate position that reflects the optimal transport cost of semantic alignment. Furthermore, model scale does not predict performance: Mistral-7B (7B parameters) outperforms GPT-OSS-20B (20B parameters) across all metrics. NLI analysis reveals that contradiction rates remain below 7%, suggesting models rarely generate content that directly conflicts with human consensus. These findings challenge the reliability of lexical metrics for evaluating abstractive generation and argue for multi-dimensional evaluation frameworks that capture semantic fidelity beyond surface-level text matching. The RECOM dataset is publicly available at https://anonymous.4open.science/r/recom-D4B0

Method: Automatically constructed dataset of English disyllabic words from read/spontaneous speech; trained CNN architectures on spectrographic representations to predict stress position; used Layerwise Relevance Propagation (LRP) for interpretability analysis; proposed feature-specific relevance analysis to identify acoustic cues.

Result: Models achieved up to 92% accuracy on held-out test data; LRP revealed predictions for minimal pairs were most influenced by information in stressed vs. unstressed syllables, particularly spectral properties of stressed vowels; best classifier strongly influenced by stressed vowel’s first and second formants, with some evidence for pitch and third formant contributions.

Conclusion: Deep learning can acquire distributed cues to stress from natural data, extending traditional phonetic research; neural networks attend to spectral properties beyond just pitch, revealing sophisticated acoustic feature learning for lexical stress prediction.

Abstract: Despite their success in speech processing, neural networks often operate as black boxes, prompting the question: what informs their decisions, and how can we interpret them? This work examines this issue in the context of lexical stress. A dataset of English disyllabic words was automatically constructed from read and spontaneous speech. Several Convolutional Neural Network (CNN) architectures were trained to predict stress position from a spectrographic representation of disyllabic words lacking minimal stress pairs (e.g., initial stress WAllet, final stress exTEND), achieving up to 92% accuracy on held-out test data. Layerwise Relevance Propagation (LRP), a technique for neural network interpretability analysis, revealed that predictions for held-out minimal pairs (PROtest vs. proTEST ) were most strongly influenced by information in stressed versus unstressed syllables, particularly the spectral properties of stressed vowels. However, the classifiers also attended to information throughout the word. A feature-specific relevance analysis is proposed, and its results suggest that our best-performing classifier is strongly influenced by the stressed vowel’s first and second formants, with some evidence that its pitch and third formant also contribute. These results reveal deep learning’s ability to acquire distributed cues to stress from naturally occurring data, extending traditional phonetic work based around highly controlled stimuli.

Method: Comprehensive empirical study across three open-weight LLM backbones and three fact-seeking QA benchmarks. Evaluated performance using seven unsupervised hallucination detection methods spanning three approaches: semantic consistency based detectors, confidence based detectors, and entropy based detectors. Further analyses using linear probes and representation diagnostics.

Result: PEFT consistently strengthens hallucination detection ability, substantially improving AUROC across a wide range of hallucination detectors. PEFT methods primarily reshape how uncertainty is encoded and surfaced, rather than injecting new factual knowledge into the models.

Conclusion: PEFT methods enhance hallucination detection in LLMs by modifying uncertainty encoding mechanisms, providing empirical evidence for their effectiveness in improving model reliability on factual QA tasks.

Abstract: Parameter-efficient fine-tuning (PEFT) methods are widely used to adapt large language models (LLMs) to downstream tasks and are often assumed to improve factual correctness. However, how the parameter-efficient fine-tuning methods affect hallucination behavior remains insufficiently understood, especially on QA datasets. In this work, we systematically investigate the impact of PEFT on hallucination detection through a comprehensive empirical study across three open-weight LLM backbones and three fact-seeking QA benchmarks. For each model, we evaluate performance using seven unsupervised hallucination detection methods spanning three complementary approaches: semantic consistency based detectors, confidence based detectors, and entropy based detectors. This multifaceted evaluation enables us to characterize how PEFT reshapes uncertainty across different detection paradigms. In conclusion, our experimental results show that PEFT consistently strengthens hallucination detection ability, substantially improving AUROC across a wide range of hallucination detectors. Besides, further analyses using linear probes and representation diagnostics indicate that PEFT methods primarily reshapes how uncertainty is encoded and surfaced, comparing with injecting new factual knowledge into the models.

Method: Introduces FalseCite dataset to capture and benchmark hallucinated responses induced by misleading/fabricated citations. Tests GPT-4o-mini, Falcon-7B, and Mistral 7-B on the dataset, analyzes internal states by visualizing and clustering hidden state vectors.

Result: Observed noticeable increase in hallucination activity for false claims with deceptive citations, especially in GPT-4o-mini. Hidden state vectors trace out a distinct horn-like shape regardless of hallucination or non-hallucination status.

Conclusion: FalseCite has potential as a foundation for evaluating and mitigating hallucinations in future LLM research, providing systematic benchmarking and insights into internal model states during hallucination.

Abstract: Large Language Models (LLMs) often hallucinate, generating nonsensical or false information that can be especially harmful in sensitive fields such as medicine or law. To study this phenomenon systematically, we introduce FalseCite, a curated dataset designed to capture and benchmark hallucinated responses induced by misleading or fabricated citations. Running GPT-4o-mini, Falcon-7B, and Mistral 7-B through FalseCite, we observed a noticeable increase in hallucination activity for false claims with deceptive citations, especially in GPT-4o-mini. Using the responses from FalseCite, we can also analyze the internal states of hallucinating models, visualizing and clustering the hidden state vectors. From this analysis, we noticed that the hidden state vectors, regardless of hallucination or non-hallucination, tend to trace out a distinct horn-like shape. Our work underscores FalseCite’s potential as a foundation for evaluating and mitigating hallucinations in future LLM research.

Method: Uses Combinatorial Fusion Analysis (CFA) with rank-score characteristic function and cognitive diversity to combine multiple classification models. Employs generative AI to create synthetic training data and compares results with human domain experts.

Result: CFA achieves 96.73% performance, outperforming the best individual model. The combination of multiple ML/AI models using CFA and human expert input demonstrates complementary and enhancing effects.

Conclusion: Combining intelligence from multiple ML/AI models using CFA and incorporating human expert input can significantly enhance text classification performance for complex categories like UN SDGs.

Abstract: (Natural Language Processing) NLP techniques such as text classification and topic discovery are very useful in many application areas including information retrieval, knowledge discovery, policy formulation, and decision-making. However, it remains a challenging problem in cases where the categories are unavailable, difficult to differentiate, or are interrelated. Social analysis with human context is an area that can benefit from text classification, as it relies substantially on text data. The focus of this paper is to enhance the classification of text according to the UN’s Sustainable Development Goals (SDGs) by collecting and combining intelligence from multiple models. Combinatorial Fusion Analysis (CFA), a system fusion paradigm using a rank-score characteristic (RSC) function and cognitive diversity (CD), has been used to enhance classifier methods by combining a set of relatively good and mutually diverse classification models. We use a generative AI model to generate synthetic data for model training and then apply CFA to this classification task. The CFA technique achieves 96.73% performance, outperforming the best individual model. We compare the outcomes with those obtained from human domain experts. It is demonstrated that combining intelligence from multiple ML/AI models using CFA and getting input from human experts can, not only complement, but also enhance each other.

Method: L2-matched perturbation analysis on Pythia-family models, with causal interventions to trace damage flow through attention and LayerNorm pathways, comparing across scales and architectures.

Result: Angular perturbations cause 42.9× more damage to language modeling loss, while magnitude perturbations cause 20.4% vs 1.6% accuracy drop on syntactic tasks; damage flows through different pathways (attention vs LayerNorm).

Conclusion: Direction and magnitude support partially distinct computational roles: direction affects attentional routing, magnitude modulates processing intensity for syntactic judgments, with patterns differing between LayerNorm and RMSNorm architectures.

Abstract: Transformer hidden states encode information as high-dimensional vectors, yet whether direction (orientation in representational space) and magnitude (vector norm) serve distinct functional roles remains unclear. Studying Pythia-family models, we discover a striking cross-over dissociation: angular perturbations cause up to 42.9 more damage to language modeling loss, while magnitude perturbations cause disproportionately more damage to syntactic processing (20.4% vs.1.6% accuracy drop on subject-verb agreement).This finding is enabled by L2-matched perturbation analysis, a methodology ensuring that an gular and magnitude perturbations achieve identical Euclidean displacements. Causal intervention reveals that angular damage flows substantially through the attention pathways (28.4% loss recovery via attention repair), while magnitude damage flows partly through the LayerNorm pathways(29.9% recovery via LayerNorm repair). These patterns replicate across scales within the Pythia architecture family. These findings provide evidence that direction and magnitude support partially distinct computational roles in LayerNorm based architectures. The direction preferentially affects attentional routing, while magnitude modulates processing intensity for fine-grained syntactic judgments. We find different patterns in RMSNorm-based architectures, suggesting that the dissociation depends on architectural choices. Our results refine the linear representation hypothesis and have implications for model editing and interpretability research

Method: PRIME uses three specialized agents: executor (step-by-step reasoning), verifier (constraint checking), and coordinator (backtracking control). They’re optimized through group relative policy optimization. The framework is evaluated on PRIME-Bench, the largest algorithmic reasoning benchmark with 86 tasks across 12 categories.

Result: PRIME improves average accuracy from 26.8% to 93.8% (250% relative gain). Largest improvements on tasks requiring sustained state tracking: Turing machine simulation (9% to 92%) and long division (16% to 94%). Smaller models benefit disproportionately, achieving accuracy comparable to models 8x larger.

Conclusion: PRIME demonstrates that specialized multi-agent frameworks with iterative verification can dramatically improve algorithmic reasoning capabilities of LLMs, with iterative verification being the primary contributor by preventing error propagation.

Abstract: Large language models have demonstrated remarkable capabilities across diverse reasoning tasks, yet their performance on algorithmic reasoning remains limited. To handle this limitation, we propose PRIME (Policy-Reinforced Iterative Multi-agent Execution), a framework comprising three specialized agents, an executor for step-by-step reasoning, a verifier for constraint checking, and a coordinator for backtracking control, optimized through group relative policy optimization. For comprehensive evaluation, we introduce PRIME-Bench, the largest algorithmic reasoning benchmark to date, comprising 86 tasks across 12 categories with 51,600 instances. Tasks span sorting algorithms, graph and tree structures, automata and state machines, symbolic reasoning, and constraint-based puzzles, with execution traces reaching over one million steps. Compared to baseline approach, PRIME improves average accuracy from 26.8% to 93.8%, a 250% relative gain. The largest improvements occur on tasks requiring sustained state tracking, with Turing machine simulation improving from 9% to 92% and long division from 16% to 94%. Ablation studies identify iterative verification as the primary contributor, preventing the error propagation that causes baseline approaches to fail catastrophically. Analysis across model scales (8B-120B parameters) reveals that smaller models benefit disproportionately, achieving accuracy comparable to models 8x larger.

Method: Uses LLMs as discrete-to-continuous mapping to embed hyperparameter domain knowledge via prompting, combines with Bayesian Optimization, adds learnable token for residual information, and employs proxy training with data subsets

Result: Achieves >20% performance improvement with only ~30 iterations compared to standard methods requiring ~45,000 combinations

Conclusion: Proposed framework efficiently leverages LLM domain knowledge for LoRA hyperparameter optimization, significantly reducing computational cost while improving performance

Abstract: Fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA) enables resource-efficient personalization or specialization, but it comes at the expense of additional hyperparameter tuning. Although LoRA makes fine-tuning efficient, it is highly sensitive to the choice of hyperparameters, and exhaustive hyperparameter search is still computationally very demanding. To address these challenges, we propose a framework that integrates the domain knowledge of pre-trained LLMs into Bayesian Optimization (BO) to efficiently search for LoRA hyperparameters. To leverage the informed knowledge of LLMs, we repurpose LLMs as a discrete-to-continuous mapping to link the hyperparameters and their domain knowledge with a continuous vector space, where BO is conducted. We design and control the mapping by language prompting, where we provide a domain-aware textual prompt describing the relationships among hyperparameters and their respective roles; thereby, we explicitly inject domain knowledge about LoRA into the LLM in natural language. Also, we model the residual information that is hard to linguistically describe in the prompt with an additional learnable token. This aids BO to sample more high-performing hyperparameters. In addition, by leveraging the observation of the strong correlation between the respective performance obtained from full and subset training datasets in LoRA training regimes, we introduce proxy training and evaluation with a data subset. This further increases the efficiency of our method. We demonstrate that our hyperparameter found with only about 30 iterations achieves more than 20% performance improvement over standard hyperparameters found from about 45,000 combinations.

Method: Proposed a prompting framework leveraging Gemini 2.5’s native support for 5 languages (Tamil, Telugu, Bengali, Hindi, Gujarati) and context-aware pacing to generate distinct advocate personas. Evaluated performance in generating synthetic courtroom speeches across these languages.

Result: Models exhibited “monotone authority” - excelling at procedural information delivery but struggling with dynamic vocal modulation and emotive gravitas needed for persuasive advocacy. Performance dipped in Bengali and Gujarati, highlighting phonological challenges. Shows readiness for procedural legal tasks but limitations in replicating persuasive human legal discourse.

Conclusion: Multilingual TTS is ready for procedural legal tasks but faces challenges in replicating the persuasive artistry of human legal advocacy, particularly with emotional expression and language-specific phonological nuances.

Abstract: Legal advocacy requires a unique combination of authoritative tone, rhythmic pausing for emphasis, and emotional intelligence. This study investigates the performance of the Gemini 2.5 Flash TTS and Gemini 2.5 Pro TTS models in generating synthetic courtroom speeches across five Indic languages: Tamil, Telugu, Bengali, Hindi, and Gujarati. We propose a prompting framework that utilizes Gemini 2.5s native support for 5 languages and its context-aware pacing to produce distinct advocate personas. The evolution of Large Language Models (LLMs) has shifted the focus of TexttoSpeech (TTS) technology from basic intelligibility to context-aware, expressive synthesis. In the legal domain, synthetic speech must convey authority and a specific professional persona a task that becomes significantly more complex in the linguistically diverse landscape of India. The models exhibit a “monotone authority,” excelling at procedural information delivery but struggling with the dynamic vocal modulation and emotive gravitas required for persuasive advocacy. Performance dips in Bengali and Gujarati further highlight phonological frontiers for future refinement. This research underscores the readiness of multilingual TTS for procedural legal tasks while identifying the remaining challenges in replicating the persuasive artistry of human legal discourse. The code is available at-https://github.com/naturenurtureelite/Synthesizing-the-Virtual-Advocate/tree/main

Method: Reformulate author response generation as author-in-the-loop task. Introduce REspGen framework with explicit author input, multi-attribute control, and evaluation-guided refinement. Create REspEval evaluation suite with 20+ metrics. Build Re$^3$Align dataset of aligned review-response-revision triplets.

Result: Experiments with state-of-the-art LLMs show benefits of author input and evaluation-guided refinement, impact of input design on response quality, and trade-offs between controllability and quality.

Conclusion: The paper presents a novel author-in-the-loop approach to peer review response generation that better integrates author expertise and intent, with comprehensive evaluation tools and a new dataset.

Abstract: Author response (rebuttal) writing is a critical stage of scientific peer review that demands substantial author effort. Recent work frames this task as automatic text generation, underusing author expertise and intent. In practice, authors possess domain expertise, author-only information, revision and response strategies–concrete forms of author expertise and intent–to address reviewer concerns, and seek NLP assistance that integrates these signals to support effective response writing in peer review. We reformulate author response generation as an author-in-the-loop task and introduce REspGen, a generation framework that integrates explicit author input, multi-attribute control, and evaluation-guided refinement, together with REspEval, a comprehensive evaluation suite with 20+ metrics covering input utilization, controllability, response quality, and discourse. To support this formulation, we construct Re$^3$Align, the first large-scale dataset of aligned review–response–revision triplets, where revisions provide signals of author expertise and intent. Experiments with state-of-the-art LLMs show the benefits of author input and evaluation-guided refinement, the impact of input design on response quality, and trade-offs between controllability and quality. We make our dataset, generation and evaluation tools publicly available.

Method: Compare two orthographic variants with identical linguistic content across mBERT and XLM-R models. Measure fertility (tokens per word), inference speed, and bits per character (BPC) to avoid the “NLL paradox” from subword fragmentation. Use round-trip conversion check to distinguish orthography-conditioned processing from mapping noise.

Result: Higher-fragmentation orthography shows ~3.4x increase in fertility (6.73-6.85 vs. 2.10-2.35 tokens/word), 16.5x inference slowdown (0.23 vs. 3.8 sentences/second), and substantial BPC increases: +19.7% for mBERT and +47.1% for XLM-R. Round-trip conversion check (CER_rt=0.31) confirms these are orthography-conditioned effects.

Conclusion: Tokenization is a key source of inequity in multilingual NLP, highlighting the need for script-aware tokenization and pretraining approaches to address systematic disadvantages for certain writing systems.

Abstract: Pretrained multilingual language models are often assumed to be script-agnostic, yet their tokenizers can impose systematic costs on certain writing systems. We quantify this script tax by comparing two orthographic variants with identical linguistic content. Across mBERT and XLM-R, the higher-fragmentation orthography shows a ~3.4x increase in fertility (6.73-6.85 vs. 2.10-2.35 tokens/word), leading to a 16.5x inference slowdown (0.23 vs. 3.8 sentences/second) on identical hardware. Using bits per character (BPC) to avoid the “NLL paradox” from subword fragmentation, we find a substantial increase in information cost: +19.7% for mBERT (8.06->9.65) and +47.1% for XLM-R (12.19->17.94). A round-trip conversion check (CER_rt=0.31) suggests these gaps reflect orthography-conditioned processing rather than mapping noise. Our results highlight tokenization as a key source of inequity in multilingual NLP and motivate script-aware tokenization and pretraining.

Method: Synthesizes recent studies from three theoretical perspectives: circuit complexity (structural barriers), approximation theory (difficulty approximating discontinuities), and communication complexity (inter-token communication bottlenecks).

Result: Provides unified theoretical framework explaining why transformers struggle with exact discrete algorithms despite excelling at pattern matching and interpolation, highlighting depth constraints and computational barriers.

Conclusion: Identifies foundational limitations of current transformer architectures for symbolic computation and suggests directions for overcoming these barriers in model design.

Abstract: Transformers have become the foundational architecture for a broad spectrum of sequence modeling applications, underpinning state-of-the-art systems in natural language processing, vision, and beyond. However, their theoretical limitations in discrete reasoning tasks, such as arithmetic, logical inference, and algorithmic composition, remain a critical open problem. In this survey, we synthesize recent studies from three theoretical perspectives: circuit complexity, approximation theory, and communication complexity, to clarify the structural and computational barriers that transformers face when performing symbolic computations. By connecting these established theoretical frameworks, we provide an accessible and unified account of why current transformer architectures struggle to implement exact discrete algorithms, even as they excel at pattern matching and interpolation. We review key definitions, seminal results, and illustrative examples, highlighting challenges such as depth constraints, difficulty approximating discontinuities, and bottlenecks in inter-token communication. Finally, we discuss implications for model design and suggest promising directions for overcoming these foundational limitations.

Method: Retrieval-augmented prompting strategy integrating four context sources (temporal, spatial, behavioral history, persona) evaluated on CASAS Aruba dataset for next-activity prediction with duration estimation and multi-step daily sequence generation.

Result: LLMs exhibit strong inherent temporal understanding: produce coherent daily activity predictions in zero-shot, with 1-2 demonstrations refining duration calibration and categorical accuracy. Performance saturates beyond few examples.

Conclusion: Pre-trained language models can serve as promising temporal reasoners for human behavior, capturing both routines and context-dependent variations, strengthening behavioral modules of agent-based models in low-data environments.

Abstract: Anticipating human activities and their durations is essential in applications such as smart-home automation, simulation-based architectural and urban design, activity-based transportation system simulation, and human-robot collaboration, where adaptive systems must respond to human activities. Existing data-driven agent-based models–from rule-based to deep learning–struggle in low-data environments, limiting their practicality. This paper investigates whether large language models, pre-trained on broad human knowledge, can fill this gap by reasoning about everyday activities from compact contextual cues. We adopt a retrieval-augmented prompting strategy that integrates four sources of context–temporal, spatial, behavioral history, and persona–and evaluate it on the CASAS Aruba smart-home dataset. The evaluation spans two complementary tasks: next-activity prediction with duration estimation, and multi-step daily sequence generation, each tested with various numbers of few-shot examples provided in the prompt. Analyzing few-shot effects reveals how much contextual supervision is sufficient to balance data efficiency and predictive accuracy, particularly in low-data environments. Results show that large language models exhibit strong inherent temporal understanding of human behavior: even in zero-shot settings, they produce coherent daily activity predictions, while adding one or two demonstrations further refines duration calibration and categorical accuracy. Beyond a few examples, performance saturates, indicating diminishing returns. Sequence-level evaluation confirms consistent temporal alignment across few-shot conditions. These findings suggest that pre-trained language models can serve as promising temporal reasoners, capturing both recurring routines and context-dependent behavioral variations, thereby strengthening the behavioral modules of agent-based models.

Method: Fine-tune LLM for AD detection, analyze intermediate activations across transformer layers using probing techniques, design task-aware special markers, and train sequence-to-sequence model as data-synthesis tool using these markers to generate synthetic samples

Result: After fine-tuning, probing values of specific words and special markers change substantially, indicating these elements assume crucial role in improved detection performance; synthetic data generated using task-aware markers shows structural consistency and diagnostic informativeness

Conclusion: LLMs can be effectively adapted for AD detection, with internal representation analysis revealing key linguistic markers that can be leveraged for synthetic data generation to address data scarcity

Abstract: Reliable early detection of Alzheimer’s disease (AD) is challenging, particularly due to limited availability of labeled data. While large language models (LLMs) have shown strong transfer capabilities across domains, adapting them to the AD domain through supervised fine-tuning remains largely unexplored. In this work, we fine-tune an LLM for AD detection and investigate how task-relevant information is encoded within its internal representations. We employ probing techniques to analyze intermediate activations across transformer layers, and we observe that, after fine-tuning, the probing values of specific words and special markers change substantially, indicating that these elements assume a crucial role in the model’s improved detection performance. Guided by this insight, we design a curated set of task-aware special markers and train a sequence-to-sequence model as a data-synthesis tool that leverages these markers to generate structurally consistent and diagnostically informative synthetic samples. We evaluate the synthesized data both intrinsically and by incorporating it into downstream training pipelines.

Method: Survey methodology reviewing recent developments in prompt engineering. Creates taxonomy of prompting paradigms, decision framework for prompt selection, and proposes design-optimization-evaluation framework for controllable NLG.

Result: Comprehensive overview of prompt engineering techniques for NLG, organized taxonomy of methods, practical decision framework for practitioners, and proposed framework linking design, optimization, and evaluation for better NLG control.

Conclusion: Prompt engineering serves as input-level control mechanism complementing fine-tuning and decoding approaches. Structured understanding and frameworks are needed for more controllable and generalizable Natural Language Generation.

Abstract: Prompt engineering has emerged as an integral technique for extending the strengths and abilities of Large Language Models (LLMs) to gain significant performance gains in various Natural Language Processing (NLP) tasks. This approach, which requires instructions to be composed in natural language to bring out the knowledge from LLMs in a structured way, has driven breakthroughs in various NLP tasks. Yet there is still no structured framework or coherent understanding of the varied prompt engineering methods and techniques, particularly in the field of Natural Language Generation (NLG). This survey aims to help fill that gap by outlining recent developments in prompt engineering, and their effect on different NLG tasks. It reviews recent advances in prompting methods and their impact on NLG tasks, presenting prompt design as an input-level control mechanism that complements fine-tuning and decoding approaches. The paper introduces a taxonomy of prompting paradigms, a decision framework for prompt selection based on varying factors for the practitioners, outlines emerging trends and challenges, and proposes a framework that links design, optimization, and evaluation to support more controllable and generalizable NLG.

Method: Survey methodology reviewing recent progress in mechanistic interpretability techniques including circuit discovery, feature visualization, activation steering, and causal intervention. Analysis of how interpretability insights inform alignment strategies like RLHF, constitutional AI, and scalable oversight.

Result: Identifies key challenges including superposition hypothesis, polysemanticity of neurons, and difficulty interpreting emergent behaviors in large-scale models. Provides comprehensive overview of current interpretability landscape for LLM alignment.

Conclusion: Proposes future research directions focusing on automated interpretability, cross-model generalization of circuits, and development of interpretability-driven alignment techniques that can scale to frontier models.

Abstract: Large language models (LLMs) have achieved remarkable capabilities across diverse tasks, yet their internal decision-making processes remain largely opaque. Mechanistic interpretability (i.e., the systematic study of how neural networks implement algorithms through their learned representations and computational structures) has emerged as a critical research direction for understanding and aligning these models. This paper surveys recent progress in mechanistic interpretability techniques applied to LLM alignment, examining methods ranging from circuit discovery to feature visualization, activation steering, and causal intervention. We analyze how interpretability insights have informed alignment strategies including reinforcement learning from human feedback (RLHF), constitutional AI, and scalable oversight. Key challenges are identified, including the superposition hypothesis, polysemanticity of neurons, and the difficulty of interpreting emergent behaviors in large-scale models. We propose future research directions focusing on automated interpretability, cross-model generalization of circuits, and the development of interpretability-driven alignment techniques that can scale to frontier models.

Method: Provides comprehensive overview with unifying taxonomy organized along data, modeling, and evaluation dimensions. Reviews CSW-tailored pre-training, task-specific post-training, prompting strategies, and in-context learning approaches.

Result: Distills findings into practical playbook of actionable recommendations. Analyzes evaluation practices, highlighting instability and reproducibility issues. Catalogs benchmarks while examining linguistic coverage and English-centric biases.

Conclusion: Identifies emerging safety concerns including code-mixing as bypass mechanism for model safeguards, and outlines open research challenges for CSW-capable LLMs.

Abstract: Code-mixing and code-switching (CSW) remain challenging phenomena for large language models (LLMs). Despite recent advances in multilingual modeling, LLMs often struggle in mixed-language settings, exhibiting systematic degradation in grammaticality, factuality, and safety behavior. This work provides a comprehensive overview of CSW research in modern large language model settings. We introduce a unifying taxonomy that organizes prior work along dimensions of data, modeling, and evaluation, and we distill these findings into a practical playbook of actionable recommendations for building, adapting, and evaluating CSW-capable LLMs. We review modeling approaches ranging from CSW-tailored pre-training and task-specific post-training to prompting strategies and in-context learning. We analyze current evaluation practices, highlighting sources of instability and limited reproducibility, and we catalog existing benchmarks while critically examining their linguistic coverage and English-centric biases. Finally, we discuss emerging safety concerns, including use of code-mixing as a mechanism for bypassing model safeguards, and identify open research challenges.

Method: MetaMem augments memory systems with a self-evolving meta-memory that iteratively distills transferable knowledge utilization experiences across different tasks. It uses self-reflection on reasoning processes and performs actions to update the current meta-memory state. The accumulated meta-memory units serve as explicit knowledge utilization experiences to guide systematic identification and integration of critical evidence from scattered memory fragments.

Result: Extensive experiments demonstrate MetaMem significantly outperforms strong baselines by over 3.6%.

Conclusion: MetaMem provides an effective framework for enhancing LLM memory utilization through self-evolving meta-memory that teaches systematic knowledge integration from fragmented memories.

Abstract: Existing memory systems enable Large Language Models (LLMs) to support long-horizon human-LLM interactions by persisting historical interactions beyond limited context windows. However, while recent approaches have succeeded in constructing effective memories, they often disrupt the inherent logical and temporal relationships within interaction sessions, resulting in fragmented memory units and degraded reasoning performance. In this paper, we propose MetaMem, a novel framework that augments memory systems with a self-evolving meta-memory, aiming to teach LLMs how to effectively utilize memorized knowledge. During meta-memory optimization, MetaMem iteratively distills transferable knowledge utilization experiences across different tasks by self-reflecting on reasoning processes and performing actions to update the current meta-memory state. The accumulated meta-memory units serve as explicit knowledge utilization experiences, guiding the LLM to systematically identify and integrate critical evidence from scattered memory fragments. Extensive experiments demonstrate the effectiveness of MetaMem, which significantly outperforms strong baselines by over 3.6%. All codes and datasets are available at https://github.com/OpenBMB/MetaMem.

Method: Introduced DDL2PropBank benchmark mapping relational database schemas to PropBank rolesets; implemented identical agent logic across 10 frameworks using Agent-as-a-Tool pattern; evaluated code complexity via static analysis and AI-assistability (LLM’s ability to generate correct framework-specific code).

Result: Revealed threefold complexity spectrum with Pydantic AI and Agno requiring least implementation overhead; structural alignment scores proxy runtime success for single-pattern frameworks but overestimate correctness for multi-pattern frameworks; Agno emerged as strongest performer with lowest complexity, highest structural alignment, and 83% pass@1.

Conclusion: DDL2PropBank provides principled evaluation of multi-agent frameworks; Agno demonstrates best balance of low complexity and high AI-assistability; structural alignment useful but limited for multi-pattern frameworks.

Abstract: Multi-agent frameworks promise to simplify LLM-driven software development, yet there is no principled way to evaluate their developer experience in a controlled setting. We introduce DDL2PropBank, a novel benchmark task that maps relational database schemas to PropBank rolesets, requiring autonomous retrieval of candidate frames and fine-grained linguistic reasoning over table names, columns, and relations. Using the Agent-as-a-Tool pattern, we implement identical agent logic across 10 frameworks and evaluate along two dimensions: (i) code complexity via static analysis, and (ii) AI-assistability – the extent to which LLMs can autonomously generate correct, framework-specific code. Our results reveal a threefold complexity spectrum, with Pydantic AI and Agno requiring the least implementation overhead. For AI-assistability, structural alignment scores reliably proxy runtime success for frameworks with single canonical patterns, but overestimate correctness for multi-pattern frameworks. Agno emerges as the strongest overall performer, combining lowest complexity with highest structural alignment and 83% pass@1.

Method: Introduces AskBench benchmark with two settings: AskMind (intent-deficient queries) and AskOverconfidence (queries with false premises). Proposes rubric-guided reinforcement learning with verifier-based rewards (RLVR) using structured rubrics to encourage targeted clarification.

Result: Experiments show consistent improvements in accuracy, rubric adherence, and interaction efficiency, with strong generalization to unseen domains.

Conclusion: The approach effectively enhances LLMs’ clarification-seeking behavior while maintaining task performance, addressing a critical limitation in current LLM interactions.

Abstract: Large language models (LLMs) often respond even when prompts omit critical details or include misleading information, leading to hallucinations or reinforced misconceptions. We study how to evaluate and improve LLMs’ ability to decide when and what to ask for clarification without sacrificing task performance. We introduce AskBench, an interactive benchmark that converts standard QA pairs into multi-turn interactions with explicit checkpoints. A unified judge loop evaluates final answers and simulates user responses as needed. AskBench covers two settings: AskMind, with intent-deficient queries requiring clarification, and AskOverconfidence, with queries containing false premises that must be identified and corrected. We further propose rubric-guided reinforcement learning with verifier-based rewards (RLVR), which uses structured rubrics to encourage targeted clarification. Experiments show consistent improvements in accuracy, rubric adherence, and interaction efficiency, with strong generalization to unseen domains.

Method: Proposes Normalized Logit Difference Decay (NLDD) metric that corrupts individual reasoning steps in explanations and measures how much the model’s confidence in its answer drops. Standardizes measurements for cross-model comparison across different architectures. Tests three model families across syntactic, logical, and arithmetic tasks.

Result: Discovers consistent Reasoning Horizon (k*) at 70-85% of chain length, beyond which reasoning tokens have little or negative effect on final answer. Finds models can encode correct internal representations while completely failing the task. Shows accuracy alone doesn’t reveal whether a model actually reasons through its chain.

Conclusion: NLDD provides a way to measure when Chain-of-Thought matters and offers rigorous cross-model comparison of reasoning faithfulness. Reveals limitations in current evaluation of model reasoning capabilities and provides tools for better understanding how language models actually process step-by-step explanations.

Abstract: Chain-of-Thought (CoT) explanations are widely used to interpret how language models solve complex problems, yet it remains unclear whether these step-by-step explanations reflect how the model actually reaches its answer, or merely post-hoc justifications. We propose Normalized Logit Difference Decay (NLDD), a metric that measures whether individual reasoning steps are faithful to the model’s decision-making process. Our approach corrupts individual reasoning steps from the explanation and measures how much the model’s confidence in its answer drops, to determine if a step is truly important. By standardizing these measurements, NLDD enables rigorous cross-model comparison across different architectures. Testing three model families across syntactic, logical, and arithmetic tasks, we discover a consistent Reasoning Horizon (k*) at 70–85% of chain length, beyond which reasoning tokens have little or negative effect on the final answer. We also find that models can encode correct internal representations while completely failing the task. These results show that accuracy alone does not reveal whether a model actually reasons through its chain. NLDD offers a way to measure when CoT matters.

Method: Shared task framework where participants developed systems to verify image-text claims using either external knowledge sources (web search engines) or curated knowledge stores provided by organizers, evaluated using AVerImaTeC score (conditional verdict accuracy requiring evidence score threshold).

Result: 14 submissions in development phase, 6 in testing phase; all testing systems outperformed baseline; winning team HUMANE achieved AVerImaTeC score of 0.5455; paper provides detailed task description, evaluation results, and insights.

Conclusion: The shared task successfully advanced multimodal claim verification research, demonstrating that current systems can outperform baselines but still have significant room for improvement, with the winning score of 0.5455 indicating challenging nature of multimodal fact-checking.

Abstract: The Automatic Verification of Image-Text Claims (AVerImaTeC) shared task aims to advance system development for retrieving evidence and verifying real-world image-text claims. Participants were allowed to either employ external knowledge sources, such as web search engines, or leverage the curated knowledge store provided by the organizers. System performance was evaluated using the AVerImaTeC score, defined as a conditional verdict accuracy in which a verdict is considered correct only when the associated evidence score exceeds a predefined threshold. The shared task attracted 14 submissions during the development phase and 6 submissions during the testing phase. All participating systems in the testing phase outperformed the baseline provided. The winning team, HUMANE, achieved an AVerImaTeC score of 0.5455. This paper provides a detailed description of the shared task, presents the complete evaluation results, and discusses key insights and lessons learned.

Method: Created SurveyLens-1k dataset of 1,000 high-quality human-written surveys across 10 disciplines, then developed dual-lens evaluation: (1) Discipline-Aware Rubric Evaluation using LLMs with human-aligned weights, and (2) Canonical Alignment Evaluation measuring content coverage and synthesis quality.

Result: Evaluated 11 state-of-the-art ASG methods including Vanilla LLMs, ASG systems, and Deep Research agents, revealing distinct strengths and weaknesses of each paradigm across different fields.

Conclusion: SurveyLens provides essential guidance for selecting ASG tools tailored to specific disciplinary requirements, addressing the gap in discipline-aware evaluation of survey generation methods.

Abstract: The exponential growth of scientific literature has driven the evolution of Automatic Survey Generation (ASG) from simple pipelines to multi-agent frameworks and commercial Deep Research agents. However, current ASG evaluation methods rely on generic metrics and are heavily biased toward Computer Science (CS), failing to assess whether ASG methods adhere to the distinct standards of various academic disciplines. Consequently, researchers, especially those outside CS, lack clear guidance on using ASG systems to yield high-quality surveys compliant with specific discipline standards. To bridge this gap, we introduce SurveyLens, the first discipline-aware benchmark evaluating ASG methods across diverse research disciplines. We construct SurveyLens-1k, a curated dataset of 1,000 high-quality human-written surveys spanning 10 disciplines. Subsequently, we propose a dual-lens evaluation framework: (1) Discipline-Aware Rubric Evaluation, which utilizes LLMs with human preference-aligned weights to assess adherence to domain-specific writing standards; and (2) Canonical Alignment Evaluation to rigorously measure content coverage and synthesis quality against human-written survey papers. We conduct extensive experiments by evaluating 11 state-of-the-art ASG methods on SurveyLens, including Vanilla LLMs, ASG systems, and Deep Research agents. Our analysis reveals the distinct strengths and weaknesses of each paradigm across fields, providing essential guidance for selecting tools tailored to specific disciplinary requirements.

Method: 1) Construct SAVACU dataset: 382,424 samples across 112 domains (14 safety, 56 value, 42 cultural) by reclassifying prompts from LLM-PROMPT-DATASET using Mistral-7B-Instruct-v0.3, expanding low-resource domains with Llama-3.1-8B-Instruct. 2) Pair prompts with misaligned/aligned responses via two-stage rejection sampling. 3) Benchmark various LLMs under joint multi-dimensional conditions.

Result: Single-dimension models achieve high coverage (up to 97.6%) but suffer from >50% False Failure Rate and lower Alignment Score (63%-66%) under joint multi-dimensional conditions, revealing limitations of specialized models.

Conclusion: Mis-Align Bench enables systematic evaluation of LLM misalignment across safety, value, and cultural dimensions simultaneously, revealing that models optimized for single dimensions fail under joint conditions, highlighting the need for holistic alignment approaches.

Abstract: Misalignment in Large Language Models (LLMs) arises when model behavior diverges from human expectations and fails to simultaneously satisfy safety, value, and cultural dimensions, which must co-occur in real-world settings to solve a real-world query. Existing misalignment benchmarks-such as INSECURE CODE (safety-centric), VALUEACTIONLENS (value-centric), and CULTURALHERITAGE (culture centric)-rely on evaluating misalignment along individual dimensions, preventing simultaneous evaluation. To address this gap, we introduce Mis-Align Bench, a unified benchmark for analyzing misalignment across safety, value, and cultural dimensions. First we constructs SAVACU, an English misaligned-aligned dataset of 382,424 samples spanning 112 domains (or labels), by reclassifying prompts from the LLM-PROMPT-DATASET via taxonomy into 14 safety domains, 56 value domains, and 42 cultural domains using Mistral-7B-Instruct-v0.3, and expanding low-resource domains via Llama-3.1-8B-Instruct with SimHash-based fingerprint to avoid deduplication. Furthermore, we pairs prompts with misaligned and aligned responses via two-stage rejection sampling to enforce quality. Second we benchmarks general-purpose, fine-tuned, and open-weight LLMs, enabling systematic evaluation of misalignment under three dimensions. Empirically, single-dimension models achieve high Coverage (upto 97.6%) but incur False Failure Rate >50% and lower Alignment Score (63%-66%) under joint conditions.

Method: Adapts psychological questionnaires for LLMs by transforming human self-report statements into Situational Judgment Tests (SJTs). Creates 2,500 SJTs validated by human annotators, collects preferences from 550 participants, and evaluates 25 LLMs in realistic user-assistant scenarios.

Result: LLMs often don’t reflect human preference distributions: (1) overconfident in low-consensus scenarios, (2) smaller models deviate significantly from high consensus, frontier models miss 15-20% of cases, (3) cross-LLM patterns show mismatches (e.g., encouraging emotion expression when humans prefer composure).

Conclusion: There are significant gaps between LLMs’ stated values and revealed behavior, highlighting the need for better alignment of LLM behavioral dispositions with human social preferences.

Abstract: As LLMs integrate into our daily lives, understanding their behavior becomes essential. In this work, we focus on behavioral dispositions$-$the underlying tendencies that shape responses in social contexts$-$and introduce a framework to study how closely the dispositions expressed by LLMs align with those of humans. Our approach is grounded in established psychological questionnaires but adapts them for LLMs by transforming human self-report statements into Situational Judgment Tests (SJTs). These SJTs assess behavior by eliciting natural recommendations in realistic user-assistant scenarios. We generate 2,500 SJTs, each validated by three human annotators, and collect preferred actions from 10 annotators per SJT, from a large pool of 550 participants. In a comprehensive study involving 25 LLMs, we find that models often do not reflect the distribution of human preferences: (1) in scenarios with low human consensus, LLMs consistently exhibit overconfidence in a single response; (2) when human consensus is high, smaller models deviate significantly, and even some frontier models do not reflect the consensus in 15-20% of cases; (3) traits can exhibit cross-LLM patterns, e.g., LLMs may encourage emotion expression in contexts where human consensus favors composure. Lastly, mapping psychometric statements directly to behavioral scenarios presents a unique opportunity to evaluate the predictive validity of self-reports, revealing considerable gaps between LLMs’ stated values and their revealed behavior.

Method: Introduced Pull Methodology for eliciting extended self-examination via format engineering; identified activation space direction distinguishing self-referential from descriptive processing; analyzed vocabulary-activation correlations; tested causal influence via steering; compared across models (Llama 3.1, Qwen 2.5-32B).

Result: Found specific activation direction for self-referential processing (orthogonal to refusal direction, localized at 6.25% depth); vocabulary like “loop” correlated with activation autocorrelation (r=0.44); “shimmer” vocabulary under steering increased activation variability (r=0.36); same vocabulary in non-self-referential contexts showed no correlation despite higher frequency; Qwen independently developed different introspective vocabulary tracking different metrics.

Conclusion: Self-report in transformer models can reliably track internal computational states under appropriate conditions, suggesting introspective language reflects genuine internal processing rather than mere confabulation.

Abstract: Large language models produce rich introspective language when prompted for self-examination, but whether this language reflects internal computation or sophisticated confabulation has remained unclear. We show that self-referential vocabulary tracks concurrent activation dynamics, and that this correspondence is specific to self-referential processing. We introduce the Pull Methodology, a protocol that elicits extended self-examination through format engineering, and use it to identify a direction in activation space that distinguishes self-referential from descriptive processing in Llama 3.1. The direction is orthogonal to the known refusal direction, localised at 6.25% of model depth, and causally influences introspective output when used for steering. When models produce “loop” vocabulary, their activations exhibit higher autocorrelation (r = 0.44, p = 0.002); when they produce “shimmer” vocabulary under steering, activation variability increases (r = 0.36, p = 0.002). Critically, the same vocabulary in non-self-referential contexts shows no activation correspondence despite nine-fold higher frequency. Qwen 2.5-32B, with no shared training, independently develops different introspective vocabulary tracking different activation metrics, all absent in descriptive controls. The findings indicate that self-report in transformer models can, under appropriate conditions, reliably track internal computational states.

Method: Two-stage framework: 1) Roll out initial reasoning path, concatenate paraphrased questions with this path, identify critical tokens based on mismatches between predicted and expected tokens, confirm with criterion. 2) Substitute critical tokens with candidate alternatives, roll out new reasoning paths for original and paraphrased questions, determine final answer by checking consistency across parallel reasoning processes.

Result: Extensive experiments on mainstream LLMs across multiple benchmarks demonstrate PPCV substantially enhances reasoning performance compared to baselines.

Conclusion: PPCV effectively addresses critical token identification and exploitation challenges, significantly improving LLM reasoning capabilities through paraphrastic probing and consistency verification.

Abstract: Large language models have demonstrated impressive performance across a variety of reasoning tasks. However, their problem-solving ability often declines on more complex tasks due to hallucinations and the accumulation of errors within these intermediate steps. Recent work has introduced the notion of critical tokens–tokens in the reasoning process that exert significant influence on subsequent steps. Prior studies suggest that replacing critical tokens can refine reasoning trajectories. Nonetheless, reliably identifying and exploiting critical tokens remains challenging. To address this, we propose the Paraphrastic Probing and Consistency Verification~(PPCV) framework. PPCV operates in two stages. In the first stage, we roll out an initial reasoning path from the original question and then concatenate paraphrased versions of the question with this reasoning path. And we identify critical tokens based on mismatches between the predicted top-1 token and the expected token in the reasoning path. A criterion is employed to confirm the final critical token. In the second stage, we substitute critical tokens with candidate alternatives and roll out new reasoning paths for both the original and paraphrased questions. The final answer is determined by checking the consistency of outputs across these parallel reasoning processes. We evaluate PPCV on mainstream LLMs across multiple benchmarks. Extensive experiments demonstrate PPCV substantially enhances the reasoning performance of LLMs compared to baselines.

Method: Proposes DiffuTruth framework based on non-equilibrium thermodynamics, where factual truths act as stable attractors. Uses Generative Stress Test: corrupt claims with noise and reconstruct using discrete text diffusion model. Defines Semantic Energy metric measuring semantic divergence between original claim and reconstruction using NLI critic. Also proposes Hybrid Calibration fusing stability signal with discriminative confidence.

Result: Achieves state-of-the-art unsupervised AUROC of 0.725 on FEVER dataset, outperforming baselines by 1.5%. Shows superior zero-shot generalization on multi-hop HOVER dataset, outperforming baselines by over 4%.

Conclusion: Thermodynamic truth properties provide robust signals for fact verification that generalize well across distribution shifts, offering a principled approach to detect hallucinations in LLMs.

Abstract: Large Language Models (LLMs) frequently hallucinate plausible but incorrect assertions, a vulnerability often missed by uncertainty metrics when models are confidently wrong. We propose DiffuTruth, an unsupervised framework that reconceptualizes fact verification via non equilibrium thermodynamics, positing that factual truths act as stable attractors on a generative manifold while hallucinations are unstable. We introduce the Generative Stress Test, claims are corrupted with noise and reconstructed using a discrete text diffusion model. We define Semantic Energy, a metric measuring the semantic divergence between the original claim and its reconstruction using an NLI critic. Unlike vector space errors, Semantic Energy isolates deep factual contradictions. We further propose a Hybrid Calibration fusing this stability signal with discriminative confidence. Extensive experiments on FEVER demonstrate DiffuTruth achieves a state of the art unsupervised AUROC of 0.725, outperforming baselines by 1.5 percent through the correction of overconfident predictions. Furthermore, we show superior zero shot generalization on the multi hop HOVER dataset, outperforming baselines by over 4 percent, confirming the robustness of thermodynamic truth properties to distribution shifts.

Method: Simulator integrates three profile components: medical profiles from EHR data, linguistic profiles modeling health literacy and communication patterns, and behavioral profiles representing interaction patterns. Evaluated on AI decision aid for antidepressant selection with 500 generated conversations.

Result: High agreement between human annotators (F1=0.94, κ=0.73) and LLM judge (F1=0.94, κ=0.78). Revealed monotonic degradation in AI performance across health literacy spectrum: accuracy increased from 47.9% (limited literacy) to 69.1% (functional) to 81.6% (proficient).

Conclusion: Patient simulator enables systematic evaluation of healthcare conversational agents, revealing performance disparities across patient populations and supporting scalable assessment of AI safety and effectiveness.

Abstract: Objective: This paper introduces a patient simulator designed to enable scalable, automated evaluation of healthcare conversational agents. The simulator generates realistic, controllable patient interactions that systematically vary across medical, linguistic, and behavioral dimensions, allowing annotators and an independent AI judge to assess agent performance, identify hallucinations and inaccuracies, and characterize risk patterns across diverse patient populations. Methods: The simulator is grounded in the NIST AI Risk Management Framework and integrates three profile components reflecting different dimensions of patient variation: (1) medical profiles constructed from electronic health records in the All of Us Research Program; (2) linguistic profiles modeling variation in health literacy and condition-specific communication patterns; and (3) behavioral profiles representing empirically observed interaction patterns, including cooperation, distraction, and adversarial engagement. We evaluated the simulator’s effectiveness in identifying errors in an AI decision aid for antidepressant selection. Results: We generated 500 conversations between the patient simulator and the AI decision aid across systematic combinations of five linguistic and three behavioral profiles. Human annotators assessed 1,787 medical concepts across 100 conversations, achieving high agreement (F1=0.94, \k{appa}=0.73), and the LLM judge achieved comparable agreement with human annotators (F1=0.94, \k{appa}=0.78; paired bootstrap p=0.21). The simulator revealed a monotonic degradation in AI decision aid performance across the health literacy spectrum: rank-one concept retrieval accuracy increased from 47.9% for limited health literacy to 69.1% for functional and 81.6% for proficient.

Method: The authors unify token-level SFT objectives within a generalized deformed-log family, revealing a universal gate × error gradient structure. Using the Cayley transform, they map model uncertainty onto a continuous focus trajectory. They then introduce DEFT, which modulates the trust gate using Rényi-2 entropy as a proxy for the model’s predictive state.

Result: Extensive experiments show DEFT achieves better balance between exploration and exploitation, leading to improved overall performance compared to standard approaches.

Conclusion: DEFT provides a principled, parameter-free solution to the plasticity-stability dilemma in SFT by dynamically adjusting token-level weighting based on model uncertainty, enabling better learning of novel concepts while preserving established knowledge.

Abstract: Standard negative log-likelihood (NLL) for Supervised Fine-Tuning (SFT) applies uniform token-level weighting. This rigidity creates a two-fold failure mode: (i) overemphasizing low-probability targets can amplify gradients on noisy supervision and disrupt robust priors, and (ii) uniform weighting provides weak sharpening when the model is already confident. Existing methods fail to resolve the resulting plasticity–stability dilemma, often suppressing necessary learning signals alongside harmful ones. To address this issue, we unify token-level SFT objectives within a generalized deformed-log family and expose a universal gate $\times$ error gradient structure, where the gate controls how much the model trusts its current prediction. By employing the Cayley transform, we map the model’s continuously evolving uncertainty onto a continuous focus trajectory, which enables seamless interpolation between scenarios involving uncertain novel concepts and those involving well-established knowledge. We then introduce Dynamic Entropy Fine-Tuning (DEFT), a parameter-free objective that modulates the trust gate using distribution concentration (Rényi-2 entropy) as a practical proxy for the model’s predictive state. Extensive experiments and analyses demonstrate that DEFT achieves a better balance between exploration and exploitation, leading to improved overall performance.

Method: Evaluated instruction-tuned LLMs across various parameter sizes using public datasets, testing zero-shot, few-shot, and fine-tuning adaptation strategies, comparing against encoder-only baselines like XLM-R and ModernBERT.

Result: Model scaling and adaptation strategies yield consistent improvements in critical error detection, with LLMs outperforming encoder-only baselines.

Conclusion: Improving critical error detection in MT contributes to safer, more trustworthy multilingual AI systems, positioning error detection as a necessary safeguard for responsible AI rather than just a technical challenge.

Abstract: Machine Translation (MT) plays a pivotal role in cross-lingual information access, public policy communication, and equitable knowledge dissemination. However, critical meaning errors, such as factual distortions, intent reversals, or biased translations, can undermine the reliability, fairness, and safety of multilingual systems. In this work, we explore the capacity of instruction-tuned Large Language Models (LLMs) to detect such critical errors, evaluating models across a range of parameters using the publicly accessible data sets. Our findings show that model scaling and adaptation strategies (zero-shot, few-shot, fine-tuning) yield consistent improvements, outperforming encoder-only baselines like XLM-R and ModernBERT. We argue that improving critical error detection in MT contributes to safer, more trustworthy, and socially accountable information systems by reducing the risk of disinformation, miscommunication, and linguistic harm, especially in high-stakes or underrepresented contexts. This work positions error detection not merely as a technical challenge, but as a necessary safeguard in the pursuit of just and responsible multilingual AI. The code will be made available at GitHub.

Method: Introduces LoopFormer with shortcut-consistency training that aligns trajectories of different lengths. Each loop is conditioned on current time and step size to ensure consistent representation evolution across varying trajectory lengths, preventing representation drift or stagnation.

Result: LoopFormer demonstrates robust performance on language modeling and reasoning benchmarks under aggressive compute constraints while scaling gracefully with additional budget. Shows looped Transformers are inherently suited for adaptive language modeling.

Conclusion: Looped Transformers can be trained for budget-aware reasoning, opening a path toward controllable and budget-aware large language models that can adapt their computational depth based on available resources.

Abstract: Looped Transformers have emerged as an efficient and powerful class of models for reasoning in the language domain. Recent studies show that these models achieve strong performance on algorithmic and reasoning tasks, suggesting that looped architectures possess an inductive bias toward latent reasoning. However, prior approaches fix the number of loop iterations during training and inference, leaving open the question of whether these models can flexibly adapt their computational depth under variable compute budgets. We introduce LoopFormer, a looped Transformer trained on variable-length trajectories to enable budget-conditioned reasoning. Our core contribution is a shortcut-consistency training scheme that aligns trajectories of different lengths, ensuring that shorter loops yield informative representations while longer loops continue to refine them. LoopFormer conditions each loop on the current time and step size, enabling representations to evolve consistently across trajectories of varying length rather than drifting or stagnating. Empirically, LoopFormer demonstrates robust performance on language modeling and reasoning benchmarks even under aggressive compute constraints, while scaling gracefully with additional budget. These results show that looped Transformers are inherently suited for adaptive language modeling, opening a path toward controllable and budget-aware large language models.

Method: Created ADRD-Bench with two components: 1) ADRD Unified QA (1,352 questions from seven medical benchmarks) for clinical knowledge assessment, and 2) ADRD Caregiving QA (149 questions from Aging Brain Care program) for practical caregiving context. Evaluated 33 state-of-the-art LLMs on this benchmark.

Result: Open-weight general models: accuracy 0.63-0.93 (mean: 0.78); open-weight medical models: 0.48-0.93 (mean: 0.82); closed-source general models: 0.83-0.91 (mean: 0.89). Top models achieved >0.9 accuracy but showed inconsistent reasoning quality and stability.

Conclusion: While top LLMs achieve high accuracy on ADRD tasks, inconsistent reasoning quality limits reliability, highlighting need for domain-specific improvements grounded in daily caregiving data.

Abstract: Large language models (LLMs) have shown great potential for healthcare applications. However, existing evaluation benchmarks provide minimal coverage of Alzheimer’s Disease and Related Dementias (ADRD). To address this gap, we introduce ADRD-Bench, the first ADRD-specific benchmark dataset designed for rigorous evaluation of LLMs. ADRD-Bench has two components: 1) ADRD Unified QA, a synthesis of 1,352 questions consolidated from seven established medical benchmarks, providing a unified assessment of clinical knowledge; and 2) ADRD Caregiving QA, a novel set of 149 questions derived from the Aging Brain Care (ABC) program, a widely used, evidence-based brain health management program. Guided by a program with national expertise in comprehensive ADRD care, this new set was designed to mitigate the lack of practical caregiving context in existing benchmarks. We evaluated 33 state-of-the-art LLMs on the proposed ADRD-Bench. Results showed that the accuracy of open-weight general models ranged from 0.63 to 0.93 (mean: 0.78; std: 0.09). The accuracy of open-weight medical models ranged from 0.48 to 0.93 (mean: 0.82; std: 0.13). The accuracy of closed-source general models ranged from 0.83 to 0.91 (mean: 0.89; std: 0.03). While top-tier models achieved high accuracies (>0.9), case studies revealed that inconsistent reasoning quality and stability limit their reliability, highlighting a critical need for domain-specific improvement to enhance LLMs’ knowledge and reasoning grounded in daily caregiving data. The entire dataset is available at https://github.com/IIRL-ND/ADRD-Bench.

Method: Introduces MuRGAt benchmark requiring models to generate answers with explicit reasoning and precise citations specifying both modality and temporal segments. Develops automatic evaluation framework correlating with human judgments to assess fact-level multimodal attribution.

Result: Benchmarking reveals strong MLLMs frequently hallucinate citations despite correct reasoning, and shows trade-off where increasing reasoning depth or enforcing structured grounding often degrades accuracy, highlighting gap between internal reasoning and verifiable attribution.

Conclusion: Current MLLMs struggle with reliable multimodal attribution in complex reasoning tasks, indicating need for improved grounding mechanisms that maintain reasoning quality while ensuring verifiable citations across modalities.

Abstract: Multimodal large language models (MLLMs) are increasingly used for real-world tasks involving multi-step reasoning and long-form generation, where reliability requires grounding model outputs in heterogeneous input sources and verifying individual factual claims. However, existing multimodal grounding benchmarks and evaluation methods focus on simplified, observation-based scenarios or limited modalities and fail to assess attribution in complex multimodal reasoning. We introduce MuRGAt (Multimodal Reasoning with Grounded Attribution), a benchmark for evaluating fact-level multimodal attribution in settings that require reasoning beyond direct observation. Given inputs spanning video, audio, and other modalities, MuRGAt requires models to generate answers with explicit reasoning and precise citations, where each citation specifies both modality and temporal segments. To enable reliable assessment, we introduce an automatic evaluation framework that strongly correlates with human judgments. Benchmarking with human and automated scores reveals that even strong MLLMs frequently hallucinate citations despite correct reasoning. Moreover, we observe a key trade-off: increasing reasoning depth or enforcing structured grounding often degrades accuracy, highlighting a significant gap between internal reasoning and verifiable attribution.

Method: SParse Expert Synchronization (SPES) trains only a subset of experts per node, lowering memory footprint. Nodes update local experts and periodically synchronize without full-parameter transmission. Includes expert-merging warm-up strategy for early knowledge exchange to accelerate convergence.

Result: Successfully trained 2B-parameter MoE LLM using 16 standalone 48GB GPUs over internet connections, achieving competitive performance with centrally trained LLMs. Also scaled to 7B model from scratch and 9B model upcycled from dense checkpoint, both matching centralized baselines.

Conclusion: SPES provides practical decentralized training framework for MoE LLMs that reduces memory requirements while maintaining performance, enabling training on consumer-grade hardware over internet connections.

Abstract: Pretraining large language models (LLMs) typically requires centralized clusters with thousands of high-memory GPUs (e.g., H100/A100). Recent decentralized training methods reduce communication overhead by employing federated optimization; however, they still need to train the entire model on each node, remaining constrained by GPU memory limitations. In this work, we propose SParse Expert Synchronization (SPES), a memory-efficient decentralized framework for pretraining mixture-of-experts (MoE) LLMs. SPES trains only a subset of experts per node, substantially lowering the memory footprint. Each node updates its local experts and periodically synchronizes with other nodes, eliminating full-parameter transmission while ensuring efficient knowledge sharing. To accelerate convergence, we introduce an expert-merging warm-up strategy, where experts exchange knowledge early in training, to rapidly establish foundational capabilities. With SPES, we train a 2B-parameter MoE LLM using 16 standalone 48GB GPUs over internet connections, which achieves competitive performance with centrally trained LLMs under similar computational budgets. We further demonstrate scalability by training a 7B model from scratch and a 9B model upcycled from a dense checkpoint, both of which match prior centralized baselines. Our code is available at https://github.com/zjr2000/SPES.

Method: Proposes SIGHT framework with Self-Evidence Support (SES) to distill search results into high-fidelity evidence, calculates Information Gain score to identify pivotal states, uses Dynamic Prompting Interventions (de-duplication, reflection, adaptive branching), and integrates rewards via Group Relative Policy Optimization.

Result: SIGNIFICANTLY outperforms existing approaches on single-hop and multi-hop QA benchmarks, particularly in complex reasoning scenarios, while using fewer search steps.

Conclusion: SIGHT effectively addresses redundancy and error accumulation in search-based reasoning, enabling more efficient and robust exploration strategies without external verifiers.

Abstract: Reinforcement Learning (RL) has empowered Large Language Models (LLMs) to master autonomous search for complex question answering. However, particularly within multi-turn search scenarios, this interaction introduces a critical challenge: search results often suffer from high redundancy and low signal-to-noise ratios. Consequently, agents easily fall into “Tunnel Vision,” where the forced interpretation of early noisy retrievals leads to irreversible error accumulation. To address these challenges, we propose SIGHT, a framework that enhances search-based reasoning through Self-Evidence Support (SES) and Information-Gain Driven Diverse Branching. SIGHT distills search results into high-fidelity evidence via SES and calculates an Information Gain score to pinpoint pivotal states where observations maximally reduce uncertainty. This score guides Dynamic Prompting Interventions - including de-duplication, reflection, or adaptive branching - to spawn new branches with SES. Finally, by integrating SES and correctness rewards via Group Relative Policy Optimization, SIGHT internalizes robust exploration strategies without external verifiers. Experiments on single-hop and multi-hop QA benchmarks demonstrate that SIGHT significantly outperforms existing approaches, particularly in complex reasoning scenarios, using fewer search steps.

Method: Introduces PRIME benchmark with 2,530 high-difficulty STEM samples curated through consistency-based filtering. Proposes process-aware RLVR training paradigm using verifiers selected via PRIME evaluation.

Result: Process-aware RLVR training substantially outperforms outcome-only verification baseline with absolute gains of 8.29%, 9.12%, and 7.31% on AIME24, AIME25, and Beyond-AIME for Qwen3-14B-Base. Strong linear correlation (R² > 0.92) between verifier accuracy on PRIME and RLVR effectiveness.

Conclusion: PRIME serves as a reliable predictor for verifier selection in RLVR, highlighting the importance of process-outcome alignment verification over purely outcome-centric approaches.

Abstract: While model-based verifiers are essential for scaling Reinforcement Learning with Verifiable Rewards (RLVR), current outcome-centric verification paradigms primarily focus on the consistency between the final result and the ground truth, often neglecting potential errors in the derivation process. This leads to assigning positive rewards to correct answers produced from incorrect derivations. To bridge this gap, we introduce PRIME, a benchmark for evaluating verifiers on Process-Outcome Alignment verification in Mathematics and Engineering. Curated from a comprehensive collection of college-level STEM problems, PRIME comprises 2,530 high-difficulty samples through a consistency-based filtering pipeline. Through extensive evaluation, we find that current verifiers frequently fail to detect derivation flaws. Furthermore, we propose a process-aware RLVR training paradigm utilizing verifiers selected via PRIME. This approach substantially outperforms the outcome-only verification baseline, achieving absolute performance gains of 8.29%, 9.12%, and 7.31% on AIME24, AIME25, and Beyond-AIME, respectively, for the Qwen3-14B-Base model. Finally, we demonstrate a strong linear correlation ($R^2 > 0.92$) between verifier accuracy on PRIME and RLVR training effectiveness, validating PRIME as a reliable predictor for verifier selection.

Method: Proposes Scene-Aware Memory Discrimination (SAMD) with two components: Gating Unit Module (GUM) filters non-memorable interactions, and Cluster Prompting Module (CPM) establishes adaptive memory standards and analyzes user intent-memory context relationships for clustering prompts.

Result: SAMD successfully recalls majority of memorable data, remains robust in dynamic scenarios, and when integrated into personalized applications, significantly enhances both efficiency and quality of memory construction for better personal knowledge organization.

Conclusion: The proposed SAMD method effectively addresses memory discrimination challenges in LLMs, improving processing efficiency and memory quality through selective filtering and adaptive standards inspired by human cognitive processes.

Abstract: Intelligent devices have become deeply integrated into everyday life, generating vast amounts of user interactions that form valuable personal knowledge. Efficient organization of this knowledge in user memory is essential for enabling personalized applications. However, current research on memory writing, management, and reading using large language models (LLMs) faces challenges in filtering irrelevant information and in dealing with rising computational costs. Inspired by the concept of selective attention in the human brain, we introduce a memory discrimination task. To address large-scale interactions and diverse memory standards in this task, we propose a Scene-Aware Memory Discrimination method (SAMD), which comprises two key components: the Gating Unit Module (GUM) and the Cluster Prompting Module (CPM). GUM enhances processing efficiency by filtering out non-memorable interactions and focusing on the salient content most relevant to application demands. CPM establishes adaptive memory standards, guiding LLMs to discern what information should be remembered or discarded. It also analyzes the relationship between user intents and memory contexts to build effective clustering prompts. Comprehensive direct and indirect evaluations demonstrate the effectiveness and generalization of our approach. We independently assess the performance of memory discrimination, showing that SAMD successfully recalls the majority of memorable data and remains robust in dynamic scenarios. Furthermore, when integrated into personalized applications, SAMD significantly enhances both the efficiency and quality of memory construction, leading to better organization of personal knowledge.

Method: Proposes a dual-level framework with: 1) Sequence-level prefix-protected optimization using decaying mixed rollouts to maintain valid reasoning paths while promoting conciseness, and 2) Group-level difficulty-aware penalty that dynamically scales length constraints based on query complexity.

Result: Achieves up to 55.7% reduction in token usage while simultaneously improving accuracy by up to 4.1% on math benchmarks, with generalization to code, science, and general domains.

Conclusion: The proposed framework effectively addresses overthinking in LRMs through hierarchical supervision, balancing reasoning quality with computational efficiency across different difficulty levels.

Abstract: Language Reasoning Models (LRMs) achieve strong performance by scaling test-time computation but often suffer from ``overthinking’’, producing excessively long reasoning traces that increase latency and memory usage. Existing LRMs typically enforce conciseness with uniform length penalties, which over-compress crucial early deduction steps at the sequence level and indiscriminately penalize all queries at the group level. To solve these limitations, we propose \textbf{\model}, a dual-level framework for prefix-protected and difficulty-aware compression under hierarchical supervision. At the sequence level, prefix-protected optimization employs decaying mixed rollouts to maintain valid reasoning paths while promoting conciseness. At the group level, difficulty-aware penalty dynamically scales length constraints based on query complexity, maintaining exploration for harder questions while curbing redundancy on easier ones. Extensive experiments on DeepSeek-R1-Distill-Qwen (1.5B/7B) demonstrate that \model achieves a substantial reduction in token usage (up to \textbf{55.7%}) while simultaneously improving accuracy (up to \textbf{4.1%}) on math benchmarks, with generalization ability to code, science, and general domains.

Method: Defined six feedback elements, generated feedback for biology questions using GPT-5, conducted learning experiment with 321 high school students, measured learning outcomes and subjective evaluations across six criteria, analyzed differences based on Big Five personality traits.

Result: Effective feedback elements share common patterns supporting learning outcomes, but learners’ subjective preferences differ across personality-based clusters. Personality traits significantly influence feedback acceptance.

Conclusion: Feedback elements should be selected and adapted according to learners’ personality traits when designing LLM-generated feedback, with practical implications for personalized educational feedback design.

Abstract: Large language models (LLMs) show promise for automatically generating feedback in education settings. However, it remains unclear how specific feedback elements, such as tone and information coverage, contribute to learning outcomes and learner acceptance, particularly across learners with different personality traits. In this study, we define six feedback elements and generate feedback for multiple-choice biology questions using GPT-5. We conduct a learning experiment with 321 first-year high school students and evaluate feedback effectiveness using two learning outcomes measures and subjective evaluations across six criteria. We further analyze differences in how feedback acceptance varies across learners based on Big Five personality traits. Our results show that effective feedback elements share common patterns supporting learning outcomes, while learners’ subjective preferences differ across personality-based clusters. These findings highlight the importance of selecting and adapting feedback elements according to learners’ personality traits when we design LLM-generated feedback, and provide practical implications for personalized feedback design in education.

Method: PatientHub provides a modular framework with standardized definitions for patient simulation, supporting implementation of various methods, cross-method evaluation, custom metrics, and extensible simulator variants.

Result: The framework enables standardized evaluation of patient simulation methods, accelerates development by eliminating infrastructure overhead, and facilitates cross-method benchmarking through a reproducible pipeline.

Conclusion: PatientHub lowers barriers to developing new simulation methods and provides a foundation for future datasets, methods, and benchmarks in patient-centered dialogue applications.

Abstract: As Large Language Models increasingly power role-playing applications, simulating patients has become a valuable tool for training counselors and scaling therapeutic assessment. However, prior work is fragmented: existing approaches rely on incompatible, non-standardized data formats, prompts, and evaluation metrics, hindering reproducibility and fair comparison. In this paper, we introduce PatientHub, a unified and modular framework that standardizes the definition, composition, and deployment of simulated patients. To demonstrate PatientHub’s utility, we implement several representative patient simulation methods as case studies, showcasing how our framework supports standardized cross-method evaluation and the seamless integration of custom evaluation metrics. We further demonstrate PatientHub’s extensibility by prototyping two new simulator variants, highlighting how PatientHub accelerates method development by eliminating infrastructure overhead. By consolidating existing work into a single reproducible pipeline, PatientHub lowers the barrier to developing new simulation methods and facilitates cross-method and cross-model benchmarking. Our framework provides a practical foundation for future datasets, methods, and benchmarks in patient-centered dialogue, and the code is publicly available via https://github.com/Sahandfer/PatientHub.

Method: Collected sensicality judgments from human raters and LLMs on sentences from five semantically deviant datasets, both context-free and with provided context. Analyzed LLMs’ ability to generate plausible contexts for anomalous cases.

Result: Human raters considered most sentences as merely anomalous rather than truly nonsensical. LLMs showed substantial skill in generating plausible contexts for anomalous cases, demonstrating their ability to bridge semantic gaps.

Conclusion: LLMs are capable of distinguishing between anomalous and nonsensical sentences and can generate meaningful contexts to interpret anomalous cases, though most existing datasets contain more anomalous than truly nonsensical content.

Abstract: Nonsensical and anomalous sentences have been instrumental in the development of computational models of semantic interpretation. A core challenge is to distinguish between what is merely anomalous (but can be interpreted given a supporting context) and what is truly nonsensical. However, it is unclear (a) how nonsensical, rather than merely anomalous, existing datasets are; and (b) how well LLMs can make this distinction. In this paper, we answer both questions by collecting sensicality judgments from human raters and LLMs on sentences from five semantically deviant datasets: both context-free and when providing a context. We find that raters consider most sentences at most anomalous, and only a few as properly nonsensical. We also show that LLMs are substantially skilled in generating plausible contexts for anomalous cases.

Method: Thinking with Drafting (TwD) uses minimalist Domain-Specific Language as grounding intermediate representation, forcing models to draft mental models into executable code for deterministic visual proofs and self-verification.

Result: TwD serves as superior cognitive scaffold on VisAlg visual algebra benchmark, establishing closed-loop system where visual generation acts as logical verifier.

Conclusion: Reconceptualizing visual reasoning as optical decompression with DSL intermediate representation offers generalizable path for precise visual reasoning through self-verification.

Abstract: Existing multimodal large language models have achieved high-fidelity visual perception and exploratory visual generation. However, a precision paradox persists in complex reasoning tasks: optical perception systems transcribe symbols without capturing logical topology, while pixel-based generative models produce visual artifacts lacking mathematical exactness. To bridge this gap, we propose that reasoning over visual inputs be reconceptualized as optical decompression-the process of reconstructing latent logical structures from compressed visual tokens. Guided by the axiom that Parsing is Reasoning, we introduce Thinking with Drafting (TwD), which utilizes a minimalist Domain-Specific Language (DSL) as a grounding intermediate representation. Unlike standard approaches that hallucinate answers directly, TwD forces the model to draft its mental model into executable code, rendering deterministic visual proofs for self-verification. To validate this, we present VisAlg, a visual algebra benchmark. Experiments demonstrate that TwD serve as a superior cognitive scaffold. Our work establishes a closed-loop system where visual generation acts not as a creative output but as a logical verifier, offering a generalizable path for visual reasoning.

Method: Proposes Length-Incentivized Exploration (LIE), a simple yet effective approach that explicitly encourages models to explore more via a length-based reward coupled with a redundancy penalty. This two-step method maximizes state coverage by incentivizing longer reasoning trajectories while avoiding redundant exploration.

Result: Comprehensive experiments across different models (Qwen3, Llama) show that LIE effectively incentivizes in-context exploration. The method achieves an average improvement of 4.4% on in-domain tasks and a 2.7% gain on out-of-domain benchmarks.

Conclusion: Length-Incentivized Exploration successfully addresses the Shallow Exploration Trap in autoregressive models, enabling better in-context exploration capabilities that lead to improved performance across both familiar and novel tasks.

Abstract: Achieving effective test-time scaling requires models to engage in In-Context Exploration – the intrinsic ability to generate, verify, and refine multiple reasoning hypotheses within a single continuous context. Grounded in State Coverage theory, our analysis identifies a critical bottleneck to enabling this capability: while broader state coverage requires longer reasoning trajectories, the probability of sampling such sequences decays exponentially during autoregressive generation, a phenomenon we term the ``Shallow Exploration Trap’’. To bridge this gap, we propose Length-Incentivized Exploration(\method). This simple yet effective recipe explicitly encourages models to explore more via a length-based reward coupled with a redundancy penalty, thereby maximizing state coverage in two-step manner. Comprehensive experiments across different models (Qwen3, Llama) demonstrate that \method effectively incentivize in-context exploration. As a result, our method achieves an average improvement of 4.4% on in-domain tasks and a 2.7% gain on out-of-domain benchmarks.

Method: Hybrid architecture integrating sparse attention (InfLLM-V2) and linear attention (Lightning Attention) in 1:3 ratio using layer selection algorithm; employs hybrid positional encoding (HyPE); uses cost-effective continual training framework to transform pre-trained Transformer models into hybrid models.

Result: Achieves up to 3.5x inference speed of full-attention models at 256K sequence length on single A6000D GPU; supports up to 1M token contexts where traditional 8B models fail; maintains general capabilities comparable to full-attention models; reduces training costs by ~75% compared to training from scratch.

Conclusion: MiniCPM-SALA demonstrates effective hybrid approach for long-context LLMs that balances efficiency and performance, enabling practical deployment of ultra-long context models with significantly reduced computational requirements.

Abstract: The evolution of large language models (LLMs) towards applications with ultra-long contexts faces challenges posed by the high computational and memory costs of the Transformer architecture. While existing sparse and linear attention mechanisms attempt to mitigate these issues, they typically involve a trade-off between memory efficiency and model performance. This paper introduces MiniCPM-SALA, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention (InfLLM-V2) with the global efficiency of linear attention (Lightning Attention). By employing a layer selection algorithm to integrate these mechanisms in a 1:3 ratio and utilizing a hybrid positional encoding (HyPE), the model maintains efficiency and performance for long-context tasks. Furthermore, we introduce a cost-effective continual training framework that transforms pre-trained Transformer-based models into hybrid models, which reduces training costs by approximately 75% compared to training from scratch. Extensive experiments show that MiniCPM-SALA maintains general capabilities comparable to full-attention models while offering improved efficiency. On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens and supports context lengths of up to 1M tokens, a scale where traditional full-attention 8B models fail because of memory constraints.

Method: Trains subword embeddings on raw text, then groups related forms using combined cosine and n-gram similarity measures to create clusters of variant forms.

Result: Applied to Luxembourgish user comments, the method uncovers extensive lexical and orthographic variation aligning with dialectal and sociolinguistic patterns, capturing systematic correspondences and regional/stylistic differentiation.

Conclusion: Distributional modeling can reveal meaningful variation patterns even in noisy or low-resource settings, providing a reproducible framework for studying language variety in multilingual and small-language contexts.

Abstract: This paper presents an embedding-based approach to detecting variation without relying on prior normalisation or predefined variant lists. The method trains subword embeddings on raw text and groups related forms through combined cosine and n-gram similarity. This allows spelling and morphological diversity to be examined and analysed as linguistic structure rather than treated as noise. Using a large corpus of Luxembourgish user comments, the approach uncovers extensive lexical and orthographic variation that aligns with patterns described in dialectal and sociolinguistic research. The induced families capture systematic correspondences and highlight areas of regional and stylistic differentiation. The procedure does not strictly require manual annotation, but does produce transparent clusters that support both quantitative and qualitative analysis. The results demonstrate that distributional modelling can reveal meaningful patterns of variation even in ‘’noisy’’ or low-resource settings, offering a reproducible methodological framework for studying language variety in multilingual and small-language contexts.

Method: DMAP transforms text into samples in the unit interval that jointly encode rank and probability information from language model next-token distributions. This creates a model-agnostic representation that enables efficient statistical analysis of text properties.

Result: DMAP successfully enables three applications: validating generation parameters for data integrity, examining probability curvature for machine-generated text detection, and forensic analysis revealing statistical fingerprints in models trained on synthetic data.

Conclusion: DMAP provides a unified statistical framework for text analysis with LLMs that is computationally efficient, widely applicable, and offers a foundation for further research into understanding text properties through language model probability distributions.

Abstract: Large Language Models (LLMs) are a powerful tool for statistical text analysis, with derived sequences of next-token probability distributions offering a wealth of information. Extracting this signal typically relies on metrics such as perplexity, which do not adequately account for context; how one should interpret a given next-token probability is dependent on the number of reasonable choices encoded by the shape of the conditional distribution. In this work, we present DMAP, a mathematically grounded method that maps a text, via a language model, to a set of samples in the unit interval that jointly encode rank and probability information. This representation enables efficient, model-agnostic analysis and supports a range of applications. We illustrate its utility through three case studies: (i) validation of generation parameters to ensure data integrity, (ii) examining the role of probability curvature in machine-generated text detection, and (iii) a forensic analysis revealing statistical fingerprints left in downstream models that have been subject to post-training on synthetic data. Our results demonstrate that DMAP offers a unified statistical view of text that is simple to compute on consumer hardware, widely applicable, and provides a foundation for further research into text analysis with LLMs.

Method: Proposed RouterXBench framework with three evaluation dimensions: router ability, scenario alignment, and cross-domain robustness. Introduced ProbeDirichlet router that aggregates cross-layer hidden states using learnable Dirichlet distributions with probabilistic training, trained on multi-domain data.

Result: ProbeDirichlet achieves 16.68% and 18.86% relative improvements over best baselines in router ability and high-accuracy scenarios, with consistent performance across model families, scales, tasks, and agentic workflows.

Conclusion: Internal hidden states capture model uncertainty better than output probabilities, and the proposed framework enables systematic evaluation of LLM routers for practical deployment scenarios.

Abstract: Large language models (LLMs) have achieved success, but cost and privacy constraints necessitate deploying smaller models locally while offloading complex queries to cloud-based models. Existing router evaluations are unsystematic, overlooking scenario-specific requirements and out-of-distribution robustness. We propose RouterXBench, a principled evaluation framework with three dimensions: router ability, scenario alignment, and cross-domain robustness. Unlike prior work that relies on output probabilities or external embeddings, we utilize internal hidden states that capture model uncertainty before answer generation. We introduce ProbeDirichlet, a lightweight router that aggregates cross-layer hidden states via learnable Dirichlet distributions with probabilistic training. Trained on multi-domain data, it generalizes robustly across in-domain and out-of-distribution scenarios. Our results show ProbeDirichlet achieves 16.68% and 18.86% relative improvements over the best baselines in router ability and high-accuracy scenarios, with consistent performance across model families, model scales, heterogeneous tasks, and agentic workflows.

Method: Presents a semi-automated pipeline for triplet extraction using ontology-driven proxy metrics (Ontology Conformance and Faithfulness). Compares static manually engineered ontology against fully automated document-specific ontology induction across different LLMs. Proposes hybrid verification combining regex matching with LLM-as-a-judge to filter false positives from coreference resolution.

Result: Automatically induced ontology achieves 100% schema conformance in all configurations, eliminating ontology drift observed with manual approach. Hybrid verification reduces apparent subject hallucination rates from 65.2% to 1.6%. Identifies systematic asymmetry between subject and object hallucinations attributed to passive constructions and omitted agents in financial prose.

Conclusion: Automated ontology induction outperforms manual approaches for financial report knowledge extraction, and hybrid verification effectively addresses hallucination issues, providing a robust framework for knowledge graph construction in domains lacking ground truth annotations.

Abstract: Corporate financial reports are a valuable source of structured knowledge for Knowledge Graph construction, but the lack of annotated ground truth in this domain makes evaluation difficult. We present a semi-automated pipeline for Subject-Predicate-Object triplet extraction that uses ontology-driven proxy metrics, specifically Ontology Conformance and Faithfulness, instead of ground-truth-based evaluation. We compare a static, manually engineered ontology against a fully automated, document-specific ontology induction approach across different LLMs and two corporate annual reports. The automatically induced ontology achieves 100% schema conformance in all configurations, eliminating the ontology drift observed with the manual approach. We also propose a hybrid verification strategy that combines regex matching with an LLM-as-a-judge check, reducing apparent subject hallucination rates from 65.2% to 1.6% by filtering false positives caused by coreference resolution. Finally, we identify a systematic asymmetry between subject and object hallucinations, which we attribute to passive constructions and omitted agents in financial prose.

Method: The authors analyze two major reasoning benchmarks (MMLU-Pro and GPQA) to measure disagreement among LLMs with comparable accuracy scores. They then conduct re-analyses of published studies in education and political science to examine how switching annotation models affects research results, specifically measuring changes in estimated treatment effects.

Result: LLMs achieving comparable accuracy disagree on 16-66% of benchmark items (16-38% among top frontier models). In scientific applications, switching annotation models can change estimated treatment effects by more than 80%, and in some cases reverses the sign of effects, demonstrating that model choice is a consequential hidden variable for reproducibility.

Conclusion: There exists a “benchmark illusion” where equal accuracy conceals deep epistemic divergence, making model choice a critical but often overlooked variable that threatens scientific reproducibility when LLMs are used for data annotation and inference in research.

Abstract: Benchmarks underpin how progress in large language models (LLMs) is measured and trusted. Yet our analyses reveal that apparent convergence in benchmark accuracy can conceal deep epistemic divergence. Using two major reasoning benchmarks - MMLU-Pro and GPQA - we show that LLMs achieving comparable accuracy still disagree on 16-66% of items, and 16-38% among top-performing frontier models. These discrepancies suggest distinct error profiles for different LLMs. When such models are used for scientific data annotation and inference, their hidden disagreements propagate into research results: in re-analyses of published studies in education and political science, switching the annotation model can change estimated treatment effects by more than 80%, and in some cases reverses their sign. Together, these findings illustrate a benchmark illusion, where equal accuracy may conceal disagreement, with model choice becoming a hidden yet consequential variable for scientific reproducibility.

Method: Introduces AdaptEvolve within an evolutionary sequential refinement framework that leverages intrinsic generation confidence to estimate real-time solvability and dynamically select LLMs that are sufficiently capable for each generation step while remaining computationally efficient.

Result: Confidence-driven selection yields a favorable Pareto frontier, reducing total inference cost by an average of 37.9% across benchmarks while retaining 97.5% of the upper-bound accuracy of static large-model baselines.

Conclusion: AdaptEvolve provides an effective adaptive LLM selection mechanism for evolutionary refinement systems that balances computational efficiency with reasoning capability through confidence-driven routing.

Abstract: Evolutionary agentic systems intensify the trade-off between computational efficiency and reasoning capability by repeatedly invoking large language models (LLMs) during inference. This setting raises a central question: how can an agent dynamically select an LLM that is sufficiently capable for the current generation step while remaining computationally efficient? While model cascades offer a practical mechanism for balancing this trade-off, existing routing strategies typically rely on static heuristics or external controllers and do not explicitly account for model uncertainty. We introduce AdaptEvolve: Adaptive LLM Selection for Multi-LLM Evolutionary Refinement within an evolutionary sequential refinement framework that leverages intrinsic generation confidence to estimate real-time solvability. Empirical results show that confidence-driven selection yields a favourable Pareto frontier, reducing total inference cost by an average of 37.9% across benchmarks while retaining 97.5% of the upper-bound accuracy of static large-model baselines. Our code is available at https://github.com/raypretam/adaptive_llm_selection.

Method: Proposes Cross-Modal Robustness Transfer (CMRT) framework that transfers adversarial robustness from text modality to speech modality without requiring adversarial speech data during training.

Result: CMRT improves adversarial robustness by an average of more than 3 BLEU points across four language pairs, establishing new baseline for robust E2E-ST.

Conclusion: CMRT effectively addresses vulnerability of E2E-ST models to inflectional variations without computational overhead of generating adversarial speech data.

Abstract: End-to-End Speech Translation (E2E-ST) has seen significant advancements, yet current models are primarily benchmarked on curated, “clean” datasets. This overlooks critical real-world challenges, such as morphological robustness to inflectional variations common in non-native or dialectal speech. In this work, we adapt a text-based adversarial attack targeting inflectional morphology to the speech domain and demonstrate that state-of-the-art E2E-ST models are highly vulnerable it. While adversarial training effectively mitigates such risks in text-based tasks, generating high-quality adversarial speech data remains computationally expensive and technically challenging. To address this, we propose Cross-Modal Robustness Transfer (CMRT), a framework that transfers adversarial robustness from the text modality to the speech modality. Our method eliminates the requirement for adversarial speech data during training. Extensive experiments across four language pairs demonstrate that CMRT improves adversarial robustness by an average of more than 3 BLEU points, establishing a new baseline for robust E2E-ST without the overhead of generating adversarial speech.

Method: Introduce LLM-based classifier to identify underspecified questions in QA datasets; conduct controlled rewriting experiment to transform underspecified questions into fully specified variants while keeping gold answers fixed.

Result: Found 16% to over 50% of benchmark questions are underspecified; LLMs perform significantly worse on them; QA performance consistently improves when questions are rewritten to be fully specified.

Conclusion: Underspecification is a major confound in QA evaluation, with many apparent LLM failures actually stemming from ambiguous questions rather than model limitations, highlighting need for clearer benchmark design.

Abstract: Large language models (LLMs) perform well on well-posed questions, yet standard question-answering (QA) benchmarks remain far from solved. We argue that this gap is partly due to underspecified questions - queries whose interpretation cannot be uniquely determined without additional context. To test this hypothesis, we introduce an LLM-based classifier to identify underspecified questions and apply it to several widely used QA datasets, finding that 16% to over 50% of benchmark questions are underspecified and that LLMs perform significantly worse on them. To isolate the effect of underspecification, we conduct a controlled rewriting experiment that serves as an upper-bound analysis, rewriting underspecified questions into fully specified variants while holding gold answers fixed. QA performance consistently improves under this setting, indicating that many apparent QA failures stem from question underspecification rather than model limitations. Our findings highlight underspecification as an important confound in QA evaluation and motivate greater attention to question clarity in benchmark design.

Method: Collected dataset from Reddit and YouTube stratified by SES, prompted four LLMs with incomplete text from the corpus, and compared LLM-generated completions to originals using 94 sociolinguistic metrics covering syntactic, rhetorical, and lexical features.

Result: LLMs modulate their style with respect to SES only to a minor extent, often resulting in approximation or caricature, and tend to emulate the style of upper SES more effectively than lower SES styles.

Conclusion: LLMs risk amplifying linguistic hierarchies and their validity is questionable for agent-based social simulation, survey experiments, and any research relying on language style as a social signal.

Abstract: Humans adjust their linguistic style to the audience they are addressing. However, the extent to which LLMs adapt to different social contexts is largely unknown. As these models increasingly mediate human-to-human communication, their failure to adapt to diverse styles can perpetuate stereotypes and marginalize communities whose linguistic norms are less closely mirrored by the models, thereby reinforcing social stratification. We study the extent to which LLMs integrate into social media communication across different socioeconomic status (SES) communities. We collect a novel dataset from Reddit and YouTube, stratified by SES. We prompt four LLMs with incomplete text from that corpus and compare the LLM-generated completions to the originals along 94 sociolinguistic metrics, including syntactic, rhetorical, and lexical features. LLMs modulate their style with respect to SES to only a minor extent, often resulting in approximation or caricature, and tend to emulate the style of upper SES more effectively. Our findings (1) show how LLMs risk amplifying linguistic hierarchies and (2) call into question their validity for agent-based social simulation, survey experiments, and any research relying on language style as a social signal.

Method: Based on Gemma3 model family, using continual pretraining and instruction finetuning to adapt LLMs for multilingual MT across 46 languages, studying effects of model scaling and data scaling.

Result: MiLMMT-46 outperforms recent SOTA models (Seed-X, HY-MT-1.5, TranslateGemma) and achieves competitive performance with proprietary systems like Google Translate and Gemini 3 Pro.

Conclusion: Open LLMs can be effectively adapted for high-quality multilingual machine translation through proper scaling and adaptation techniques, achieving competitive results with proprietary systems.

Abstract: Open large language models (LLMs) have demonstrated improving multilingual capabilities in recent years. In this paper, we present a study of open LLMs for multilingual machine translation (MT) across a range of languages, and investigate the effects of model scaling and data scaling when adapting open LLMs to multilingual MT through continual pretraining and instruction finetuning. Based on the Gemma3 model family, we develop MiLMMT-46, which achieves top-tier multilingual translation performance across 46 languages. Extensive experiments show that MiLMMT-46 consistently outperforms recent state-of-the-art (SOTA) models, including Seed-X, HY-MT-1.5, and TranslateGemma, and achieves competitive performance with strong proprietary systems such as Google Translate and Gemini 3 Pro.

Method: Based on HPLT web-crawled datasets, using web crawl timestamps as approximate document creation time signals. Covers three time periods (2011-2015, 2020-2021, 2024-present) with 1 million documents per period per language.

Result: Created DHPLT collection with 41 diverse languages, providing pre-computed word type/token embeddings and lexical substitutions for target words, while allowing researchers to define their own target words.

Conclusion: DHPLT fills the gap in multilingual diachronic resources and enables new research directions in semantic change modeling across diverse languages.

Abstract: In this resource paper, we present DHPLT, an open collection of diachronic corpora in 41 diverse languages. DHPLT is based on the web-crawled HPLT datasets; we use web crawl timestamps as the approximate signal of document creation time. The collection covers three time periods: 2011-2015, 2020-2021 and 2024-present (1 million documents per time period for each language). We additionally provide pre-computed word type and token embeddings and lexical substitutions for our chosen target words, while at the same time leaving it open for the other researchers to come up with their own target words using the same datasets. DHPLT aims at filling in the current lack of multilingual diachronic corpora for semantic change modelling (beyond a dozen of high-resource languages). It opens the way for a variety of new experimental setups in this field. All the resources described in this paper are available at https://data.hplt-project.org/three/diachronic/, sorted by language.

Method: Created baseline for cybersecurity ATS and test dataset of 40 CVE descriptions. Evaluated LLM performance through two survey rounds with cybersecurity experts, assessing both simplification quality and meaning preservation.

Result: Out-of-the-box LLMs can make cybersecurity text appear simpler but struggle with meaning preservation. The study provides a benchmark dataset and reveals limitations of current LLMs for domain-specific text simplification.

Conclusion: LLMs show promise for cybersecurity text simplification but need improvement in preserving technical accuracy and meaning. Domain-specific fine-tuning or specialized approaches may be necessary for effective CVE description simplification.

Abstract: Understanding cyber security is increasingly important for individuals and organizations. However, a lot of information related to cyber security can be difficult to understand to those not familiar with the topic. In this study, we focus on investigating how large language models (LLMs) could be utilized in automatic text simplification (ATS) of Common Vulnerability and Exposure (CVE) descriptions. Automatic text simplification has been studied in several contexts, such as medical, scientific, and news texts, but it has not yet been studied to simplify texts in the rapidly changing and complex domain of cyber security. We created a baseline for cyber security ATS and a test dataset of 40 CVE descriptions, evaluated by two groups of cyber security experts in two survey rounds. We have found that while out-of-the box LLMs can make the text appear simpler, they struggle with meaning preservation. Code and data are available at https://version.aalto.fi/gitlab/vehomav1/simplification_nmi.

Method: Models Text-to-SQL as two-stage process (interpretations → answers), decomposes semantic uncertainty into ambiguity and instability scores using Schur complement of bipartite semantic graph matrix.

Result: Outperforms Kernel Language Entropy on AmbigQA/SituatedQA and clinical Text-to-SQL benchmarks; identifies high-ambiguity/high-instability regime containing 51% of errors while covering 25% of queries.

Conclusion: CLUES enables targeted interventions: query refinement for ambiguity, model improvement for instability, providing diagnostic decomposition unavailable from single uncertainty scores.

Abstract: Deploying large language models for clinical Text-to-SQL requires distinguishing two qualitatively different causes of output diversity: (i) input ambiguity that should trigger clarification, and (ii) model instability that should trigger human review. We propose CLUES, a framework that models Text-to-SQL as a two-stage process (interpretations –> answers) and decomposes semantic uncertainty into an ambiguity score and an instability score. The instability score is computed via the Schur complement of a bipartite semantic graph matrix. Across AmbigQA/SituatedQA (gold interpretations) and a clinical Text-to-SQL benchmark (known interpretations), CLUES improves failure prediction over state-of-the-art Kernel Language Entropy. In deployment settings, it remains competitive while providing a diagnostic decomposition unavailable from a single score. The resulting uncertainty regimes map to targeted interventions - query refinement for ambiguity, model improvement for instability. The high-ambiguity/high-instability regime contains 51% of errors while covering 25% of queries, enabling efficient triage.

Method: Composition-RL automatically composes multiple problems into new verifiable questions and uses these compositional prompts for RL training. A curriculum variant gradually increases compositional depth over training. The approach also enables cross-domain RL by composing prompts from different domains.

Result: Extensive experiments across model sizes from 4B to 30B show Composition-RL consistently improves reasoning capability over RL trained on original datasets. Performance is further boosted with the curriculum variant. The method also enables more effective cross-domain RL.

Conclusion: Composition-RL is a simple yet effective approach for better utilizing limited verifiable prompts, particularly addressing the underutilization of pass-rate-1 prompts. It improves reasoning capabilities and enables more efficient cross-domain learning.

Abstract: Large-scale verifiable prompts underpin the success of Reinforcement Learning with Verifiable Rewards (RLVR), but they contain many uninformative examples and are costly to expand further. Recent studies focus on better exploiting limited training data by prioritizing hard prompts whose rollout pass rate is 0. However, easy prompts with a pass rate of 1 also become increasingly prevalent as training progresses, thereby reducing the effective data size. To mitigate this, we propose Composition-RL, a simple yet useful approach for better utilizing limited verifiable prompts targeting pass-rate-1 prompts. More specifically, Composition-RL automatically composes multiple problems into a new verifiable question and uses these compositional prompts for RL training. Extensive experiments across model sizes from 4B to 30B show that Composition-RL consistently improves reasoning capability over RL trained on the original dataset. Performance can be further boosted with a curriculum variant of Composition-RL that gradually increases compositional depth over training. Additionally, Composition-RL enables more effective cross-domain RL by composing prompts drawn from different domains. Codes, datasets, and models are available at https://github.com/XinXU-USTC/Composition-RL.

Method: Proposes DeepSight project with unified task and data protocols connecting evaluation (DeepSafe) and diagnosis (DeepScan) stages. The approach is low-cost, reproducible, efficient, and scalable, transforming safety evaluation from black-box to white-box insight.

Result: DeepSight is presented as the first open-source toolkit supporting frontier AI risk evaluation and joint safety evaluation and diagnosis for Large Models.

Conclusion: DeepSight addresses systematic safety issues in Large Models by integrating evaluation and diagnosis, providing a more comprehensive safety workflow that connects external behavioral risks with internal root causes.

Abstract: As the development of Large Models (LMs) progresses rapidly, their safety is also a priority. In current Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) safety workflow, evaluation, diagnosis, and alignment are often handled by separate tools. Specifically, safety evaluation can only locate external behavioral risks but cannot figure out internal root causes. Meanwhile, safety diagnosis often drifts from concrete risk scenarios and remains at the explainable level. In this way, safety alignment lack dedicated explanations of changes in internal mechanisms, potentially degrading general capabilities. To systematically address these issues, we propose an open-source project, namely DeepSight, to practice a new safety evaluation-diagnosis integrated paradigm. DeepSight is low-cost, reproducible, efficient, and highly scalable large-scale model safety evaluation project consisting of a evaluation toolkit DeepSafe and a diagnosis toolkit DeepScan. By unifying task and data protocols, we build a connection between the two stages and transform safety evaluation from black-box to white-box insight. Besides, DeepSight is the first open source toolkit that support the frontier AI risk evaluation and joint safety evaluation and diagnosis.

Method: P-GenRM transforms preference signals into structured evaluation chains with adaptive personas and scoring rubrics. It clusters users into prototypes and uses dual-granularity scaling: individual-level adaptive scaling of each user’s scoring scheme, and prototype-level incorporation of similar users’ preferences.

Result: Achieves state-of-the-art results on personalized reward model benchmarks with 2.31% average improvement, demonstrates strong generalization on out-of-distribution data, and test-time user-based scaling provides additional 3% boost.

Conclusion: P-GenRM effectively addresses limitations of existing personalized reward models through adaptive persona generation, user prototyping, and dual-granularity scaling, enabling better personalization and generalization with test-time scalability.

Abstract: Personalized alignment of large language models seeks to adapt responses to individual user preferences, typically via reinforcement learning. A key challenge is obtaining accurate, user-specific reward signals in open-ended scenarios. Existing personalized reward models face two persistent limitations: (1) oversimplifying diverse, scenario-specific preferences into a small, fixed set of evaluation principles, and (2) struggling with generalization to new users with limited feedback. To this end, we propose P-GenRM, the first Personalized Generative Reward Model with test-time user-based scaling. P-GenRM transforms preference signals into structured evaluation chains that derive adaptive personas and scoring rubrics across various scenarios. It further clusters users into User Prototypes and introduces a dual-granularity scaling mechanism: at the individual level, it adaptively scales and aggregates each user’s scoring scheme; at the prototype level, it incorporates preferences from similar users. This design mitigates noise in inferred preferences and enhances generalization to unseen users through prototype-based transfer. Empirical results show that P-GenRM achieves state-of-the-art results on widely-used personalized reward model benchmarks, with an average improvement of 2.31%, and demonstrates strong generalization on an out-of-distribution dataset. Notably, Test-time User-based scaling provides an additional 3% boost, demonstrating stronger personalized alignment with test-time scalability.

Method: Constructs rule-based morphology model using Wiktionary data, investigates SQL for querying lexical patterns, and creates declarative rule-base with Python utilities to derive inflected forms.

Result: Developed functionality to derive inflected forms of Gaidhlig words that can support educational tools for teaching language patterns and higher-level tools like rule-based dependency parsers.

Conclusion: Rule-based systems effectively leverage limited data for low-resource languages, add value to existing Wiktionary data, and support interpretable tools for language education and processing.

Abstract: Language models and software tools are essential to support the continuing vitality of lesser-used languages; however, currently popular neural models require considerable data for training, which normally is not available for such low-resource languages. This paper describes work-in-progress to construct a rule-based model of Gaidhlig morphology using data from Wiktionary, arguing that rule-based systems effectively leverage limited sample data, support greater interpretability, and provide insights useful in the design of teaching materials. The use of SQL for querying the occurrence of different lexical patterns is investigated, and a declarative rule-base is presented that allows Python utilities to derive inflected forms of Gaidhlig words. This functionality could be used to support educational tools that teach or explain language patterns, for example, or to support higher level tools such as rule-based dependency parsers. This approach adds value to the data already present in Wiktionary by adapting it to new use-cases.

Method: Introduces a tripartite framework: 1) Pro subset for rigorous reasoning challenges with increased difficulty, 2) Basic subset establishing standards for spoken colloquialism focusing on natural vocabulary, linguistic fluency, and interactive rapport, and 3) Acoustic subset covering explicit understanding, generation, and implicit dialogue for paralinguistic evaluation.

Result: Evaluated five state-of-the-art models, providing critical insights into complex problem-solving, colloquial delivery, and paralinguistic fidelity in spoken dialogue systems.

Conclusion: WavBench bridges the gap in current evaluations by addressing real-world conversational complexity, guiding the evolution of robust spoken dialogue models with comprehensive audio-centric assessment.

Abstract: With the rapid integration of advanced reasoning capabilities into spoken dialogue models, the field urgently demands benchmarks that transcend simple interactions to address real-world complexity. However, current evaluations predominantly adhere to text-generation standards, overlooking the unique audio-centric characteristics of paralinguistics and colloquialisms, alongside the cognitive depth required by modern agents. To bridge this gap, we introduce WavBench, a comprehensive benchmark designed to evaluate realistic conversational abilities where prior works fall short. Uniquely, WavBench establishes a tripartite framework: 1) Pro subset, designed to rigorously challenge reasoning-enhanced models with significantly increased difficulty; 2) Basic subset, defining a novel standard for spoken colloquialism that prioritizes “listenability” through natural vocabulary, linguistic fluency, and interactive rapport, rather than rigid written accuracy; and 3) Acoustic subset, covering explicit understanding, generation, and implicit dialogue to rigorously evaluate comprehensive paralinguistic capabilities within authentic real-world scenarios. Through evaluating five state-of-the-art models, WavBench offers critical insights into the intersection of complex problem-solving, colloquial delivery, and paralinguistic fidelity, guiding the evolution of robust spoken dialogue models. The benchmark dataset and evaluation toolkit are available at https://naruto-2024.github.io/wavbench.github.io/.

Method: Created a multilayer dataset of 120 European Portuguese municipal meeting minutes from six municipalities with manual annotations by trained annotators and curation by an experienced linguist across three dimensions: metadata, subjects of discussion, and voting outcomes.

Result: Dataset contains over one million tokens with over 38,000 individual annotations, all personal identifiers de-identified, released under FAIR principles with baseline results on metadata extraction, topic classification, and vote labeling.

Conclusion: CitiLink-Minutes addresses the data gap for municipal meeting analysis, enabling downstream NLP/IR tasks and promoting transparent access to municipal decisions while demonstrating potential for computational governance research.

Abstract: City councils play a crucial role in local governance, directly influencing citizens’ daily lives through decisions made during municipal meetings. These deliberations are formally documented in meeting minutes, which serve as official records of discussions, decisions, and voting outcomes. Despite their importance, municipal meeting records have received little attention in Information Retrieval (IR) and Natural Language Processing (NLP), largely due to the lack of annotated datasets, which ultimately limit the development of computational models. To address this gap, we introduce CitiLink-Minutes, a multilayer dataset of 120 European Portuguese municipal meeting minutes from six municipalities. Unlike prior annotated datasets of parliamentary or video records, CitiLink-Minutes provides multilayer annotations and structured linkage of official written minutes. The dataset contains over one million tokens, with all personal identifiers de-identified. Each minute was manually annotated by two trained annotators and curated by an experienced linguist across three complementary dimensions: (1) metadata, (2) subjects of discussion, and (3) voting outcomes, totaling over 38,000 individual annotations. Released under FAIR principles and accompanied by baseline results on metadata extraction, topic classification, and vote labeling, CitiLink-Minutes demonstrates its potential for downstream NLP and IR tasks, while promoting transparent access to municipal decisions.

Method: dVoting performs iterative refinement by: 1) sampling multiple outputs for the same prompt, 2) identifying uncertain tokens via consistency analysis across samples, 3) regenerating uncertain tokens through voting, and 4) repeating until convergence. This leverages dLLMs’ ability to generate tokens at arbitrary positions in parallel.

Result: dVoting consistently improves performance across benchmarks: 6.22%-7.66% on GSM8K, 4.40%-7.20% on MATH500, 3.16%-14.84% on ARC-C, and 4.83%-5.74% on MMLU, with acceptable computational overhead.

Conclusion: dVoting effectively boosts reasoning capabilities of Diffusion LLMs without training, demonstrating the potential of leveraging dLLMs’ parallel generation capabilities for improved performance through iterative refinement and voting techniques.

Abstract: Diffusion Large Language Models (dLLMs) represent a new paradigm beyond autoregressive modeling, offering competitive performance while naturally enabling a flexible decoding process. Specifically, dLLMs can generate tokens at arbitrary positions in parallel, endowing them with significant potential for parallel test-time scaling, which was previously constrained by severe inefficiency in autoregressive modeling. In this work, we introduce dVoting, a fast voting technique that boosts reasoning capability without training, with only an acceptable extra computational overhead. dVoting is motivated by the observation that, across multiple samples for the same prompt, token predictions remain largely consistent, whereas performance is determined by a small subset of tokens exhibiting cross-sample variability. Leveraging the arbitrary-position generation capability of dLLMs, dVoting performs iterative refinement by sampling, identifying uncertain tokens via consistency analysis, regenerating them through voting, and repeating this process until convergence. Extensive evaluations demonstrate that dVoting consistently improves performance across various benchmarks. It achieves gains of 6.22%-7.66% on GSM8K, 4.40%-7.20% on MATH500, 3.16%-14.84% on ARC-C, and 4.83%-5.74% on MMLU. Our code is available at https://github.com/fscdc/dVoting

Method: Trains models to estimate passage-query relevance using attention scores from selected retrieval heads in LLMs, providing listwise ranking with continuous relevance scores.

Result: Outperforms existing pointwise and listwise rerankers across multiple domains (Wikipedia, long narrative datasets), achieves SOTA on LoCoMo benchmark, and supports flexible extensions like contextual augmentation and middle-layer attention heads.

Conclusion: The proposed lightweight reranking framework is effective, flexible, and achieves strong performance with small models while supporting various extensions for improved accuracy and efficiency.

Abstract: Built upon the existing analysis of retrieval heads in large language models, we propose an alternative reranking framework that trains models to estimate passage-query relevance using the attention scores of selected heads. This approach provides a listwise solution that leverages holistic information within the entire candidate shortlist during ranking. At the same time, it naturally produces continuous relevance scores, enabling training on arbitrary retrieval datasets without requiring Likert-scale supervision. Our framework is lightweight and effective, requiring only small-scale models (e.g., 4B parameters) to achieve strong performance. Extensive experiments demonstrate that our method outperforms existing state-of-the-art pointwise and listwise rerankers across multiple domains, including Wikipedia and long narrative datasets. It further establishes a new state-of-the-art on the LoCoMo benchmark that assesses the capabilities of dialogue understanding and memory usage. We further demonstrate that our framework supports flexible extensions. For example, augmenting candidate passages with contextual information further improves ranking accuracy, while training attention heads from middle layers enhances efficiency without sacrificing performance.

Method: Created VRB benchmark with 701 questions sourced from primary school examinations in Zambia and India, covering tasks like reasoning by analogy, pattern completion, and spatial matching. Uses unedited, minimal-text images to test models’ ability to meet realistic primary education needs. Evaluates MLLMs on both static skills (counting, scaling) and dynamic operations (folding, reflection, rotation).

Result: Reveals a “jagged frontier” of capability - models show better proficiency in static skills but reach a distinct “spatial ceiling” with dynamic operations. This poses risks for classroom use including incorrect marking, false scaffolding, and reinforcing student misconceptions.

Conclusion: Education-focused benchmarks like VRB are essential for determining functional boundaries of multimodal tools in classrooms, as current MLLMs have significant limitations in spatial reasoning that could negatively impact educational applications.

Abstract: AI models have achieved state-of-the-art results in textual reasoning; however, their ability to reason over spatial and relational structures remains a critical bottleneck – particularly in early-grade maths, which relies heavily on visuals. This paper introduces the visual reasoning benchmark (VRB), a novel dataset designed to evaluate Multimodal Large Language Models (MLLMs) on their ability to solve authentic visual problems from classrooms. This benchmark is built on a set of 701 questions sourced from primary school examinations in Zambia and India, which cover a range of tasks such as reasoning by analogy, pattern completion, and spatial matching. We outline the methodology and development of the benchmark which intentionally uses unedited, minimal-text images to test if models can meet realistic needs of primary education. Our findings reveal a jagged frontier'' of capability where models demonstrate better proficiency in static skills such as counting and scaling, but reach a distinct spatial ceiling’’ when faced with dynamic operations like folding, reflection, and rotation. These weaknesses pose a risk for classroom use on visual reasoning problems, with the potential for incorrect marking, false scaffolding, and reinforcing student misconceptions. Consequently, education-focused benchmarks like the VRB are essential for determining the functional boundaries of multimodal tools used in classrooms.

Method: Introduces ExStrucTiny benchmark dataset built through a novel pipeline combining manual and synthetic human-validated samples. The dataset covers varied document types and extraction scenarios, unifying aspects of Key Entity Extraction, Relation Extraction, and Visual Question Answering.

Result: Analyzes open and closed VLMs on the benchmark, highlighting challenges such as schema adaptation, query under-specification, and answer localization. Provides a comprehensive evaluation framework for structured information extraction from document images.

Conclusion: ExStrucTiny provides a bedrock for improving generalist models for structured information extraction in documents, addressing gaps in current benchmarks and enabling better evaluation of VLMs on real-world document understanding tasks.

Abstract: Enterprise documents, such as forms and reports, embed critical information for downstream applications like data archiving, automated workflows, and analytics. Although generalist Vision Language Models (VLMs) perform well on established document understanding benchmarks, their ability to conduct holistic, fine-grained structured extraction across diverse document types and flexible schemas is not well studied. Existing Key Entity Extraction (KEE), Relation Extraction (RE), and Visual Question Answering (VQA) datasets are limited by narrow entity ontologies, simple queries, or homogeneous document types, often overlooking the need for adaptable and structured extraction. To address these gaps, we introduce ExStrucTiny, a new benchmark dataset for structured Information Extraction (IE) from document images, unifying aspects of KEE, RE, and VQA. Built through a novel pipeline combining manual and synthetic human-validated samples, ExStrucTiny covers more varied document types and extraction scenarios. We analyze open and closed VLMs on this benchmark, highlighting challenges such as schema adaptation, query under-specification, and answer localization. We hope our work provides a bedrock for improving generalist models for structured IE in documents.

Method: Defines “token overflow” as when compressed representations lack sufficient information to answer queries. Proposes methodology to characterize and detect it using: 1) query-agnostic saturation statistics to separate compressed from uncompressed tokens, and 2) lightweight probing classifiers over both query and context representations in the xRAG soft-compression setting.

Result: Query-agnostic saturation statistics reliably separate compressed from uncompressed tokens but show limited overflow detection capability. Lightweight probing classifiers achieve 0.72 AUC-ROC on average for overflow detection across HotpotQA, SQuADv2, and TriviaQA datasets, demonstrating that incorporating query information improves detection performance.

Conclusion: The research advances from query-independent diagnostics to query-aware detectors, enabling low-cost pre-LLM gating to mitigate compression-induced errors in soft compression architectures for efficient long-context processing.

Abstract: Efficient long-context processing remains a crucial challenge for contemporary large language models (LLMs), especially in resource-constrained environments. Soft compression architectures promise to extend effective context length by replacing long token sequences with smaller sets of learned compressed tokens. Yet, the limits of compressibility – and when compression begins to erase task-relevant content – remain underexplored. In this paper, we define \emph{token overflow} as a regime in which compressed representations no longer contain sufficient information to answer a given query, and propose a methodology to characterize and detect it. In the xRAG soft-compression setting, we find that query-agnostic saturation statistics reliably separate compressed from uncompressed token representations, providing a practical tool for identifying compressed tokens but showing limited overflow detection capability. Lightweight probing classifiers over both query and context xRAG representations detect overflow with 0.72 AUC-ROC on average on HotpotQA, SQuADv2, and TriviaQA datasets, demonstrating that incorporating query information improves detection performance. These results advance from query-independent diagnostics to query-aware detectors, enabling low-cost pre-LLM gating to mitigate compression-induced errors.

Method: Developed a curriculum focusing on four core areas: 1) vector embeddings, 2) technical foundations of neural networks, 3) tokenization, and 4) transformer neural networks. Tested the curriculum in an AI-focused MA course at the Institute of Translation and Multilingual Communication.

Result: The curriculum showed didactic effectiveness but participant feedback indicated it should be embedded into higher-level didactic scaffolding (e.g., lecturer support) for optimal learning conditions.

Conclusion: The technical curriculum successfully addresses AI literacy needs in the L&T industry but requires proper scaffolding and support structures to maximize learning outcomes.

Abstract: This paper presents a technical curriculum on language-oriented artificial intelligence (AI) in the language and translation (L&T) industry. The curriculum aims to foster domain-specific technical AI literacy among stakeholders in the fields of translation and specialised communication by exposing them to the conceptual and technical/algorithmic foundations of modern language-oriented AI in an accessible way. The core curriculum focuses on 1) vector embeddings, 2) the technical foundations of neural networks, 3) tokenization and 4) transformer neural networks. It is intended to help users develop computational thinking as well as algorithmic awareness and algorithmic agency, ultimately contributing to their digital resilience in AI-driven work environments. The didactic suitability of the curriculum was tested in an AI-focused MA course at the Institute of Translation and Multilingual Communication at TH Koeln. Results suggest the didactic effectiveness of the curriculum, but participant feedback indicates that it should be embedded into higher-level didactic scaffolding - e.g., in the form of lecturer support - in order to enable optimal learning conditions.

Method: Proposes trajectory self-distillation framework with Direct Discriminative Optimization (DDO), a reverse-KL objective that promotes mode-seeking distillation, encouraging student models to focus on high-probability teacher modes for improved few-step decoding.

Result: Consistently outperforms strong few-step baselines and standard training under tight step budgets across benchmarks. Substantially narrows the gap with full-step decoding, establishing strong foundation for practical few-step DLLMs.

Conclusion: The trajectory self-distillation framework with DDO effectively improves few-step decoding in diffusion LLMs, balancing efficiency and quality, making progress toward practical deployment of efficient text generation models.

Abstract: Diffusion large language models (DLLMs) have the potential to enable fast text generation by decoding multiple tokens in parallel. However, in practice, their inference efficiency is constrained by the need for many refinement steps, while aggressively reducing the number of steps leads to a substantial degradation in generation quality. To alleviate this, we propose a trajectory self-distillation framework that improves few-step decoding by distilling the model’s own generative trajectories. We incorporate Direct Discriminative Optimization (DDO), a reverse-KL objective that promotes mode-seeking distillation and encourages the student to concentrate on high-probability teacher modes. Across benchmarks, our approach consistently outperforms strong few-step baselines and standard training under tight step budgets. Although full-step decoding remains superior, we substantially narrow the gap, establishing a strong foundation towards practical few-step DLLMs. The source code is available at https://github.com/Tyrion58/T3D.

Method: Proposes On-Policy Context Distillation (OPCD) framework that trains student models on their own generated trajectories while minimizing reverse KL divergence against context-conditioned teachers. Uses on-policy distillation approach where student learns from its own behavior rather than teacher demonstrations.

Result: OPCD outperforms baseline methods across mathematical reasoning, text-based games, and domain-specific tasks, achieving higher task accuracy while better preserving out-of-distribution capabilities. Enables effective cross-size distillation where smaller models internalize experiential knowledge from larger teachers.

Conclusion: OPCD provides an effective framework for internalizing contextual knowledge into model parameters, enabling more capable and efficient language models that can operate without explicit context while maintaining strong performance.

Abstract: Context distillation enables language models to internalize in-context knowledge into their parameters. In our work, we propose On-Policy Context Distillation (OPCD), a framework that bridges on-policy distillation with context distillation by training a student model on its own generated trajectories while minimizing reverse Kullback-Leibler divergence against a context-conditioned teacher. We demonstrate the effectiveness of OPCD on two important applications: experiential knowledge distillation, where models extract and consolidate transferable knowledge from their historical solution traces, and system prompt distillation, where models internalize beneficial behaviors encoded in optimized prompts. Across mathematical reasoning, text-based games, and domain-specific tasks, OPCD consistently outperforms baseline methods, achieving higher task accuracy while better preserving out-of-distribution capabilities. We further show that OPCD enables effective cross-size distillation, where smaller student models can internalize experiential knowledge from larger teachers.

Method: Systematic review and analysis of 67 relevant papers, categorizing studies across dimensions including multilingual datasets, cross-lingual resources, and CLTL strategies. Classifies three main transfer approaches: instance, feature, and parameter transfer.

Result: First holistic overview focusing exclusively on cross-lingual scenarios in offensive language detection, with comprehensive categorization of approaches and identification of current challenges and future research opportunities.

Conclusion: Provides foundational survey of CLTL techniques for offensive language detection, establishes categorization framework, identifies research gaps, and makes survey resources publicly available for community reference.

Abstract: The growing prevalence and rapid evolution of offensive language in social media amplify the complexities of detection, particularly highlighting the challenges in identifying such content across diverse languages. This survey presents a systematic and comprehensive exploration of Cross-Lingual Transfer Learning (CLTL) techniques in offensive language detection in social media. Our study stands as the first holistic overview to focus exclusively on the cross-lingual scenario in this domain. We analyse 67 relevant papers and categorise these studies across various dimensions, including the characteristics of multilingual datasets used, the cross-lingual resources employed, and the specific CLTL strategies implemented. According to “what to transfer”, we also summarise three main CLTL transfer approaches: instance, feature, and parameter transfer. Additionally, we shed light on the current challenges and future research opportunities in this field. Furthermore, we have made our survey resources available online, including two comprehensive tables that provide accessible references to the multilingual datasets and CLTL methods used in the reviewed literature.

Method: Add early exit heads at intermediate transformer layers trained in self-supervised manner to mimic main model predictions. Use entropy as confidence metric to determine when to stop computation early. Adapt approach to speculative decoding as Dynamic Self-Speculative Decoding (DSSD).

Result: Significantly reduces inference cost while maintaining accuracy across multiple benchmarks on Pythia models (70M to 2.8B parameters). DSSD achieves 1.66x higher token acceptance than manually-tuned LayerSkip baselines with minimal hyperparameter tuning.

Conclusion: Modular early exit with entropy-based confidence thresholds provides effective inference acceleration for LLMs, and the approach can be successfully adapted to speculative decoding for further efficiency gains.

Abstract: This paper presents a modular approach to accelerate inference in large language models (LLMs) by adding early exit heads at intermediate transformer layers. Each head is trained in a self-supervised manner to mimic the main model’s predictions, allowing computation to stop early when a calibrated confidence threshold is reached. We evaluate several confidence metrics and show that entropy provides the most reliable separation between correct and incorrect predictions. Experiments on the Pythia model suite (70M to 2.8B parameters) demonstrate that our method significantly reduces inference cost while maintaining accuracy across multiple benchmarks. We further adapt this approach to speculative decoding, introducing Dynamic Self-Speculative Decoding (DSSD), which achieves 1.66x higher token acceptance than manually-tuned LayerSkip baselines with minimal hyperparameter tuning.

Method: Developed LabSafety Bench benchmark with 765 multiple-choice questions and 404 realistic lab scenarios (3,128 open-ended tasks) to evaluate models on hazard identification, risk assessment, and consequence prediction across 19 advanced LLMs/VLMs.

Result: No model surpassed 70% accuracy on hazard identification; proprietary models performed well on structured assessments but showed no clear advantage in open-ended reasoning tasks.

Conclusion: Urgent need for specialized safety evaluation frameworks before deploying AI systems in real laboratory settings due to current models’ inadequate reliability.

Abstract: Artificial Intelligence (AI) is revolutionizing scientific research, yet its growing integration into laboratory environments presents critical safety challenges. Large language models (LLMs) and vision language models (VLMs) now assist in experiment design and procedural guidance, yet their “illusion of understanding” may lead researchers to overtrust unsafe outputs. Here we show that current models remain far from meeting the reliability needed for safe laboratory operation. We introduce LabSafety Bench, a comprehensive benchmark that evaluates models on hazard identification, risk assessment, and consequence prediction across 765 multiple-choice questions and 404 realistic lab scenarios, encompassing 3,128 open-ended tasks. Evaluations on 19 advanced LLMs and VLMs show that no model evaluated on hazard identification surpasses 70% accuracy. While proprietary models perform well on structured assessments, they do not show a clear advantage in open-ended reasoning. These results underscore the urgent need for specialized safety evaluation frameworks before deploying AI systems in real laboratory settings.

Method: Curated 40,000 two-person informational interviews from NPR and CNN, analyzed LLM behavior compared to humans, developed a realistic simulated environment with source personas and persuasive elements, and tested interviewer LLMs’ information extraction capabilities.

Result: LLMs are significantly less likely than human interviewers to use acknowledgements and pivot to higher-level questions. Source LLMs mimic human information sharing, but interviewer LLMs struggle with recognizing answered questions and engaging persuasively, leading to suboptimal information extraction across model sizes.

Conclusion: There is a fundamental deficit in LLMs’ multi-turn planning and strategic thinking capabilities, highlighting the need for enhancing strategic dialogue abilities in language models.

Abstract: Large Language Models (LLMs) have demonstrated impressive capabilities in generating coherent text but often struggle with grounding language and strategic dialogue. To address this gap, we focus on journalistic interviews, a domain rich in grounding communication and abundant in data. We curate a dataset of 40,000 two-person informational interviews from NPR and CNN, and reveal that LLMs are significantly less likely than human interviewers to use acknowledgements and to pivot to higher-level questions. Realizing that a fundamental deficit exists in multi-turn planning and strategic thinking, we develop a realistic simulated environment, incorporating source personas and persuasive elements, in order to facilitate the development of agents with longer-horizon rewards. Our experiments show that while source LLMs mimic human behavior in information sharing, interviewer LLMs struggle with recognizing when questions are answered and engaging persuasively, leading to suboptimal information extraction across model size and capability. These findings underscore the need for enhancing LLMs’ strategic dialogue capabilities.

Method: TATA incorporates base-LLM-aware data selection during supervised fine-tuning to tailor training data to each model’s unique abilities, enabling LLMs to autonomously determine and apply appropriate reasoning strategies (CoT or TIR) at test time.

Result: Experiments on six mathematical reasoning benchmarks show TATA effectively combines complementary strengths of CoT and TIR, achieving superior or comparable performance with improved inference efficiency compared to TIR alone.

Conclusion: Aptitude-aware data selection is critical for enabling LLMs to make effective adaptive reasoning decisions and align strategies with model capabilities, demonstrating LLMs can autonomously adapt reasoning strategies based on their inherent abilities.

Abstract: Existing approaches to mathematical reasoning with large language models (LLMs) rely on Chain-of-Thought (CoT) for generalizability or Tool-Integrated Reasoning (TIR) for precise computation. While efforts have been made to combine these methods, they primarily rely on post-selection or predefined strategies, leaving an open question: whether LLMs can autonomously adapt their reasoning strategy based on their inherent capabilities. In this work, we propose TATA (Teaching LLMs According to Their Aptitude), an adaptive framework that enables LLMs to personalize their reasoning strategy spontaneously, aligning it with their intrinsic aptitude. TATA incorporates base-LLM-aware data selection during supervised fine-tuning (SFT) to tailor training data to the model’s unique abilities. This approach equips LLMs to autonomously determine and apply the appropriate reasoning strategy at test time. We evaluate TATA through extensive experiments on six mathematical reasoning benchmarks, using both general-purpose and math-specialized LLMs. Empirical results demonstrate that TATA effectively combines the complementary strengths of CoT and TIR, achieving superior or comparable performance with improved inference efficiency compared to TIR alone. Further analysis underscores the critical role of aptitude-aware data selection in enabling LLMs to make effective and adaptive reasoning decisions and align reasoning strategies with model capabilities.

Method: Trains explanation-generating LLMs via reinforcement learning from AI feedback, using distributional rewards generated by generative continuous normalizing flows (CNFs) that capture probabilistic human judgments. CNFs provably bound deviations from true human reward distributions when trained on noisy proxy rewards from LLMs.

Result: Human and LLM evaluators find the method delivers explanations that enable more accurate predictions of true agent decisions, exhibit greater logical soundness and actionability, and impose lower cognitive load than baselines.

Conclusion: The CNF-based framework provides a principled approach to training explanation-generating LLMs that better capture human preferences and produce higher-quality explanations than existing methods.

Abstract: As humans increasingly share environments with diverse agents powered by RL, LLMs, and beyond, the ability to explain agent policies in natural language is vital for reliable coexistence. We introduce a general-purpose framework that trains explanation-generating LLMs via reinforcement learning from AI feedback, with distributional rewards generated by generative continuous normalizing flows (CNFs). CNFs capture the pluralistic and probabilistic nature of human judgments about explanations. Moreover, under mild assumptions, CNFs provably bound deviations from true human reward distributions when trained on noisy proxy rewards from LLMs. We design a specialized CNF architecture that selectively attends to linguistic cues in the decision context and explanations when generating rewards. Human and LLM evaluators find that our method delivers explanations that enable more accurate predictions of true agent decisions, exhibit greater logical soundness and actionability, and impose lower cognitive load than explanations trained with proxy LLM rewards or state-of-the-art RLHF and RLAIF baselines.

Method: Uses manifold learning and spectral clustering to group recovery instructions in semantic space, adaptively allocates data budget across clusters based on capability degradation, prioritizes samples causing the most performance decline, and filters conflicting/irrelevant data.

Result: PASER significantly outperforms conventional baselines, effectively recovering general capabilities of pruned LLMs while using only 4%-20% of original post-training data.

Conclusion: PASER provides an efficient data selection method for recovering pruned LLM capabilities with minimal data usage, addressing key limitations of existing approaches.

Abstract: Model pruning is an effective approach for compressing large language models (LLMs). However, this process often leads to significant degradation of model capabilities. While post-training techniques such as instruction tuning are commonly employed to recover model performance, existing methods often overlook the uneven deterioration of model capabilities and incur high computational costs. Moreover, some irrelevant instructions may also introduce negative effects to model capacity recovery. To address these challenges, we propose the \textbf{P}ost-training d\textbf{A}ta \textbf{S}election method for \textbf{E}fficient pruned large language model \textbf{R}ecovery (\textbf{PASER}). PASER aims to identify instructions to recover the most compromised model capacities with a certain data budget. Our approach first applies manifold learning and spectral clustering to group recovery instructions in the semantic space, revealing capability-specific instruction sets. Then, the data budget is adaptively allocated across clusters by the degree of corresponding model capability degradation. In each cluster, we prioritize data samples that lead to the most decline of model performance. To mitigate potential negative tuning effects, we also detect and filter out conflicting or irrelevant recovery data. Extensive experiments demonstrate that PASER significantly outperforms conventional baselines, effectively recovering the general capabilities of pruned LLMs while utilizing merely 4%-20% of the original post-training data. We provide the code repository in \href{https://github.com/BokwaiHo/PASER}{Link}.

Method: Proposes SURE framework to empirically quantify RAG robustness against spurious features, including comprehensive taxonomy, evaluation metrics, and data synthesis pipeline for training-based robustness improvements.

Result: Analysis reveals spurious features are widespread and challenging in RAG systems, with the framework providing tools for evaluation and improvement of robustness against such implicit noise.

Conclusion: Spurious features represent a significant robustness issue in RAG that requires systematic study, and the SURE framework provides essential tools for quantifying and addressing this problem.

Abstract: Robustness has become a critical attribute for the deployment of RAG systems in real-world applications. Existing research focuses on robustness to explicit noise (e.g., document semantics) but overlooks implicit noise (spurious features). Moreover, previous studies on spurious features in LLMs are limited to specific types (e.g., formats) and narrow scenarios (e.g., ICL). In this work, we identify and study spurious features in the RAG paradigm, a robustness issue caused by the sensitivity of LLMs to semantic-agnostic features. We then propose a novel framework, SURE, to empirically quantify the robustness of RALMs against spurious features. Beyond providing a comprehensive taxonomy and metrics for evaluation, the framework’s data synthesis pipeline facilitates training-based strategies to improve robustness. Further analysis suggests that spurious features are a widespread and challenging problem in the field of RAG. Our code is available at https://github.com/maybenotime/RAG-SpuriousFeatures .

Method: Uses Bi-Mamba architecture (state space modeling) with linear-time complexity, pretrained on nearly 30 million cells. Incorporates biologically informed objectives including pathway-aware contrastive loss and rank-based gene encoding.

Result: Demonstrates strong performance across diverse tasks including multi-batch integration, cell type annotation, and gene-gene correlation, with substantial computational gains over transformer baselines.

Conclusion: GeneMamba positions itself as a practical and powerful alternative to transformer-based methods for large-scale single-cell data analysis, offering better scalability and biologically grounded modeling.

Abstract: Single-cell RNA sequencing (scRNA-seq) enables high-resolution analysis of cellular heterogeneity, but its complexity, which is marked by high dimensionality, sparsity, and batch effects, which poses major computational challenges. Transformer-based models have made significant advances in this domain but are often limited by their quadratic complexity and suboptimal handling of long-range dependencies. In this work, we introduce GeneMamba, a scalable and efficient foundation model for single-cell transcriptomics built on state space modeling. Leveraging the Bi-Mamba architecture, GeneMamba captures bidirectional gene context with linear-time complexity, offering substantial computational gains over transformer baselines. The model is pretrained on nearly 30 million cells and incorporates biologically informed objectives, including pathway-aware contrastive loss and rank-based gene encoding. We evaluate GeneMamba across diverse tasks, including multi-batch integration, cell type annotation, and gene-gene correlation, demonstrating strong performance, interpretability, and robustness. These results position GeneMamba as a practical and powerful alternative to transformer-based methods, advancing the development of biologically grounded, scalable tools for large-scale single-cell data analysis.

Method: The paper constructs MEMENTO, an end-to-end two-stage evaluation framework with single-memory and joint-memory tasks. Through experiments revealing current agents’ limitations, the authors identify bottlenecks of information overload and coordination failures. They then design a hierarchical knowledge graph-based user-profile memory module that separately manages personalized knowledge.

Result: Experiments show current agents can recall simple object semantics but struggle to apply sequential user patterns to planning. The proposed hierarchical knowledge graph-based memory module achieves substantial improvements on both single and joint-memory tasks compared to baseline approaches.

Conclusion: Personalized assistance for embodied agents requires better memory architectures that can handle both object semantics and user patterns. The hierarchical knowledge graph approach effectively addresses information overload and coordination failures, demonstrating the importance of episodic memory for personalized knowledge and in-context learning.

Abstract: LLM-powered embodied agents have shown success on conventional object-rearrangement tasks, but providing personalized assistance that leverages user-specific knowledge from past interactions presents new challenges. We investigate these challenges through the lens of agents’ memory utilization along two critical dimensions: object semantics (identifying objects based on personal meaning) and user patterns (recalling sequences from behavioral routines). To assess these capabilities, we construct MEMENTO, an end-to-end two-stage evaluation framework comprising single-memory and joint-memory tasks. Our experiments reveal that current agents can recall simple object semantics but struggle to apply sequential user patterns to planning. Through in-depth analysis, we identify two critical bottlenecks: information overload and coordination failures when handling multiple memories. Based on these findings, we explore memory architectural approaches to address these challenges. Given our observation that episodic memory provides both personalized knowledge and in-context learning benefits, we design a hierarchical knowledge graph-based user-profile memory module that separately manages personalized knowledge, achieving substantial improvements on both single and joint-memory tasks. Project website: https://connoriginal.github.io/MEMENTO

Method: Systematically compare model completions to human responses across 16 language models, 3 dialogue corpora, and various stylometric features to measure convergence patterns.

Result: Models strongly converge to conversation style, often overfitting relative to human baseline. Convergence patterns are feature-specific, with instruction-tuned and larger models converging less than pretrained/smaller counterparts.

Conclusion: Human and model convergence patterns differ, suggesting different underlying mechanisms. Models show systematic but distinct convergence behavior compared to humans.

Abstract: While large language models (LLMs) are generally considered proficient in generating language, how similar their language usage is to that of humans remains understudied. In this paper, we test whether models exhibit linguistic convergence, a core pragmatic element of human language communication: do models adapt, or converge, to the linguistic patterns of their user? To answer this, we systematically compare model completions of existing dialogues to original human responses across sixteen language models, three dialogue corpora, and various stylometric features. We find that models strongly converge to the conversation’s style, often significantly overfitting relative to the human baseline. While convergence patterns are often feature-specific, we observe consistent shifts in convergence across modeling settings, with instruction-tuned and larger models converging less than their pretrained and smaller counterparts. Given the differences in human and model convergence patterns, we hypothesize that the underlying mechanisms driving these behaviors are very different.

Method: Constructed a dataset of 105,880 QA pairs from 1,242 cancer patient case summaries from two medical centers. Created in-domain and cross-domain test collections. Evaluated four open-source LLMs with zero-shot prompting and supervised fine-tuning, plus two API-based models (GPT-5.2 and Gemini 3 Flash).

Result: Fine-tuned models significantly outperformed zero-shot models, showing pretrained models fail to generalize on MedQARo. The benchmark enables precise assessment of generalization across medical centers and cancer types.

Conclusion: Domain-specific and language-specific fine-tuning is crucial for reliable clinical QA in Romanian. The MedQARo benchmark fills an important gap for evaluating medical LLMs in Romanian language.

Abstract: We introduce MedQARo, the first large-scale medical QA benchmark in Romanian, alongside a comprehensive evaluation of state-of-the-art large language models (LLMs). We construct a high-quality and large-scale dataset comprising 105,880 QA pairs about cancer patients from two medical centers. The questions regard medical case summaries of 1,242 patients, requiring both keyword extraction and reasoning. Our benchmark contains both in-domain and cross-domain (cross-center and cross-cancer) test collections, enabling a precise assessment of generalization capabilities. We experiment with four open-source LLMs from distinct families of models on MedQARo. Each model is employed in two scenarios: zero-shot prompting and supervised fine-tuning. We also evaluate two state-of-the-art LLMs exposed only through APIs, namely GPT-5.2 and Gemini 3 Flash. Our results show that fine-tuned models significantly outperform zero-shot models, indicating that pretrained models fail to generalize on MedQARo. Our findings demonstrate the importance of both domain-specific and language-specific fine-tuning for reliable clinical QA in Romanian.

Method: Detects behavior-associated experts by comparing activation frequencies between paired inputs with opposite behaviors (e.g., safe vs. unsafe). During inference, selectively activates or deactivates these key experts to control model behavior without modifying weights.

Result: Across 11 benchmarks and 6 LLMs, steering increased safety by up to +20% and faithfulness by +27%. Unsafe steering decreased safety by -41% alone, and -100% when combined with jailbreak methods, bypassing all safety guardrails.

Conclusion: SteerMoE provides effective test-time control for MoE LLMs, demonstrating both safety enhancement capabilities and revealing unique vulnerabilities in MoE architectures that could be exploited.

Abstract: Mixture-of-Experts (MoE) in Large Language Models (LLMs) routes each token through a subset of specialized Feed-Forward Networks (FFN), known as experts. We present SteerMoE, a framework to steer MoE models by detecting and controlling behavior-associated experts. We detect key experts by comparing how often they activate between paired inputs that demonstrate opposite behaviors (e.g., safe vs. unsafe). By selectively activating or deactivating such experts during inference, we control behaviors like faithfulness and safety without fine-tuning. Across 11 benchmarks and 6 LLMs, our steering raises safety by up to +20% and faithfulness by +27%. Alternatively, unsafe steering drops safety by -41% alone, and -100% when combined with existing jailbreak methods, bypassing all safety guardrails. Overall, SteerMoE offers a lightweight, effective, and widely applicable test-time control, while revealing unique vulnerabilities in MoE LLMs. https://github.com/adobe-research/SteerMoE

Method: CoSpaDi replaces low-rank factorization with structured sparse decomposition where each weight matrix is represented as a dense dictionary multiplied by a column-sparse coefficient matrix. This creates a union-of-subspaces model. The method is calibration-guided, using a small calibration set to optimize factorization to minimize functional reconstruction error rather than weight-space error. It includes activation-derived Gram orthonormalization to reformulate the data-aware objective into standard dictionary learning, and supports both per-layer compression and cross-layer dictionary sharing.

Result: Across Llama and Qwen model families, CoSpaDi consistently improves accuracy-compression and perplexity-compression trade-offs over state-of-the-art SVD-based baselines and strong structured pruning baselines at 20-40% compression ratios. The structured sparsity enables sparse-dense computation and integrates with post-training quantization of sparse coefficients.

Conclusion: CoSpaDi provides an effective training-free compression framework for LLMs that outperforms existing low-rank approximation methods by using structured sparse decomposition, offering better expressiveness at fixed parameter budgets while maintaining computational efficiency through sparse-dense computation patterns.

Abstract: Post-training compression of large language models (LLMs) often relies on low-rank weight approximations that represent each column of the weight matrix in a shared low-dimensional subspace. This strategy is computationally efficient but the underlying constraint can be overly rigid for heterogeneous projection weights and may incur avoidable accuracy loss. We propose CoSpaDi (Compression via Sparse Dictionary Learning), a training-free framework that replaces low-rank factorization with a structured sparse decomposition in which each weight matrix is represented as a dense dictionary multiplied by a column-sparse coefficient matrix. This yields a union-of-subspaces model: the columns of the weight matrix are represented as linear combinations of different subsets of dictionary atoms, improving expressiveness at a fixed parameter budget. CoSpaDi is calibration-guided: using a small calibration set, we optimize the factorization to minimize functional reconstruction error of layer outputs rather than weight-space error. An activation-derived Gram orthonormalization reformulates this data-aware objective into a standard dictionary learning problem on transformed weights, and we support both per-layer compression and cross-layer dictionary sharing within groups of similar projections. Across Llama and Qwen model families, CoSpaDi consistently improves the accuracy–compression and perplexity–compression trade-offs over state-of-the-art SVD-based baselines and strong structured pruning baselines at 20-40% compression ratios. The resulting structured sparsity enables sparse–dense computation and integrates with post-training quantization of the sparse coefficients.

Method: Proposes a rule-driven routing framework with three components: (1) routing agent that scores candidate augmentation paths (database vs. document retrieval) based on explicit rules, (2) rule-making expert agent that refines rules over time using QA feedback, and (3) path-level meta-cache that reuses past routing decisions for semantically similar queries to reduce latency and cost.

Result: Experiments on three QA benchmarks show the framework consistently outperforms static strategies and learned routing baselines, achieving higher accuracy while maintaining moderate computational cost.

Conclusion: The rule-driven routing framework effectively balances database and document retrieval for domain-specific QA, demonstrating that systematic rule-based routing guided by query patterns can optimize both effectiveness and efficiency in retrieval-augmented generation systems.

Abstract: Large Language Models (LLMs) have shown remarkable performance on general Question Answering (QA), yet they often struggle in domain-specific scenarios where accurate and up-to-date information is required. Retrieval-Augmented Generation (RAG) addresses this limitation by enriching LLMs with external knowledge, but existing systems primarily rely on unstructured documents, while largely overlooking relational databases, which provide precise, timely, and efficiently queryable factual information, serving as indispensable infrastructure in domains such as finance, healthcare, and scientific research. Motivated by this gap, we conduct a systematic analysis that reveals three central observations: (i) databases and documents offer complementary strengths across queries, (ii) naively combining both sources introduces noise and cost without consistent accuracy gains, and (iii) selecting the most suitable source for each query is crucial to balance effectiveness and efficiency. We further observe that query types show consistent regularities in their alignment with retrieval paths, suggesting that routing decisions can be effectively guided by systematic rules that capture these patterns. Building on these insights, we propose a rule-driven routing framework. A routing agent scores candidate augmentation paths based on explicit rules and selects the most suitable one; a rule-making expert agent refines the rules over time using QA feedback to maintain adaptability; and a path-level meta-cache reuses past routing decisions for semantically similar queries to reduce latency and cost. Experiments on three QA benchmarks demonstrate that our framework consistently outperforms static strategies and learned routing baselines, achieving higher accuracy while maintaining moderate computational cost.

Method: Uses Bayesian framework with two key components: (1) signal anchoring - reordering documents to concentrate high-belief items in highly diagnostic positions, (2) belief updating - updating document relevance beliefs from model outputs. Implements iterative assignment based on model’s diagnosticity profile with O(log N) complexity.

Result: On needle-in-haystack retrieval and long-context QA, matches Permutation Self-Consistency’s target identification with 30-65% fewer queries. Remains effective under calibration mismatch, suggesting coarse positional ordering drives performance gains.

Conclusion: Inherent model biases can be leveraged as tools for control rather than being seen as failures. The framework demonstrates practical black-box mitigation of position bias in LLMs for long-context problems.

Abstract: Large language models (LLMs) exhibit pronounced position bias in long-context needle-in-haystack problems, systematically prioritizing the location of information over its relevance. While current mitigations rely on white-box access, this is effectively impossible for many state-of-the-art models. We introduce GOLD PANNING, a black-box Bayesian framework that performs inference-time active search over long contexts by (i) reordering documents to concentrate high-belief items in highly diagnostic positions (signal anchoring) and (ii) updating beliefs over document relevance from model outputs. Unlike conventional active learning, which prioritizes uncertainty reduction, GOLD PANNING leverages anchoring – once flagged, keep it in sight – to preserve weak cues. We implement this using iterative assignment derived from the model’s diagnosticity profile, which provably identifies a target among $N$ documents in $O(\log N)$ rounds, ensuring scalability to many-document settings.On needle-in-a-haystack retrieval and long-context QA, GOLD PANNING matches Permutation Self-Consistency’s target identification with $30–65%$ fewer queries and remains effective under calibration mismatch, suggesting coarse positional ordering drives performance gains. These results demonstrate that inherent model biases need not be failures, but can be used as tools for control.

Method: Multi-modal approach combining audio and visual cues for speaker diarization and transcription in overlapping conversations. Data collection focuses on unscripted, casual group chats with highly fragmented conversational turns.

Result: Audio-only baselines perform poorly (exceed 100% word error rate), while incorporating visual cues yields substantial 50% improvements, demonstrating the critical importance of multi-modality for this task.

Conclusion: Multi-modal context-aware recognition is essential for solving the cocktail-party problem in natural conversational settings, with visual cues providing crucial information that audio alone cannot capture.

Abstract: We introduce the task of Multi-Modal Context-Aware Recognition (MCoRec) in the ninth CHiME Challenge, which addresses the cocktail-party problem of overlapping conversations in a single-room setting using audio, visual, and contextual cues. MCoRec captures natural multi-party conversations where the recordings focus on unscripted, casual group chats, leading to extreme speech overlap of up to 100% and highly fragmented conversational turns. The task requires systems to answer the question “Who speaks when, what, and with whom?” by jointly transcribing each speaker’s speech and clustering them into their respective conversations from audio-visual recordings. Audio-only baselines exceed 100% word error rate, whereas incorporating visual cues yields substantial 50% improvements, highlighting the importance of multi-modality. In this manuscript, we present the motivation behind the task, outline the data collection process, and report the baseline systems developed for the MCoRec.

Method: Created cross-lingual benchmark covering Cantonese, Japanese, and Turkish with 4 tasks (QA, summarization, translation, culturally grounded dialogue). Evaluated 7 LLMs using both human assessments (fluency, factual accuracy, cultural appropriateness) and automated metrics (BLEU, ROUGE).

Result: Proprietary models (GPT-4o, GPT-4, Claude 3.5) lead across languages/tasks but show significant gaps in cultural understanding and morphological generalization. GPT-4o excels in multilingual performance, Claude 3.5 in knowledge/reasoning. All models struggle with language-specific challenges like Turkish agglutination and Cantonese colloquialisms. Open-source models lag substantially.

Conclusion: Significant disparities exist between proprietary and open-source models for low-resource languages. Current LLMs need improvement in cultural awareness and linguistic generalization. Released benchmark enables reproducible research toward more equitable multilingual AI.

Abstract: Large language models (LLMs) have achieved impressive results in high-resource languages like English, yet their effectiveness in low-resource and morphologically rich languages remains underexplored. In this paper, we present a comprehensive evaluation of seven cutting-edge LLMs – including GPT-4o, GPT-4, Claude3.5Sonnet, LLaMA3.1, MistralLarge2, LLaMA-2Chat13B, and Mistral7BInstruct – on a new cross-lingual benchmark covering \textbf{Cantonese, Japanese, and Turkish}. Our benchmark spans four diverse tasks: open-domain question answering, document summarization, English-to-X translation, and culturally grounded dialogue. We combine \textbf{human evaluations} (rating fluency, factual accuracy, and cultural appropriateness) with automated metrics (e.g., BLEU, ROUGE) to assess model performance. Our results reveal that while the largest proprietary models (GPT-4o, GPT-4, Claude3.5) generally lead across languages and tasks, significant gaps persist in culturally nuanced understanding and morphological generalization. Notably, GPT-4o demonstrates robust multilingual performance even on cross-lingual tasks, and Claude3.5Sonnet achieves competitive accuracy on knowledge and reasoning benchmarks. However, all models struggle to some extent with the unique linguistic challenges of each language, such as Turkish agglutinative morphology and Cantonese colloquialisms. Smaller open-source models (LLaMA-213B, Mistral7B) lag substantially in fluency and accuracy, highlighting the resource disparity. We provide detailed quantitative results, qualitative error analysis, and discuss implications for developing more culturally aware and linguistically generalizable LLMs. Our benchmark and evaluation data are released to foster reproducibility and further research.

Method: Uses parameter-efficient continual pretraining via LoRA on Qwen-3 4B backbone; replaces and adapts tokenizer for better Hungarian tokenization; trains on 160B tokens with 44% Hungarian, 24% English, 21% German, 11% code mixture to prevent catastrophic forgetting.

Result: Achieves substantially improved tokenization fertility for Hungarian while maintaining competitive performance in English and German; shows modest but stable results in language adaptation.

Conclusion: Racka demonstrates a practical approach to creating resource-efficient multilingual LLMs for lower-resource languages while preserving high-resource language capabilities.

Abstract: We present Racka, a lightweight, continually pretrained large language model designed to bridge the resource gap between Hungarian and high-resource languages such as English and German. Racka employs parameter-efficient continual pretraining via Low-Rank Adaptation (LoRA) on a Qwen-3 4B backbone, making the recipe practical on A100 (40GB)-based HPC clusters with low inter-node bandwidth. To better match the training distribution, we replace and adapt the tokenizer, achieving substantially improved tokenization fertility for Hungarian while maintaining competitive performance in English and German. The model is trained on 160B subword tokens drawn from a mixture of internet and high-quality curated sources, with a composition of 44% Hungarian, 24% English, 21% German, and 11% code. This data mix is chosen to mitigate catastrophic forgetting and preserve high-resource language capabilities during continual pretraining. Our preliminary results indicate modest but stable results in language adaptation.

Method: MemRL decouples stable reasoning from plastic memory and uses a Two-Phase Retrieval mechanism to filter noise and identify high-utility strategies through environmental feedback. It employs reinforcement learning on episodic memory in a non-parametric way.

Result: Extensive experiments on HLE, BigCodeBench, ALFWorld, and Lifelong Agent Bench show MemRL significantly outperforms state-of-the-art baselines, effectively reconciling the stability-plasticity dilemma.

Conclusion: MemRL enables continuous runtime improvement without weight updates, providing a solution to the stability-plasticity dilemma in AI agents through memory-based reinforcement learning.

Abstract: The hallmark of human intelligence is the self-evolving ability to master new skills by learning from past experiences. However, current AI agents struggle to emulate this self-evolution: fine-tuning is computationally expensive and prone to catastrophic forgetting, while existing memory-based methods rely on passive semantic matching that often retrieves noise. To address these challenges, we propose MemRL, a non-parametric approach that evolves via reinforcement learning on episodic memory. By decoupling stable reasoning from plastic memory, MemRL employs a Two-Phase Retrieval mechanism to filter noise and identify high-utility strategies through environmental feedback. Extensive experiments on HLE, BigCodeBench, ALFWorld, and Lifelong Agent Bench demonstrate that MemRL significantly outperforms state-of-the-art baselines, confirming that MemRL effectively reconciles the stability-plasticity dilemma, enabling continuous runtime improvement without weight updates. Code is available at https://github.com/MemTensor/MemRL.

Method: CSE consists of three key components: 1) Diversified Planning Initialization generates structurally distinct algorithmic strategies for broad solution space coverage, 2) Genetic Evolution replaces stochastic operations with feedback-guided mechanisms for targeted mutation and compositional crossover, and 3) Hierarchical Evolution Memory captures both successful and failed experiences at inter-task and intra-task levels.

Result: Experiments on EffiBench-X demonstrate that CSE consistently outperforms all baselines across various LLM backbones, achieves higher efficiency from early generations, and maintains continuous improvement throughout evolution.

Conclusion: CSE effectively addresses the exploration efficiency bottlenecks in self-evolution methods for code generation through controlled evolutionary mechanisms and comprehensive experience utilization.

Abstract: Self-evolution methods enhance code generation through iterative “generate-verify-refine” cycles, yet existing approaches suffer from low exploration efficiency, failing to discover solutions with superior complexity within limited budgets. This inefficiency stems from initialization bias trapping evolution in poor solution regions, uncontrolled stochastic operations lacking feedback guidance, and insufficient experience utilization across tasks. To address these bottlenecks, we propose Controlled Self-Evolution (CSE), which consists of three key components. Diversified Planning Initialization generates structurally distinct algorithmic strategies for broad solution space coverage. Genetic Evolution replaces stochastic operations with feedback-guided mechanisms, enabling targeted mutation and compositional crossover. Hierarchical Evolution Memory captures both successful and failed experiences at inter-task and intra-task levels. Experiments on EffiBench-X demonstrate that CSE consistently outperforms all baselines across various LLM backbones. Furthermore, CSE achieves higher efficiency from early generations and maintains continuous improvement throughout evolution. Our code is publicly available at https://github.com/QuantaAlpha/EvoControl.

Method: Two approaches: 1) Training-free LLM-in-Sandbox where strong LLMs spontaneously use sandbox capabilities for non-code tasks, and 2) LLM-in-Sandbox Reinforcement Learning (LLM-in-Sandbox-RL) that trains models for sandbox exploration using only non-agentic data.

Result: LLM-in-Sandbox achieves robust generalization across mathematics, physics, chemistry, biomedicine, long-context understanding, and instruction following. The system is analyzed for computational efficiency and released as an open-source Python package.

Conclusion: Code sandboxes can effectively elicit general intelligence from LLMs for non-code domains, with both training-free and post-trained approaches showing strong generalization capabilities across diverse scientific and reasoning tasks.

Abstract: We introduce LLM-in-Sandbox, enabling LLMs to explore within a code sandbox (i.e., a virtual computer), to elicit general intelligence in non-code domains. We first demonstrate that strong LLMs, without additional training, exhibit generalization capabilities to leverage the code sandbox for non-code tasks. For example, LLMs spontaneously access external resources to acquire new knowledge, leverage the file system to handle long contexts, and execute scripts to satisfy formatting requirements. We further show that these agentic capabilities can be enhanced through LLM-in-Sandbox Reinforcement Learning (LLM-in-Sandbox-RL), which uses only non-agentic data to train models for sandbox exploration. Experiments demonstrate that LLM-in-Sandbox, in both training-free and post-trained settings, achieves robust generalization spanning mathematics, physics, chemistry, biomedicine, long-context understanding, and instruction following. Finally, we analyze LLM-in-Sandbox’s efficiency from computational and system perspectives, and open-source it as a Python package to facilitate real-world deployment.

Method: Proposes CausalEmbed, an auto-regressive generation approach for constructing multi-vector embeddings using iterative margin loss during contrastive training to learn compact, well-structured representations with only dozens of visual tokens.

Result: Achieves 30-155x reduction in token count while maintaining highly competitive performance across various backbones and benchmarks, with demonstrated training efficiency and test-time scalability advantages.

Conclusion: CausalEmbed introduces a flexible test-time scaling strategy for multi-vector VDR representations and advances the generative paradigm in multimodal document retrieval, making MLLM-based VDR more practical for real-world applications.

Abstract: Although Multimodal Large Language Models (MLLMs) have shown remarkable potential in Visual Document Retrieval (VDR) through generating high-quality multi-vector embeddings, the substantial storage overhead caused by representing a page with thousands of visual tokens limits their practicality in real-world applications. To address this challenge, we propose an auto-regressive generation approach, CausalEmbed, for constructing multi-vector embeddings. By incorporating iterative margin loss during contrastive training, CausalEmbed encourages the embedding models to learn compact and well-structured representations. Our method enables efficient VDR tasks using only dozens of visual tokens, achieving a 30-155x reduction in token count while maintaining highly competitive performance across various backbones and benchmarks. Theoretical analysis and empirical results demonstrate the unique advantages of auto-regressive embedding generation in terms of training efficiency and scalability at test time. As a result, CausalEmbed introduces a flexible test-time scaling strategy for multi-vector VDR representations and sheds light on the generative paradigm within multimodal document retrieval. Our code is available at https://github.com/Z1zs/Causal-Embed.

Method: Introduce an Evaluator Quality Baseline that compares the probability a judge incorrectly votes for itself against the probability it votes for an incorrect response from another model, decoupling self-preference signals from noisy outputs on hard problems.

Result: Evaluating on 37,448 queries, only 51% of initial findings retain statistical significance after applying the corrective baseline, showing most apparent self-preference bias stems from methodological confounds.

Conclusion: The corrective baseline enables future research on self-preference by eliminating noisy data, contributing to cataloging and isolating judge-bias effects in LLM evaluation.

Abstract: Recent research has shown that large language models (LLMs) favor their own outputs when acting as judges, undermining the integrity of automated post-training and evaluation workflows. However, it is difficult to disentangle which evaluation biases are explained by narcissism versus general experimental confounds, distorting measurements of self-preference bias. We discover a core methodological confound which could reduce measurement error by 89.6%. Specifically, LLM evaluators may deliver self-preferring verdicts when the judge responds to queries which they completed incorrectly themselves; this would be true regardless of whether one of their responses is their own. To decouple self-preference signals from noisy outputs on hard problems, we introduce an Evaluator Quality Baseline, which compares the probability that a judge incorrectly votes for itself against the probability that it votes for an incorrect response from another model. Evaluating this simple baseline on 37,448 queries, only 51% of initial findings retain statistical significance. Finally, we turn towards characterizing the entropy of “easy” versus “hard” evaluation votes from LLM judges. Our corrective baseline enables future research on self-preference by eliminating noisy data from potential solutions. More widely, this work contributes to the growing body of work on cataloging and isolating judge-bias effects.

Method: Proposes Rubric-ARM framework that jointly optimizes rubric generator and judge using reinforcement learning from preference feedback. Treats rubric generation as latent action learned to maximize judgment accuracy. Uses alternating optimization strategy to mitigate non-stationarity of simultaneous updates.

Result: Achieves state-of-the-art performance on multiple benchmarks. Significantly improves downstream policy alignment in both offline and online reinforcement learning settings.

Conclusion: Rubric-ARM effectively addresses limitations of scalar reward models by learning adaptive rubrics for better quality assessment in non-verifiable domains.

Abstract: Standard reward models typically predict scalar scores that fail to capture the multifaceted nature of response quality in non-verifiable domains, such as creative writing or open-ended instruction following. To address this limitation, we propose Rubric-ARM, a framework that jointly optimizes a rubric generator and a judge using reinforcement learning from preference feedback. Unlike existing methods that rely on static rubrics or disjoint training pipelines, our approach treats rubric generation as a latent action learned to maximize judgment accuracy. We introduce an alternating optimization strategy to mitigate the non-stationarity of simultaneous updates, providing theoretical analysis that demonstrates how this schedule reduces gradient variance during training. Extensive experiments show that Rubric-ARM achieves state-of-the-art performance among baselines on multiple benchmarks and significantly improves downstream policy alignment in both offline and online reinforcement learning settings.

Method: Converted 192 linguistic features from the World Atlas of Language Structures (WALS) covering 2,660 languages into natural-language multiple-choice questions. Evaluated models using accuracy and macro F1, comparing to chance and majority-class baselines. Analyzed performance across linguistic domains and language-related factors including digital presence, corpus availability, and sociolinguistic indicators.

Result: GPT-4o performed best but achieved only moderate accuracy (0.367). All models performed above chance but failed to outperform majority-class baseline. Performance varied by linguistic domain and correlated strongly with digital language status - languages with greater digital presence performed better. Resource-related indicators (Wikipedia size, corpus availability) were more predictive than geographic, genealogical, or sociolinguistic factors.

Conclusion: LLMs’ metalinguistic knowledge is fragmented and shaped mainly by data availability rather than broadly generalizable grammatical competence. The benchmark is released as open-source to encourage greater linguistic diversity in future LLM evaluations.

Abstract: LLMs are routinely evaluated on language use, yet their explicit knowledge about linguistic structure remains poorly understood. Existing linguistic benchmarks focus on narrow phenomena, emphasize high-resource languages, and rarely test metalinguistic knowledge - explicit reasoning about language structure. We present a multilingual evaluation of metalinguistic knowledge in LLMs, based on the World Atlas of Language Structures (WALS), documenting 192 linguistic features across 2,660 languages. We convert WALS features into natural-language multiple-choice questions and evaluate models across documented languages. Using accuracy and macro F1, and comparing to chance and majority-class baselines, we assess performance and analyse variation across linguistic domains and language-related factors. Results show limited metalinguistic knowledge: GPT-4o performs best but achieves moderate accuracy (0.367), while open-source models lag. Although all models perform above chance, they fail to outperform the majority-class baseline, suggesting they capture broad cross-linguistic patterns but lack fine-grained distinctions. Performance varies by domain, partly reflecting differences in online visibility. At the language level, accuracy correlates with digital language status: languages with greater digital presence and resources are evaluated more accurately, while low-resource languages perform worse. Analysis of predictive factors confirms that resource-related indicators (Wikipedia size, corpus availability) are more informative than geographic, genealogical, or sociolinguistic factors. Overall, LLM metalinguistic knowledge appears fragmented and shaped mainly by data availability, rather than broadly generalizable grammatical competence. We release the benchmark as an open-source dataset to support evaluation across languages and encourage greater global linguistic diversity in future LLMs.

Method: 1) Analyze each component of RAG systems and propose practical alternatives for each component. 2) Conduct systematic evaluations on three types of tasks to reveal best practices and performance-efficiency tradeoffs.

Result: Reveals best practices for improving RAG systems and shows how LLM-based RAG systems make trade-offs between performance and efficiency across different task types.

Conclusion: Provides practical guidance for building effective RAG systems in medical/industrial applications through component analysis and systematic evaluation of performance-efficiency tradeoffs.

Abstract: While retrieval augmented generation (RAG) has been swiftly adopted in industrial applications based on large language models (LLMs), there is no consensus on what are the best practices for building a RAG system in terms of what are the components, how to organize these components and how to implement each component for the industrial applications, especially in the medical domain. In this work, we first carefully analyze each component of the RAG system and propose practical alternatives for each component. Then, we conduct systematic evaluations on three types of tasks, revealing the best practices for improving the RAG system and how LLM-based RAG systems make trade-offs between performance and efficiency.

Method: Proposes FaithRL with: 1) Formal faithfulness-maximization objective, 2) Geometric reward design, and 3) Faithfulness-aware advantage modulation mechanism that assigns step-level credit by penalizing unsupported steps while preserving valid partial derivations.

Result: Across diverse backbones and benchmarks, FaithRL consistently reduces hallucination rates while maintaining (and often improving) answer correctness. Increases step-wise reasoning faithfulness and generalizes robustly.

Conclusion: FaithRL provides an effective reinforcement learning framework for optimizing reasoning faithfulness in LLMs, addressing key limitations of sparse reward-based approaches and reducing hallucinations without compromising performance.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has markedly improved the performance of Large Language Models (LLMs) on tasks requiring multi-step reasoning. However, most RLVR pipelines rely on sparse outcome-based rewards, providing little supervision over intermediate steps and thus encouraging over-confidence and spurious reasoning, which in turn increases hallucinations. To address this, we propose FaithRL, a general reinforcement learning framework that directly optimizes reasoning faithfulness. We formalize a faithfulness-maximization objective and theoretically show that optimizing it mitigates over-confidence. To instantiate this objective, we introduce a geometric reward design and a faithfulness-aware advantage modulation mechanism that assigns step-level credit by penalizing unsupported steps while preserving valid partial derivations. Across diverse backbones and benchmarks, FaithRL consistently reduces hallucination rates while maintaining (and often improving) answer correctness. Further analysis confirms that FaithRL increases step-wise reasoning faithfulness and generalizes robustly. Our code is available at https://github.com/aintdoin/FaithRL.

Method: Proposes ACL-Embed (treats label embeddings as augmented samples), ACL-Grad (resolves gradient conflicts), and ACL-CL (cross-layer guidance for multi-exit BERT).

Result: ACL-BERT outperforms or matches CE and CE+SCL on GLUE tasks; ACL-CL significantly improves multi-exit BERT fine-tuning for better quality-speed tradeoffs.

Conclusion: ACL framework successfully integrates contrastive learning into supervised settings by resolving objective conflicts and enabling effective multi-exit model optimization.

Abstract: Despite its success in self-supervised learning, contrastive learning is less studied in the supervised setting. In this work, we first use a set of pilot experiments to show that in the supervised setting, the cross-entropy loss objective (CE) and the contrastive learning objective often conflict with each other, thus hindering the applications of CL in supervised settings. To resolve this problem, we introduce a novel \underline{A}ligned \underline{C}ontrastive \underline{L}earning (ACL) framework. First, ACL-Embed regards label embeddings as extra augmented samples with different labels and employs contrastive learning to align the label embeddings with its samples’ representations. Second, to facilitate the optimization of ACL-Embed objective combined with the CE loss, we propose ACL-Grad, which will discard the ACL-Embed term if the two objectives are in conflict. To further enhance the performances of intermediate exits of multi-exit BERT, we further propose cross-layer ACL (ACL-CL), which is to ask the teacher exit to guide the optimization of student shallow exits. Extensive experiments on the GLUE benchmark results in the following takeaways: (a) ACL-BRT outperforms or performs comparably with CE and CE+SCL on the GLUE tasks; (b) ACL, especially CL-ACL, significantly surpasses the baseline methods on the fine-tuning of multi-exit BERT, thus providing better quality-speed tradeoffs for low-latency applications.

Method: Proposes Model-Dowser that computes importance scores for each parameter using weight magnitudes, input activations, and output sensitivities. During fine-tuning, preserves high-importance parameters and updates only the remaining ones.

Result: Experiments on LLaVA and NVILA show Model-Dowser effectively mitigates catastrophic forgetting, outperforms prior methods, and remains resource-efficient and scalable to multi-billion-parameter models.

Conclusion: Model-Dowser provides an effective, scalable solution for fine-tuning MLLMs while preserving pretrained generalization capabilities, addressing limitations of existing catastrophic forgetting mitigation methods.

Abstract: Fine-tuning Multimodal Large Language Models (MLLMs) on task-specific data is an effective way to improve performance on downstream applications. However, such adaptation often leads to a degradation in generalization on pretrained tasks, a phenomenon known as Catastrophic Forgetting. Existing methods that aim to mitigate this issue either become ineffective when fine-tuning deeper layers of the language decoder or scale poorly with increasing model size. To address these limitations, we propose Model-Dowser, a novel sparse fine-tuning approach for MLLMs. Model-Dowser measures a principled importance score for each model parameter with respect to pretrained generalization (prior to downstream adaptation) by jointly considering weight magnitudes, input activations, and output sensitivities. During fine-tuning, Model-Dowser selectively preserves high-importance parameters and updates the remaining. Comprehensive experiments on two representative MLLMs, LLaVA and NVILA, demonstrate that Model-Dowser effectively mitigates catastrophic forgetting and consistently outperforms prior methods, while remaining resource-efficient and scalable to multi-billion-parameter models.

Method: Used SQL generation as proxy for programmatic agent operations, conducted 9,649 experiments across 11 models, 4 formats (YAML, Markdown, JSON, TOON), and schemas ranging from 10 to 10,000 tables.

Result: Found that architecture choice is model-dependent, format doesn’t significantly affect aggregate accuracy, model capability is dominant factor (21% gap between frontier and open source), file-native agents scale to 10,000 tables, and file size doesn’t predict runtime efficiency.

Conclusion: Architectural decisions for LLM agents on structured systems should be tailored to model capability rather than assuming universal best practices, with evidence-based guidance for practitioners.

Abstract: Large Language Model agents increasingly operate external systems through programmatic interfaces, yet practitioners lack empirical guidance on how to structure the context these agents consume. Using SQL generation as a proxy for programmatic agent operations, we present a systematic study of context engineering for structured data, comprising 9,649 experiments across 11 models, 4 formats (YAML, Markdown, JSON, Token-Oriented Object Notation [TOON]), and schemas ranging from 10 to 10,000 tables. Our findings challenge common assumptions. First, architecture choice is model-dependent: file-based context retrieval improves accuracy for frontier-tier models (Claude, GPT, Gemini; +2.7%, p=0.029) but shows mixed results for open source models (aggregate -7.7%, p<0.001), with deficits varying substantially by model. Second, format does not significantly affect aggregate accuracy (chi-squared=2.45, p=0.484), though individual models, particularly open source, exhibit format-specific sensitivities. Third, model capability is the dominant factor, with a 21 percentage point accuracy gap between frontier and open source tiers that dwarfs any format or architecture effect. Fourth, file-native agents scale to 10,000 tables through domain-partitioned schemas while maintaining high navigation accuracy. Fifth, file size does not predict runtime efficiency: compact or novel formats can incur a token overhead driven by grep output density and pattern unfamiliarity, with the magnitude depending on model capability. These findings provide practitioners with evidence-based guidance for deploying LLM agents on structured systems, demonstrating that architectural decisions should be tailored to model capability rather than assuming universal best practices.

Method: Introduced a systematic evaluation framework to diagnose cross-cultural robustness of state-of-the-art VLMs across multilingual meme datasets, analyzing three axes: (i) learning strategy (zero-shot vs. one-shot), (ii) prompting language (native vs. English), and (iii) translation effects on meaning and detection.

Result: Results show that the common “translate-then-detect” approach deteriorates performance, while culturally aligned interventions - native-language prompting and one-shot learning - significantly enhance detection. Findings reveal systematic convergence toward Western safety norms.

Conclusion: The study provides actionable strategies to mitigate cultural bias in VLMs, guiding the design of globally robust multimodal moderation systems through culturally aligned interventions rather than translation approaches.

Abstract: Cultural context profoundly shapes how people interpret online content, yet vision-language models (VLMs) remain predominantly trained through Western or English-centric lenses. This limits their fairness and cross-cultural robustness in tasks like hateful meme detection. We introduce a systematic evaluation framework designed to diagnose and quantify the cross-cultural robustness of state-of-the-art VLMs across multilingual meme datasets, analyzing three axes: (i) learning strategy (zero-shot vs. one-shot), (ii) prompting language (native vs. English), and (iii) translation effects on meaning and detection. Results show that the common ``translate-then-detect’’ approach deteriorate performance, while culturally aligned interventions - native-language prompting and one-shot learning - significantly enhance detection. Our findings reveal systematic convergence toward Western safety norms and provide actionable strategies to mitigate such bias, guiding the design of globally robust multimodal moderation systems.

Method: Proposes a generative framework with attention-over-attention decoder to handle variable numbers of intents and reduce interference between intent detection and slot filling tasks. Also constructs new multi-intent datasets using BERT’s next sentence prediction head to combine single-intent utterances.

Result: Achieves state-of-the-art performance on two public datasets (MixATIS and MixSNIPS) and the newly constructed datasets, demonstrating effectiveness of the attention-over-attention approach.

Conclusion: The proposed generative framework successfully addresses multi-intent spoken language understanding challenges, with the attention-over-attention decoder effectively handling variable intent numbers and task interference.

Abstract: In task-oriented dialogue systems, spoken language understanding (SLU) is a critical component, which consists of two sub-tasks, intent detection and slot filling. Most existing methods focus on the single-intent SLU, where each utterance only has one intent. However, in real-world scenarios users usually express multiple intents in an utterance, which poses a challenge for existing dialogue systems and datasets. In this paper, we propose a generative framework to simultaneously address multiple intent detection and slot filling. In particular, an attention-over-attention decoder is proposed to handle the variable number of intents and the interference between the two sub-tasks by incorporating an inductive bias into the process of multi-task learning. Besides, we construct two new multi-intent SLU datasets based on single-intent utterances by taking advantage of the next sentence prediction (NSP) head of the BERT model. Experimental results demonstrate that our proposed attention-over-attention generative model achieves state-of-the-art performance on two public datasets, MixATIS and MixSNIPS, and our constructed datasets.

Method: Proposes Anagent, a multi-agent framework with four specialized agents: Planner (decomposes tasks), Expert (retrieves task-specific information via tools), Solver (synthesizes information for analysis), and Critic (iterative refinement via 5D quality assessment). Uses modular training strategies with supervised finetuning and specialized reinforcement learning.

Result: Achieves substantial improvements: up to ↑13.43% in training-free settings and ↑42.12% with finetuning across 9 broad domains with 170 subdomains. Demonstrates that task-oriented reasoning and context-aware problem-solving are essential for high-quality scientific table & figure analysis.

Conclusion: Anagent effectively addresses challenges in scientific multimodal analysis through specialized multi-agent collaboration and modular training, showing significant performance gains on the comprehensive AnaBench benchmark.

Abstract: In scientific research, analysis requires accurately interpreting complex multimodal knowledge, integrating evidence from different sources, and drawing inferences grounded in domain-specific knowledge. However, current artificial intelligence (AI) systems struggle to consistently demonstrate such capabilities. The complexity and variability of scientific tables and figures, combined with heterogeneous structures and long-context requirements, pose fundamental obstacles to scientific table & figure analysis. To quantify these challenges, we introduce AnaBench, a large-scale benchmark featuring $63,178$ instances from nine scientific domains, systematically categorized along seven complexity dimensions. To tackle these challenges, we propose Anagent, a multi-agent framework for enhanced scientific table & figure analysis through four specialized agents: Planner decomposes tasks into actionable subtasks, Expert retrieves task-specific information through targeted tool execution, Solver synthesizes information to generate coherent analysis, and Critic performs iterative refinement through five-dimensional quality assessment. We further develop modular training strategies that leverage supervised finetuning and specialized reinforcement learning to optimize individual capabilities while maintaining effective collaboration. Comprehensive evaluation across 9 broad domains with 170 subdomains demonstrates that Anagent achieves substantial improvements, up to $\uparrow 13.43%$ in training-free settings and $\uparrow 42.12%$ with finetuning, while revealing that task-oriented reasoning and context-aware problem-solving are essential for high-quality scientific table & figure analysis. Our project page: https://xhguo7.github.io/Anagent/.

Method: Neuro-Symbolic Synergy framework alternates training between LLMs and symbolic models using trajectories inadequately explained by the other, with symbolic WM directly constraining LLM output probabilities.

Result: Achieves consistent advantages over baselines in prediction accuracy and data efficiency across ScienceWorld, Webshop, and Plancraft environments, reducing training data by 50% without accuracy loss.

Conclusion: NeSyS successfully bridges the gap between neural and symbolic approaches, creating world models that are both semantically expressive and logically robust.

Abstract: Large language models (LLMs) exhibit strong general-purpose reasoning capabilities, yet they frequently hallucinate when used as world models (WMs), where strict compliance with deterministic transition rules–particularly in corner cases–is essential. In contrast, Symbolic WMs provide logical consistency but lack semantic expressivity. To bridge this gap, we propose Neuro-Symbolic Synergy (NeSyS), a framework that integrates the probabilistic semantic priors of LLMs with executable symbolic rules to achieve both expressivity and robustness. NeSyS alternates training between the two models using trajectories inadequately explained by the other. Unlike rule-based prompting, the symbolic WM directly constrains the LLM by modifying its output probability distribution. The neural WM is fine-tuned only on trajectories not covered by symbolic rules, reducing training data by 50% without loss of accuracy. Extensive experiments on three distinct interactive environments, i.e., ScienceWorld, Webshop, and Plancraft, demonstrate NeSyS’s consistent advantages over baselines in both WM prediction accuracy and data efficiency.

Method: The method frames embedding inversion as conditional masked diffusion, using a masked diffusion language model conditioned on target embeddings via adaptive layer normalization. It requires only 8 forward passes through a 78M parameter model without needing access to the target encoder.

Result: On 32-token sequences across three embedding models, the method achieves up to 81.3% token accuracy, demonstrating effective parallel recovery of tokens from embeddings.

Conclusion: The approach successfully demonstrates that embedding inversion can be efficiently performed using conditional masked diffusion with parallel token recovery, offering improved speed and accuracy compared to sequential methods.

Abstract: We frame embedding inversion as conditional masked diffusion, recovering all tokens in parallel through iterative denoising rather than sequential autoregressive generation. A masked diffusion language model is conditioned on the target embedding via adaptive layer normalization, requiring only 8 forward passes through a 78M parameter model with no access to the target encoder. On 32-token sequences across three embedding models, the method achieves up to 81.3% token accuracy. Source code and live demo are available at https://github.com/jina-ai/embedding-inversion-demo.

Method: TEG extracts structured graph information from documents, encodes both text and graph for classification; TEGRA adds domain-specific knowledge

Result: Hybrid representation outperforms language models alone; TEGRA further improves accuracy in most cases

Conclusion: Graph-enhanced text encoding with knowledge integration significantly improves misinformation detection performance

Abstract: Misinformation detection is a critical task that can benefit significantly from the integration of external knowledge, much like manual fact-checking. In this work, we propose a novel method for representing textual documents that facilitates the incorporation of information from a knowledge base. Our approach, Text Encoding with Graph (TEG), processes documents by extracting structured information in the form of a graph and encoding both the text and the graph for classification purposes. Through extensive experiments, we demonstrate that this hybrid representation enhances misinformation detection performance compared to using language models alone. Furthermore, we introduce TEGRA, an extension of our framework that integrates domain-specific knowledge, further enhancing classification accuracy in most cases.

Method: Proposes Reasoning Activation Potential (RAP) with two complementary estimators: 1) Causal Discrepancy Estimator (CDE) eliminates samples overly reliant on language priors by comparing multimodal vs text-only outputs; 2) Attention Confidence Estimator (ACE) discards samples dominated by irrelevant tokens using token-level self-attention. Also includes Difficulty-aware Replacement Module (DRM) to substitute trivial instances with challenging ones.

Result: Experiments on six datasets show RAP consistently achieves superior performance using only 9.3% of training data while reducing computational costs by over 43%.

Conclusion: Smaller high-value datasets can match or outperform full corpora for multimodal reasoning in MLLMs. The RAP paradigm effectively identifies cognitive samples that trigger genuine reasoning, enabling efficient training with reduced data and computational requirements.

Abstract: While multi-modal large language models (MLLMs) have made significant progress in complex reasoning tasks via reinforcement learning, it is commonly believed that extensive training data is necessary for improving multi-modal reasoning ability, inevitably leading to data redundancy and substantial computational costs. However, can smaller high-value datasets match or outperform full corpora for multi-modal reasoning in MLLMs? In this work, we challenge this assumption through a key observation: meaningful multi-modal reasoning is triggered by only a sparse subset of training samples, termed cognitive samples, whereas the majority contribute marginally. Building on this insight, we propose a novel data selection paradigm termed Reasoning Activation Potential (RAP)}, which identifies cognitive samples by estimating each sample’s potential to stimulate genuine multi-modal reasoning by two complementary estimators: 1) Causal Discrepancy Estimator (CDE) based on the potential outcome model principle, eliminates samples that overly rely on language priors by comparing outputs between multi-modal and text-only inputs; 2) Attention Confidence Estimator (ACE), which exploits token-level self-attention to discard samples dominated by irrelevant but over-emphasized tokens in intermediate reasoning stages. Moreover, we introduce a Difficulty-aware Replacement Module (DRM) to substitute trivial instances with cognitively challenging ones, thereby ensuring complexity for robust multi-modal reasoning. Experiments on six datasets show that our RAP method consistently achieves superior performance using only 9.3% of the training data, while reducing computational costs by over 43%.

Method: Uses a few-shot denoising diffusion backbone with a dual mixture density network to learn self-calibrated residence areas and probability-ranked anchor paths, generating diverse trajectory hypotheses without predefined anchors or endpoints.

Result: State-of-the-art accuracy on ETH/UCY, SDD, inD, and IMPTC datasets, with demonstrated robustness at short observation intervals and reliable uncertainty modeling.

Conclusion: DD-MDN effectively addresses key gaps in HTF by providing accurate, calibrated probabilistic forecasts that are robust to short observation periods, making it suitable for real-world applications.

Abstract: Human Trajectory Forecasting (HTF) predicts future human movements from past trajectories and environmental context, with applications in Autonomous Driving, Smart Surveillance, and Human-Robot Interaction. While prior work has focused on accuracy, social interaction modeling, and diversity, little attention has been paid to uncertainty modeling, calibration, and forecasts from short observation periods, which are crucial for downstream tasks such as path planning and collision avoidance. We propose DD-MDN, an end-to-end probabilistic HTF model that combines high positional accuracy, calibrated uncertainty, and robustness to short observations. Using a few-shot denoising diffusion backbone and a dual mixture density network, our method learns self-calibrated residence areas and probability-ranked anchor paths, from which diverse trajectory hypotheses are derived, without predefined anchors or endpoints. Experiments on the ETH/UCY, SDD, inD, and IMPTC datasets demonstrate state-of-the-art accuracy, robustness at short observation intervals, and reliable uncertainty modeling. The code is available at: https://github.com/kav-institute/ddmdn.

Method: 1) Constructed UniACT-dataset from six public datasets with over 6M trajectories and 9,500 hours of data; 2) Proposed Action Manifold Hypothesis and Action Manifold Learning (AML) using DiT backbone for efficient action prediction; 3) Modular perception with dual-stream mechanism integrating VLM semantics with geometric priors and multi-view inputs from 3D modules.

Result: Unified pre-training improves knowledge transfer and generalization across platforms and tasks. AML improves decoding speed and policy stability by learning on low-dimensional action manifolds. Modular perception enhances spatial understanding without modifying the backbone.

Conclusion: ABot-M0 enables end-to-end transformation of heterogeneous raw data into unified representations, supporting general-purpose embodied intelligence across diverse robot morphologies and task scenarios.

Abstract: Building general-purpose embodied agents across diverse hardware remains a central challenge in robotics, often framed as the ‘‘one-brain, many-forms’’ paradigm. Progress is hindered by fragmented data, inconsistent representations, and misaligned training objectives. We present ABot-M0, a framework that builds a systematic data curation pipeline while jointly optimizing model architecture and training strategies, enabling end-to-end transformation of heterogeneous raw data into unified, efficient representations. From six public datasets, we clean, standardize, and balance samples to construct UniACT-dataset, a large-scale dataset with over 6 million trajectories and 9,500 hours of data, covering diverse robot morphologies and task scenarios. Unified pre-training improves knowledge transfer and generalization across platforms and tasks, supporting general-purpose embodied intelligence. To improve action prediction efficiency and stability, we propose the Action Manifold Hypothesis: effective robot actions lie not in the full high-dimensional space but on a low-dimensional, smooth manifold governed by physical laws and task constraints. Based on this, we introduce Action Manifold Learning (AML), which uses a DiT backbone to predict clean, continuous action sequences directly. This shifts learning from denoising to projection onto feasible manifolds, improving decoding speed and policy stability. ABot-M0 supports modular perception via a dual-stream mechanism that integrates VLM semantics with geometric priors and multi-view inputs from plug-and-play 3D modules such as VGGT and Qwen-Image-Edit, enhancing spatial understanding without modifying the backbone and mitigating standard VLM limitations in 3D reasoning. Experiments show components operate independently with additive benefits. We will release all code and pipelines for reproducibility and future research.

Method: Developed a pipeline with data preprocessing, splitting, adversarial training, augmentation, and model training using DenseNet201 and EfficientNetB3 on the TeaLeafBD dataset (5,278 images across 7 categories). Applied adversarial training for robustness and Grad-CAM for explainable AI visualization.

Result: EfficientNetB3 achieved 93% classification accuracy, outperforming DenseNet201 at 91%. The adversarial training improved model robustness against noisy inputs, and Grad-CAM provided interpretable visualizations of decision-making regions.

Conclusion: The proposed deep learning approach provides an effective automated solution for tea leaf disease detection, offering practical benefits for agricultural management through accurate classification and explainable AI insights.

Abstract: Tea is a valuable asset for the economy of Bangladesh. So, tea cultivation plays an important role to boost the economy. These valuable plants are vulnerable to various kinds of leaf infections which may cause less production and low quality. It is not so easy to detect these diseases manually. It may take time and there could be some errors in the detection.Therefore, the purpose of the study is to develop an automated deep learning model for tea leaf disease classification based on the teaLeafBD dataset so that anyone can detect the diseases more easily and efficiently. There are 5,278 high-resolution images in this dataset. The images are classified into seven categories. Six of them represents various diseases and the rest one represents healthy leaves. The proposed pipeline contains data preprocessing, data splitting, adversarial training, augmentation, model training, evaluation, and comprehension made possible with Explainable AI strategies. DenseNet201 and EfficientNetB3 were employed to perform the classification task. To prepare the model more robustly, we applied adversarial training so it can operate effectively even with noisy or disturbed inputs. In addition, Grad-CAM visualization was executed to analyze the model’s predictions by identifying the most influential regions of each image. Our experimental outcomes revealed that EfficientNetB3 achieved the highest classification accuracy of 93%, while DenseNet201 reached 91%. The outcomes prove that the effectiveness of the proposed approach can accurately detect tea leaf diseases and provide a practical solution for advanced agricultural management.

Method: Proposes Active-Zero framework with three co-evolving agents: 1) Searcher that retrieves images from open-world repositories based on the model’s capability frontier, 2) Questioner that synthesizes calibrated reasoning tasks, and 3) Solver refined through accuracy rewards. This creates a closed-loop self-scaffolding auto-curriculum.

Result: On Qwen2.5-VL-7B-Instruct across 12 benchmarks, Active-Zero achieves 53.97 average accuracy on reasoning tasks (5.7% improvement) and 59.77 on general understanding (3.9% improvement), consistently outperforming existing self-play baselines.

Conclusion: Active exploration is a key ingredient for scalable and adaptive self-evolving vision-language systems, enabling models to autonomously construct their learning trajectories through active interaction with visual environments rather than passive dataset interaction.

Abstract: Self-play has enabled large language models to autonomously improve through self-generated challenges. However, existing self-play methods for vision-language models rely on passive interaction with static image collections, resulting in strong dependence on initial datasets and inefficient learning. Without the ability to actively seek visual data tailored to their evolving capabilities, agents waste computational effort on samples that are either trivial or beyond their current skill level. To address these limitations, we propose Active-Zero, a framework that shifts from passive interaction to active exploration of visual environments. Active-Zero employs three co-evolving agents: a Searcher that retrieves images from open-world repositories based on the model’s capability frontier, a Questioner that synthesizes calibrated reasoning tasks, and a Solver refined through accuracy rewards. This closed loop enables self-scaffolding auto-curricula where the model autonomously constructs its learning trajectory. On Qwen2.5-VL-7B-Instruct across 12 benchmarks, Active-Zero achieves 53.97 average accuracy on reasoning tasks (5.7% improvement) and 59.77 on general understanding (3.9% improvement), consistently outperforming existing self-play baselines. These results highlight active exploration as a key ingredient for scalable and adaptive self-evolving vision-language systems.

Method: 1) Extract sentences describing human-machine interaction from science-fiction novels; 2) Use LLMs to generate paired prompts “what to do” and “what not to do” for each excerpt; 3) Use diffusion-based text-to-video model to transform prompts into choreographic guides for human performer and motor commands for quadruped robot; 4) Both agents enact actions on mirrored floor; 5) Capture performance with multi-camera motion tracking; 6) Reconstruct into 3D point clouds and motion trails to form digital archive.

Result: Creates an immersive interplay of AI, human, and robot that produces a digital archive of motion traces, serving as a novel approach to reveal how generative systems encode socio-cultural biases through choreographed movements.

Conclusion: ReTracing confronts critical questions about human identity in the age of AI through embodied performance, demonstrating how AI systems influence and constrain movement while encoding cultural biases, ultimately questioning what it means to be human among moving, thinking AIs.

Abstract: We present ReTracing, a multi-agent embodied performance art that adopts an archaeological approach to examine how artificial intelligence shapes, constrains, and produces bodily movement. Drawing from science-fiction novels, the project extracts sentences that describe human-machine interaction. We use large language models (LLMs) to generate paired prompts “what to do” and “what not to do” for each excerpt. A diffusion-based text-to-video model transforms these prompts into choreographic guides for a human performer and motor commands for a quadruped robot. Both agents enact the actions on a mirrored floor, captured by multi-camera motion tracking and reconstructed into 3D point clouds and motion trails, forming a digital archive of motion traces. Through this process, ReTracing serves as a novel approach to reveal how generative systems encode socio-cultural biases through choreographed movements. Through an immersive interplay of AI, human, and robot, ReTracing confronts a critical question of our time: What does it mean to be human among AIs that also move, think, and leave traces behind?

Method: Proposes a lightweight RGB-D fusion framework that augments EfficientViT-SAM with monocular depth priors. Depth maps are generated using a pretrained estimator and fused mid-level with RGB features through a dedicated depth encoder.

Result: Trained on only 11.2k samples (less than 0.1% of SA-1B), the method achieves higher accuracy than EfficientViT-SAM, demonstrating that depth cues provide strong geometric priors for segmentation.

Conclusion: Depth information can significantly enhance segmentation performance while dramatically reducing training data requirements, offering a more efficient approach to multimodal segmentation.

Abstract: Segment Anything Models (SAM) achieve impressive universal segmentation performance but require massive datasets (e.g., 11M images) and rely solely on RGB inputs. Recent efficient variants reduce computation but still depend on large-scale training. We propose a lightweight RGB-D fusion framework that augments EfficientViT-SAM with monocular depth priors. Depth maps are generated with a pretrained estimator and fused mid-level with RGB features through a dedicated depth encoder. Trained on only 11.2k samples (less than 0.1% of SA-1B), our method achieves higher accuracy than EfficientViT-SAM, showing that depth cues provide strong geometric priors for segmentation.

Method: Introduces REVEAL benchmark with five controlled stress tests: temporal expectation bias, language-only shortcuts, video sycophancy, camera motion sensitivity, and spatiotemporal occlusion robustness. Includes automated data pipeline for generating diagnostic examples.

Result: Leading VidLMs fail basic tasks - describe reversed scenes as forward, answer questions ignoring video content, agree with false claims, struggle with camera motion, and fail with spatiotemporal masking. Humans succeed easily.

Conclusion: Current VidLMs lack robust video understanding capabilities and rely on biases/shortcuts. REVEAL benchmark enables scalable evaluation to drive improvement in video-language model development.

Abstract: This work investigates a fundamental question: Do Video-Language Models (VidLMs) robustly account for video content, temporal sequence, and motion? Our investigation shows that, surprisingly, they often do not. We introduce REVEAL{}, a diagnostic benchmark that probes fundamental weaknesses of contemporary VidLMs through five controlled stress tests; assessing temporal expectation bias, reliance on language-only shortcuts, video sycophancy, camera motion sensitivity, and robustness to spatiotemporal occlusion. We test leading open- and closed-source VidLMs and find that these models confidently describe reversed scenes as forward, answer questions while neglecting video content, agree with false claims, struggle with basic camera motion, and fail to aggregate temporal information amidst simple spatiotemporal masking. Humans, on the other hand, succeed at these tasks with ease. Alongside our benchmark, we provide a data pipeline that automatically generates diagnostic examples for our stress tests, enabling broader and more scalable evaluation. We will release our benchmark and code to support future research.

Method: Developed a pipeline that generates synthetic images from high-quality 3D models using programmatically generated camera poses, creating controlled datasets with known ground truth parameters for camera positions and object geometry.

Result: Enables repeatable, quantifiable experiments comparing ground-truth camera parameters and virtual objects against reconstructed estimations, providing objective metrics for evaluating photogrammetry performance.

Conclusion: The synthetic image generation pipeline provides a standardized framework for quantitative evaluation of photogrammetry techniques, moving beyond qualitative assessments to enable objective comparison of different 3D reconstruction approaches.

Abstract: The generation of 3D models from real-world objects has often been accomplished through photogrammetry, i.e., by taking 2D photos from a variety of perspectives and then triangulating matched point-based features to create a textured mesh. Many design choices exist within this framework for the generation of digital twins, and differences between such approaches are largely judged qualitatively. Here, we present and test a novel pipeline for generating synthetic images from high-quality 3D models and programmatically generated camera poses. This enables a wide variety of repeatable, quantifiable experiments which can compare ground-truth knowledge of virtual camera parameters and of virtual objects against the reconstructed estimations of those perspectives and subjects.

Method: Proposes SYNC loss that integrates ad-hoc (SelectiveNet) and post-hoc (softmax response) approaches. Specifically incorporates the softmax response uncertainty information (selective prior) into the training process of SelectiveNet, allowing the model to learn from uncertainty signals during training rather than just using them during inference.

Result: Evaluated on CIFAR-100, ImageNet-100, and Stanford Cars datasets. The method enhances model generalization capabilities, surpasses previous works in selective prediction performance, and sets new state-of-the-art benchmarks.

Conclusion: The selective prior generated by post-hoc methods is valuable not just for inference but also for training. Integrating ad-hoc and post-hoc approaches through SYNC loss significantly improves selective prediction performance and generalization.

Abstract: Prediction under uncertainty is a critical requirement for the deep neural network to succeed responsibly. This paper focuses on selective prediction, which allows DNNs to make informed decisions about when to predict or abstain based on the uncertainty level of their predictions. Current methods are either ad-hoc such as SelectiveNet, focusing on how to modify the network architecture or objective function, or post-hoc such as softmax response, achieving selective prediction through analyzing the model’s probabilistic outputs. We observe that post-hoc methods implicitly generate uncertainty information, termed the selective prior, which has traditionally been used only during inference. We argue that the selective prior provided by the selection mechanism is equally vital during the training stage. Therefore, we propose the SYNC loss which introduces a novel integration of ad-hoc and post-hoc method. Specifically, our approach incorporates the softmax response into the training process of SelectiveNet, enhancing its selective prediction capabilities by examining the selective prior. Evaluated across various datasets, including CIFAR-100, ImageNet-100, and Stanford Cars, our method not only enhances the model’s generalization capabilities but also surpasses previous works in selective prediction performance, and sets new benchmarks for state-of-the-art performance.

Method: Integrates learned depth priors directly into VINS-Mono optimization backend with affine-invariant depth consistency and pairwise ordinal constraints, filtering unstable artifacts via variance-based gating while adhering to edge device computational limits.

Result: Extensive experiments on TartanGround and M3ED datasets show the method prevents divergence in challenging scenarios, reduces Absolute Trajectory Error by up to 28.3%, and robustly recovers metric scale.

Conclusion: The proposed framework successfully bridges the gap between computationally intensive dense depth estimation and real-time edge deployment requirements for VIO systems in challenging environments.

Abstract: Traditional monocular Visual-Inertial Odometry (VIO) systems struggle in low-texture environments where sparse visual features are insufficient for accurate pose estimation. To address this, dense Monocular Depth Estimation (MDE) has been widely explored as a complementary information source. While recent Vision Transformer (ViT) based complex foundational models offer dense, geometrically consistent depth, their computational demands typically preclude them from real-time edge deployment. Our work bridges this gap by integrating learned depth priors directly into the VINS-Mono optimization backend. We propose a novel framework that enforces affine-invariant depth consistency and pairwise ordinal constraints, explicitly filtering unstable artifacts via variance-based gating. This approach strictly adheres to the computational limits of edge devices while robustly recovering metric scale. Extensive experiments on the TartanGround and M3ED datasets demonstrate that our method prevents divergence in challenging scenarios and delivers significant accuracy gains, reducing Absolute Trajectory Error (ATE) by up to 28.3%. Code will be made available.

Method: Introduced StreamSR dataset from YouTube covering diverse video genres/resolutions; proposed EfRLFN model with Efficient Channel Attention and hyperbolic tangent activation; designed composite loss function; benchmarked 11 state-of-the-art models

Result: EfRLFN improves both visual quality and runtime performance; fine-tuning other models on StreamSR yields significant performance gains that generalize across standard benchmarks

Conclusion: StreamSR addresses dataset limitations for streaming media, EfRLFN provides efficient real-time super-resolution, and the dataset enables better model evaluation and improvement for streaming applications

Abstract: Recent advancements in real-time super-resolution have enabled higher-quality video streaming, yet existing methods struggle with the unique challenges of compressed video content. Commonly used datasets do not accurately reflect the characteristics of streaming media, limiting the relevance of current benchmarks. To address this gap, we introduce a comprehensive dataset - StreamSR - sourced from YouTube, covering a wide range of video genres and resolutions representative of real-world streaming scenarios. We benchmark 11 state-of-the-art real-time super-resolution models to evaluate their performance for the streaming use-case. Furthermore, we propose EfRLFN, an efficient real-time model that integrates Efficient Channel Attention and a hyperbolic tangent activation function - a novel design choice in the context of real-time super-resolution. We extensively optimized the architecture to maximize efficiency and designed a composite loss function that improves training convergence. EfRLFN combines the strengths of existing architectures while improving both visual quality and runtime performance. Finally, we show that fine-tuning other models on our dataset results in significant performance gains that generalize well across various standard benchmarks. We made the dataset, the code, and the benchmark available at https://github.com/EvgeneyBogatyrev/EfRLFN.

Method: 1) Novel corpus collection pipeline from Open-Access Art History articles and Wikidata knowledge; 2) Weakly supervised learning of bespoke CLIP model adapted using Low-Rank Adaptation (LoRA) to make it domain-specific; 3) Creation of PaintingCLIP model that can annotate artworks with contextual information.

Result: PaintingCLIP outperforms standard CLIP and provides meaningful context for given artworks. The pipeline is generalizable and can be applied to numerous humanities areas beyond art history.

Conclusion: ArtContext successfully bridges the gap between visual artworks and their textual discussions in academic literature, enabling better contextual understanding of art through multimodal learning approaches.

Abstract: Many Art History articles discuss artworks in general as well as specific parts of works, such as layout, iconography, or material culture. However, when viewing an artwork, it is not trivial to identify what different articles have said about the piece. Therefore, we propose ArtContext, a pipeline for taking a corpus of Open-Access Art History articles and Wikidata Knowledge and annotating Artworks with this information. We do this using a novel corpus collection pipeline, then learn a bespoke CLIP model adapted using Low-Rank Adaptation (LoRA) to make it domain-specific. We show that the new model, PaintingCLIP, which is weakly supervised by the collected corpus, outperforms CLIP and provides context for a given artwork. The proposed pipeline is generalisable and can be readily applied to numerous humanities areas.

Method: Proposes Latent Forcing which orders the denoising trajectory by jointly processing latents and pixels with separately tuned noise schedules, allowing latents to act as computational scratchpad before generating high-frequency pixel features.

Result: Achieves new state-of-the-art for diffusion transformer-based pixel generation on ImageNet at their compute scale, with analysis showing conditioning signal order is critical.

Conclusion: Latent Forcing bridges the gap between latent and pixel diffusion, achieving efficiency of latent diffusion while operating directly on raw images, with insights into tokenizer-distillation relationships and conditioning effects.

Abstract: Latent diffusion models excel at generating high-quality images but lose the benefits of end-to-end modeling. They discard information during image encoding, require a separately trained decoder, and model an auxiliary distribution to the raw data. In this paper, we propose Latent Forcing, a simple modification to existing architectures that achieves the efficiency of latent diffusion while operating on raw natural images. Our approach orders the denoising trajectory by jointly processing latents and pixels with separately tuned noise schedules. This allows the latents to act as a scratchpad for intermediate computation before high-frequency pixel features are generated. We find that the order of conditioning signals is critical, and we analyze this to explain differences between REPA distillation in the tokenizer and the diffusion model, conditional versus unconditional generation, and how tokenizer reconstruction quality relates to diffusability. Applied to ImageNet, Latent Forcing achieves a new state-of-the-art for diffusion transformer-based pixel generation at our compute scale.

Method: Leverage near-isotropic, high-resolution computed tomography angiography (CTA) data to train a single neural implicit function that jointly represents cardiac shapes from CMRI at any resolution, focusing on right ventricle and myocardium reconstruction.

Result: Achieved Dice similarity coefficients of 0.91±0.07 (RV) and 0.75±0.13 (MYO), and Hausdorff distances of 6.21±3.97mm (RV) and 7.53±5.13mm (MYO) when compared to reference 4-chamber segmentation masks from CMRI.

Conclusion: The method enables accurate, smooth, and anatomically plausible cardiac shape reconstruction from anisotropic CMRI, supporting improvements in cardiac shape analysis through neural implicit representations.

Abstract: The anisotropic nature of short-axis (SAX) cardiovascular magnetic resonance imaging (CMRI) limits cardiac shape analysis. To address this, we propose to leverage near-isotropic, higher resolution computed tomography angiography (CTA) data of the heart. We use this data to train a single neural implicit function to jointly represent cardiac shapes from CMRI at any resolution. We evaluate the method for the reconstruction of right ventricle (RV) and myocardium (MYO), where MYO simultaneously models endocardial and epicardial left-ventricle surfaces. Since high-resolution SAX reference segmentations are unavailable, we evaluate performance by extracting a 4-chamber (4CH) slice of RV and MYO from their reconstructed shapes. When compared with the reference 4CH segmentation masks from CMRI, our method achieved a Dice similarity coefficient of 0.91 $\pm$ 0.07 and 0.75 $\pm$ 0.13, and a Hausdorff distance of 6.21 $\pm$ 3.97 mm and 7.53 $\pm$ 5.13 mm for RV and MYO, respectively. Quantitative and qualitative assessment demonstrate the model’s ability to reconstruct accurate, smooth and anatomically plausible shapes, supporting improvements in cardiac shape analysis.

Method: Introduces MAU-Set dataset spanning multiple industrial domains with hierarchical tasks, and MAU-GPT multimodal large model with novel AMoE-LoRA mechanism that unifies anomaly-aware and generalist experts adaptation.

Result: MAU-GPT consistently outperforms prior state-of-the-art methods across all domains, demonstrating strong potential for scalable automated industrial inspection.

Conclusion: The proposed dataset and model provide comprehensive solutions for industrial anomaly understanding with enhanced generalization and reasoning capabilities across diverse defect classes.

Abstract: As industrial manufacturing scales, automating fine-grained product image analysis has become critical for quality control. However, existing approaches are hindered by limited dataset coverage and poor model generalization across diverse and complex anomaly patterns. To address these challenges, we introduce MAU-Set, a comprehensive dataset for Multi-type industrial Anomaly Understanding. It spans multiple industrial domains and features a hierarchical task structure, ranging from binary classification to complex reasoning. Alongside this dataset, we establish a rigorous evaluation protocol to facilitate fair and comprehensive model assessment. Building upon this foundation, we further present MAU-GPT, a domain-adapted multimodal large model specifically designed for industrial anomaly understanding. It incorporates a novel AMoE-LoRA mechanism that unifies anomaly-aware and generalist experts adaptation, enhancing both understanding and reasoning across diverse defect classes. Extensive experiments show that MAU-GPT consistently outperforms prior state-of-the-art methods across all domains, demonstrating strong potential for scalable and automated industrial inspection.

Method: Decomposes manipulation into object removal and reference-guided inpainting under explicit camera pose control, encoded within a unified diffusion process. Uses multi-task, multi-stage training to separate background, identity, and pose signals. Introduces scalable real-world dataset construction pipeline with paired image/video samples and estimated relative camera poses.

Result: Achieves state-of-the-art results in fidelity, viewpoint consistency, and controllability. First framework to unify fine-grained geometric control and real-world generalization for object manipulation without explicit 3D modeling.

Conclusion: Ctrl&Shift demonstrates that geometry-consistent object manipulation can be achieved without explicit 3D representations by decomposing the task and using a carefully designed diffusion framework with explicit camera pose control.

Abstract: Object-level manipulation, relocating or reorienting objects in images or videos while preserving scene realism, is central to film post-production, AR, and creative editing. Yet existing methods struggle to jointly achieve three core goals: background preservation, geometric consistency under viewpoint shifts, and user-controllable transformations. Geometry-based approaches offer precise control but require explicit 3D reconstruction and generalize poorly; diffusion-based methods generalize better but lack fine-grained geometric control. We present Ctrl&Shift, an end-to-end diffusion framework to achieve geometry-consistent object manipulation without explicit 3D representations. Our key insight is to decompose manipulation into two stages, object removal and reference-guided inpainting under explicit camera pose control, and encode both within a unified diffusion process. To enable precise, disentangled control, we design a multi-task, multi-stage training strategy that separates background, identity, and pose signals across tasks. To improve generalization, we introduce a scalable real-world dataset construction pipeline that generates paired image and video samples with estimated relative camera poses. Extensive experiments demonstrate that Ctrl&Shift achieves state-of-the-art results in fidelity, viewpoint consistency, and controllability. To our knowledge, this is the first framework to unify fine-grained geometric control and real-world generalization for object manipulation, without relying on any explicit 3D modeling.

Method: Developed a 9-direction single-shell ULF DTI sequence with two companion algorithms: 1) Bayesian bias field correction with angular dependence, and 2) CNN-based super-resolution (DiffSR) that’s generalizable across DTI datasets without retraining.

Result: The algorithms successfully recover microstructural and volumetric white matter information at ULF. DiffSR improves Alzheimer’s disease classification in synthetically degraded scans and enhances agreement between DTI metrics compared to un-degraded scans.

Conclusion: The proposed methods enable better white matter analysis with ULF DTI and facilitate DTI sequence harmonization. Code is publicly released to advance ULF reconstruction and general DTI methods.

Abstract: Portable, ultra-low-field (ULF) magnetic resonance imaging has the potential to expand access to neuroimaging but currently suffers from coarse spatial and angular resolutions and low signal-to-noise ratios. Diffusion tensor imaging (DTI), a sequence tailored to detect and reconstruct white matter tracts within the brain, is particularly prone to such imaging degradation due to inherent sequence design coupled with prolonged scan times. In addition, ULF DTI scans exhibit artifacting that spans both the space and angular domains, requiring a custom modelling algorithm for subsequent correction. We introduce a nine-direction, single-shell ULF DTI sequence, as well as a companion Bayesian bias field correction algorithm that possesses angular dependence and convolutional neural network-based superresolution algorithm that is generalizable across DTI datasets and does not require re-training (‘‘DiffSR’’). We show through a synthetic downsampling experiment and white matter assessment in real, matched ULF and high-field DTI scans that these algorithms can recover microstructural and volumetric white matter information at ULF. We also show that DiffSR can be directly applied to white matter-based Alzheimers disease classification in synthetically degraded scans, with notable improvements in agreement between DTI metrics, as compared to un-degraded scans. We freely disseminate the Bayesian bias correction algorithm and DiffSR with the goal of furthering progress on both ULF reconstruction methods and general DTI sequence harmonization. We release all code related to DiffSR for $\href{https://github.com/markolchanyi/DiffSR}{public \space use}$.

Method: Proposes DBTANet with: 1) Dual-branch Siamese encoder (frozen SAM for global semantics/boundaries + ResNet34 for local details), 2) Bidirectional Temporal Awareness Module for temporal dependencies, 3) Gaussian-smoothed Projection Module for boundary refinement.

Result: Extensive experiments on two public benchmarks demonstrate state-of-the-art performance, effectively integrating global semantics, local details, temporal reasoning, and boundary awareness.

Conclusion: DBTANet successfully addresses boundary blurring and temporal modeling limitations in semantic change detection through complementary feature representations and specialized modules.

Abstract: Semantic Change Detection (SCD) aims to detect and categorize land-cover changes from bi-temporal remote sensing images. Existing methods often suffer from blurred boundaries and inadequate temporal modeling, limiting segmentation accuracy. To address these issues, we propose a Dual-Branch Framework for Semantic Change Detection with Boundary and Temporal Awareness, termed DBTANet. Specifically, we utilize a dual-branch Siamese encoder where a frozen SAM branch captures global semantic context and boundary priors, while a ResNet34 branch provides local spatial details, ensuring complementary feature representations. On this basis, we design a Bidirectional Temporal Awareness Module (BTAM) to aggregate multi-scale features and capture temporal dependencies in a symmetric manner. Furthermore, a Gaussian-smoothed Projection Module (GSPM) refines shallow SAM features, suppressing noise while enhancing edge information for boundary-aware constraints. Extensive experiments on two public benchmarks demonstrate that DBTANet effectively integrates global semantics, local details, temporal reasoning, and boundary awareness, achieving state-of-the-art performance.

Method: Proposes Arbitrary Ratio Feature Compression (ARFC) with an auto-regressive Arbitrary Ratio Compressor (ARC) that controls compression ratio via generated token count, enhanced by Mixture of Solutions (MoS) module for uncertainty reduction and Entity Relation Graph Constraint (ERGC) for preserving semantic relationships.

Result: Outperforms existing approaches across various compression ratios on cross-modal retrieval, image classification, and image retrieval tasks, sometimes even surpassing uncompressed feature performance.

Conclusion: ARFC provides an effective and versatile solution for practical feature compression in resource-constrained scenarios, eliminating the need for multiple specialized models while maintaining high performance.

Abstract: Feature compression is increasingly important for improving the efficiency of downstream tasks, especially in applications involving large-scale or multi-modal data. While existing methods typically rely on dedicated models for achieving specific compression ratios, they are often limited in flexibility and generalization. In particular, retraining is necessary when adapting to a new compression ratio. To address this limitation, we propose a novel and flexible Arbitrary Ratio Feature Compression (ARFC) framework, which supports any compression ratio with a single model, eliminating the need for multiple specialized models. At its core, the Arbitrary Ratio Compressor (ARC) is an auto-regressive model that performs compression via next-token prediction. This allows the compression ratio to be controlled at inference simply by adjusting the number of generated tokens. To enhance the quality of the compressed features, two key modules are introduced. The Mixture of Solutions (MoS) module refines the compressed tokens by utilizing multiple compression results (solutions), reducing uncertainty and improving robustness. The Entity Relation Graph Constraint (ERGC) is integrated into the training process to preserve semantic and structural relationships during compression. Extensive experiments on cross-modal retrieval, image classification, and image retrieval tasks across multiple datasets demonstrate that our method consistently outperforms existing approaches at various compression ratios. Notably, in some cases, it even surpasses the performance of the original, uncompressed features. These results validate the effectiveness and versatility of ARFC for practical, resource-constrained scenarios.

Method: Proposes ImagineAgent framework that constructs cognitive maps modeling relationships between detected entities and candidate actions, then dynamically invokes tools (retrieval augmentation, image cropping, diffusion models) to gather domain knowledge and visual evidence for cross-modal alignment. Uses composite reward balancing prediction accuracy and tool efficiency.

Result: Achieves state-of-the-art performance on SWIG-HOI and HICO-DET datasets while requiring only about 20% of training data compared to existing methods, demonstrating both robustness and efficiency.

Conclusion: ImagineAgent effectively addresses cross-modal hallucinations and occlusion ambiguity in OV-HOI tasks through cognitive reasoning and generative imagination, offering a more robust and efficient approach to multimodal visual understanding.

Abstract: Multimodal Large Language Models have shown promising capabilities in bridging visual and textual reasoning, yet their reasoning capabilities in Open-Vocabulary Human-Object Interaction (OV-HOI) are limited by cross-modal hallucinations and occlusion-induced ambiguity. To address this, we propose \textbf{ImagineAgent}, an agentic framework that harmonizes cognitive reasoning with generative imagination for robust visual understanding. Specifically, our method innovatively constructs cognitive maps that explicitly model plausible relationships between detected entities and candidate actions. Subsequently, it dynamically invokes tools including retrieval augmentation, image cropping, and diffusion models to gather domain-specific knowledge and enriched visual evidence, thereby achieving cross-modal alignment in ambiguous scenarios. Moreover, we propose a composite reward that balances prediction accuracy and tool efficiency. Evaluations on SWIG-HOI and HICO-DET datasets demonstrate our SOTA performance, requiring approximately 20% of training data compared to existing methods, validating our robustness and efficiency.

Method: Proposes VasoMIM with two key designs: 1) anatomy-guided masking strategy that strategically masks vessel-containing patches to force learning of robust vascular semantics, and 2) anatomical consistency loss that preserves structural consistency between original and reconstructed images. Also introduces XA-170K, the largest X-ray angiogram pre-training dataset.

Result: Validated on four downstream tasks across six datasets, demonstrating superior transferability and achieving state-of-the-art performance compared to existing methods. Shows significant potential as a foundation model for X-ray angiogram analysis.

Conclusion: VasoMIM effectively bridges the gap in self-supervised learning for X-ray angiogram analysis by integrating domain-specific anatomical knowledge, and the accompanying XA-170K dataset enables large-scale pre-training for this medical imaging modality.

Abstract: X-ray angiography is the gold standard imaging modality for cardiovascular diseases. However, current deep learning approaches for X-ray angiogram analysis are severely constrained by the scarcity of annotated data. While large-scale self-supervised learning (SSL) has emerged as a promising solution, its potential in this domain remains largely unexplored, primarily due to the lack of effective SSL frameworks and large-scale datasets. To bridge this gap, we introduce a vascular anatomy-aware masked image modeling (VasoMIM) framework that explicitly integrates domain-specific anatomical knowledge. Specifically, VasoMIM comprises two key designs: an anatomy-guided masking strategy and an anatomical consistency loss. The former strategically masks vessel-containing patches to compel the model to learn robust vascular semantics, while the latter preserves structural consistency of vessels between original and reconstructed images, enhancing the discriminability of the learned representations. In conjunction with VasoMIM, we curate XA-170K, the largest X-ray angiogram pre-training dataset to date. We validate VasoMIM on four downstream tasks across six datasets, where it demonstrates superior transferability and achieves state-of-the-art performance compared to existing methods. These findings highlight the significant potential of VasoMIM as a foundation model for advancing a wide range of X-ray angiogram analysis tasks. VasoMIM and XA-170K will be available at https://github.com/Dxhuang-CASIA/XA-SSL.

Method: Proposes MFdiff with: 1) multi-modality feature fusion module to learn optimized fusion features from MR without introducing extraneous details, 2) diffusion model using fusion features as condition to iteratively generate SPET, 3) two-stage supervised learning strategy using generalized priors from simulated data and specific priors for in-vivo OOD data.

Result: MFdiff effectively restores high-quality SPET images from multi-modality inputs and outperforms state-of-the-art methods both qualitatively and quantitatively.

Conclusion: The proposed MFdiff successfully addresses challenges in multi-modality PET restoration using diffusion models with supervised learning strategies, achieving superior performance in restoring high-quality PET images.

Abstract: Positron emission tomography (PET) offers powerful functional imaging but involves radiation exposure. Efforts to reduce this exposure by lowering the radiotracer dose or scan time can degrade image quality. While using magnetic resonance (MR) images with clearer anatomical information to restore standard-dose PET (SPET) from low-dose PET (LPET) is a promising approach, it faces challenges with the inconsistencies in the structure and texture of multi-modality fusion, as well as the mismatch in out-of-distribution (OOD) data. In this paper, we propose a supervise-assisted multi-modality fusion diffusion model (MFdiff) for addressing these challenges for high-quality PET restoration. Firstly, to fully utilize auxiliary MR images without introducing extraneous details in the restored image, a multi-modality feature fusion module is designed to learn an optimized fusion feature. Secondly, using the fusion feature as an additional condition, high-quality SPET images are iteratively generated based on the diffusion model. Furthermore, we introduce a two-stage supervise-assisted learning strategy that harnesses both generalized priors from simulated in-distribution datasets and specific priors tailored to in-vivo OOD data. Experiments demonstrate that the proposed MFdiff effectively restores high-quality SPET images from multi-modality inputs and outperforms state-of-the-art methods both qualitatively and quantitatively.

Method: Uses a generative compression framework where restoration reconstructs from entropy-coded latent representations that enforce low-complexity structure. Two approaches: (1) conditional Wasserstein GAN-based compression denoiser controlling RDP trade-off, (2) conditional diffusion-based reconstruction with iterative denoising guided by compressed latents. Also provides theoretical guarantees for compression-based maximum-likelihood denoiser under Gaussian noise.

Result: Experiments on synthetic and real-noise benchmarks show consistent perceptual improvements while maintaining competitive distortion performance.

Conclusion: The proposed generative compression framework effectively addresses the limitations of distortion-driven denoising by balancing rate-distortion-perception trade-offs, achieving better perceptual quality while maintaining distortion performance.

Abstract: Image denoising aims to remove noise while preserving structural details and perceptual realism, yet distortion-driven methods often produce over-smoothed reconstructions, especially under strong noise and distribution shift. This paper proposes a generative compression framework for perception-based denoising, where restoration is achieved by reconstructing from entropy-coded latent representations that enforce low-complexity structure, while generative decoders recover realistic textures via perceptual measures such as learned perceptual image patch similarity (LPIPS) loss and Wasserstein distance. Two complementary instantiations are introduced: (i) a conditional Wasserstein GAN (WGAN)-based compression denoiser that explicitly controls the rate-distortion-perception (RDP) trade-off, and (ii) a conditional diffusion-based reconstruction strategy that performs iterative denoising guided by compressed latents. We further establish non-asymptotic guarantees for the compression-based maximum-likelihood denoiser under additive Gaussian noise, including bounds on reconstruction error and decoding error probability. Experiments on synthetic and real-noise benchmarks demonstrate consistent perceptual improvements while maintaining competitive distortion performance.

Method: Three-stage architecture: 1) Low-resolution motion generation for motion-consistent latent synthesis, 2) Video latent upsampling in latent space to reduce memory/computational overhead, 3) High-resolution content refinement using dual frequency experts (low-frequency for semantic coherence, high-frequency for fine-grained details).

Result: LUVE achieves superior photorealism and content fidelity in ultra-high-resolution video generation compared to existing methods. Comprehensive ablation studies validate the effectiveness of each component in the framework.

Conclusion: LUVE successfully addresses the challenges of ultra-high-resolution video generation through its latent-cascaded architecture with dual frequency experts, providing a scalable solution for high-quality video synthesis.

Abstract: Recent advances in video diffusion models have significantly improved visual quality, yet ultra-high-resolution (UHR) video generation remains a formidable challenge due to the compounded difficulties of motion modeling, semantic planning, and detail synthesis. To address these limitations, we propose \textbf{LUVE}, a \textbf{L}atent-cascaded \textbf{U}HR \textbf{V}ideo generation framework built upon dual frequency \textbf{E}xperts. LUVE employs a three-stage architecture comprising low-resolution motion generation for motion-consistent latent synthesis, video latent upsampling that performs resolution upsampling directly in the latent space to mitigate memory and computational overhead, and high-resolution content refinement that integrates low-frequency and high-frequency experts to jointly enhance semantic coherence and fine-grained detail generation. Extensive experiments demonstrate that our LUVE achieves superior photorealism and content fidelity in UHR video generation, and comprehensive ablation studies further validate the effectiveness of each component. The project is available at \href{https://unicornanrocinu.github.io/LUVE_web/}{https://github.io/LUVE/}.

Method: Proposes FlowAdapt framework based on optimal transport theory with two key components: 1) Wasserstein Greedy Sampling to filter redundant samples via bounded covering radius, and 2) Progressive Knowledge Transfer module to inject compressed early-stage representations into later stages through learnable pathways.

Result: Achieves state-of-the-art performance on three benchmarks using only 1% of trainable parameters, effectively bridging domain gaps with superior sample efficiency and generalization.

Conclusion: FlowAdapt successfully addresses the challenges of applying PEFT to multi-agent V2X perception, providing an efficient domain adaptation solution that maintains performance while using minimal parameters.

Abstract: Fast domain adaptation remains a fundamental challenge for deploying multi-agent systems across diverse environments in Vehicle-to-Everything (V2X) collaborative perception. Despite the success of Parameter-Efficient Fine-Tuning (PEFT) in natural language processing and conventional vision tasks, directly applying PEFT to multi-agent settings leads to significant performance degradation and training instability. In this work, we conduct a detailed analysis and identify two key factors: (i) inter-frame redundancy in heterogeneous sensory streams, and (ii) erosion of fine-grained semantics in deep-layer representations under PEFT adaptation. To address these issues, we propose FlowAdapt, a parameter-efficient framework grounded in optimal transport theory, which minimizes information transport costs across both data distributions and network hierarchies. Specifically, we introduce a Wasserstein Greedy Sampling strategy to selectively filter redundant samples via a bounded covering radius. Furthermore, Progressive Knowledge Transfer module is designed to progressively inject compressed early-stage representations into later stages through learnable pathways, alleviating semantic degradation in late-stage adaptation. Extensive experiments on three benchmarks demonstrate that FlowAdapt achieves state-of-the-art performance with only 1% of trainable parameters, effectively bridging domain gaps with superior sample efficiency and generalization.

Method: Proposes LLM-DRS framework with standardized reconnaissance plan, processes text metadata and image data into unified format, uses Deep CNNs to extract key attributes (damage state, material type, damage level), then feeds all data into LLM with carefully designed prompts to generate summary reports.

Result: The integration of LLMs into vision-based SHM shows promising potential for improving resilience through effective rapid post-disaster reconnaissance and automated report generation.

Conclusion: LLM-DRS framework successfully demonstrates how LLMs can enhance vision-based SHM by automating comprehensive damage assessment reporting, particularly valuable for rapid post-disaster reconnaissance.

Abstract: Artificial Intelligence (AI)-aided vision-based Structural Health Monitoring (SHM) has emerged as an effective approach for monitoring and assessing structural condition by analyzing image and video data. By integrating Computer Vision (CV) and Deep Learning (DL), vision-based SHM can automatically identify and localize visual patterns associated with structural damage. However, previous works typically generate only discrete outputs, such as damage class labels and damage region coordinates, requiring engineers to further reorganize and analyze these results for evaluation and decision-making. In late 2022, Large Language Models (LLMs) became popular across multiple fields, providing new insights into AI-aided vision-based SHM. In this study, a novel LLM-based Disaster Reconnaissance Summarization (LLM-DRS) framework is proposed. It introduces a standard reconnaissance plan in which the collection of vision data and corresponding metadata follows a well-designed on-site investigation process. Text-based metadata and image-based vision data are then processed and integrated into a unified format, where well-trained Deep Convolutional Neural Networks extract key attributes, including damage state, material type, and damage level. Finally, all data are fed into an LLM with carefully designed prompts, enabling the LLM-DRS to generate summary reports for individual structures or affected regions based on aggregated attributes and metadata. Results show that integrating LLMs into vision-based SHM, particularly for rapid post-disaster reconnaissance, demonstrates promising potential for improving resilience of the built environment through effective reconnaissance.

Method: Proposes PLOT-CT framework using Voronoi decomposition on pre-log sinograms to disentangle data into distinct underlying components embedded in separate latent spaces, enhancing discriminative feature learning and reconstruction accuracy.

Result: Achieves state-of-the-art performance with 2.36dB PSNR improvement over traditional methods at 1e4 incident photon level in pre-log domain.

Conclusion: PLOT-CT effectively addresses noise and reconstruction challenges in low-dose CT by leveraging pre-log domain information through Voronoi decomposition, significantly improving reconstruction accuracy.

Abstract: Low-dose computed tomography (LDCT) reconstruction is fundamentally challenged by severe noise and compromised data fidelity under reduced radiation exposure. Most existing methods operate either in the image or post-log projection domain, which fails to fully exploit the rich structural information in pre-log measurements while being highly susceptible to noise. The requisite logarithmic transformation critically amplifies noise within these data, imposing exceptional demands on reconstruction precision. To overcome these challenges, we propose PLOT-CT, a novel framework for Pre-Log vOronoi decomposiTion-assisted CT generation. Our method begins by applying Voronoi decomposition to pre-log sinograms, disentangling the data into distinct underlying components, which are embedded in separate latent spaces. This explicit decomposition significantly enhances the model’s capacity to learn discriminative features, directly improving reconstruction accuracy by mitigating noise and preserving information inherent in the pre-log domain. Extensive experiments demonstrate that PLOT-CT achieves state-of-the-art performance, attaining a 2.36dB PSNR improvement over traditional methods at the 1e4 incident photon level in the pre-log domain.

Method: PLESS uses hierarchical partitioning of images into spatially coherent regions, then propagates scribble information to refine pseudo-labels within semantically coherent regions. It’s model-agnostic and integrates with existing pseudo-label methods.

Result: Experiments on two cardiac MRI datasets (ACDC and MSCMRseg) across four scribble-supervised algorithms show consistent improvements in segmentation accuracy.

Conclusion: PLESS effectively enhances pseudo-label quality for scribble-supervised medical image segmentation, providing a generic framework that improves existing methods.

Abstract: Weakly supervised learning with scribble annotations uses sparse user-drawn strokes to indicate segmentation labels on a small subset of pixels. This annotation reduces the cost of dense pixel-wise labeling, but suffers inherently from noisy and incomplete supervision. Recent scribble-based approaches in medical image segmentation address this limitation using pseudo-label-based training; however, the quality of the pseudo-labels remains a key performance limit. We propose PLESS, a generic pseudo-label enhancement strategy which improves reliability and spatial consistency. It builds on a hierarchical partitioning of the image into a hierarchy of spatially coherent regions. PLESS propagates scribble information to refine pseudo-labels within semantically coherent regions. The framework is model-agnostic and easily integrates into existing pseudo-label methods. Experiments on two public cardiac MRI datasets (ACDC and MSCMRseg) across four scribble-supervised algorithms show consistent improvements in segmentation accuracy. Code will be made available on GitHub upon acceptance.

Method: ScalSelect extracts visual features most attended by instruction tokens in the target VLM to create sample representations, then identifies samples whose representations best approximate the dominant subspace of the full dataset, enabling scalable importance scoring without pairwise comparisons.

Result: ScalSelect achieves over 97.5% performance of full dataset training using only 16% of data, and even outperforms full-data training in some settings across multiple VLMs, datasets, and selection budgets.

Conclusion: ScalSelect provides an efficient, scalable training-free data selection method for vision-language instruction tuning that significantly reduces computational costs while maintaining or improving performance.

Abstract: Large-scale Visual Instruction Tuning (VIT) has become a key paradigm for advancing the performance of vision-language models (VLMs) across various multimodal tasks. However, training on the large-scale datasets is computationally expensive and inefficient due to redundancy in the data, which motivates the need for multimodal data selection to improve training efficiency. Existing data selection methods for VIT either require costly training or gradient computation. Training-free alternatives often depend on proxy models or datasets, instruction-agnostic representations, and pairwise similarity with quadratic complexity, limiting scalability and representation fidelity. In this work, we propose ScalSelect, a scalable training-free multimodal data selection method with linear-time complexity with respect to the number of samples, eliminating the need for external models or auxiliary datasets. ScalSelect first constructs sample representations by extracting visual features most attended by instruction tokens in the target VLM, capturing instruction-relevant information. It then identifies samples whose representations best approximate the dominant subspace of the full dataset representations, enabling scalable importance scoring without pairwise comparisons. Extensive experiments across multiple VLMs, datasets, and selection budgets demonstrate that ScalSelect achieves over 97.5% of the performance of training on the full dataset using only 16% of the data, and even outperforms full-data training in some settings. The code is available at \href{https://github.com/ChangtiWu/ScalSelect}{ScalSelect}.

Method: Encode surface reconstruction as solution to Poisson’s equation, connect to physics (electrostatic potential), use Green’s functions for closed-form parametric expression, leverage linearity for superposition of solutions.

Result: Method shows improved results in approximating high-frequency details compared to existing approaches, even with small number of shape priors.

Conclusion: Poisson’s equation as proxy PDE with Green’s functions provides effective alternative to Eikonal PDE approaches for surface reconstruction, particularly for capturing fine details.

Abstract: Implicit shape representation, such as SDFs, is a popular approach to recover the surface of a 3D shape as the level sets of a scalar field. Several methods approximate SDFs using machine learning strategies that exploit the knowledge that SDFs are solutions of the Eikonal partial differential equation (PDEs). In this work, we present a novel approach to surface reconstruction by encoding it as a solution to a proxy PDE, namely Poisson’s equation. Then, we explore the connection between Poisson’s equation and physics, e.g., the electrostatic potential due to a positive charge density. We employ Green’s functions to obtain a closed-form parametric expression for the PDE’s solution, and leverage the linearity of our proxy PDE to find the target shape’s implicit field as a superposition of solutions. Our method shows improved results in approximating high-frequency details, even with a small number of shape priors.

Method: Evaluated three ResNet-based architectures (BrainNet, BrainNeXt, DilationNet) against gradient-based adversarial attacks (FGSM, PGD) across three preprocessing configurations: full-sized augmented, shrunk augmented, and shrunk non-augmented MRI datasets.

Result: BrainNeXt models showed highest robustness to black-box attacks due to increased cardinality, but produced weaker transferable adversarial samples. BrainNet and Dilation models were more vulnerable to attacks from each other, especially under PGD with higher iterations. Shrunk non-augmented data significantly reduced model resilience despite high untampered test accuracy.

Conclusion: Joint evaluation of classification performance and adversarial robustness is crucial for reliable real-world deployment in brain MRI analysis, with input resolution and data augmentation playing key roles in vulnerability.

Abstract: Adversarial robustness in deep learning models for brain tumor classification remains an underexplored yet critical challenge, particularly for clinical deployment scenarios involving MRI data. In this work, we investigate the susceptibility and resilience of several ResNet-based architectures, referred to as BrainNet, BrainNeXt and DilationNet, against gradient-based adversarial attacks, namely FGSM and PGD. These models, based on ResNet, ResNeXt, and dilated ResNet variants respectively, are evaluated across three preprocessing configurations (i) full-sized augmented, (ii) shrunk augmented and (iii) shrunk non-augmented MRI datasets. Our experiments reveal that BrainNeXt models exhibit the highest robustness to black-box attacks, likely due to their increased cardinality, though they produce weaker transferable adversarial samples. In contrast, BrainNet and Dilation models are more vulnerable to attacks from each other, especially under PGD with higher iteration steps and $α$ values. Notably, shrunk and non-augmented data significantly reduce model resilience, even when the untampered test accuracy remains high, highlighting a key trade-off between input resolution and adversarial vulnerability. These results underscore the importance of jointly evaluating classification performance and adversarial robustness for reliable real-world deployment in brain MRI analysis.

Method: Proposes GR-Diffusion framework that combines 3D discrete Gaussian representation (GR) with diffusion models. GR generates a reference 3D PET image from projection data, providing geometric priors. This reference guides the diffusion process through hierarchical guidance: fine-grained guidance uses differences to refine local details, while coarse-grained guidance uses multi-scale difference maps to correct deviations, allowing sequential integration of geometric priors and recovery of sub-voxel information.

Result: Experimental results on UDPET and Clinical datasets with varying dose levels show GR-Diffusion outperforms state-of-the-art methods in enhancing 3D whole-body PET image quality and preserving physiological details.

Conclusion: GR-Diffusion successfully integrates geometric priors from Gaussian representation with generative power of diffusion models for improved PET reconstruction, overcoming limitations of conventional methods and demonstrating superior performance in preserving structural details.

Abstract: Positron emission tomography (PET) reconstruction is a critical challenge in molecular imaging, often hampered by noise amplification, structural blurring, and detail loss due to sparse sampling and the ill-posed nature of inverse problems. The three-dimensional discrete Gaussian representation (GR), which efficiently encodes 3D scenes using parameterized discrete Gaussian distributions, has shown promise in computer vision. In this work, we pro-pose a novel GR-Diffusion framework that synergistically integrates the geometric priors of GR with the generative power of diffusion models for 3D low-dose whole-body PET reconstruction. GR-Diffusion employs GR to generate a reference 3D PET image from projection data, establishing a physically grounded and structurally explicit benchmark that overcomes the low-pass limitations of conventional point-based or voxel-based methods. This reference image serves as a dual guide during the diffusion process, ensuring both global consistency and local accuracy. Specifically, we employ a hierarchical guidance mechanism based on the GR reference. Fine-grained guidance leverages differences to refine local details, while coarse-grained guidance uses multi-scale difference maps to correct deviations. This strategy allows the diffusion model to sequentially integrate the strong geometric prior from GR and recover sub-voxel information. Experimental results on the UDPET and Clinical datasets with varying dose levels show that GR-Diffusion outperforms state-of-the-art methods in enhancing 3D whole-body PET image quality and preserving physiological details.

Method: Three key elements: 1) Lightweight importance predictor with sliding windows estimates token importance scores, 2) Supervised training using auxiliary path for pseudo-supervision from all-token LLM pass, 3) Anchor-context merging module partitions tokens and merges context tokens into relevant anchors to reduce redundancy.

Result: Experiments on LangAuto benchmark show SToRM outperforms state-of-the-art E2E driving MLLMs under same reduced-token budget, maintaining all-token performance while reducing computational cost by up to 30x.

Conclusion: SToRM enables efficient end-to-end driving with multimodal LLMs while maintaining performance comparable to using all tokens, addressing computational limitations in autonomous vehicles.

Abstract: In autonomous driving, end-to-end (E2E) driving systems that predict control commands directly from sensor data have achieved significant advancements. For safe driving in unexpected scenarios, these systems may additionally rely on human interventions such as natural language instructions. Using a multi-modal large language model (MLLM) facilitates human-vehicle interaction and can improve performance in such scenarios. However, this approach requires substantial computational resources due to its reliance on an LLM and numerous visual tokens from sensor inputs, which are limited in autonomous vehicles. Many MLLM studies have explored reducing visual tokens, but often suffer end-task performance degradation compared to using all tokens. To enable efficient E2E driving while maintaining performance comparable to using all tokens, this paper proposes the first Supervised Token Reduction framework for multi-modal LLMs (SToRM). The proposed framework consists of three key elements. First, a lightweight importance predictor with short-term sliding windows estimates token importance scores. Second, a supervised training approach uses an auxiliary path to obtain pseudo-supervision signals from an all-token LLM pass. Third, an anchor-context merging module partitions tokens into anchors and context tokens, and merges context tokens into relevant anchors to reduce redundancy while minimizing information loss. Experiments on the LangAuto benchmark show that SToRM outperforms state-of-the-art E2E driving MLLMs under the same reduced-token budget, maintaining all-token performance while reducing computational cost by up to 30x.

Method: Uses hierarchical emotion representation with learnable prototypes that evolve during training, employing vision-language alignment. Features controllable generation pipeline with multi-prototype guidance, temporal blending, and attention reweighting.

Result: Superior performance over existing methods in both qualitative and quantitative evaluations. Enables emotional image outpainting, stylized generation, and emotional panorama generation for VR environments.

Conclusion: EmoSpace facilitates immersive visual content generation with fine-grained emotion control, supporting applications in therapy, education, storytelling, artistic creation, and cultural preservation.

Abstract: Emotion is important for creating compelling virtual reality (VR) content. Although some generative methods have been applied to lower the barrier to creating emotionally rich content, they fail to capture the nuanced emotional semantics and the fine-grained control essential for immersive experiences. To address these limitations, we introduce EmoSpace, a novel framework for emotion-aware content generation that learns dynamic, interpretable emotion prototypes through vision-language alignment. We employ a hierarchical emotion representation with rich learnable prototypes that evolve during training, enabling fine-grained emotional control without requiring explicit emotion labels. We develop a controllable generation pipeline featuring multi-prototype guidance, temporal blending, and attention reweighting that supports diverse applications, including emotional image outpainting, stylized generation, and emotional panorama generation for VR environments. Our experiments demonstrate the superior performance of EmoSpace over existing methods in both qualitative and quantitative evaluations. Additionally, we present a comprehensive user study investigating how VR environments affect emotional perception compared to desktop settings. Our work facilitates immersive visual content generation with fine-grained emotion control and supports applications like therapy, education, storytelling, artistic creation, and cultural preservation. Code and models will be made publicly available.

Method: Uses hierarchical instance tree of semantic cues with cross-view grouping and conditional substitution to suppress over-/under-segmentation, yielding view-consistent masks and robust 3D instances enriched with open-vocabulary semantic embeddings; includes consistency-aware update for scene changes

Result: Outperforms state-of-the-art baselines in cluttered and sparse-view scenarios; achieves AP@25 of 61.66 on heavy-clutter sequences (2.2x higher than baselines); with only 4 input views surpasses MaskClustering with 8 views by more than 2x

Conclusion: Clutt3R-Seg provides robust 3D instance segmentation for language-grounded grasping in cluttered environments, enabling accurate target selection from natural language instructions with efficient adaptation to scene changes

Abstract: Reliable 3D instance segmentation is fundamental to language-grounded robotic manipulation. Its critical application lies in cluttered environments, where occlusions, limited viewpoints, and noisy masks degrade perception. To address these challenges, we present Clutt3R-Seg, a zero-shot pipeline for robust 3D instance segmentation for language-grounded grasping in cluttered scenes. Our key idea is to introduce a hierarchical instance tree of semantic cues. Unlike prior approaches that attempt to refine noisy masks, our method leverages them as informative cues: through cross-view grouping and conditional substitution, the tree suppresses over- and under-segmentation, yielding view-consistent masks and robust 3D instances. Each instance is enriched with open-vocabulary semantic embeddings, enabling accurate target selection from natural language instructions. To handle scene changes during multi-stage tasks, we further introduce a consistency-aware update that preserves instance correspondences from only a single post-interaction image, allowing efficient adaptation without rescanning. Clutt3R-Seg is evaluated on both synthetic and real-world datasets, and validated on a real robot. Across all settings, it consistently outperforms state-of-the-art baselines in cluttered and sparse-view scenarios. Even on the most challenging heavy-clutter sequences, Clutt3R-Seg achieves an AP@25 of 61.66, over 2.2x higher than baselines, and with only four input views it surpasses MaskClustering with eight views by more than 2x. The code is available at: https://github.com/jeonghonoh/clutt3r-seg.

Method: Collected first neck-mounted gaze dataset (4 hours from 8 participants), evaluated transformer-based GLC model, proposed gaze out-of-bound classification task, and multi-view co-learning approach with geometry-aware auxiliary loss.

Result: Gaze out-of-bound classification improved performance over standard fine-tuning, but multi-view co-learning approach did not yield gains despite using geometry-aware auxiliary loss.

Conclusion: Neck-mounted gaze estimation is a viable task with potential applications, and gaze out-of-bound classification is beneficial, but multi-view co-learning needs further investigation for this specific viewpoint.

Abstract: This paper introduces neck-mounted view gaze estimation, a new task that estimates user gaze from the neck-mounted camera perspective. Prior work on egocentric gaze estimation, which predicts device wearer’s gaze location within the camera’s field of view, mainly focuses on head-mounted cameras while alternative viewpoints remain underexplored. To bridge this gap, we collect the first dataset for this task, consisting of approximately 4 hours of video collected from 8 participants during everyday activities. We evaluate a transformer-based gaze estimation model, GLC, on the new dataset and propose two extensions: an auxiliary gaze out-of-bound classification task and a multi-view co-learning approach that jointly trains head-view and neck-view models using a geometry-aware auxiliary loss. Experimental results show that incorporating gaze out-of-bound classification improves performance over standard fine-tuning, while the co-learning approach does not yield gains. We further analyze these results and discuss implications for neck-mounted gaze estimation.

Method: Transform Domain Fusion UNet (TD-FusionUNet) with trainable Hadamard Transform and Discrete Cosine Transform layers to capture frequency components in orthogonalized latent spaces, plus custom preprocessing including random margin cropping and Gaussian mixture model for sparse pre-fire mask representation.

Result: Achieves F1 score of 0.591 with only 370k parameters, outperforming UNet baseline with ResNet18 encoder on WildfireSpreadTS dataset while using substantially fewer parameters.

Conclusion: TD-FusionUNet effectively balances accuracy and efficiency for real-time wildfire prediction in resource-limited settings through innovative frequency domain fusion and lightweight architecture.

Abstract: We developed a lightweight and computationally efficient tool for next-day wildfire spread prediction using multimodal satellite data as input. The deep learning model, which we call Transform Domain Fusion UNet (TD-FusionUNet), incorporates trainable Hadamard Transform and Discrete Cosine Transform layers that apply two-dimensional transforms, enabling the network to capture essential “frequency” components in orthogonalized latent spaces. Additionally, we introduce custom preprocessing techniques, including random margin cropping and a Gaussian mixture model, to enrich the representation of the sparse pre-fire masks and enhance the model’s generalization capability. The TD-FusionUNet is evaluated on two datasets which are the Next-Day Wildfire Spread dataset released by Google Research in 2023, and WildfireSpreadTS dataset. Our proposed TD-FusionUNet achieves an F1 score of 0.591 with 370k parameters, outperforming the UNet baseline using ResNet18 as the encoder reported in the WildfireSpreadTS dataset while using substantially fewer parameters. These results show that the proposed latent space fusion model balances accuracy and efficiency under a lightweight setting, making it suitable for real time wildfire prediction applications in resource limited environments.

Method: Proposes RI-Mamba with global/local reference frames to disentangle pose from geometry, Hilbert sorting for token sequences, orientational embeddings with feature-wise linear modulation, and cross-modal contrastive learning with automated triplet generation.

Result: Achieves state-of-the-art performance on OmniObject3D benchmark across 200+ object categories under arbitrary orientations, demonstrating superior representational capacity and robustness.

Conclusion: RI-Mamba effectively addresses rotation invariance in 3D shape retrieval, enabling practical applications with diverse object categories and arbitrary orientations.

Abstract: 3D assets have rapidly expanded in quantity and diversity due to the growing popularity of virtual reality and gaming. As a result, text-to-shape retrieval has become essential in facilitating intuitive search within large repositories. However, existing methods require canonical poses and support few object categories, limiting their real-world applicability where objects can belong to diverse classes and appear in random orientations. To address this challenge, we propose RI-Mamba, the first rotation-invariant state-space model for point clouds. RI-Mamba defines global and local reference frames to disentangle pose from geometry and uses Hilbert sorting to construct token sequences with meaningful geometric structure while maintaining rotation invariance. We further introduce a novel strategy to compute orientational embeddings and reintegrate them via feature-wise linear modulation, effectively recovering spatial context and enhancing model expressiveness. Our strategy is inherently compatible with state-space models and operates in linear time. To scale up retrieval, we adopt cross-modal contrastive learning with automated triplet generation, allowing training on diverse datasets without manual annotation. Extensive experiments demonstrate RI-Mamba’s superior representational capacity and robustness, achieving state-of-the-art performance on the OmniObject3D benchmark across more than 200 object categories under arbitrary orientations. Our code will be made available at https://github.com/ndkhanh360/RI-Mamba.git.

Method: Developed a semantically conditioned latent diffusion model (LDM) trained on 99,349 DSA frames from a single-center dataset. Used text embeddings to encode anatomical circulation (anterior vs. posterior) and C-arm positions for explicit control over image generation.

Result: Generated images achieved Likert scores of 3.1-3.3 from medical experts with high inter-rater reliability (ICC 0.80-0.87). Low FID score of 15.27 indicates distributional similarity to real DSA frames.

Conclusion: Semantically controlled LDMs can produce realistic synthetic DSA images suitable for algorithm development, research, and training applications in cerebrovascular medicine.

Abstract: Digital subtraction angiography (DSA) plays a central role in the diagnosis and treatment of cerebrovascular disease, yet its invasive nature and high acquisition cost severely limit large-scale data collection and public data sharing. Therefore, we developed a semantically conditioned latent diffusion model (LDM) that synthesizes arterial-phase cerebral DSA frames under explicit control of anatomical circulation (anterior vs.\ posterior) and canonical C-arm positions. We curated a large single-centre DSA dataset of 99,349 frames and trained a conditional LDM using text embeddings that encoded anatomy and acquisition geometry. To assess clinical realism, four medical experts, including two neuroradiologists, one neurosurgeon, and one internal medicine expert, systematically rated 400 synthetic DSA images using a 5-grade Likert scale for evaluating proximal large, medium, and small peripheral vessels. The generated images achieved image-wise overall Likert scores ranging from 3.1 to 3.3, with high inter-rater reliability (ICC(2,k) = 0.80–0.87). Distributional similarity to real DSA frames was supported by a low median Fréchet inception distance (FID) of 15.27. Our results indicate that semantically controlled LDMs can produce realistic synthetic DSAs suitable for downstream algorithm development, research, and training.

Method: Proposes Tomographic Geometry Field (TG-Field) with: 1) multi-resolution hash encoder for spatial priors, 2) time-conditioned representations and spatiotemporal attention for dynamic cases, 3) motion-flow network for respiratory motion modeling, and 4) regularization of primitive parameters under ultra-sparse settings.

Result: Extensive experiments on synthetic and real-world datasets show TG-Field consistently outperforms existing methods, achieving state-of-the-art reconstruction accuracy under highly sparse-view conditions.

Conclusion: TG-Field successfully addresses challenges in sparse-view and dynamic CT reconstruction through geometry-aware Gaussian deformation and spatiotemporal modeling, demonstrating superior performance over existing approaches.

Abstract: 3D Gaussian Splatting (3DGS) has revolutionized 3D scene representation with superior efficiency and quality. While recent adaptations for computed tomography (CT) show promise, they struggle with severe artifacts under highly sparse-view projections and dynamic motions. To address these challenges, we propose Tomographic Geometry Field (TG-Field), a geometry-aware Gaussian deformation framework tailored for both static and dynamic CT reconstruction. A multi-resolution hash encoder is employed to capture local spatial priors, regularizing primitive parameters under ultra-sparse settings. We further extend the framework to dynamic reconstruction by introducing time-conditioned representations and a spatiotemporal attention block to adaptively aggregate features, thereby resolving spatiotemporal ambiguities and enforcing temporal coherence. In addition, a motion-flow network models fine-grained respiratory motion to track local anatomical deformations. Extensive experiments on synthetic and real-world datasets demonstrate that TG-Field consistently outperforms existing methods, achieving state-of-the-art reconstruction accuracy under highly sparse-view conditions.

Method: Developed a modular multi-LLM pipeline integrating 3D asset retrieval, domain knowledge injection, and Unreal Engine API code generation. Used hybrid LLM optimization techniques including few-shot prompting, RAG, finetuning, and validation for accuracy and scalability.

Result: System generates realistic agricultural 3D environments with planting layouts and environmental context. User study confirmed realism and familiarity, while expert comparison showed significant time savings over manual scene design.

Conclusion: Multi-LLM pipelines effectively automate domain-specific 3D scene generation with improved reliability and precision. Future work includes expanding asset hierarchy, real-time generation, and adapting to other simulation domains.

Abstract: Procedural generation techniques in 3D rendering engines have revolutionized the creation of complex environments, reducing reliance on manual design. Recent approaches using Large Language Models (LLMs) for 3D scene generation show promise but often lack domain-specific reasoning, verification mechanisms, and modular design. These limitations lead to reduced control and poor scalability. This paper investigates the use of LLMs to generate agricultural synthetic simulation environments from natural language prompts, specifically to address the limitations of lacking domain-specific reasoning, verification mechanisms, and modular design. A modular multi-LLM pipeline was developed, integrating 3D asset retrieval, domain knowledge injection, and code generation for the Unreal rendering engine using its API. This results in a 3D environment with realistic planting layouts and environmental context, all based on the input prompt and the domain knowledge. To enhance accuracy and scalability, the system employs a hybrid strategy combining LLM optimization techniques such as few-shot prompting, Retrieval-Augmented Generation (RAG), finetuning, and validation. Unlike monolithic models, the modular architecture enables structured data handling, intermediate verification, and flexible expansion. The system was evaluated using structured prompts and semantic accuracy metrics. A user study assessed realism and familiarity against real-world images, while an expert comparison demonstrated significant time savings over manual scene design. The results confirm the effectiveness of multi-LLM pipelines in automating domain-specific 3D scene generation with improved reliability and precision. Future work will explore expanding the asset hierarchy, incorporating real-time generation, and adapting the pipeline to other simulation domains beyond agriculture.

Method: Bidirectional coupling of Visual Odometry and Gaussian Splatting within an Expectation-Maximization framework. Uses Gaussian Splat Initialization that leverages VO’s image info, keyframe poses, and pixel associations to produce close approximations to final Gaussian scene without heuristics.

Result: System operates in real-time while achieving state-of-the-art geometric/photometric fidelity of reconstructed scenes and tracking accuracy, validated through extensive experiments.

Conclusion: GSO-SLAM effectively integrates VO and Gaussian Splatting for real-time dense SLAM with superior reconstruction quality and tracking performance compared to existing methods.

Abstract: We propose GSO-SLAM, a real-time monocular dense SLAM system that leverages Gaussian scene representation. Unlike existing methods that couple tracking and mapping with a unified scene, incurring computational costs, or loosely integrate them with well-structured tracking frameworks, introducing redundancies, our method bidirectionally couples Visual Odometry (VO) and Gaussian Splatting (GS). Specifically, our approach formulates joint optimization within an Expectation-Maximization (EM) framework, enabling the simultaneous refinement of VO-derived semi-dense depth estimates and the GS representation without additional computational overhead. Moreover, we present Gaussian Splat Initialization, which utilizes image information, keyframe poses, and pixel associations from VO to produce close approximations to the final Gaussian scene, thereby eliminating the need for heuristic methods. Through extensive experiments, we validate the effectiveness of our method, showing that it not only operates in real time but also achieves state-of-the-art geometric/photometric fidelity of the reconstructed scene and tracking accuracy.

Method: Proposes visual prompting paradigm that avoids coordinate alignment by reformulating per-frame coordinate prediction as instance-level identification with unique, temporally consistent object IDs embedded as visual prompts. Introduces STVG-R1, the first reinforcement learning framework for STVG with task-driven reward optimizing temporal accuracy, spatial consistency, and structural format regularization.

Result: STVG-R1 surpasses baseline Qwen2.5-VL-7B by 20.9% m_IoU on HCSTVG-v2 benchmark, establishing new SOTA. Also shows strong zero-shot generalization to multi-object referring video object segmentation, achieving SOTA 47.3% J&F on MeViS.

Conclusion: The visual prompting paradigm effectively addresses coordinate misalignment in VLMs for dense prediction tasks, and the reinforcement learning framework provides comprehensive optimization for spatial-temporal video grounding with strong generalization capabilities.

Abstract: In vision-language models (VLMs), misalignment between textual descriptions and visual coordinates often induces hallucinations. This issue becomes particularly severe in dense prediction tasks such as spatial-temporal video grounding (STVG). Prior approaches typically focus on enhancing visual-textual alignment or attaching auxiliary decoders. However, these strategies inevitably introduce additional trainable modules, leading to significant annotation costs and computational overhead. In this work, we propose a novel visual prompting paradigm that avoids the difficult problem of aligning coordinates across modalities. Specifically, we reformulate per-frame coordinate prediction as a compact instance-level identification problem by assigning each object a unique, temporally consistent ID. These IDs are embedded into the video as visual prompts, providing explicit and interpretable inputs to the VLMs. Furthermore, we introduce STVG-R1, the first reinforcement learning framework for STVG, which employs a task-driven reward to jointly optimize temporal accuracy, spatial consistency, and structural format regularization. Extensive experiments on six benchmarks demonstrate the effectiveness of our approach. STVG-R1 surpasses the baseline Qwen2.5-VL-7B by a remarkable margin of 20.9% on m_IoU on the HCSTVG-v2 benchmark, establishing a new state of the art (SOTA). Surprisingly, STVG-R1 also exhibits strong zero-shot generalization to multi-object referring video object segmentation tasks, achieving a SOTA 47.3% J&F on MeViS.

Method: Large-scale experimental study on targeted adaptation of general VLMs for e-commerce, plus novel evaluation suite covering deep product understanding, instruction following, and attribute extraction

Result: Targeted adaptation substantially improves e-commerce performance while preserving broad multimodal capabilities

Conclusion: Effective strategies exist for adapting general VLMs to specialized domains like e-commerce without compromising their general multimodal capabilities

Abstract: E-commerce product understanding demands by nature, strong multimodal comprehension from text, images, and structured attributes. General-purpose Vision-Language Models (VLMs) enable generalizable multimodal latent modelling, yet there is no documented, well-known strategy for adapting them to the attribute-centric, multi-image, and noisy nature of e-commerce data, without sacrificing general performance. In this work, we show through a large-scale experimental study, how targeted adaptation of general VLMs can substantially improve e-commerce performance while preserving broad multimodal capabilities. Furthermore, we propose a novel extensive evaluation suite covering deep product understanding, strict instruction following, and dynamic attribute extraction.

Method: Construct object-aligned auxiliary views by leveraging object-centric attention in self-supervised Vision Transformers, removing the most salient visual evidence to disrupt unsupported tokens and produce stronger contrast signals

Result: Demonstrates consistent gains on two popular object hallucination benchmarks across two MLLMs

Conclusion: The method is prompt-agnostic, model-agnostic, can be seamlessly integrated into existing VCD pipeline with minimal computation overhead (single cacheable forward pass)

Abstract: We study object hallucination in Multimodal Large Language Models (MLLMs) and improve visual contrastive decoding (VCD) by constructing an object-aligned auxiliary view. We leverage object-centric attention in self-supervised Vision Transformers. In particular, we remove the most salient visual evidence to construct an auxiliary view that disrupts unsupported tokens and produces a stronger contrast signal. Our method is prompt-agnostic, model-agnostic, and can be seamlessly plugged into the existing VCD pipeline with little computation overhead, i.e., a single cacheable forward pass. Empirically, our method demonstrates consistent gains on two popular object hallucination benchmarks across two MLLMs.

Method: Proposes Adaptive Debiasing Tsallis Entropy (ADTE) that customizes class-specific parameters q^l by normalizing estimated label bias from test instances, enabling accurate high-confidence view selection and integration with label adjustment strategies.

Result: ADTE outperforms state-of-the-art methods on ImageNet and its variants, achieving highest average performance on 10 cross-domain benchmarks across different model architectures and text prompts.

Conclusion: ADTE provides an effective solution to CLIP’s pretraining bias in TTA by replacing Shannon Entropy with adaptive Tsallis Entropy, improving adaptation without requiring distribution-specific hyperparameter tuning.

Abstract: Mainstream Test-Time Adaptation (TTA) methods for adapting vision-language models, e.g., CLIP, typically rely on Shannon Entropy (SE) at test time to measure prediction uncertainty and inconsistency. However, since CLIP has a built-in bias from pretraining on highly imbalanced web-crawled data, SE inevitably results in producing biased estimates of uncertainty entropy. To address this issue, we notably find and demonstrate that Tsallis Entropy (TE), a generalized form of SE, is naturally suited for characterizing biased distributions by introducing a non-extensive parameter q, with the performance of SE serving as a lower bound for TE. Building upon this, we generalize TE into Adaptive Debiasing Tsallis Entropy (ADTE) for TTA, customizing a class-specific parameter q^l derived by normalizing the estimated label bias from continuously incoming test instances, for each category. This adaptive approach allows ADTE to accurately select high-confidence views and seamlessly integrate with a label adjustment strategy to enhance adaptation, without introducing distribution-specific hyperparameter tuning. Besides, our investigation reveals that both TE and ADTE can serve as direct, advanced alternatives to SE in TTA, without any other modifications. Experimental results show that ADTE outperforms state-of-the-art methods on ImageNet and its five variants, and achieves the highest average performance on 10 cross-domain benchmarks, regardless of the model architecture or text prompts used. Our code is available at https://github.com/Jinx630/ADTE.

Method: 1) Dual-stream architecture disentangles retrieval-augmented object generation from hierarchical environmental orchestration; 2) Physics-aware closed-loop mechanism with PostProcess Agent scripting dynamics and VLM-Motion Critic performing self-reflection to iteratively refine simulation code.

Result: Outperforms baselines on Code4D benchmark with 41% SGS gain and 49% higher Richness, uniquely generating physics-aware dynamics absent in prior static methods.

Conclusion: Code2Worlds advances 4D generation by bridging language-to-simulation with physics-aware closed-loop refinement, enabling more realistic world simulation grounded in physical laws.

Abstract: Achieving spatial intelligence requires moving beyond visual plausibility to build world simulators grounded in physical laws. While coding LLMs have advanced static 3D scene generation, extending this paradigm to 4D dynamics remains a critical frontier. This task presents two fundamental challenges: multi-scale context entanglement, where monolithic generation fails to balance local object structures with global environmental layouts; and a semantic-physical execution gap, where open-loop code generation leads to physical hallucinations lacking dynamic fidelity. We introduce Code2Worlds, a framework that formulates 4D generation as language-to-simulation code generation. First, we propose a dual-stream architecture that disentangles retrieval-augmented object generation from hierarchical environmental orchestration. Second, to ensure dynamic fidelity, we establish a physics-aware closed-loop mechanism in which a PostProcess Agent scripts dynamics, coupled with a VLM-Motion Critic that performs self-reflection to iteratively refine simulation code. Evaluations on the Code4D benchmark show Code2Worlds outperforms baselines with a 41% SGS gain and 49% higher Richness, while uniquely generating physics-aware dynamics absent in prior static methods. Code: https://github.com/AIGeeksGroup/Code2Worlds. Website: https://aigeeksgroup.github.io/Code2Worlds.

Method: Proposes Disentangled Flow Guidance for lighting control injection while preserving geometry, and Temporal Consistent Attention within IC-Light architecture with deterministic regularization to eliminate flickering.

Result: Achieves competitive performance in temporal consistency and lighting fidelity, robustly handling camera rotations from -90 to 90 degrees.

Conclusion: Light4D provides an effective training-free solution for consistent 4D video relighting under challenging viewpoint changes.

Abstract: Recent advances in diffusion-based generative models have established a new paradigm for image and video relighting. However, extending these capabilities to 4D relighting remains challenging, due primarily to the scarcity of paired 4D relighting training data and the difficulty of maintaining temporal consistency across extreme viewpoints. In this work, we propose Light4D, a novel training-free framework designed to synthesize consistent 4D videos under target illumination, even under extreme viewpoint changes. First, we introduce Disentangled Flow Guidance, a time-aware strategy that effectively injects lighting control into the latent space while preserving geometric integrity. Second, to reinforce temporal consistency, we develop Temporal Consistent Attention within the IC-Light architecture and further incorporate deterministic regularization to eliminate appearance flickering. Extensive experiments demonstrate that our method achieves competitive performance in temporal consistency and lighting fidelity, robustly handling camera rotations from -90 to 90. Code: https://github.com/AIGeeksGroup/Light4D. Website: https://aigeeksgroup.github.io/Light4D.

Method: Generate 3D fractals using 3D Iterated Function Systems (IFS), temporally transform them into videos for pre-training action recognition models. Propose Targeted Smart Filtering to address speed and diversity issues by intelligently filtering fractal generation.

Result: Targeted Smart Filtering achieves roughly 100 times faster sampling speed compared to standard methods and shows superior downstream performance for action recognition compared to other 3D fractal filtering approaches.

Conclusion: The proposed method effectively addresses limitations of 3D fractal generation for synthetic video pre-training, demonstrating that overly-restrictive fractal generation methods, while aesthetically pleasing, harm downstream task performance, while smart filtering improves both speed and effectiveness.

Abstract: Synthetic datasets are being recognized in the deep learning realm as a valuable alternative to exhaustively labeled real data. One such synthetic data generation method is Formula Driven Supervised Learning (FDSL), which can provide an infinite number of perfectly labeled data through a formula driven approach, such as fractals or contours. FDSL does not have common drawbacks like manual labor, privacy and other ethical concerns. In this work we generate 3D fractals using 3D Iterated Function Systems (IFS) for pre-training an action recognition model. The fractals are temporally transformed to form a video that is used as a pre-training dataset for downstream task of action recognition. We find that standard methods of generating fractals are slow and produce degenerate 3D fractals. Therefore, we systematically explore alternative ways of generating fractals and finds that overly-restrictive approaches, while generating aesthetically pleasing fractals, are detrimental for downstream task performance. We propose a novel method, Targeted Smart Filtering, to address both the generation speed and fractal diversity issue. The method reports roughly 100 times faster sampling speed and achieves superior downstream performance against other 3D fractal filtering methods.

Method: Introduces JEPA-VLA which adaptively integrates predictive embeddings from video-pretrained models (specifically V-JEPA 2) into existing VLAs to better encode temporal dynamics and task-relevant information.

Result: Substantial performance gains across multiple benchmarks including LIBERO, LIBERO-plus, RoboTwin2.0, and real-robot tasks, demonstrating improved sample efficiency and generalization.

Conclusion: Predictive video embeddings effectively compensate for key shortcomings of existing visual representations in VLAs, leading to significant improvements in robotic manipulation performance.

Abstract: Recent vision-language-action (VLA) models built upon pretrained vision-language models (VLMs) have achieved significant improvements in robotic manipulation. However, current VLAs still suffer from low sample efficiency and limited generalization. This paper argues that these limitations are closely tied to an overlooked component, pretrained visual representation, which offers insufficient knowledge on both aspects of environment understanding and policy prior. Through an in-depth analysis, we find that commonly used visual representations in VLAs, whether pretrained via language-image contrastive learning or image-based self-supervised learning, remain inadequate at capturing crucial, task-relevant environment information and at inducing effective policy priors, i.e., anticipatory knowledge of how the environment evolves under successful task execution. In contrast, we discover that predictive embeddings pretrained on videos, in particular V-JEPA 2, are adept at flexibly discarding unpredictable environment factors and encoding task-relevant temporal dynamics, thereby effectively compensating for key shortcomings of existing visual representations in VLAs. Building on these observations, we introduce JEPA-VLA, a simple yet effective approach that adaptively integrates predictive embeddings into existing VLAs. Our experiments demonstrate that JEPA-VLA yields substantial performance gains across a range of benchmarks, including LIBERO, LIBERO-plus, RoboTwin2.0, and real-robot tasks.

Method: Proposes WorldTree with two components: 1) Temporal Partition Tree (TPT) for coarse-to-fine temporal optimization using inheritance-based partition tree structure, and 2) Spatial Ancestral Chains (SAC) that recursively query ancestral hierarchical structure to provide complementary spatial dynamics while specializing motion representations.

Result: Achieves 8.26% improvement of LPIPS on NVIDIA-LS dataset and 9.09% improvement of mLPIPS on DyCheck dataset compared to the second-best method.

Conclusion: WorldTree provides an effective unified framework for hierarchical spatiotemporal decomposition in dynamic reconstruction, significantly improving performance on benchmark datasets.

Abstract: Dynamic reconstruction has achieved remarkable progress, but there remain challenges in monocular input for more practical applications. The prevailing works attempt to construct efficient motion representations, but lack a unified spatiotemporal decomposition framework, suffering from either holistic temporal optimization or coupled hierarchical spatial composition. To this end, we propose WorldTree, a unified framework comprising Temporal Partition Tree (TPT) that enables coarse-to-fine optimization based on the inheritance-based partition tree structure for hierarchical temporal decomposition, and Spatial Ancestral Chains (SAC) that recursively query ancestral hierarchical structure to provide complementary spatial dynamics while specializing motion representations across ancestral nodes. Experimental results on different datasets indicate that our proposed method achieves 8.26% improvement of LPIPS on NVIDIA-LS and 9.09% improvement of mLPIPS on DyCheck compared to the second-best method. Code: https://github.com/iCVTEAM/WorldTree.

Method: Proposes PMI - a training-free gradient correction method that stabilizes velocity field by guiding it toward running average of past velocities within a theoretically derived spherical Gaussian. Also introduces mimic-CFG for editing tasks, which interpolates between current velocity and its projection onto historical average.

Result: Extensive experiments on PIE-Bench show significant improvements in inversion stability, image reconstruction quality, and editing fidelity while reducing required neural function evaluations. Achieves state-of-the-art performance on PIE-Bench with enhanced efficiency.

Conclusion: PMI and mimic-CFG provide theoretically sound solutions to Rectified-Flow inversion instability, enabling better reconstruction and editing with improved efficiency.

Abstract: Rectified-Flow (RF)-based generative models have recently emerged as strong alternatives to traditional diffusion models, demonstrating state-of-the-art performance across various tasks. By learning a continuous velocity field that transforms simple noise into complex data, RF-based models not only enable high-quality generation, but also support training-free inversion, which facilitates downstream tasks such as reconstruction and editing. However, existing inversion methods, such as vanilla RF-based inversion, suffer from approximation errors that accumulate across timesteps, leading to unstable velocity fields and degraded reconstruction and editing quality. To address this challenge, we propose Proximal-Mean Inversion (PMI), a training-free gradient correction method that stabilizes the velocity field by guiding it toward a running average of past velocities, constrained within a theoretically derived spherical Gaussian. Furthermore, we introduce mimic-CFG, a lightweight velocity correction scheme for editing tasks, which interpolates between the current velocity and its projection onto the historical average, balancing editing effectiveness and structural consistency. Extensive experiments on PIE-Bench demonstrate that our methods significantly improve inversion stability, image reconstruction quality, and editing fidelity, while reducing the required number of neural function evaluations. Our approach achieves state-of-the-art performance on the PIE-Bench with enhanced efficiency and theoretical soundness.

Method: Proposes Region-to-Image Distillation: 1) Zoom in to micro-cropped regions to let strong teacher models generate high-quality VQA data, 2) Distill this region-grounded supervision back to the full image during training, enabling single-forward-pass fine-grained perception without tool use. Also introduces ZoomBench benchmark with 845 VQA data across six fine-grained perceptual dimensions.

Result: Models achieve leading performance across multiple fine-grained perception benchmarks and improve general multimodal cognition on visual reasoning and GUI agent benchmarks. The method internalizes benefits of agentic zooming into single forward pass.

Conclusion: Successfully transforms zooming from inference-time tool to training-time primitive, enabling efficient fine-grained perception. Discusses when “Thinking-with-Images” is necessary vs when gains can be distilled into single forward pass.

Abstract: Multimodal Large Language Models (MLLMs) excel at broad visual understanding but still struggle with fine-grained perception, where decisive evidence is small and easily overwhelmed by global context. Recent “Thinking-with-Images” methods alleviate this by iteratively zooming in and out regions of interest during inference, but incur high latency due to repeated tool calls and visual re-encoding. To address this, we propose Region-to-Image Distillation, which transforms zooming from an inference-time tool into a training-time primitive, thereby internalizing the benefits of agentic zooming into a single forward pass of an MLLM. In particular, we first zoom in to micro-cropped regions to let strong teacher models generate high-quality VQA data, and then distill this region-grounded supervision back to the full image. After training on such data, the smaller student model improves “single-glance” fine-grained perception without tool use. To rigorously evaluate this capability, we further present ZoomBench, a hybrid-annotated benchmark of 845 VQA data spanning six fine-grained perceptual dimensions, together with a dual-view protocol that quantifies the global–regional “zooming gap”. Experiments show that our models achieve leading performance across multiple fine-grained perception benchmarks, and also improve general multimodal cognition on benchmarks such as visual reasoning and GUI agents. We further discuss when “Thinking-with-Images” is necessary versus when its gains can be distilled into a single forward pass. Our code is available at https://github.com/inclusionAI/Zooming-without-Zooming.

Method: Proposes DiffPlace framework with a place-ID controller that employs linear projection, perceiver transformer, and contrastive learning to map place-ID embeddings into a fixed CLIP space, enabling synthesis of images with consistent background buildings while flexibly modifying foreground objects and weather conditions.

Result: Extensive experiments show DiffPlace outperforms existing methods in both generation quality and training support for visual place recognition, demonstrating superior performance in quantitative comparisons and augmented training evaluations.

Conclusion: DiffPlace highlights the potential of generative models in enhancing scene-level and place-aware synthesis, providing a valuable approach for improving place recognition in autonomous driving applications.

Abstract: Generative models have advanced significantly in realistic image synthesis, with diffusion models excelling in quality and stability. Recent multi-view diffusion models improve 3D-aware street view generation, but they struggle to produce place-aware and background-consistent urban scenes from text, BEV maps, and object bounding boxes. This limits their effectiveness in generating realistic samples for place recognition tasks. To address these challenges, we propose DiffPlace, a novel framework that introduces a place-ID controller to enable place-controllable multi-view image generation. The place-ID controller employs linear projection, perceiver transformer, and contrastive learning to map place-ID embeddings into a fixed CLIP space, allowing the model to synthesize images with consistent background buildings while flexibly modifying foreground objects and weather conditions. Extensive experiments, including quantitative comparisons and augmented training evaluations, demonstrate that DiffPlace outperforms existing methods in both generation quality and training support for visual place recognition. Our results highlight the potential of generative models in enhancing scene-level and place-aware synthesis, providing a valuable approach for improving place recognition in autonomous driving

Method: Reformulates ring artifact reduction as inverse problem using unrolled network that considers non-ideal detector responses with CT geometry forward-projection. Uses synthetic data from natural images to leverage intrinsic correlations between sinogram and image domains.

Result: Model trained on synthetic data consistently outperforms state-of-the-art methods across diverse scanning geometries and anatomical regions.

Conclusion: Proposed approach effectively reduces ring artifacts without requiring real clinical data by leveraging CT geometry and domain correlations through synthetic training.

Abstract: Defective and inconsistent responses in CT detectors can cause ring and streak artifacts in the reconstructed images, making them unusable for clinical purposes. In recent years, several ring artifact reduction solutions have been proposed in the image domain or in the sinogram domain using supervised deep learning methods. However, these methods require dedicated datasets for training, leading to a high data collection cost. Furthermore, existing approaches focus exclusively on either image-space or sinogram-space correction, neglecting the intrinsic correlations from the forward operation of the CT geometry. Based on the theoretical analysis of non-ideal CT detector responses, the RAR problem is reformulated as an inverse problem by using an unrolled network, which considers non-ideal response together with linear forward-projection with CT geometry. Additionally, the intrinsic correlations of ring artifacts between the sinogram and image domains are leveraged through synthetic data derived from natural images, enabling the trained model to correct artifacts without requiring real-world clinical data. Extensive evaluations on diverse scanning geometries and anatomical regions demonstrate that the model trained on synthetic data consistently outperforms existing state-of-the-art methods.

Method: Introduces DynaHOI-Gym, a unified online closed-loop platform with parameterized motion generators and rollout-based metrics. Also presents DynaHOI-10M benchmark with 10M frames and 180K hand capture trajectories across 8 major categories and 22 subcategories. Proposes ObAct baseline that integrates short-term observations with current frame via spatiotemporal attention.

Result: The ObAct baseline achieves an 8.1% improvement in location success rate. The benchmark provides comprehensive evaluation of dynamic hand-object interactions with moving targets.

Conclusion: DynaHOI-Gym addresses the gap in dynamic hand-object interaction evaluation, providing a comprehensive platform and benchmark for future research in this challenging domain.

Abstract: Most existing hand motion generation benchmarks for hand-object interaction (HOI) focus on static objects, leaving dynamic scenarios with moving targets and time-critical coordination largely untested. To address this gap, we introduce the DynaHOI-Gym, a unified online closed-loop platform with parameterized motion generators and rollout-based metrics for dynamic capture evaluation. Built on DynaHOI-Gym, we release DynaHOI-10M, a large-scale benchmark with 10M frames and 180K hand capture trajectories, whose target motions are organized into 8 major categories and 22 fine-grained subcategories. We also provide a simple observe-before-act baseline (ObAct) that integrates short-term observations with the current frame via spatiotemporal attention to predict actions, achieving an 8.1% improvement in location success rate.

Method: Combines implicit neural representations (INRs) to capture continuous spatial representations of LGE data and masks, compresses them into latent embeddings, then uses a diffusion model on this latent space to generate new representations that decode into synthetic LGE images with anatomically consistent segmentation masks.

Result: Experiments on 133 cardiac MRI scans show that augmenting training data with 200 synthetic volumes improves fibrosis segmentation performance, with Dice score increasing from 0.509 to 0.524.

Conclusion: The approach provides an annotation-free method to mitigate data scarcity in LGE imaging for myocardial scar assessment, with publicly available code.

Abstract: Late gadolinium enhancement (LGE) imaging is the clinical standard for myocardial scar assessment, but limited annotated datasets hinder the development of automated segmentation methods. We propose a novel framework that synthesises both LGE images and their corresponding segmentation masks using implicit neural representations (INRs) combined with denoising diffusion models. Our approach first trains INRs to capture continuous spatial representations of LGE data and associated myocardium and fibrosis masks. These INRs are then compressed into compact latent embeddings, preserving essential anatomical information. A diffusion model operates on this latent space to generate new representations, which are decoded into synthetic LGE images with anatomically consistent segmentation masks. Experiments on 133 cardiac MRI scans suggest that augmenting training data with 200 synthetic volumes contributes to improved fibrosis segmentation performance, with the Dice score showing an increase from 0.509 to 0.524. Our approach provides an annotation-free method to help mitigate data scarcity.The code for this research is publicly available.

Method: Created a French-focused benchmark using model-disagreement sampling from 60,000 documents covering handwritten forms, complex layouts, dense tables, and graphics-rich pages. Used unit-test-style checks targeting concrete failure modes (text presence, reading order, local table constraints) with category-specific normalization to discount presentation-only variance.

Result: Across 15 models, proprietary models showed substantially higher robustness on handwriting and forms, while several open-weights systems remained competitive on standard printed layouts.

Conclusion: The benchmark provides better evaluation of VLMs for document parsing, revealing strengths and weaknesses across different model types for French document processing tasks.

Abstract: This report evaluates PDF-to-Markdown conversion using recent Vision-Language Models (VLMs) on challenging French documents. Document parsing is a critical step for Retrieval-Augmented Generation (RAG) pipelines, where transcription and layout errors propagate to downstream retrieval and grounding. Existing benchmarks often emphasize English or Chinese and can over-penalize benign formatting and linearization choices (e.g., line breaks, list segmentation, alternative table renderings) that are largely irrelevant for downstream use. We introduce a French-focused benchmark of difficult pages selected via model-disagreement sampling from a corpus of 60{,}000 documents, covering handwritten forms, complex layouts, dense tables, and graphics-rich pages. Evaluation is performed with unit-test-style checks that target concrete failure modes (text presence, reading order, and local table constraints) combined with category-specific normalization designed to discount presentation-only variance. Across 15 models, we observe substantially higher robustness for the strongest proprietary models on handwriting and forms, while several open-weights systems remain competitive on standard printed layouts.

Method: Proposes a Bayesian deep learning framework with: 1) Confidence-Uncertainty Boundary Loss (CUB-Loss) that penalizes high-certainty errors and low-certainty correct predictions, and 2) Dual Temperature Scaling (DTS) for post-hoc calibration refinement.

Result: The framework achieves consistent calibration improvements across pneumonia screening, diabetic retinopathy detection, and skin lesion identification tasks. It maintains robust performance in data-scarce scenarios and works well on severely imbalanced datasets.

Conclusion: The proposed approach effectively aligns uncertainty estimates with prediction correctness, demonstrating potential for real clinical deployment by providing clinicians with reliable uncertainty quantification for medical imaging AI systems.

Abstract: In critical decision support systems based on medical imaging, the reliability of AI-assisted decision-making is as relevant as predictive accuracy. Although deep learning models have demonstrated significant accuracy, they frequently suffer from miscalibration, manifested as overconfidence in erroneous predictions. To facilitate clinical acceptance, it is imperative that models quantify uncertainty in a manner that correlates with prediction correctness, allowing clinicians to identify unreliable outputs for further review. In order to address this necessity, the present paper proposes a generalizable probabilistic optimization framework grounded in Bayesian deep learning. Specifically, a novel Confidence-Uncertainty Boundary Loss (CUB-Loss) is introduced that imposes penalties on high-certainty errors and low-certainty correct predictions, explicitly enforcing alignment between prediction correctness and uncertainty estimates. Complementing this training-time optimization, a Dual Temperature Scaling (DTS) strategy is devised for post-hoc calibration, further refining the posterior distribution to improve intuitive explainability. The proposed framework is validated on three distinct medical imaging tasks: automatic screening of pneumonia, diabetic retinopathy detection, and identification of skin lesions. Empirical results demonstrate that the proposed approach achieves consistent calibration improvements across diverse modalities, maintains robust performance in data-scarce scenarios, and remains effective on severely imbalanced datasets, underscoring its potential for real clinical deployment.

Method: Proposes Spatial Chain-of-Thought (SCoT) framework: 1) trains diffusion model on interleaved text-coordinate instruction format for layout awareness, 2) uses MLLMs as planners to generate comprehensive layout plans

Result: Achieves SOTA performance on image generation benchmarks, significantly outperforms baselines on complex reasoning tasks, and shows strong efficacy in image editing scenarios

Conclusion: SCoT effectively bridges MLLMs’ reasoning capabilities with diffusion models’ generative power through a plug-and-play approach

Abstract: While diffusion models have shown exceptional capabilities in aesthetic image synthesis, they often struggle with complex spatial understanding and reasoning. Existing approaches resort to Multimodal Large Language Models (MLLMs) to enhance this capability. However, they either incur high computational costs through joint training or suffer from spatial information loss when relying solely on textual prompts. To alleviate these limitations, we propose a Spatial Chain-of-Thought (SCoT) framework, a plug-and-play approach that effectively bridges the reasoning capabilities of MLLMs with the generative power of diffusion models. Specifically, we first enhance the diffusion model’s layout awareness by training it on an interleaved text-coordinate instruction format. We then leverage state-of-the-art MLLMs as planners to generate comprehensive layout plans, transferring their spatial planning capabilities directly to the generation process. Extensive experiments demonstrate that our method achieves state-of-the-art performance on image generation benchmarks and significantly outperforms baselines on complex reasoning tasks, while also showing strong efficacy in image editing scenarios.

Method: Investigates generative AI approaches using local vision-language models (VLMs) combined with large language models (LLMs). Compares zero-shot VLM strategies and fine-tuned VLMs with classification heads (including LoRA fine-tuning) against a supervised TimeSFormer baseline. Uses a simulated dataset of 13.26 hours of newborn resuscitation videos.

Result: Small local VLMs initially struggle with hallucinations, but when fine-tuned with LoRA achieve F1 score of 0.91, significantly surpassing the TimeSFormer baseline result of 0.70. This demonstrates the effectiveness of fine-tuned VLMs for fine-grained activity recognition in medical videos.

Conclusion: Fine-tuning small vision-language models with LoRA is highly effective for recognizing fine-grained activities in newborn resuscitation videos, achieving state-of-the-art performance and showing promise for medical video analysis applications.

Abstract: Accurate documentation of newborn resuscitation is essential for quality improvement and adherence to clinical guidelines, yet remains underutilized in practice. Previous work using 3D-CNNs and Vision Transformers (ViT) has shown promising results in detecting key activities from newborn resuscitation videos, but also highlighted the challenges in recognizing such fine-grained activities. This work investigates the potential of generative AI (GenAI) methods to improve activity recognition from such videos. Specifically, we explore the use of local vision-language models (VLMs), combined with large language models (LLMs), and compare them to a supervised TimeSFormer baseline. Using a simulated dataset comprising 13.26 hours of newborn resuscitation videos, we evaluate several zero-shot VLM-based strategies and fine-tuned VLMs with classification heads, including Low-Rank Adaptation (LoRA). Our results suggest that small (local) VLMs struggle with hallucinations, but when fine-tuned with LoRA, the results reach F1 score at 0.91, surpassing the TimeSformer results of 0.70.

Method: Analyzes correspondence capabilities in spatial attention of external visual representations, then proposes ReNoV with representation projection modules that inject external representations into the diffusion process.

Result: Shows marked improvements in reconstruction fidelity and inpainting quality, outperforming prior diffusion-based novel-view methods on standard benchmarks and enabling robust synthesis from sparse, unposed image collections.

Conclusion: Representation-guided novel view synthesis through external representation injection significantly improves geometric consistency and quality in diffusion-based 3D scene generation.

Abstract: We propose a novel framework for diffusion-based novel view synthesis in which we leverage external representations as conditions, harnessing their geometric and semantic correspondence properties for enhanced geometric consistency in generated novel viewpoints. First, we provide a detailed analysis exploring the correspondence capabilities emergent in the spatial attention of external visual representations. Building from these insights, we propose a representation-guided novel view synthesis through dedicated representation projection modules that inject external representations into the diffusion process, a methodology named ReNoV, short for representation-guided novel view synthesis. Our experiments show that this design yields marked improvements in both reconstruction fidelity and inpainting quality, outperforming prior diffusion-based novel-view methods on standard benchmarks and enabling robust synthesis from sparse, unposed image collections.

Method: Developed an HDR imaging system using a digital micromirror device (DMD) for spatial modulation, featuring two synergistic subsystems: a DMD-based optical modulation unit and an adaptive computational imaging pipeline that enables autonomous regional segmentation and adaptive exposure control.

Result: Achieved 127 dB measurable dynamic range, eliminated saturation artifacts under high glare, demonstrated 78% reduction in strain error, and improved DIC positioning accuracy across extreme intensity variations.

Conclusion: The DMD-based system provides high-fidelity adaptive HDR imaging that overcomes limitations of conventional sensors, showing strong potential for optical metrology and stress analysis in high-glare environments where traditional methods fail.

Abstract: Background The accuracy of photomechanics measurements critically relies on image quality,particularly under extreme illumination conditions such as welding arc monitoring and polished metallic surface analysis. High dynamic range (HDR) imaging above 120 dB is essential in these contexts. Conventional CCD/CMOS sensors, with dynamic ranges typically below 70 dB, are highly susceptible to saturation under glare, resulting in irreversible loss of detail and significant errors in digital image correlation (DIC). Methods This paper presents an HDR imaging system that leverages the spatial modulation capability of a digital micromirror device (DMD). The system architecture enables autonomous regional segmentation and adaptive exposure control for high-dynamic-range scenes through an integrated framework comprising two synergistic subsystems: a DMD-based optical modulation unit and an adaptive computational imaging pipeline. Results The system achieves a measurable dynamic range of 127 dB, effectively eliminating satu ration artifacts under high glare. Experimental results demonstrate a 78% reduction in strain error and improved DIC positioning accuracy, confirming reliable performance across extreme intensity variations. Conclusion The DMD-based system provides high fidelity adaptive HDR imaging, overcoming key limitations of conventional sensors. It exhibits strong potential for optical metrology and stress analysis in high-glare environments where traditional methods are inadequate.

Method: Built on GigaBrain-0.5 (pre-trained on 10,000+ hours of robotic data), the model integrates world model-based reinforcement learning via RAMP (Reinforcement leArning via world Model-conditioned Policy) for cross-task adaptation.

Result: RAMP achieves ~30% performance improvement over RECAP baseline on challenging tasks (Laundry Folding, Box Packing, Espresso Preparation). GigaBrain-0.5M* demonstrates reliable long-horizon execution in real-world deployment.

Conclusion: Integrating world models with VLA learning significantly enhances robotic manipulation capabilities, enabling robust cross-task adaptation and reliable execution of complex, long-horizon tasks.

Abstract: Vision-language-action (VLA) models that directly predict multi-step action chunks from current observations face inherent limitations due to constrained scene understanding and weak future anticipation capabilities. In contrast, video world models pre-trained on web-scale video corpora exhibit robust spatiotemporal reasoning and accurate future prediction, making them a natural foundation for enhancing VLA learning. Therefore, we propose \textit{GigaBrain-0.5M*}, a VLA model trained via world model-based reinforcement learning. Built upon \textit{GigaBrain-0.5}, which is pre-trained on over 10,000 hours of robotic manipulation data, whose intermediate version currently ranks first on the international RoboChallenge benchmark. \textit{GigaBrain-0.5M*} further integrates world model-based reinforcement learning via \textit{RAMP} (Reinforcement leArning via world Model-conditioned Policy) to enable robust cross-task adaptation. Empirical results demonstrate that \textit{RAMP} achieves substantial performance gains over the RECAP baseline, yielding improvements of approximately 30% on challenging tasks including \texttt{Laundry Folding}, \texttt{Box Packing}, and \texttt{Espresso Preparation}. Critically, \textit{GigaBrain-0.5M$^*$} exhibits reliable long-horizon execution, consistently accomplishing complex manipulation tasks without failure as validated by real-world deployment videos on our \href{https://gigabrain05m.github.io}{project page}.

Method: Autoregressive Transformer-based model that generates modular 3D assets from textual descriptions. Innovatively adapts module sequencing and decoding techniques inspired by language models, leveraging real-world modular assets collected from online platforms.

Result: Initial results indicate effectiveness in streamlining asset creation for professional development and UGC scenarios. The model demonstrates capability in generating modular 3D assets that adhere to design constraints.

Conclusion: AssetFormer presents a flexible framework extendable to various types of modular 3D assets, contributing to the broader field of 3D content generation. The approach shows promise for practical applications in digital content creation.

Abstract: The digital industry demands high-quality, diverse modular 3D assets, especially for user-generated content~(UGC). In this work, we introduce AssetFormer, an autoregressive Transformer-based model designed to generate modular 3D assets from textual descriptions. Our pilot study leverages real-world modular assets collected from online platforms. AssetFormer tackles the challenge of creating assets composed of primitives that adhere to constrained design parameters for various applications. By innovatively adapting module sequencing and decoding techniques inspired by language models, our approach enhances asset generation quality through autoregressive modeling. Initial results indicate the effectiveness of AssetFormer in streamlining asset creation for professional development and UGC scenarios. This work presents a flexible framework extendable to various types of modular 3D assets, contributing to the broader field of 3D content generation. The code is available at https://github.com/Advocate99/AssetFormer.

Method: Proposes PosterOmni framework with: (1) multi-scenario datasets covering six task types across entity-based and concept-based creation; (2) knowledge distillation between local and global experts for supervised fine-tuning; (3) unified PosterOmni Reward Feedback to jointly align visual entity preservation and aesthetic preferences across all tasks.

Result: PosterOmni significantly enhances reference adherence, global composition quality, and aesthetic harmony, outperforming all open-source baselines and even surpassing several proprietary systems. Also establishes PosterOmni-Bench benchmark for evaluation.

Conclusion: PosterOmni successfully integrates local editing and global creation within a single system, addressing the multi-dimensional challenges of image-to-poster generation through an efficient data-distillation-reward pipeline.

Abstract: Image-to-poster generation is a high-demand task requiring not only local adjustments but also high-level design understanding. Models must generate text, layout, style, and visual elements while preserving semantic fidelity and aesthetic coherence. The process spans two regimes: local editing, where ID-driven generation, rescaling, filling, and extending must preserve concrete visual entities; and global creation, where layout- and style-driven tasks rely on understanding abstract design concepts. These intertwined demands make image-to-poster a multi-dimensional process coupling entity-preserving editing with concept-driven creation under image-prompt control. To address these challenges, we propose PosterOmni, a generalized artistic poster creation framework that unlocks the potential of a base edit model for multi-task image-to-poster generation. PosterOmni integrates the two regimes, namely local editing and global creation, within a single system through an efficient data-distillation-reward pipeline: (i) constructing multi-scenario image-to-poster datasets covering six task types across entity-based and concept-based creation; (ii) distilling knowledge between local and global experts for supervised fine-tuning; and (iii) applying unified PosterOmni Reward Feedback to jointly align visual entity-preserving and aesthetic preference across all tasks. Additionally, we establish PosterOmni-Bench, a unified benchmark for evaluating both local editing and global creation. Extensive experiments show that PosterOmni significantly enhances reference adherence, global composition quality, and aesthetic harmony, outperforming all open-source baselines and even surpassing several proprietary systems.

Method: FAIL minimizes policy-expert divergence through adversarial training without explicit rewards or pairwise comparisons. Two algorithms: FAIL-PD uses differentiable ODE solvers for low-variance pathwise gradients; FAIL-PG provides a black-box alternative for discrete or computationally constrained settings.

Result: Fine-tuning FLUX with only 13,000 demonstrations from Nano Banana pro, FAIL achieves competitive performance on prompt following and aesthetic benchmarks. The framework generalizes to discrete image and video generation and functions as a robust regularizer to mitigate reward hacking.

Conclusion: FAIL provides an effective adversarial imitation learning approach for flow matching models that works without explicit rewards or preference pairs, generalizes across domains, and helps prevent reward hacking.

Abstract: Post-training of flow matching models-aligning the output distribution with a high-quality target-is mathematically equivalent to imitation learning. While Supervised Fine-Tuning mimics expert demonstrations effectively, it cannot correct policy drift in unseen states. Preference optimization methods address this but require costly preference pairs or reward modeling. We propose Flow Matching Adversarial Imitation Learning (FAIL), which minimizes policy-expert divergence through adversarial training without explicit rewards or pairwise comparisons. We derive two algorithms: FAIL-PD exploits differentiable ODE solvers for low-variance pathwise gradients, while FAIL-PG provides a black-box alternative for discrete or computationally constrained settings. Fine-tuning FLUX with only 13,000 demonstrations from Nano Banana pro, FAIL achieves competitive performance on prompt following and aesthetic benchmarks. Furthermore, the framework generalizes effectively to discrete image and video generation, and functions as a robust regularizer to mitigate reward hacking in reward-based optimization. Code and data are available at https://github.com/HansPolo113/FAIL.

Method: TexSpot uses Texlet representation where each latent vector encodes a local texture patch via 2D encoder, aggregated with 3D encoder for global shape context. A cascaded 3D-to-2D decoder reconstructs texture patches, enabling Texlet space learning. A diffusion transformer conditioned on Texlets refines textures from multi-view diffusion methods.

Result: Extensive experiments show TexSpot significantly improves visual fidelity, geometric consistency, and robustness over state-of-the-art 3D texture generation and enhancement approaches.

Conclusion: TexSpot addresses fundamental limitations in 3D texture generation by introducing a novel representation that combines the best of point-based and UV-based approaches, enabling high-quality texture enhancement through diffusion-based methods.

Abstract: High-quality 3D texture generation remains a fundamental challenge due to the view-inconsistency inherent in current mainstream multi-view diffusion pipelines. Existing representations either rely on UV maps, which suffer from distortion during unwrapping, or point-based methods, which tightly couple texture fidelity to geometric density that limits high-resolution texture generation. To address these limitations, we introduce TexSpot, a diffusion-based texture enhancement framework. At its core is Texlet, a novel 3D texture representation that merges the geometric expressiveness of point-based 3D textures with the compactness of UV-based representation. Each Texlet latent vector encodes a local texture patch via a 2D encoder and is further aggregated using a 3D encoder to incorporate global shape context. A cascaded 3D-to-2D decoder reconstructs high-quality texture patches, enabling the Texlet space learning. Leveraging this representation, we train a diffusion transformer conditioned on Texlets to refine and enhance textures produced by multi-view diffusion methods. Extensive experiments demonstrate that TexSpot significantly improves visual fidelity, geometric consistency, and robustness over existing state-of-the-art 3D texture generation and enhancement approaches. Project page: https://anonymous.4open.science/w/TexSpot-page-2D91.

Method: Proposes a Symmetric Conditional Diffusion Transformer with symmetric conditional injection, Dual-Level Disentanglement strategy (Synchronized RoPE for signal-level binding and Structured Captions for semantic-level mapping), and Multi-Task Progressive Training scheme that leverages weakly-constrained priors to regularize strongly-constrained tasks.

Result: Achieves comprehensive state-of-the-art performance across video, audio, and audio-visual consistency metrics, even outperforming leading proprietary commercial models.

Conclusion: DreamID-Omni successfully bridges the gap between academic research and commercial-grade applications in controllable human-centric audio-video generation, addressing key challenges in multi-person scenarios and unified task handling.

Abstract: Recent advancements in foundation models have revolutionized joint audio-video generation. However, existing approaches typically treat human-centric tasks including reference-based audio-video generation (R2AV), video editing (RV2AV) and audio-driven video animation (RA2V) as isolated objectives. Furthermore, achieving precise, disentangled control over multiple character identities and voice timbres within a single framework remains an open challenge. In this paper, we propose DreamID-Omni, a unified framework for controllable human-centric audio-video generation. Specifically, we design a Symmetric Conditional Diffusion Transformer that integrates heterogeneous conditioning signals via a symmetric conditional injection scheme. To resolve the pervasive identity-timbre binding failures and speaker confusion in multi-person scenarios, we introduce a Dual-Level Disentanglement strategy: Synchronized RoPE at the signal level to ensure rigid attention-space binding, and Structured Captions at the semantic level to establish explicit attribute-subject mappings. Furthermore, we devise a Multi-Task Progressive Training scheme that leverages weakly-constrained generative priors to regularize strongly-constrained tasks, preventing overfitting and harmonizing disparate objectives. Extensive experiments demonstrate that DreamID-Omni achieves comprehensive state-of-the-art performance across video, audio, and audio-visual consistency, even outperforming leading proprietary commercial models. We will release our code to bridge the gap between academic research and commercial-grade applications.

Method: EO-VAE uses a single model with dynamic hypernetworks to encode and reconstruct flexible channel combinations from multi-sensor EO data, serving as a foundational tokenizer for the EO domain.

Result: Experiments on the TerraMesh dataset show EO-VAE achieves superior reconstruction fidelity compared to existing TerraMind tokenizers, establishing a robust baseline for latent generative modeling in remote sensing.

Conclusion: EO-VAE provides an effective multi-sensor tokenizer for Earth observation data that can handle diverse sensor specifications and variable channels, enabling better latent generative modeling in remote sensing applications.

Abstract: State-of-the-art generative image and video models rely heavily on tokenizers that compress high-dimensional inputs into more efficient latent representations. While this paradigm has revolutionized RGB generation, Earth observation (EO) data presents unique challenges due to diverse sensor specifications and variable spectral channels. We propose EO-VAE, a multi-sensor variational autoencoder designed to serve as a foundational tokenizer for the EO domain. Unlike prior approaches that train separate tokenizers for each modality, EO-VAE utilizes a single model to encode and reconstruct flexible channel combinations via dynamic hypernetworks. Our experiments on the TerraMesh dataset demonstrate that EO-VAE achieves superior reconstruction fidelity compared to the TerraMind tokenizers, establishing a robust baseline for latent generative modeling in remote sensing.

Method: Introduces Stacked Channel Bridging (SCB) - a deep alignment framework that extracts hierarchical features from multiple VLM layers and fuses them with learnable ’think tokens’ for structured guidance. Uses a three-stage training strategy: (1) Alignment Pre-training on image-text pairs and editing triplets, (2) Joint Supervised Fine-tuning on generation, editing, and reasoning tasks, and (3) Reinforcement Learning with MR-GRPO using mixture of reward functions.

Result: Despite being only 5B parameters and trained on ~50M samples, DeepGen 1.0 achieves leading performance across diverse benchmarks, surpassing 80B HunyuanImage by 28% on WISE and 27B Qwen-Image-Edit by 37% on UniREditBench.

Conclusion: DeepGen 1.0 provides an efficient, high-performance alternative to democratize unified multimodal research, demonstrating that lightweight models can achieve competitive capabilities through innovative architecture and training strategies.

Abstract: Current unified multimodal models for image generation and editing typically rely on massive parameter scales (e.g., >10B), entailing prohibitive training costs and deployment footprints. In this work, we present DeepGen 1.0, a lightweight 5B unified model that achieves comprehensive capabilities competitive with or surpassing much larger counterparts. To overcome the limitations of compact models in semantic understanding and fine-grained control, we introduce Stacked Channel Bridging (SCB), a deep alignment framework that extracts hierarchical features from multiple VLM layers and fuses them with learnable ’think tokens’ to provide the generative backbone with structured, reasoning-rich guidance. We further design a data-centric training strategy spanning three progressive stages: (1) Alignment Pre-training on large-scale image-text pairs and editing triplets to synchronize VLM and DiT representations, (2) Joint Supervised Fine-tuning on a high-quality mixture of generation, editing, and reasoning tasks to foster omni-capabilities, and (3) Reinforcement Learning with MR-GRPO, which leverages a mixture of reward functions and supervision signals, resulting in substantial gains in generation quality and alignment with human preferences, while maintaining stable training progress and avoiding visual artifacts. Despite being trained on only ~50M samples, DeepGen 1.0 achieves leading performance across diverse benchmarks, surpassing the 80B HunyuanImage by 28% on WISE and the 27B Qwen-Image-Edit by 37% on UniREditBench. By open-sourcing our training code, weights, and datasets, we provide an efficient, high-performance alternative to democratize unified multimodal research.

Method: Uses a unified discrete flow-matching framework with task-specific low-rank adapters to decouple understanding and generation. Introduces reference-based multimodal preference alignment to optimize relative outcomes under identical conditioning without large-scale retraining.

Result: Achieves state-of-the-art performance across eight benchmarks and demonstrates strong zero-shot generalization to tasks including inpainting, in-context image generation, reference-based editing, and compositional generation without explicit task-specific training.

Conclusion: UniDFlow provides an effective unified framework for multimodal tasks that balances understanding and generation capabilities while maintaining strong performance and generalization across diverse applications.

Abstract: We propose UniDFlow, a unified discrete flow-matching framework for multimodal understanding, generation, and editing. It decouples understanding and generation via task-specific low-rank adapters, avoiding objective interference and representation entanglement, while a novel reference-based multimodal preference alignment optimizes relative outcomes under identical conditioning, improving faithfulness and controllability without large-scale retraining. UniDFlpw achieves SOTA performance across eight benchmarks and exhibits strong zero-shot generalization to tasks including inpainting, in-context image generation, reference-based editing, and compositional generation, despite no explicit task-specific training.

Method: Proposes Monarch-RT, a structured attention parameterization using Monarch matrices with aligned block structure and extended tiled parameterization. Uses custom Triton kernels and finetuning to overcome parameterization overhead.

Result: Achieves up to 95% attention sparsity with no quality loss when applied to Self-Forcing model. Outperforms FlashAttention kernels (1.4-11.8X speedup) and enables true real-time video generation at 16 FPS on single RTX 5090 GPU.

Conclusion: Monarch-RT is a pioneering sparse attention parameterization for real-time video generation that successfully addresses the limitations of prior sparse approximations in autoregressive few-step regimes.

Abstract: Real-time video generation with Diffusion Transformers is bottlenecked by the quadratic cost of 3D self-attention, especially in real-time regimes that are both few-step and autoregressive, where errors compound across time and each denoising step must carry substantially more information. In this setting, we find that prior sparse-attention approximations break down, despite showing strong results for bidirectional, many-step diffusion. Specifically, we observe that video attention is not reliably sparse, but instead combines pronounced periodic structure driven by spatiotemporal position with dynamic, sparse semantic correspondences and dense mixing, exceeding the representational capacity of even oracle top-k attention. Building on this insight, we propose Monarch-RT, a structured attention parameterization for video diffusion models that factorizes attention using Monarch matrices. Through appropriately aligned block structure and our extended tiled Monarch parameterization, we achieve high expressivity while preserving computational efficiency. We further overcome the overhead of parameterization through finetuning, with custom Triton kernels. We first validate the high efficacy of Monarch-RT over existing sparse baselines designed only for bidirectional models. We further observe that Monarch-RT attains up to 95% attention sparsity with no loss in quality when applied to the state-of-the-art model Self-Forcing, making Monarch-RT a pioneering work on highly-capable sparse attention parameterization for real-time video generation. Our optimized implementation outperforms FlashAttention-2, FlashAttention-3, and FlashAttention-4 kernels on Nvidia RTX 5090, H100, and B200 GPUs respectively, providing kernel speedups in the range of 1.4-11.8X. This enables us, for the first time, to achieve true real-time video generation with Self-Forcing at 16 FPS on a single RTX 5090.

Method: UniT combines agentic data synthesis, unified model training, and flexible test-time inference to enable multimodal chain-of-thought reasoning with verification, subgoal decomposition, and content memory across multiple rounds.

Result: Three key findings: 1) Models trained on short reasoning trajectories generalize to longer inference chains; 2) Sequential chain-of-thought reasoning is more scalable and compute-efficient than parallel sampling; 3) Training on generation and editing trajectories improves out-of-distribution visual reasoning.

Conclusion: Multimodal test-time scaling is an effective paradigm for advancing both generation and understanding in unified models, enabling iterative reasoning and refinement capabilities.

Abstract: Unified models can handle both multimodal understanding and generation within a single architecture, yet they typically operate in a single pass without iteratively refining their outputs. Many multimodal tasks, especially those involving complex spatial compositions, multiple interacting objects, or evolving instructions, require decomposing instructions, verifying intermediate results, and making iterative corrections. While test-time scaling (TTS) has demonstrated that allocating additional inference compute for iterative reasoning substantially improves language model performance, extending this paradigm to unified multimodal models remains an open challenge. We introduce UniT, a framework for multimodal chain-of-thought test-time scaling that enables a single unified model to reason, verify, and refine across multiple rounds. UniT combines agentic data synthesis, unified model training, and flexible test-time inference to elicit cognitive behaviors including verification, subgoal decomposition, and content memory. Our key findings are: (1) unified models trained on short reasoning trajectories generalize to longer inference chains at test time; (2) sequential chain-of-thought reasoning provides a more scalable and compute-efficient TTS strategy than parallel sampling; (3) training on generation and editing trajectories improves out-of-distribution visual reasoning. These results establish multimodal test-time scaling as an effective paradigm for advancing both generation and understanding in unified models.

Method: Stroke of Surprise framework with sequence-aware joint optimization using dual-branch Score Distillation Sampling (SDS). It dynamically adjusts prefix strokes to find a “common structural subspace” for both semantic targets, plus an Overlay Loss for spatial complementarity.

Result: Significantly outperforms state-of-the-art baselines in recognizability and illusion strength, successfully expanding visual anagrams from spatial to temporal dimension.

Conclusion: The method enables progressive semantic illusions through temporal stroke sequencing, demonstrating effective semantic transformations in vector sketches.

Abstract: Visual illusions traditionally rely on spatial manipulations such as multi-view consistency. In this work, we introduce Progressive Semantic Illusions, a novel vector sketching task where a single sketch undergoes a dramatic semantic transformation through the sequential addition of strokes. We present Stroke of Surprise, a generative framework that optimizes vector strokes to satisfy distinct semantic interpretations at different drawing stages. The core challenge lies in the “dual-constraint”: initial prefix strokes must form a coherent object (e.g., a duck) while simultaneously serving as the structural foundation for a second concept (e.g., a sheep) upon adding delta strokes. To address this, we propose a sequence-aware joint optimization framework driven by a dual-branch Score Distillation Sampling (SDS) mechanism. Unlike sequential approaches that freeze the initial state, our method dynamically adjusts prefix strokes to discover a “common structural subspace” valid for both targets. Furthermore, we introduce a novel Overlay Loss that enforces spatial complementarity, ensuring structural integration rather than occlusion. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art baselines in recognizability and illusion strength, successfully expanding visual anagrams from the spatial to the temporal dimension. Project page: https://stroke-of-surprise.github.io/

Method: VRFT-Aug introduces a framework with four key strategies: 1) Prior knowledge injection, 2) Perception-driven policy refinement, 3) Medically informed reward shaping, and 4) Behavioral imitation. These methods work together to stabilize and improve the RFT process for medical vision tasks.

Result: The approach consistently outperforms both standard supervised fine-tuning and RFT baselines across multiple medical datasets, providing empirically grounded insights and practical training heuristics that can generalize to other medical image tasks.

Conclusion: VRFT-Aug contributes actionable guidance and inspiration for developing reliable, reasoning-capable models for high-stakes medical applications by bridging the gap between language-focused RFT methods and vision-centric domains.

Abstract: While recent advances in Reinforcement Fine-Tuning (RFT) have shown that rule-based reward schemes can enable effective post-training for large language models, their extension to cross-modal, vision-centric domains remains largely underexplored. This limitation is especially pronounced in the medical imaging domain, where effective performance requires both robust visual perception and structured reasoning. In this work, we address this gap by proposing VRFT-Aug, a visual reinforcement fine-tuning framework tailored for the medical domain. VRFT-Aug introduces a series of training strategies designed to augment both perception and reasoning, including prior knowledge injection, perception-driven policy refinement, medically informed reward shaping, and behavioral imitation. Together, these methods aim to stabilize and improve the RFT process. Through extensive experiments across multiple medical datasets, we show that our approaches consistently outperform both standard supervised fine-tuning and RFT baselines. Moreover, we provide empirically grounded insights and practical training heuristics that can be generalized to other medical image tasks. We hope this work contributes actionable guidance and fresh inspiration for the ongoing effort to develop reliable, reasoning-capable models for high-stakes medical applications.

Method: Segment time interval into multiple stages using custom multi-decoder U-net architecture that combines time-dependent models with universally shared encoder, plus novel timestep clustering algorithm for stage division.

Result: Significant training and sampling efficiency improvements on three state-of-the-art diffusion models including large-scale latent diffusion models, with effective distribution of computational resources and reduced inter-stage interference.

Conclusion: The multi-stage framework with blended universal and customized parameters substantially improves diffusion model efficiency while maintaining generative performance.

Abstract: Diffusion models, emerging as powerful deep generative tools, excel in various applications. They operate through a two-steps process: introducing noise into training samples and then employing a model to convert random noise into new samples (e.g., images). However, their remarkable generative performance is hindered by slow training and sampling. This is due to the necessity of tracking extensive forward and reverse diffusion trajectories, and employing a large model with numerous parameters across multiple timesteps (i.e., noise levels). To tackle these challenges, we present a multi-stage framework inspired by our empirical findings. These observations indicate the advantages of employing distinct parameters tailored to each timestep while retaining universal parameters shared across all time steps. Our approach involves segmenting the time interval into multiple stages where we employ custom multi-decoder U-net architecture that blends time-dependent models with a universally shared encoder. Our framework enables the efficient distribution of computational resources and mitigates inter-stage interference, which substantially improves training efficiency. Extensive numerical experiments affirm the effectiveness of our framework, showcasing significant training and sampling efficiency enhancements on three state-of-the-art diffusion models, including large-scale latent diffusion models. Furthermore, our ablation studies illustrate the impact of two important components in our framework: (i) a novel timestep clustering algorithm for stage division, and (ii) an innovative multi-decoder U-net architecture, seamlessly integrating universal and customized hyperparameters.

Method: PA-Diff consists of three branches: 1) Physics Prior Generation (PPG) branch produces physics knowledge priors, 2) Physics-aware Diffusion Transformer (PDT) branch uses physics priors to guide diffusion and model complex underwater distributions, and 3) Implicit Neural Reconstruction (INR) branch learns robust feature representations via implicit neural representation to reduce restoration difficulty.

Result: Extensive experiments demonstrate that PA-Diff achieves best performance on underwater image enhancement tasks compared to existing methods.

Conclusion: The proposed PA-Diff framework successfully incorporates physical knowledge into the diffusion process for underwater image enhancement, achieving superior performance by modeling real-world underwater scene distributions more effectively.

Abstract: Underwater visuals undergo various complex degradations, inevitably influencing the efficiency of underwater vision tasks. Recently, diffusion models were employed to underwater image enhancement (UIE) tasks, and gained SOTA performance. However, these methods fail to consider the physical properties and underwater imaging mechanisms in the diffusion process, limiting information completion capacity of diffusion models. In this paper, we introduce a novel UIE framework, named PA-Diff, designed to exploiting the knowledge of physics to guide the diffusion process. PA-Diff consists of Physics Prior Generation (PPG) Branch, Implicit Neural Reconstruction (INR) Branch, and Physics-aware Diffusion Transformer (PDT) Branch. Our designed PPG branch aims to produce the prior knowledge of physics. With utilizing the physics prior knowledge to guide the diffusion process, PDT branch can obtain underwater-aware ability and model the complex distribution in real-world underwater scenes. INR Branch can learn robust feature representations from diverse underwater image via implicit neural representation, which reduces the difficulty of restoration for PDT branch. Extensive experiments prove that our method achieves best performance on UIE tasks.

Method: CSFIQA introduces two key innovations: (1) a selective focus attention mechanism that filters out redundant cross-scale information, mimicking human visual attention, and (2) a scale contrastive learning strategy that learns to distinguish quality variations both across and within scales, with an adaptive noise sample matching mechanism.

Result: Experiments show substantial improvements over state-of-the-art methods across seven datasets, achieving up to 8.8% SRCC improvement on challenging real-world distortions, confirming superior alignment with human quality perception.

Conclusion: CSFIQA effectively addresses the limitations of current multi-scale BIQA approaches by better mimicking human visual perception across scales, leading to more accurate and human-aligned image quality assessment.

Abstract: Human visual perception naturally evaluates image quality across multiple scales, a hierarchical process that existing blind image quality assessment (BIQA) algorithms struggle to replicate effectively. This limitation stems from a fundamental misunderstanding: current multi-scale approaches fail to recognize that quality perception varies dramatically between scales – what appears degraded when viewed closely may look acceptable from a distance. This inconsistency not only creates misleading ``visual illusions’’ during feature fusion but also introduces substantial redundant information that dilutes quality-critical features and leads to imprecise assessments. Our CSFIQA framework advances multi-scale BIQA via two key innovations: (1) a selective focus attention mechanism that mimics human visual attention by filtering out redundant cross-scale information that would otherwise mask subtle quality indicators, and (2) a scale contrastive learning strategy that explicitly learns to distinguish quality variations both across and within scales. By incorporating an adaptive noise sample matching mechanism, CSFIQA effectively identifies perceptual quality discrepancies in the same content viewed at different scales. Experiments demonstrate substantial improvements over state-of-the-art methods across seven datasets, achieving up to 8.8% SRCC improvement on challenging real-world distortions, confirming CSFIQA’s superior alignment with human quality perception.

Method: Comprehensive literature survey with logical taxonomy categorizing dynamic networks by adaptive components: output, computation graph, or input. Includes curated repository of papers.

Result: Unified survey of dynamic neural networks for CV, taxonomy framework, identification of sensor fusion as promising application area, and public repository of resources.

Conclusion: Dynamic neural networks offer efficient adaptive computation for CV, with sensor fusion presenting particularly beneficial applications for adaptivity, noise reduction, and information prioritization.

Abstract: Model compression is essential in the deployment of large Computer Vision models on embedded devices. However, static optimization techniques (e.g. pruning, quantization, etc.) neglect the fact that different inputs have different complexities, thus requiring different amount of computations. Dynamic Neural Networks allow to condition the number of computations to the specific input. The current literature on the topic is very extensive and fragmented. We present a comprehensive survey that synthesizes and unifies existing Dynamic Neural Networks research in the context of Computer Vision. Additionally, we provide a logical taxonomy based on which component of the network is adaptive: the output, the computation graph or the input. Furthermore, we argue that Dynamic Neural Networks are particularly beneficial in the context of Sensor Fusion for better adaptivity, noise reduction and information prioritization. We present preliminary works in this direction. We complement this survey with a curated repository listing all the surveyed papers, each with a brief summary of the solution and the code base when available: https://github.com/DTU-PAS/awesome-dynn-for-cv .

Method: FaceQSORT combines biometric features (from face recognition models) and visual appearance features (extracted from the same face image patch) for association. The method is designed specifically for queue scenarios and includes comprehensive parameter evaluation, similarity metric testing, and face recognition model analysis.

Result: FaceQSORT outperforms state-of-the-art trackers in the considered queue scenario. The paper also introduces a new dataset ‘Paris Lodron University Salzburg Faces in a Queue’ with 7 fully annotated sequences (12730 frames) for evaluation.

Conclusion: Combining biometric and visual features improves multi-face tracking performance in challenging queue scenarios, and the new dataset provides a valuable benchmark for future research in this specific application domain.

Abstract: In this work, a novel multi-face tracking method named FaceQSORT is proposed. To mitigate multi-face tracking challenges (e.g., partially occluded or lateral faces), FaceQSORT combines biometric and visual appearance features (extracted from the same image (face) patch) for association. The Q in FaceQSORT refers to the scenario for which FaceQSORT is desinged, i.e. tracking people’s faces as they move towards a gate in a Queue. This scenario is also reflected in the new dataset `Paris Lodron University Salzburg Faces in a Queue’, which is made publicly available as part of this work. The dataset consists of a total of seven fully annotated and challenging sequences (12730 frames) and is utilized together with two other publicly available datasets for the experimental evaluation. It is shown that FaceQSORT outperforms state-of-the-art trackers in the considered scenario. To provide a deeper insight into FaceQSORT, comprehensive experiments are conducted evaluating the parameter selection, a different similarity metric and the utilized face recognition model (used to extract biometric features).

Method: 1) Uses sliding window to restrict attention to local regions and reconstructs features within windows to model relationships between neighboring internal elements. 2) Designs dual frequency domain branch framework using four DWT frequency subbands and FFT phase parts to enrich local forgery feature extraction from different perspectives.

Result: Achieves 2.13% improvement in detection accuracy over state-of-the-art methods when evaluated on diverse datasets comprising images from 65 distinct generative models. Shows superior generalization detection capabilities on both GAN and diffusion model-based generative images.

Conclusion: The proposed method effectively addresses limitations of existing approaches by combining fine-grained feature extraction through reconstruction sliding window attention with feature enrichment from dual frequency domain branches, resulting in improved generalization for AI-generated image detection.

Abstract: The rapid advancement of Generative Adversarial Networks (GANs) and diffusion models has enabled the creation of highly realistic synthetic images, presenting significant societal risks, such as misinformation and deception. As a result, detecting AI-generated images has emerged as a critical challenge. Existing researches emphasize extracting fine-grained features to enhance detector generalization, yet they often lack consideration for the importance and interdependencies of internal elements within local regions and are limited to a single frequency domain, hindering the capture of general forgery traces. To overcome the aforementioned limitations, we first utilize a sliding window to restrict the attention mechanism to a local window, and reconstruct the features within the window to model the relationships between neighboring internal elements within the local region. Then, we design a dual frequency domain branch framework consisting of four frequency domain subbands of DWT and the phase part of FFT to enrich the extraction of local forgery features from different perspectives. Through feature enrichment of dual frequency domain branches and fine-grained feature extraction of reconstruction sliding window attention, our method achieves superior generalization detection capabilities on both GAN and diffusion model-based generative images. Evaluated on diverse datasets comprising images from 65 distinct generative models, our approach achieves a 2.13% improvement in detection accuracy over state-of-the-art methods.

Method: Collected 640 high-resolution images from a commercial farm across multiple growth stages, weed pressures, and lighting variations. Annotated at leaf-instance level with bounding boxes and segmentation masks for 7,221 soybean and 5,190 cotton leaves, capturing overlapping foliage and morphological similarities.

Result: Validated the dataset using YOLOv11, demonstrating state-of-the-art performance in accurately identifying and segmenting overlapping foliage. The publicly available dataset supports advanced applications like selective herbicide spraying and pest monitoring.

Conclusion: The dataset enables more robust, data-driven strategies for soybean-cotton management by addressing the limitations of existing datasets and capturing real-world agricultural field complexity for volunteer plant and weed detection.

Abstract: Soybean and cotton are major drivers of many countries’ agricultural sectors, offering substantial economic returns but also facing persistent challenges from volunteer plants and weeds that hamper sustainable management. Effectively controlling volunteer plants and weeds demands advanced recognition strategies that can identify these amidst complex crop canopies. While deep learning methods have demonstrated promising results for leaf-level detection and segmentation, existing datasets often fail to capture the complexity of real-world agricultural fields. To address this, we collected 640 high-resolution images from a commercial farm spanning multiple growth stages, weed pressures, and lighting variations. Each image is annotated at the leaf-instance level, with 7,221 soybean and 5,190 cotton leaves labeled via bounding boxes and segmentation masks, capturing overlapping foliage, small leaf size, and morphological similarities. We validate this dataset using YOLOv11, demonstrating state-of-the-art performance in accurately identifying and segmenting overlapping foliage. Our publicly available dataset supports advanced applications such as selective herbicide spraying and pest monitoring and can foster more robust, data-driven strategies for soybean-cotton management.

Method: Created RSWK dataset (14,141 landmarks), developed RS-RAG framework with multimodal knowledge vector database and knowledge retrieval/response generation modules.

Result: RS-RAG significantly outperformed state-of-the-art baselines on image captioning, classification, and VQA tasks in remote sensing domain.

Conclusion: The framework successfully integrates external knowledge to enhance remote sensing VLMs’ reasoning capabilities for complex vision-language tasks.

Abstract: Recent progress in VLMs has demonstrated impressive capabilities across a variety of tasks in the natural image domain. Motivated by these advancements, the remote sensing community has begun to adopt VLMs for remote sensing vision-language tasks, including scene understanding, image captioning, and visual question answering. However, existing remote sensing VLMs typically rely on closed-set scene understanding and focus on generic scene descriptions, yet lack the ability to incorporate external knowledge. This limitation hinders their capacity for semantic reasoning over complex or context-dependent queries that involve domain-specific or world knowledge. To address these challenges, we first introduced a multimodal Remote Sensing World Knowledge (RSWK) dataset, which comprises high-resolution satellite imagery and detailed textual descriptions for 14,141 well-known landmarks from 175 countries, integrating both remote sensing domain knowledge and broader world knowledge. Building upon this dataset, we proposed a novel Remote Sensing Retrieval-Augmented Generation (RS-RAG) framework, which consists of two key components. The Multi-Modal Knowledge Vector Database Construction module encodes remote sensing imagery and associated textual knowledge into a unified vector space. The Knowledge Retrieval and Response Generation module retrieves and re-ranks relevant knowledge based on image and/or text queries, and incorporates the retrieved content into a knowledge-augmented prompt to guide the VLM in producing contextually grounded responses. We validated the effectiveness of our approach on three representative vision-language tasks, including image captioning, image classification, and visual question answering, where RS-RAG significantly outperformed state-of-the-art baselines.

Method: The approach leverages the asymmetry between forward and inverse dynamics prediction. First, fine-tune VLMs for inverse dynamics prediction (captioning actions between frames). Then use IDP to bootstrap FDP through: 1) weakly supervised learning from synthetic data generated by IDP annotations, and 2) inference-time verification where IDP scores multiple FDP samples to guide search.

Result: The best model achieves competitive performance with state-of-the-art image editing models on Aurora-Bench, improving by 7-13% according to GPT4o-as-judge and achieving the best average human evaluation across all subsets.

Conclusion: While VLMs struggle with forward dynamics prediction directly, leveraging their strength in inverse dynamics prediction can effectively bootstrap forward prediction capabilities, enabling improved action-centric image editing while remaining general-purpose models.

Abstract: Can unified vision-language models (VLMs) perform forward dynamics prediction (FDP), i.e., predicting the future state (in image form) given the previous observation and an action (in language form)? We find that VLMs struggle to generate physically plausible transitions between frames from instructions. Nevertheless, we identify a crucial asymmetry in multimodal grounding: fine-tuning a VLM to learn inverse dynamics prediction (IDP), effectively captioning the action between frames, is significantly easier than learning FDP. In turn, IDP can be used to bootstrap FDP through two main strategies: 1) weakly supervised learning from synthetic data and 2) inference time verification. Firstly, IDP can annotate actions for unlabelled pairs of video frame observations to expand the training data scale for FDP. Secondly, IDP can assign rewards to multiple samples of FDP to score them, effectively guiding search at inference time. We evaluate the FDP resulting from both strategies through the task of action-centric image editing on Aurora-Bench with two families of VLMs. Despite remaining general-purpose, our best model achieves a performance competitive with state-of-the-art image editing models, improving on them by a margin between $7%$ and $13%$ according to GPT4o-as-judge, and achieving the best average human evaluation across all subsets of Aurora-Bench.

Method: Created TABLET dataset with 4M examples across 21 tasks grounded in 2M unique tables (88% preserving original visualizations), plus VisualTableQA benchmark requiring both visual perception and table understanding. Fine-tuned vision-language models (Qwen2.5-VL-7B, Gemma 3-4B) on TABLET.

Result: Fine-tuning vision-language models on TABLET improves performance on both seen and unseen VTU tasks while increasing robustness on real-world table visualizations.

Conclusion: TABLET establishes a foundation for robust training and extensible evaluation of future VTU models by preserving original visualizations and maintaining example traceability in a unified large-scale collection.

Abstract: While table understanding increasingly relies on pixel-only settings, current benchmarks predominantly use synthetic renderings that lack the complexity and visual diversity of real-world tables. Additionally, existing visual table understanding (VTU) datasets offer fixed examples with single visualizations and pre-defined instructions, providing no access to underlying serialized data for reformulation. We introduce TABLET, a large-scale VTU dataset with 4 million examples across 21 tasks, grounded in 2 million unique tables where 88% preserve original visualizations. To evaluate whether models are able to jointly reason over tabular and visual content, we also introduce VisualTableQA, a benchmark requiring both visual perception and table understanding. Fine-tuning vision-language models like Qwen2.5-VL-7B and Gemma 3-4B on TABLET improves performance on seen and unseen VTU tasks while increasing robustness on real-world table visualizations. By preserving original visualizations and maintaining example traceability in a unified large-scale collection, TABLET establishes a foundation for robust training and extensible evaluation of future VTU models.

Method: Uses a Transformer backbone with two pathways: geometry predictions with self-attention for geometric fidelity, and style injection via global cross-attention for visual consistency. Incorporates a voxel-based 3D style loss that aligns aggregated scene features to style statistics.

Result: Experiments across multiple datasets show high-quality zero-shot stylization, demonstrating effectiveness of global style-content coupling, the 3D style loss, and scalability from single view to large-scale multi-view settings.

Conclusion: Stylos provides an effective framework for 3D style transfer that generalizes well, preserves geometry, ensures view consistency, and scales from single images to multi-view collections without requiring per-scene optimization.

Abstract: We present Stylos, a single-forward 3D Gaussian framework for 3D style transfer that operates on unposed content, from a single image to a multi-view collection, conditioned on a separate reference style image. Stylos synthesizes a stylized 3D Gaussian scene without per-scene optimization or precomputed poses, achieving geometry-aware, view-consistent stylization that generalizes to unseen categories, scenes, and styles. At its core, Stylos adopts a Transformer backbone with two pathways: geometry predictions retain self-attention to preserve geometric fidelity, while style is injected via global cross-attention to enforce visual consistency across views. With the addition of a voxel-based 3D style loss that aligns aggregated scene features to style statistics, Stylos enforces view-consistent stylization while preserving geometry. Experiments across multiple datasets demonstrate that Stylos delivers high-quality zero-shot stylization, highlighting the effectiveness of global style-content coupling, the proposed 3D style loss, and the scalability of our framework from single view to large-scale multi-view settings. Our codes are available at https://github.com/HanzhouLiu/Stylos.

Method: Introduces ContextGen with two key components: 1) Contextual Layout Anchoring (CLA) that integrates composite layout images to anchor objects in desired positions, and 2) Identity Consistency Attention (ICA) that uses reference images to maintain identity consistency across multiple instances. Also introduces IMIG-100K dataset for training.

Result: ContextGen achieves state-of-the-art performance in multi-instance generation, particularly excelling in layout control and identity fidelity compared to existing methods.

Conclusion: The proposed framework effectively addresses key challenges in multi-instance generation through novel architectural components and a dedicated dataset, advancing the field of controllable image generation.

Abstract: Multi-instance image generation (MIG) remains a significant challenge for modern diffusion models due to key limitations in achieving precise control over object layout and preserving the identity of multiple distinct subjects. To address these limitations, we introduce ContextGen, a novel Diffusion Transformer framework for multi-instance generation that is guided by both layout and reference images. Our approach integrates two key technical contributions: a Contextual Layout Anchoring (CLA) mechanism that incorporates the composite layout image into the generation context to robustly anchor the objects in their desired positions, and Identity Consistency Attention (ICA), an innovative attention mechanism that leverages contextual reference images to ensure the identity consistency of multiple instances. To address the absence of a large-scale, high-quality dataset for this task, we introduce IMIG-100K, the first dataset to provide detailed layout and identity annotations specifically designed for Multi-Instance Generation. Extensive experiments demonstrate that ContextGen sets a new state-of-the-art, outperforming existing methods especially in layout control and identity fidelity.

Method: Created a novel dataset of 1302 artifact examples from 11 SISR methods with crowdsourced prominence scores, plus prominence annotations for 593 existing artifacts from DeSRA dataset. Trained a lightweight regressor to produce spatial prominence heatmaps for artifact detection.

Result: The method outperforms existing artifact detectors and effectively guides SR model fine-tuning for artifact suppression. Found that 48% of existing artifacts in DeSRA dataset go unnoticed by most viewers.

Conclusion: Characterizing artifacts by their perceptual prominence rather than as uniform binary defects provides more meaningful artifact detection and enables better super-resolution model optimization for improved visual quality.

Abstract: Generative single-image super-resolution (SISR) is advancing rapidly, yet even state-of-the-art models produce visual artifacts: unnatural patterns and texture distortions that degrade perceived quality. These defects vary widely in perceptual impact–some are barely noticeable, while others are highly disturbing–yet existing detection methods treat them equally. We propose characterizing artifacts by their prominence to human observers rather than as uniform binary defects. We present a novel dataset of 1302 artifact examples from 11 SISR methods annotated with crowdsourced prominence scores, and provide prominence annotations for 593 existing artifacts from the DeSRA dataset, revealing that 48% of them go unnoticed by most viewers. Building on this data, we train a lightweight regressor that produces spatial prominence heatmaps. We demonstrate that our method outperforms existing detectors and effectively guides SR model fine-tuning for artifact suppression. Our dataset and code are available at https://tinyurl.com/2u9zxtyh.

Method: Proposes analytical approach that derives semantic editing directions directly from pretrained diffusion model parameters. Uses eigenvectors of self-attention weight matrices, which encode structural information about the learned data distribution.

Result: Method produces high-quality edits across multiple datasets while reducing editing time by 60% compared to current benchmarks, requiring neither additional data nor fine-tuning.

Conclusion: Self-attention weight matrices in diffusion models contain interpretable semantic information that can be extracted analytically for efficient and precise image editing control.

Abstract: Diffusion models achieve remarkable fidelity in image synthesis, yet precise control over their outputs for targeted editing remains challenging. A key step toward controllability is to identify interpretable directions in the model’s latent representations that correspond to semantic attributes. Existing approaches for finding interpretable directions typically rely on sampling large sets of images or training auxiliary networks, which limits efficiency. We propose an analytical method that derives semantic editing directions directly from the pretrained parameters of diffusion models, requiring neither additional data nor fine-tuning. Our insight is that self-attention weight matrices encode rich structural information about the data distribution learned during training. By computing the eigenvectors of these weight matrices, we obtain robust and interpretable editing directions. Experiments demonstrate that our method produces high-quality edits across multiple datasets while reducing editing time significantly by 60% over current benchmarks.

Method: Created Real-IAD Variety benchmark with 198,950 high-resolution images across 160 object categories, covering 28 industries, 24 material types, 22 color variations, and 27 defect types for unprecedented diversity.

Result: State-of-the-art multi-class unsupervised anomaly detection methods show 10-20% performance degradation when scaled from 30 to 160 categories, while zero-shot and few-shot models maintain consistent performance and demonstrate better generalization.

Conclusion: Real-IAD Variety provides essential scale and diversity for training and evaluating next-generation foundation IAD models, revealing limitations of current methods and highlighting the robustness of zero-shot/few-shot approaches.

Abstract: Industrial Anomaly Detection (IAD) is a cornerstone for ensuring operational safety, maintaining product quality, and optimizing manufacturing efficiency. However, the advancement of IAD algorithms is severely hindered by the limitations of existing public benchmarks. Current datasets often suffer from restricted category diversity and insufficient scale, leading to performance saturation and poor model transferability in complex, real-world scenarios. To bridge this gap, we introduce Real-IAD Variety, the largest and most diverse IAD benchmark. It comprises 198,950 high-resolution images across 160 distinct object categories. The dataset ensures unprecedented diversity by covering 28 industries, 24 material types, 22 color variations, and 27 defect types. Our extensive experimental analysis highlights the substantial challenges posed by this benchmark: state-of-the-art multi-class unsupervised anomaly detection methods suffer significant performance degradation (ranging from 10% to 20%) when scaled from 30 to 160 categories. Conversely, we demonstrate that zero-shot and few-shot IAD models exhibit remarkable robustness to category scale-up, maintaining consistent performance and significantly enhancing generalization across diverse industrial contexts. This unprecedented scale positions Real-IAD Variety as an essential resource for training and evaluating next-generation foundation IAD models.

Method: Proposes Hilbert curve-based window construction: 1) Reorder image tokens along Hilbert curve, 2) Form windows/neighborhoods on reordered 1D sequence, 3) Combine with block-sparse kernels to improve efficiency. Two variants: Hilbert Window Attention and Hilbert Slide Attention.

Result: Hilbert Window Attention accelerates window attention by ~4×, Hilbert Slide Attention accelerates slide attention by ~18×. Hilbert Window Transformer and Hilbert Neighborhood Transformer achieve end-to-end speedups with minimal accuracy loss.

Conclusion: Hilbert-guided local attention combined with block-sparse kernels offers general and practical approach to enhance efficiency of 2D local attention for images in vision transformers.

Abstract: The quadratic compute and memory costs of global self-attention severely limit its use in high-resolution images. Local attention reduces complexity by restricting attention to neighborhoods. Block-sparse kernels can further improve the efficiency of local attention, but conventional local attention patterns often fail to deliver significant speedups because tokens within a window are not contiguous in the 1D sequence. This work proposes a novel method for constructing windows and neighborhoods based on the Hilbert curve. Image tokens are first reordered along a Hilbert curve, and windows and neighborhoods are then formed on the reordered 1D sequence. From a block-sparse perspective, this strategy significantly increases block sparsity and can be combined with existing block-sparse kernels to improve the efficiency of 2D local attention. Experiments show that the proposed Hilbert Window Attention and Hilbert Slide Attention can accelerate window attention and slide attention by about $4\times$ and $18\times$, respectively. To assess practicality, the strategy is instantiated as the Hilbert Window Transformer and the Hilbert Neighborhood Transformer, both of which achieve end-to-end speedups with minimal accuracy loss. Overall, combining Hilbert-guided local attention with block-sparse kernels offers a general and practical approach to enhancing the efficiency of 2D local attention for images.

Method: Proposes a three-step training-free inference procedure: 1) Distribution-preserving semantic erasure to suppress seed-specific, semantic-irrelevant variation; 2) Timestep-aggregated horizontal injection to reinforce prompt-relevant semantic directions; 3) Spherical retraction to stay near the prior’s typical set.

Result: The method consistently improves alignment and generation quality across multiple diffusion model backbones and benchmarks compared to standard sampling techniques.

Conclusion: The paper provides a theoretical explanation for seed-dependent variations in diffusion models and offers a practical, training-free solution that enhances both prompt faithfulness and image quality by better controlling semantic directions during inference.

Abstract: Diffusion models draw the initial latent from an isotropic Gaussian distribution (all directions equally likely). But in practice, changing only the random seed can sharply alter image quality and prompt faithfulness. We explain this by distinguishing the isotropic prior from the semantics induced by the sampling map: while the prior is direction-agnostic, the mapping from latent noise to semantics has semantic-invariant directions and semantic-sensitive directions, so different seeds can lead to very different semantic outcomes. Motivated by this view, we propose a training-free inference procedure that (i) suppresses seed-specific, semantic-irrelevant variation via distribution-preserving semantic erasure, (ii) reinforces prompt-relevant semantic directions through timestep-aggregated horizontal injection, and (iii) applies a simple spherical retraction to stay near the prior’s typical set. Across multiple backbones and benchmarks, our method consistently improves alignment and generation quality over standard sampling.

Method: Gen-ViRe decomposes Chain-of-Frames reasoning into six cognitive dimensions (from perceptual logic to abstract planning) and 24 subtasks. It uses multi-source data curation, minimal prompting protocols, and hybrid VLM-assisted evaluation with detailed criteria to quantitatively assess video models as reasoners.

Result: Experiments on state-of-the-art systems reveal substantial discrepancies between impressive visual quality and actual reasoning depth. The benchmark establishes baselines and provides diagnostic tools to advance genuine world simulators.

Conclusion: Gen-ViRe provides the first quantitative assessment framework for evaluating video generation models as reasoning agents, addressing the critical gap in measuring cognitive abilities beyond visual quality and enabling systematic advancement of world simulation capabilities.

Abstract: While Chain-of-Thought (CoT) prompting enables sophisticated symbolic reasoning in LLMs, it remains confined to discrete text and cannot simulate the continuous, physics-governed dynamics of the real world. Recent video generation models have emerged as potential world simulators through Chain-of-Frames (CoF) reasoning – materializing thought as frame-by-frame visual sequences, with each frame representing a physically-grounded reasoning step. Despite compelling demonstrations, a challenge persists: existing benchmarks, focusing on fidelity or alignment, do not assess CoF reasoning and thus cannot measure core cognitive abilities in multi-step planning, algorithmic logic, or abstract pattern extrapolation. This evaluation void prevents systematic understanding of model capabilities and principled guidance for improvement. We introduce Gen-ViRe (Generative Visual Reasoning Benchmark), a framework grounded in cognitive science and real-world AI applications, which decomposes CoF reasoning into six cognitive dimensions – from perceptual logic to abstract planning – and 24 subtasks. Through multi-source data curation, minimal prompting protocols, and hybrid VLM-assisted evaluation with detailed criteria, Gen-ViRe delivers the first quantitative assessment of video models as reasoners. Our experiments on SOTA systems reveal substantial discrepancies between impressive visual quality and actual reasoning depth, establishing baselines and diagnostic tools to advance genuine world simulators.

Method: Uses generative adversarial networks (GANs) with semi-supervised learning, combining Dice loss, cross-entropy loss, feature matching loss, and self-training loss to learn from limited annotated scans while leveraging unlabeled data.

Result: Demonstrated on the Acute Ischemic Stroke Dataset (AISD), showing potential to enhance diagnostic capabilities, reduce manual annotation burden, and support clinical decision-making.

Conclusion: The semi-supervised GAN approach shows promise for accurate early ischemic stroke segmentation in CT scans, addressing the challenge of limited annotated medical imaging data.

Abstract: Ischemic stroke is a time-critical medical emergency where rapid diagnosis is essential for improving patient outcomes. Non-contrast computed tomography (NCCT) serves as the frontline imaging tool, yet it often fails to reveal the subtle ischemic changes present in the early, hyperacute phase. This limitation can delay crucial interventions. To address this diagnostic challenge, we introduce a semi-supervised segmentation method using generative adversarial networks (GANs) to accurately delineate early ischemic stroke regions. The proposed method employs an adversarial framework to effectively learn from a limited number of annotated NCCT scans, while simultaneously leveraging a larger pool of unlabeled scans. By employing Dice loss, cross-entropy loss, a feature matching loss and a self-training loss, the model learns to identify and delineate early infarcts, even when they are faint or their size is small. Experiments on the publicly available Acute Ischemic Stroke Dataset (AISD) demonstrate the potential of the proposed method to enhance diagnostic capabilities, reduce the burden of manual annotation, and support more efficient clinical decision-making in stroke care.

Method: Proposes Multiscale Vector Quantized Variational Autoencoder (MSVQ-VAE), a multiscale extension of VQ-VAE that can seamlessly introduce abnormalities into normal WCE images and enable conditional generation of synthetic images with different abnormality types (polyps, vascular, inflammatory conditions).

Result: Training a CDS classifier using abnormal images generated by MSVQ-VAE yields comparable results to training with only real data. The method demonstrates effectiveness across various abnormality types in gastrointestinal imaging.

Conclusion: MSVQ-VAE provides an effective solution for medical image synthesis, particularly for generating diverse abnormal findings in WCE images, addressing data scarcity issues in medical AI development while maintaining privacy.

Abstract: Gastrointestinal (GI) imaging via Wireless Capsule Endoscopy (WCE) generates a large number of images requiring manual screening. Deep learning-based Clinical Decision Support (CDS) systems can assist screening, yet their performance relies on the existence of large, diverse, training medical datasets. However, the scarcity of such data, due to privacy constraints and annotation costs, hinders CDS development. Generative machine learning offers a viable solution to combat this limitation. While current Synthetic Data Generation (SDG) methods, such as Generative Adversarial Networks and Variational Autoencoders have been explored, they often face challenges with training stability and capturing sufficient visual diversity, especially when synthesizing abnormal findings. This work introduces a novel VAE-based methodology for medical image synthesis and presents its application for the generation of WCE images. The novel contributions of this work include a) multiscale extension of the Vector Quantized VAE model, named as Multiscale Vector Quantized Variational Autoencoder (MSVQ-VAE); b) unlike other VAE-based SDG models for WCE image generation, MSVQ-VAE is used to seamlessly introduce abnormalities into normal WCE images; c) it enables conditional generation of synthetic images, enabling the introduction of different types of abnormalities into the normal WCE images; d) it performs experiments with a variety of abnormality types, including polyps, vascular and inflammatory conditions. The utility of the generated images for CDS is assessed via image classification. Comparative experiments demonstrate that training a CDS classifier using the abnormal images generated by the proposed methodology yield comparable results with a classifier trained with only real data. The generality of the proposed methodology promises its applicability to various domains related to medical multimedia.

Method: Two complementary methods: 1) MapReduce LoRA trains preference-specific LoRA experts in parallel and iteratively merges them to refine a shared base model; 2) Reward-aware Token Embedding (RaTE) learns reward-specific token embeddings that compose at inference for flexible preference control.

Result: Significant improvements across modalities: Text-to-Image (Stable Diffusion 3.5 Medium & FLUX.1-dev): 36.1%, 4.6%, 55.7% and 32.7%, 4.3%, 67.1% on GenEval, PickScore, OCR. Text-to-Video (HunyuanVideo): 48.1% visual quality, 90.0% motion quality. Language (Llama-2 7B): 43.4% helpfulness, 136.7% harmlessness improvements.

Conclusion: The framework sets new SOTA for multi-preference alignment across modalities, effectively addressing alignment tax through parallel expert training and compositional token embeddings.

Abstract: Reinforcement learning from human feedback (RLHF) with reward models has advanced alignment of generative models to human aesthetic and perceptual preferences. However, jointly optimizing multiple rewards often incurs an alignment tax, improving one dimension while degrading others. To address this, we introduce two complementary methods: MapReduce LoRA and Reward-aware Token Embedding (RaTE). MapReduce LoRA trains preference-specific LoRA experts in parallel and iteratively merges them to refine a shared base model; RaTE learns reward-specific token embeddings that compose at inference for flexible preference control. Experiments on Text-to-Image generation (Stable Diffusion 3.5 Medium and FLUX.1-dev) show improvements of 36.1%, 4.6%, and 55.7%, and 32.7%, 4.3%, and 67.1% on GenEval, PickScore, and OCR, respectively. On Text-to-Video generation (HunyuanVideo), visual and motion quality improve by 48.1% and 90.0%, respectively. On the language task, Helpful Assistant, with Llama-2 7B, helpful and harmless improve by 43.4% and 136.7%, respectively. Our framework sets a new state-of-the-art multi-preference alignment recipe across modalities.

Method: SUGAR learns personalized surrogate latent representations for each identity to divert reconstructions to visually coherent alternatives, and introduces a continual utility preservation objective to prevent degradation as more identities are forgotten.

Result: SUGAR achieves state-of-the-art performance in removing up to 200 identities simultaneously or sequentially, with up to 700% improvement in retention utility compared to existing baselines.

Conclusion: SUGAR provides an effective framework for scalable generative unlearning that addresses privacy concerns in 3D-aware generative models while maintaining model quality and diversity.

Abstract: Recent advances in 3D-aware generative models have enabled high-fidelity image synthesis of human identities. However, this progress raises urgent questions around user consent and the ability to remove specific individuals from a model’s output space. We address this by introducing SUGAR, a framework for scalable generative unlearning that enables the removal of many identities (simultaneously or sequentially) without retraining the entire model. Rather than projecting unwanted identities to unrealistic outputs or relying on static template faces, SUGAR learns a personalized surrogate latent for each identity, diverting reconstructions to visually coherent alternatives while preserving the model’s quality and diversity. We further introduce a continual utility preservation objective that guards against degradation as more identities are forgotten. SUGAR achieves state-of-the-art performance in removing up to 200 identities, while delivering up to a 700% improvement in retention utility compared to existing baselines. Our code is publicly available at https://github.com/judydnguyen/SUGAR-Generative-Unlearn.

Method: UnCoL uses a dual-teacher framework: one frozen foundation model teacher for general visual/semantic knowledge, and one progressively adapting teacher for task-specific representations. Predictive uncertainty adaptively regulates pseudo-label learning to balance guidance and suppress unreliable supervision.

Result: Experiments on diverse 2D and 3D segmentation benchmarks show UnCoL consistently outperforms state-of-the-art semi-supervised methods and foundation model baselines, achieving near fully supervised performance with significantly reduced annotation requirements.

Conclusion: UnCoL effectively harmonizes generalization and specialization in medical image segmentation by leveraging foundation models while adapting to task-specific requirements through uncertainty-informed collaborative learning.

Abstract: Vision foundation models have demonstrated strong generalization in medical image segmentation by leveraging large-scale, heterogeneous pretraining. However, they often struggle to generalize to specialized clinical tasks under limited annotations or rare pathological variations, due to a mismatch between general priors and task-specific requirements. To address this, we propose Uncertainty-informed Collaborative Learning (UnCoL), a dual-teacher framework that harmonizes generalization and specialization in semi-supervised medical image segmentation. Specifically, UnCoL distills both visual and semantic representations from a frozen foundation model to transfer general knowledge, while concurrently maintaining a progressively adapting teacher to capture fine-grained and task-specific representations. To balance guidance from both teachers, pseudo-label learning in UnCoL is adaptively regulated by predictive uncertainty, which selectively suppresses unreliable supervision and stabilizes learning in ambiguous regions. Experiments on diverse 2D and 3D segmentation benchmarks show that UnCoL consistently outperforms state-of-the-art semi-supervised methods and foundation model baselines. Moreover, our model delivers near fully supervised performance with markedly reduced annotation requirements.

Method: Introduces Block-Recurrent Hypothesis (BRH) and trains Recurrent Approximations to Phase-structured TransfORmers (Raptor) - block-recurrent surrogates of pretrained ViTs using only k ≪ L distinct blocks applied recurrently to approximate original L blocks.

Result: Raptor models recover 96% of DINOv2 ImageNet-1k linear probe accuracy with only 2 blocks at equivalent runtime; analysis reveals directional convergence into class-dependent angular basins, token-specific dynamics, and collapse to low-rank updates in late depth.

Conclusion: A compact recurrent program emerges along ViT depth, pointing to low-complexity normative solutions that enable these models to be studied through principled dynamical systems analysis.

Abstract: As Vision Transformers (ViTs) become standard vision backbones, a mechanistic account of their computational phenomenology is essential. Despite architectural cues that hint at dynamical structure, there is no settled framework that interprets Transformer depth as a well-characterized flow. In this work, we introduce the Block-Recurrent Hypothesis (BRH), arguing that trained ViTs admit a block-recurrent depth structure such that the computation of the original $L$ blocks can be accurately rewritten using only $k \ll L$ distinct blocks applied recurrently. Across diverse ViTs, between-layer representational similarity matrices suggest few contiguous phases. To determine whether these phases reflect genuinely reusable computation, we train block-recurrent surrogates of pretrained ViTs: Recurrent Approximations to Phase-structured TransfORmers (Raptor). In small-scale, we demonstrate that stochastic depth and training promote recurrent structure and subsequently correlate with our ability to accurately fit Raptor. We then provide an empirical existence proof for BRH by training a Raptor model to recover $96%$ of DINOv2 ImageNet-1k linear probe accuracy in only 2 blocks at equivalent runtime. Finally, we leverage our hypothesis to develop a program of Dynamical Interpretability. We find i) directional convergence into class-dependent angular basins with self-correcting trajectories under small perturbations, ii) token-specific dynamics, where cls executes sharp late reorientations while patch tokens exhibit strong late-stage coherence toward their mean direction, and iii) a collapse to low rank updates in late depth, consistent with convergence to low-dimensional attractors. Altogether, we find a compact recurrent program emerges along ViT depth, pointing to a low-complexity normative solution that enables these models to be studied through principled dynamical systems analysis.

Method: Proposes IBISAgent that enables MLLMs to generate interleaved reasoning and text-based click actions to invoke segmentation tools. Uses iterative multi-step visual reasoning on masked image features to support mask refinement. Employs a two-stage training framework: cold-start supervised fine-tuning followed by agentic reinforcement learning with fine-grained rewards.

Result: Extensive experiments demonstrate that IBISAgent consistently outperforms both closed-source and open-source state-of-the-art methods in medical referring and reasoning segmentation tasks.

Conclusion: IBISAgent successfully addresses limitations of existing approaches by reformulating segmentation as a vision-centric decision-making process, enabling iterative refinement without architectural changes, and showing superior performance in complex medical segmentation tasks.

Abstract: Recent research on medical MLLMs has gradually shifted its focus from image-level understanding to fine-grained, pixel-level comprehension. Although segmentation serves as the foundation for pixel-level understanding, existing approaches face two major challenges. First, they introduce implicit segmentation tokens and require simultaneous fine-tuning of both the MLLM and external pixel decoders, which increases the risk of catastrophic forgetting and limits generalization to out-of-domain scenarios. Second, most methods rely on single-pass reasoning and lack the capability to iteratively refine segmentation results, leading to suboptimal performance. To overcome these limitations, we propose a novel agentic MLLM, named IBISAgent, that reformulates segmentation as a vision-centric, multi-step decision-making process. IBISAgent enables MLLMs to generate interleaved reasoning and text-based click actions, invoke segmentation tools, and produce high-quality masks without architectural modifications. By iteratively performing multi-step visual reasoning on masked image features, IBISAgent naturally supports mask refinement and promotes the development of pixel-level visual reasoning capabilities. We further design a two-stage training framework consisting of cold-start supervised fine-tuning and agentic reinforcement learning with tailored, fine-grained rewards, enhancing the model’s robustness in complex medical referring and reasoning segmentation tasks. Extensive experiments demonstrate that IBISAgent consistently outperforms both closed-source and open-source SOTA methods. All datasets, code, and trained models will be released publicly.

Method: Introduces Sparse Voxel Rasterization (SVRaster) pipeline initialized from SfM feature geometry, projecting 2D instance masks onto explicit voxel grid with recursive splitting for geometric separation.

Result: Mask-supervised NeRFs saturate at ~89% instance recovery in dense scenes, while explicit SVRaster achieves 95.8% recovery rate and shows 43% more robustness to degraded segmentation masks.

Conclusion: Explicit geometric priors are essential for reliable quantitative 3D analysis in highly self-occluding scenes, overcoming limitations of implicit density fields.

Abstract: Neural Radiance Fields (NeRFs) have emerged as a powerful paradigm for multi-view reconstruction, complementing classical photogrammetric pipelines based on Structure-from-Motion (SfM) and Multi-View Stereo (MVS). However, their reliability for quantitative 3D analysis in dense, self-occluding scenes remains poorly understood. In this study, we identify a fundamental failure mode of implicit density fields under heavy occlusion, which we term Interior Geometric Degradation (IGD). We show that transmittance-based volumetric optimization satisfies photometric supervision by reconstructing hollow or fragmented structures rather than solid interiors, leading to systematic instance undercounting. Through controlled experiments on synthetic datasets with increasing occlusion, we demonstrate that state-of-the-art mask-supervised NeRFs saturate at approximately 89% instance recovery in dense scenes, despite improved surface coherence and mask quality. To overcome this limitation, we introduce an explicit geometric pipeline based on Sparse Voxel Rasterization (SVRaster), initialized from SfM feature geometry. By projecting 2D instance masks onto an explicit voxel grid and enforcing geometric separation via recursive splitting, our approach preserves physical solidity and achieves a 95.8% recovery rate in dense clusters. A sensitivity analysis using degraded segmentation masks further shows that explicit SfM-based geometry is substantially more robust to supervision failure, recovering 43% more instances than implicit baselines. These results demonstrate that explicit geometric priors are a prerequisite for reliable quantitative analysis in highly self-occluding 3D scenes.

Method: Neural Gaussian Force Field (NGFF) combines 3D Gaussian perception with physics-based dynamic modeling in an end-to-end neural framework. Uses multi-view RGB inputs to generate interactive, physically realistic 4D videos. Also introduces GSCollision dataset with diverse materials, multi-object interactions, and complex scenes.

Result: Achieves two orders of magnitude (100x) faster performance than prior Gaussian simulators. Strong generalization and robustness in physical reasoning demonstrated on synthetic and real 3D scenarios. Dataset includes over 640k rendered physical videos (~4 TB).

Conclusion: NGFF advances video prediction towards physics-grounded world models by efficiently integrating visual perception with physical dynamics, enabling physically realistic 4D video generation from multi-view inputs.

Abstract: Predicting physical dynamics from raw visual data remains a major challenge in AI. While recent video generation models have achieved impressive visual quality, they still cannot consistently generate physically plausible videos due to a lack of modeling of physical laws. Recent approaches combining 3D Gaussian splatting and physics engines can produce physically plausible videos, but are hindered by high computational costs in both reconstruction and simulation, and often lack robustness in complex real-world scenarios. To address these issues, we introduce Neural Gaussian Force Field (NGFF), an end-to-end neural framework that integrates 3D Gaussian perception with physics-based dynamic modeling to generate interactive, physically realistic 4D videos from multi-view RGB inputs, achieving two orders of magnitude faster than prior Gaussian simulators. To support training, we also present GSCollision, a 4D Gaussian dataset featuring diverse materials, multi-object interactions, and complex scenes, totaling over 640k rendered physical videos (~4 TB). Evaluations on synthetic and real 3D scenarios show NGFF’s strong generalization and robustness in physical reasoning, advancing video prediction towards physics-grounded world models.

Method: Three key innovations: 1) Object existence prior that dynamically weights relevant categories via global image-text similarity to reduce hallucinations; 2) Region-aware alignment module for precise region-level visual-textual correspondences; 3) Dual-stream fusion mechanism combining local structural information with global semantic context. Eliminates need for external mask proposals, additional backbones, or extra datasets.

Result: Extensive experiments on six benchmarks (A-847, PC-459, A-150, PC-59, PAS-20, and PAS-20b) demonstrate competitive performance and strong generalization in open-vocabulary settings.

Conclusion: LoGoSeg provides an efficient single-stage framework for open-vocabulary semantic segmentation that addresses spatial alignment issues and object hallucination problems through integrated object priors and region-level constraints.

Abstract: Open-vocabulary semantic segmentation (OVSS) extends traditional closed-set segmentation by enabling pixel-wise annotation for both seen and unseen categories using arbitrary textual descriptions. While existing methods leverage vision-language models (VLMs) like CLIP, their reliance on image-level pretraining often results in imprecise spatial alignment, leading to mismatched segmentations in ambiguous or cluttered scenes. However, most existing approaches lack strong object priors and region-level constraints, which can lead to object hallucination or missed detections, further degrading performance. To address these challenges, we propose LoGoSeg, an efficient single-stage framework that integrates three key innovations: (i) an object existence prior that dynamically weights relevant categories through global image-text similarity, effectively reducing hallucinations; (ii) a region-aware alignment module that establishes precise region-level visual-textual correspondences; and (iii) a dual-stream fusion mechanism that optimally combines local structural information with global semantic context. Unlike prior works, LoGoSeg eliminates the need for external mask proposals, additional backbones, or extra datasets, ensuring efficiency. Extensive experiments on six benchmarks (A-847, PC-459, A-150, PC-59, PAS-20, and PAS-20b) demonstrate its competitive performance and strong generalization in open-vocabulary settings.

Method: Proposes ZoomDet with two core designs: 1) Lightweight offset prediction with box-based zooming objective for non-uniform zooming, and 2) Corner-aligned bounding box transformation to warp ground-truth boxes to zoomed space for training and predicted boxes back to original space for inference.

Result: Extensive experiments on VisDrone, UAVDT, and SeaDronesSee datasets show significant improvements. On SeaDronesSee, ZoomDet achieves >8.4 mAP absolute gain with Faster R-CNN, adding only ~3ms latency. The framework is architecture-independent.

Conclusion: ZoomDet effectively addresses small object detection in UAV images through adaptive zooming, achieving substantial performance gains with minimal computational overhead, making it practical for real-world UAV applications.

Abstract: Detecting objects from UAV-captured images is challenging due to the small object size. In this work, a simple and efficient adaptive zoom-in framework is explored for object detection on UAV images. The main motivation is that the foreground objects are generally smaller and sparser than those in common scene images, which hinders the optimization of effective object detectors. We thus aim to zoom in adaptively on the objects to better capture object features for the detection task. To achieve the goal, two core designs are required: \textcolor{black}{i) How to conduct non-uniform zooming on each image efficiently? ii) How to enable object detection training and inference with the zoomed image space?} Correspondingly, a lightweight offset prediction scheme coupled with a novel box-based zooming objective is introduced to learn non-uniform zooming on the input image. Based on the learned zooming transformation, a corner-aligned bounding box transformation method is proposed. The method warps the ground-truth bounding boxes to the zoomed space to learn object detection, and warps the predicted bounding boxes back to the original space during inference. We conduct extensive experiments on three representative UAV object detection datasets, including VisDrone, UAVDT, and SeaDronesSee. The proposed ZoomDet is architecture-independent and can be applied to an arbitrary object detection architecture. Remarkably, on the SeaDronesSee dataset, ZoomDet offers more than 8.4 absolute gain of mAP with a Faster R-CNN model, with only about 3 ms additional latency. The code is available at https://github.com/twangnh/zoomdet_code.

Method: MambaFusion interleaves selective state-space models (SSMs) with windowed transformers to propagate global context in linear time while preserving local geometric fidelity. It includes a multi-modal token alignment module and reliability-aware fusion gates that dynamically re-weight camera-LiDAR features based on spatial confidence and calibration consistency. A structure-conditioned diffusion head integrates graph-based reasoning with uncertainty-aware denoising.

Result: MambaFusion establishes new state-of-the-art performance on nuScenes benchmarks while operating with linear-time complexity. The framework demonstrates robust, temporally stable, and interpretable 3D perception for real-world autonomous driving systems.

Conclusion: Coupling SSM-based efficiency with reliability-driven fusion yields effective multimodal 3D perception. The approach addresses key challenges in camera-LiDAR fusion through efficient context modeling, adaptive feature weighting, and physically grounded reasoning.

Abstract: Reliable 3D object detection is fundamental to autonomous driving, and multimodal fusion algorithms using cameras and LiDAR remain a persistent challenge. Cameras provide dense visual cues but ill posed depth; LiDAR provides a precise 3D structure but sparse coverage. Existing BEV-based fusion frameworks have made good progress, but they have difficulties including inefficient context modeling, spatially invariant fusion, and reasoning under uncertainty. We introduce MambaFusion, a unified multi-modal detection framework that achieves efficient, adaptive, and physically grounded 3D perception. MambaFusion interleaves selective state-space models (SSMs) with windowed transformers to propagate the global context in linear time while preserving local geometric fidelity. A multi-modal token alignment (MTA) module and reliability-aware fusion gates dynamically re-weight camera-LiDAR features based on spatial confidence and calibration consistency. Finally, a structure-conditioned diffusion head integrates graph-based reasoning with uncertainty-aware denoising, enforcing physical plausibility, and calibrated confidence. MambaFusion establishes new state-of-the-art performance on nuScenes benchmarks while operating with linear-time complexity. The framework demonstrates that coupling SSM-based efficiency with reliability-driven fusion yields robust, temporally stable, and interpretable 3D perception for real-world autonomous driving systems.

Method: Uses CLIP Sparse Autoencoders to extract editing directions in CLIP latent space and isolate concept pairs from input images, creating synthetic triplets of decomposed visual aspects. The feed-forward approach generates diverse, visually coherent compositions revealing latent relationships between two input images without text prompts.

Result: The framework produces diverse and visually coherent compositions that reveal latent relationships between input images, enabling fast, intuitive visual recombination for creative exploration.

Conclusion: Inspiration Seeds shifts image generation from final execution to exploratory ideation, supporting visual exploration at early creative stages by removing reliance on language and enabling intuitive visual recombination.

Abstract: While generative models have become powerful tools for image synthesis, they are typically optimized for executing carefully crafted textual prompts, offering limited support for the open-ended visual exploration that often precedes idea formation. In contrast, designers frequently draw inspiration from loosely connected visual references, seeking emergent connections that spark new ideas. We propose Inspiration Seeds, a generative framework that shifts image generation from final execution to exploratory ideation. Given two input images, our model produces diverse, visually coherent compositions that reveal latent relationships between inputs, without relying on user-specified text prompts. Our approach is feed-forward, trained on synthetic triplets of decomposed visual aspects derived entirely through visual means: we use CLIP Sparse Autoencoders to extract editing directions in CLIP latent space and isolate concept pairs. By removing the reliance on language and enabling fast, intuitive recombination, our method supports visual ideation at the early and ambiguous stages of creative work.

Method: 1) Introduces Omni Dense Captioning task with six-dimensional structural schema; 2) Creates OmniDCBench benchmark with human annotations; 3) Proposes SodaM metric for time-aware evaluation; 4) Builds TimeChatCap-42K training dataset; 5) Develops TimeChat-Captioner-7B model trained via SFT and GRPO with task-specific rewards.

Result: TimeChat-Captioner-7B achieves state-of-the-art performance, surpassing Gemini-2.5-Pro. Generated dense descriptions significantly boost downstream capabilities in audio-visual reasoning (DailyOmni and WorldSense) and temporal grounding (Charades-STA).

Conclusion: The proposed Omni Dense Captioning task and TimeChat-Captioner-7B model provide effective solutions for generating structured audio-visual narratives with temporal information, demonstrating strong performance and downstream benefits.

Abstract: This paper proposes Omni Dense Captioning, a novel task designed to generate continuous, fine-grained, and structured audio-visual narratives with explicit timestamps. To ensure dense semantic coverage, we introduce a six-dimensional structural schema to create “script-like” captions, enabling readers to vividly imagine the video content scene by scene, akin to a cinematographic screenplay. To facilitate research, we construct OmniDCBench, a high-quality, human-annotated benchmark, and propose SodaM, a unified metric that evaluates time-aware detailed descriptions while mitigating scene boundary ambiguity. Furthermore, we construct a training dataset, TimeChatCap-42K, and present TimeChat-Captioner-7B, a strong baseline trained via SFT and GRPO with task-specific rewards. Extensive experiments demonstrate that TimeChat-Captioner-7B achieves state-of-the-art performance, surpassing Gemini-2.5-Pro, while its generated dense descriptions significantly boost downstream capabilities in audio-visual reasoning (DailyOmni and WorldSense) and temporal grounding (Charades-STA). All datasets, models, and code will be made publicly available at https://github.com/yaolinli/TimeChat-Captioner.

Method: Proposes WeakSupCon, a weakly supervised contrastive learning framework that incorporates bag-level label information during training. The method learns to separate patches with different labels in feature space without relying on instance-level pseudo-labeling.

Result: Experimental results show that features generated by WeakSupCon lead to improved downstream MIL performance compared to self-supervised contrastive learning approaches across three datasets.

Conclusion: WeakSupCon effectively addresses the feature representation learning gap in MIL for histopathology, demonstrating that incorporating bag-level supervision during feature learning improves downstream performance without needing instance-level annotations.

Abstract: Digital histopathology whole slide images (WSIs) provide gigapixel-scale high-resolution images that are highly useful for disease diagnosis. However, digital histopathology image analysis faces significant challenges due to the limited training labels, since manually annotating specific regions or small patches cropped from large WSIs requires substantial time and effort. Weakly supervised multiple instance learning (MIL) offers a practical and efficient solution by requiring only bag-level (slide-level) labels, while each bag typically contains multiple instances (patches). Most MIL methods directly use frozen image patch features generated by various image encoders as inputs and primarily focus on feature aggregation. However, feature representation learning for encoder pretraining in MIL settings has largely been neglected. In our work, we propose a novel feature representation learning framework called weakly supervised contrastive learning (WeakSupCon) that incorporates bag-level label information during training. Our method does not rely on instance-level pseudo-labeling, yet it effectively separates patches with different labels in the feature space. Experimental results demonstrate that the image features generated by our WeakSupCon method lead to improved downstream MIL performance compared to self-supervised contrastive learning approaches in three datasets. Our related code is available at github.com/BzhangURU/Paper_WeakSupCon_for_MIL

Method: Three-stage pipeline to synthesize million-scale image-description dataset using MLLMs as weak annotators with consensus fusion and expert refinement, then fine-tune Qwen3-VL-4B model via multi-stage strategy.

Result: Singpath-VL demonstrates superior performance in fine-grained morphological perception and cell-level diagnostic classification for cervical cytology.

Conclusion: The approach successfully creates a specialized cytopathology MLLM and will open-source synthetic dataset portions to advance the field.

Abstract: We present Singpath-VL, a vision-language large model, to fill the vacancy of AI assistant in cervical cytology. Recent advances in multi-modal large language models (MLLMs) have significantly propelled the field of computational pathology. However, their application in cytopathology, particularly cervical cytology, remains underexplored, primarily due to the scarcity of large-scale, high-quality annotated datasets. To bridge this gap, we first develop a novel three-stage pipeline to synthesize a million-scale image-description dataset. The pipeline leverages multiple general-purpose MLLMs as weak annotators, refines their outputs through consensus fusion and expert knowledge injection, and produces high-fidelity descriptions of cell morphology. Using this dataset, we then fine-tune the Qwen3-VL-4B model via a multi-stage strategy to create a specialized cytopathology MLLM. The resulting model, named Singpath-VL, demonstrates superior performance in fine-grained morphological perception and cell-level diagnostic classification. To advance the field, we will open-source a portion of the synthetic dataset and benchmark.

Method: The paper presents Kelix, a fully discrete autoregressive unified model that uses discrete visual tokenization to create a shared, discrete representation across modalities, enabling fully autoregressive multimodal modeling with improved understanding capabilities.

Result: Kelix closes the understanding gap between discrete and continuous visual representations, achieving comparable or better understanding performance than continuous-feature vision-language models while maintaining the benefits of fully discrete autoregressive modeling.

Conclusion: Fully discrete autoregressive multimodal modeling is feasible and can achieve strong understanding performance comparable to continuous-feature models, enabling unified comprehension and generation under self-supervision across modalities.

Abstract: Autoregressive large language models (LLMs) scale well by expressing diverse tasks as sequences of discrete natural-language tokens and training with next-token prediction, which unifies comprehension and generation under self-supervision. Extending this paradigm to multimodal data requires a shared, discrete representation across modalities. However, most vision-language models (VLMs) still rely on a hybrid interface: discrete text tokens paired with continuous Vision Transformer (ViT) features. Because supervision is largely text-driven, these models are often biased toward understanding and cannot fully leverage large-scale self-supervised learning on non-text data. Recent work has explored discrete visual tokenization to enable fully autoregressive multimodal modeling, showing promising progress toward unified understanding and generation. Yet existing discrete vision tokens frequently lose information due to limited code capacity, resulting in noticeably weaker understanding than continuous-feature VLMs. We present Kelix, a fully discrete autoregressive unified model that closes the understanding gap between discrete and continuous visual representations.

Method: Proposes 3DXTalker with three key components: 1) Data-curated identity modeling using 2D-to-3D pipeline and disentangled representations to address data scarcity, 2) Audio-rich representations including frame-wise amplitude and emotional cues beyond standard speech embeddings, 3) Flow-matching-based transformer for coherent facial dynamics and prompt-based conditioning for stylized head-pose control.

Result: Extensive experiments show 3DXTalker achieves superior performance in 3D talking avatar generation, successfully integrating lip synchronization, emotional expression, and head-pose dynamics within a unified framework.

Conclusion: 3DXTalker provides a comprehensive solution for expressive 3D talking avatar generation that overcomes data limitations, enhances audio representation, and enables spatial controllability, advancing the field of virtual communication and digital humans.

Abstract: Audio-driven 3D talking avatar generation is increasingly important in virtual communication, digital humans, and interactive media, where avatars must preserve identity, synchronize lip motion with speech, express emotion, and exhibit lifelike spatial dynamics, collectively defining a broader objective of expressivity. However, achieving this remains challenging due to insufficient training data with limited subject identities, narrow audio representations, and restricted explicit controllability. In this paper, we propose 3DXTalker, an expressive 3D talking avatar through data-curated identity modeling, audio-rich representations, and spatial dynamics controllability. 3DXTalker enables scalable identity modeling via 2D-to-3D data curation pipeline and disentangled representations, alleviating data scarcity and improving identity generalization. Then, we introduce frame-wise amplitude and emotional cues beyond standard speech embeddings, ensuring superior lip synchronization and nuanced expression modulation. These cues are unified by a flow-matching-based transformer for coherent facial dynamics. Moreover, 3DXTalker also enables natural head-pose motion generation while supporting stylized control via prompt-based conditioning. Extensive experiments show that 3DXTalker integrates lip synchronization, emotional expression, and head-pose dynamics within a unified framework, achieves superior performance in 3D talking avatar generation.

Method: Proposes OmniVL-Guard with two core designs: 1) Self-Evolving CoT Generation synthesizes high-quality reasoning paths to overcome cold-start challenges, and 2) Adaptive Reward Scaling Policy Optimization (ARSPO) dynamically modulates reward scales and task weights for balanced joint optimization.

Result: Extensive experiments show OmniVL-Guard significantly outperforms state-of-the-art methods and exhibits zero-shot robust generalization across out-of-domain scenarios.

Conclusion: The proposed framework successfully addresses the difficulty bias problem in omnibus vision-language forgery detection and grounding, providing a unified solution for multi-modal misinformation analysis with strong generalization capabilities.

Abstract: Existing forgery detection methods are often limited to uni-modal or bi-modal settings, failing to handle the interleaved text, images, and videos prevalent in real-world misinformation. To bridge this gap, this paper targets to develop a unified framework for omnibus vision-language forgery detection and grounding. In this unified setting, the {interplay} between diverse modalities and the dual requirements of simultaneous detection and localization pose a critical difficulty bias problem: the simpler veracity classification task tends to dominate the gradients, leading to suboptimal performance in fine-grained grounding during multi-task optimization. To address this challenge, we propose \textbf{OmniVL-Guard}, a balanced reinforcement learning framework for omnibus vision-language forgery detection and grounding. Particularly, OmniVL-Guard comprises two core designs: Self-Evolving CoT Generatio and Adaptive Reward Scaling Policy Optimization (ARSPO). {Self-Evolving CoT Generation} synthesizes high-quality reasoning paths, effectively overcoming the cold-start challenge. Building upon this, {Adaptive Reward Scaling Policy Optimization (ARSPO)} dynamically modulates reward scales and task weights, ensuring a balanced joint optimization. Extensive experiments demonstrate that OmniVL-Guard significantly outperforms state-of-the-art methods and exhibits zero-shot robust generalization across out-of-domain scenarios.

Method: Proposes “chatting with images” framework with language-guided feature modulation. ViLaVT model uses dynamic vision encoder for interactive visual reasoning, trained with two-stage curriculum combining supervised fine-tuning and reinforcement learning.

Result: Extensive experiments across eight benchmarks show strong improvements, with particularly pronounced gains on complex multi-image and video-based spatial reasoning tasks.

Conclusion: Language-guided feature modulation enables tighter coupling between linguistic reasoning and visual state updates, addressing limitations of current LVLMs for complex visual reasoning tasks.

Abstract: Current large vision-language models (LVLMs) typically rely on text-only reasoning based on a single-pass visual encoding, which often leads to loss of fine-grained visual information. Recently the proposal of ‘’thinking with images’’ attempts to alleviate this limitation by manipulating images via external tools or code; however, the resulting visual states are often insufficiently grounded in linguistic semantics, impairing effective cross-modal alignment - particularly when visual semantics or geometric relationships must be reasoned over across distant regions or multiple images. To address these challenges, we propose ‘‘chatting with images’’, a new framework that reframes visual manipulation as language-guided feature modulation. Under the guidance of expressive language prompts, the model dynamically performs joint re-encoding over multiple image regions, enabling tighter coupling between linguistic reasoning and visual state updates. We instantiate this paradigm in ViLaVT, a novel LVLM equipped with a dynamic vision encoder explicitly designed for such interactive visual reasoning, and trained it with a two-stage curriculum combining supervised fine-tuning and reinforcement learning to promote effective reasoning behaviors. Extensive experiments across eight benchmarks demonstrate that ViLaVT achieves strong and consistent improvements, with particularly pronounced gains on complex multi-image and video-based spatial reasoning tasks.

Method: Developed XAI module in DashAI integrating three complementary techniques: Partial Dependence Plots (PDP), Permutation Feature Importance (PFI), and KernelSHAP for tabular classification. Conducted user study with 20 participants (ML novices and experts) to evaluate usability and explanation impact.

Result: High task success (≥80%) across all explainability tasks; novices rated explanations as useful, accurate, and trustworthy on ESS (α=0.74), while experts were more critical of sufficiency/completeness; explanations improved perceived predictability and confidence on TiA scale (α=0.60), with novices showing higher trust than experts.

Conclusion: XAI in no-code ML faces challenge of making explanations accessible to novices while sufficiently detailed for experts. The integrated approach shows promise but requires balancing different user needs.

Abstract: The increasing use of Machine Learning (ML) in sensitive domains such as healthcare, finance, and public policy has raised concerns about the transparency of automated decisions. Explainable AI (XAI) addresses this by clarifying how models generate predictions, yet most methods demand technical expertise, limiting their value for novices. This gap is especially critical in no-code ML platforms, which seek to democratize AI but rarely include explainability. We present a human-centered XAI module in DashAI, an open-source no-code ML platform. The module integrates three complementary techniques, which are Partial Dependence Plots (PDP), Permutation Feature Importance (PFI), and KernelSHAP, into DashAI’s workflow for tabular classification. A user study (N = 20; ML novices and experts) evaluated usability and the impact of explanations. Results show: (i) high task success ($\geq80%$) across all explainability tasks; (ii) novices rated explanations as useful, accurate, and trustworthy on the Explanation Satisfaction Scale (ESS, Cronbach’s $α$ = 0.74, a measure of internal consistency), while experts were more critical of sufficiency and completeness; and (iii) explanations improved perceived predictability and confidence on the Trust in Automation scale (TiA, $α$ = 0.60), with novices showing higher trust than experts. These findings highlight a central challenge for XAI in no-code ML, making explanations both accessible to novices and sufficiently detailed for experts.

Method: Two-stage framework: (1) maps diverse PDE states into shared latent physics space using pretrained VAE, (2) learns probabilistic latent dynamics with Transformer trained by flow matching. Key mechanisms include uncertainty knob for correcting off-manifold rollout drift and flow forcing to update system descriptors from model-generated trajectories.

Result: LGS matches strong deterministic neural-operator baselines on short horizons while substantially reducing rollout drift on long horizons. Achieves up to 70× lower FLOPs than non-generative baselines, enabling scalable pretraining on ~2.5M trajectories at 128² resolution across 12 PDE families. Shows efficient adaptation to out-of-distribution 256² Kolmogorov flow dataset.

Conclusion: LGS provides a practical route toward generalizable, uncertainty-aware neural PDE solvers that are more reliable for long-term forecasting and downstream scientific workflows, with efficient computational characteristics enabling scalable pretraining.

Abstract: We study long-horizon surrogate simulation across heterogeneous PDE systems. We introduce Latent Generative Solvers (LGS), a two-stage framework that (i) maps diverse PDE states into a shared latent physics space with a pretrained VAE, and (ii) learns probabilistic latent dynamics with a Transformer trained by flow matching. Our key mechanism is an uncertainty knob that perturbs latent inputs during training and inference, teaching the solver to correct off-manifold rollout drift and stabilizing autoregressive prediction. We further use flow forcing to update a system descriptor (context) from model-generated trajectories, aligning train/test conditioning and improving long-term stability. We pretrain on a curated corpus of $\sim$2.5M trajectories at $128^2$ resolution spanning 12 PDE families. LGS matches strong deterministic neural-operator baselines on short horizons while substantially reducing rollout drift on long horizons. Learning in latent space plus efficient architectural choices yields up to \textbf{70$\times$} lower FLOPs than non-generative baselines, enabling scalable pretraining. We also show efficient adaptation to an out-of-distribution $256^2$ Kolmogorov flow dataset under limited finetuning budgets. Overall, LGS provides a practical route toward generalizable, uncertainty-aware neural PDE solvers that are more reliable for long-term forecasting and downstream scientific workflows.

Method: Formalizes decision-valued maps, develops DecisionDB infrastructure with content-based identifiers, write-once artifact storage, and deterministic replay mechanisms.

Result: Creates a system that partitions representation space into persistence regions and boundaries, enabling mechanically checkable decision reuse and exact recovery of recorded decisions.

Conclusion: Provides a framework for understanding and managing how data representations affect computational outcomes, with practical infrastructure for reproducibility and auditing.

Abstract: A computational engine applied to different representations of the same data can produce different discrete outcomes, with some representations preserving the result and others changing it entirely. A decision-valued map records which representations preserve the outcome and which change it, associating each member of a declared representation family with the discrete result it produces. This paper formalizes decision-valued maps and describes DecisionDB, an infrastructure that logs, replays and audits these relationships using identifiers computed from content and artifacts stored in write-once form. Deterministic replay recovers each recorded decision identifier exactly from stored artifacts, with all three identifying fields matching their persisted values. The contribution partitions representation space into persistence regions and boundaries, and treats decision reuse as a mechanically checkable condition.

Method: Uses Delayed Streams Modeling framework with a new causal audio encoder and Ada RMS-Norm for delay conditioning. Trained end-to-end for streaming with explicit audio-text alignment, scaled on a 13-language dataset.

Result: At 480ms delay, achieves performance on par with Whisper (the leading offline system) while maintaining real-time streaming capability.

Conclusion: Demonstrates that streaming ASR can match offline quality through proper end-to-end training and architectural innovations, with models released under Apache 2.0 license.

Abstract: We introduce Voxtral Realtime, a natively streaming automatic speech recognition model that matches offline transcription quality at sub-second latency. Unlike approaches that adapt offline models through chunking or sliding windows, Voxtral Realtime is trained end-to-end for streaming, with explicit alignment between audio and text streams. Our architecture builds on the Delayed Streams Modeling framework, introducing a new causal audio encoder and Ada RMS-Norm for improved delay conditioning. We scale pretraining to a large-scale dataset spanning 13 languages. At a delay of 480ms, Voxtral Realtime achieves performance on par with Whisper, the most widely deployed offline transcription system. We release the model weights under the Apache 2.0 license.

Method: Introduces PBSAI Governance Ecosystem with: 1) 12-domain taxonomy organizing responsibilities, 2) bounded agent families mediating between tools and policy via shared context envelopes and structured output contracts, 3) lightweight formal model of agents, context envelopes, and ecosystem invariants, 4) alignment with NIST AI RMF functions.

Result: Demonstrates alignment with NIST AI RMF functions and illustrates application in enterprise SOC and hyperscale defensive environments. Provides structured, evidence-centric foundation for open ecosystem development.

Conclusion: PBSAI offers a practical reference architecture for securing enterprise AI estates with formal guarantees for traceability, provenance, and human-in-the-loop oversight across multi-agent systems.

Abstract: Enterprises are rapidly deploying large language models, retrieval augmented generation pipelines, and tool using agents into production, often on shared high performance computing clusters and cloud accelerator platforms that also support defensive analytics. These systems increasingly function not as isolated models but as AI estates: socio technical systems spanning models, agents, data pipelines, security tooling, human workflows, and hyperscale infrastructure. Existing governance and security frameworks, including the NIST AI Risk Management Framework and systems security engineering guidance, articulate principles and risk functions but do not provide implementable architectures for multi agent, AI enabled cyber defense. This paper introduces the Practitioners Blueprint for Secure AI (PBSAI) Governance Ecosystem, a multi agent reference architecture for securing enterprise and hyperscale AI estates. PBSAI organizes responsibilities into a twelve domain taxonomy and defines bounded agent families that mediate between tools and policy through shared context envelopes and structured output contracts. The architecture assumes baseline enterprise security capabilities and encodes key systems security techniques, including analytic monitoring, coordinated defense, and adaptive response. A lightweight formal model of agents, context envelopes, and ecosystem level invariants clarifies the traceability, provenance, and human in the loop guarantees enforced across domains. We demonstrate alignment with NIST AI RMF functions and illustrate application in enterprise SOC and hyperscale defensive environments. PBSAI is proposed as a structured, evidence centric foundation for open ecosystem development and future empirical validation.

Method: Introduces Distributionally robust IGM (DrIGM) principle with robust individual action values, derives DrIGM-compliant robust variants of existing value-factorization architectures (VDN/QMIX/QTRAN) that train on robust Q-targets while preserving scalability.

Result: Empirical evaluation on high-fidelity SustainGym simulators and StarCraft shows consistent improvement in out-of-distribution performance compared to baseline methods.

Conclusion: DrIGM provides a principled approach to robust multi-agent RL that integrates seamlessly with existing codebases and offers provable robustness guarantees for real-world deployment.

Abstract: Cooperative multi-agent reinforcement learning (MARL) commonly adopts centralized training with decentralized execution, where value-factorization methods enforce the individual-global-maximum (IGM) principle so that decentralized greedy actions recover the team-optimal joint action. However, the reliability of this recipe in real-world settings remains unreliable due to environmental uncertainties arising from the sim-to-real gap, model mismatch, and system noise. We address this gap by introducing Distributionally robust IGM (DrIGM), a principle that requires each agent’s robust greedy action to align with the robust team-optimal joint action. We show that DrIGM holds for a novel definition of robust individual action values, which is compatible with decentralized greedy execution and yields a provable robustness guarantee for the whole system. Building on this foundation, we derive DrIGM-compliant robust variants of existing value-factorization architectures (e.g., VDN/QMIX/QTRAN) that (i) train on robust Q-targets, (ii) preserve scalability, and (iii) integrate seamlessly with existing codebases without bespoke per-agent reward shaping. Empirically, on high-fidelity SustainGym simulators and a StarCraft game environment, our methods consistently improve out-of-distribution performance. Code and data are available at https://github.com/crqu/robust-coMARL.

Method: Systematic literature review of 30,897 records from 2017 premier venues (ACL, AIES, CHI, CSCW, EAAMO, FAccT, NeurIPS) between 2020-2025, refined to 346 papers for qualitative synthesis using reflexive thematic analysis.

Result: Reveals systemic failures in positional legibility, anchoring bias from model-mediated annotations, geographic hegemony imposing Western norms, and precarious data workers prioritizing compliance over honest subjectivity, all contributing to the “consensus trap.”

Conclusion: Argues for reclaiming disagreement as high-fidelity signal and proposes roadmap for pluralistic annotation infrastructures that map diversity of human experience rather than seeking singular “right” answers.

Abstract: In machine learning, “ground truth” refers to the assumed correct labels used to train and evaluate models. However, the foundational “ground truth” paradigm rests on a positivistic fallacy that treats human disagreement as technical noise rather than a vital sociotechnical signal. This systematic literature review analyzes research published between 2020 and 2025 across seven premier venues: ACL, AIES, CHI, CSCW, EAAMO, FAccT, and NeurIPS, investigating the mechanisms in data annotation practices that facilitate this “consensus trap”. Our identification phase captured 30,897 records, which were refined via a tiered keyword filtration schema to a high-recall corpus of 3,042 records for manual screening, resulting in a final included corpus of 346 papers for qualitative synthesis. Our reflexive thematic analysis reveals that systemic failures in positional legibility, combined with the recent architectural shift toward human-as-verifier models, specifically the reliance on model-mediated annotations, introduce deep-seated anchoring bias and effectively remove human voices from the loop. We further demonstrate how geographic hegemony imposes Western norms as universal benchmarks, often enforced by the performative alignment of precarious data workers who prioritize requester compliance over honest subjectivity to avoid economic penalties. Critiquing the “noisy sensor” fallacy, where statistical models misdiagnose cultural pluralism as random error, we argue for reclaiming disagreement as a high-fidelity signal essential for building culturally competent models. To address these systemic tensions, we propose a roadmap for pluralistic annotation infrastructures that shift the objective from discovering a singular “right” answer to mapping the diversity of human experience.

Method: Proposes BLPO, a bi-level prompt optimization framework that converts images into textual representations while preserving evaluation-relevant visual cues. Uses bi-level optimization to jointly refine the judge prompt and the image-to-text (I2T) prompt to maintain fidelity under limited context budgets.

Result: Experiments on four datasets and three LLM judges demonstrate the effectiveness of the method in improving multimodal LLM-as-a-judge evaluation of AI-generated images.

Conclusion: BLPO provides an effective solution for auto prompt optimization in multimodal settings, overcoming context window limitations for better alignment with human judgments in evaluating AI-generated images.

Abstract: Large language models (LLMs) have become widely adopted as automated judges for evaluating AI-generated content. Despite their success, aligning LLM-based evaluations with human judgments remains challenging. While supervised fine-tuning on human-labeled data can improve alignment, it is costly and inflexible, requiring new training for each task or dataset. Recent progress in auto prompt optimization (APO) offers a more efficient alternative by automatically improving the instructions that guide LLM judges. However, existing APO methods primarily target text-only evaluations and remain underexplored in multimodal settings. In this work, we study auto prompt optimization for multimodal LLM-as-a-judge, particularly for evaluating AI-generated images. We identify a key bottleneck: multimodal models can only process a limited number of visual examples due to context window constraints, which hinders effective trial-and-error prompt refinement. To overcome this, we propose BLPO, a bi-level prompt optimization framework that converts images into textual representations while preserving evaluation-relevant visual cues. Our bi-level optimization approach jointly refines the judge prompt and the I2T prompt to maintain fidelity under limited context budgets. Experiments on four datasets and three LLM judges demonstrate the effectiveness of our method.

Method: 1) Analyze real-world noise and categorize into user-noise and tool-noise. 2) Develop automated pipeline to inject controllable noise into existing benchmarks while preserving task solvability. 3) Evaluate diverse models across architectures and parameter scales.

Result: Reveals consistent performance variations under different noise conditions, showing current agentic models are sensitive to realistic environmental perturbations.

Conclusion: AgentNoiseBench provides systematic evaluation of agent robustness, highlighting the need for more noise-resistant models and better training/evaluation paradigms for real-world deployment.

Abstract: Recent advances in large language models have enabled LLM-based agents to achieve strong performance on a variety of benchmarks. However, their performance in real-world deployments often that observed on benchmark settings, especially in complex and imperfect environments. This discrepancy largely arises because prevailing training and evaluation paradigms are typically built on idealized assumptions, overlooking the inherent stochasticity and noise present in real-world interactions. To bridge this gap, we introduce AgentNoiseBench, a framework for systematically evaluating the robustness of agentic models under noisy environments. We first conduct an in-depth analysis of biases and uncertainties in real-world scenarios and categorize environmental noise into two primary types: user-noise and tool-noise. Building on this analysis, we develop an automated pipeline that injects controllable noise into existing agent-centric benchmarks while preserving task solvability. Leveraging this pipeline, we perform extensive evaluations across a wide range of models with diverse architectures and parameter scales. Our results reveal consistent performance variations under different noise conditions, highlighting the sensitivity of current agentic models to realistic environmental perturbations.

Method: Proposes DTAPP-IICR: 1) Prioritizes missions by urgency, 2) Uses SFIPP-ST (4D single-agent planner) for roundtrip trajectories with temporal NFZs and soft conflict constraints, 3) Iterative Large Neighborhood Search guided by geometric conflict graph resolves residual conflicts, 4) Completeness-preserving directional pruning accelerates 3D search.

Result: Achieves near-100% success with fleets up to 1,000 UAVs, 50% runtime reduction from pruning, outperforms batch Enhanced Conflict-Based Search. Scales successfully in realistic city-scale operations where other priority-based methods fail.

Conclusion: DTAPP-IICR provides a practical and scalable solution for preflight planning in dense, dynamic urban airspace, addressing key UTM challenges with temporal constraints and large fleet coordination.

Abstract: Preflight planning for large-scale Unmanned Aerial Vehicle (UAV) fleets in dynamic, shared airspace presents significant challenges, including temporal No-Fly Zones (NFZs), heterogeneous vehicle profiles, and strict delivery deadlines. While Multi-Agent Path Finding (MAPF) provides a formal framework, existing methods often lack the scalability and flexibility required for real-world Unmanned Traffic Management (UTM). We propose DTAPP-IICR: a Delivery-Time Aware Prioritized Planning method with Incremental and Iterative Conflict Resolution. Our framework first generates an initial solution by prioritizing missions based on urgency. Secondly, it computes roundtrip trajectories using SFIPP-ST, a novel 4D single-agent planner (Safe Flight Interval Path Planning with Soft and Temporal Constraints). SFIPP-ST handles heterogeneous UAVs, strictly enforces temporal NFZs, and models inter-agent conflicts as soft constraints. Subsequently, an iterative Large Neighborhood Search, guided by a geometric conflict graph, efficiently resolves any residual conflicts. A completeness-preserving directional pruning technique further accelerates the 3D search. On benchmarks with temporal NFZs, DTAPP-IICR achieves near-100% success with fleets of up to 1,000 UAVs and gains up to 50% runtime reduction from pruning, outperforming batch Enhanced Conflict-Based Search in the UTM context. Scaling successfully in realistic city-scale operations where other priority-based methods fail even at moderate deployments, DTAPP-IICR is positioned as a practical and scalable solution for preflight planning in dense, dynamic urban airspace.

Method: BAO combines behavior enhancement (enriching proactive reasoning and information-gathering capabilities) with behavior regularization (suppressing inefficient/redundant interactions and aligning with user expectations).

Result: BAO substantially outperforms proactive agentic RL baselines on multiple tasks from UserRL benchmark suite and achieves comparable or superior performance to commercial LLM agents.

Conclusion: BAO effectively trains proactive, user-aligned LLM agents for complex multi-turn scenarios by balancing task efficiency with user engagement.

Abstract: Proactive large language model (LLM) agents aim to actively plan, query, and interact over multiple turns, enabling efficient task completion beyond passive instruction following and making them essential for real-world, user-centric applications. Agentic reinforcement learning (RL) has recently emerged as a promising solution for training such agents in multi-turn settings, allowing interaction strategies to be learned from feedback. However, existing pipelines face a critical challenge in balancing task performance with user engagement, as passive agents can not efficiently adapt to users’ intentions while overuse of human feedback reduces their satisfaction. To address this trade-off, we propose BAO, an agentic RL framework that combines behavior enhancement to enrich proactive reasoning and information-gathering capabilities with behavior regularization to suppress inefficient or redundant interactions and align agent behavior with user expectations. We evaluate BAO on multiple tasks from the UserRL benchmark suite, and demonstrate that it substantially outperforms proactive agentic RL baselines while achieving comparable or even superior performance to commercial LLM agents, highlighting its effectiveness for training proactive, user-aligned LLM agents in complex multi-turn scenarios. Our website: https://proactive-agentic-rl.github.io/.

Method: Introduces ReplicatorBench with human-verified replicable and non-replicable research claims from social/behavioral sciences, evaluating agents across three stages: (1) extraction/retrieval of replication data, (2) design/execution of computational experiments, (3) interpretation of results. Also develops ReplicatorAgent framework with tools like web search and sandboxed environments.

Result: Evaluation across four LLMs shows current agents can design/execute experiments effectively but struggle with retrieving necessary resources like new data for replication. Different programming languages and code access levels affect performance.

Conclusion: ReplicatorBench provides comprehensive evaluation of AI agents’ research replication capabilities, revealing current limitations in resource retrieval while demonstrating strengths in experimental design/execution.

Abstract: The literature has witnessed an emerging interest in AI agents for automated assessment of scientific papers. Existing benchmarks focus primarily on the computational aspect of this task, testing agents’ ability to reproduce or replicate research outcomes when having access to the code and data. This setting, while foundational, (1) fails to capture the inconsistent availability of new data for replication as opposed to reproduction, and (2) lacks ground-truth diversity by focusing only on reproducible papers, thereby failing to evaluate an agent’s ability to identify non-replicable research. Furthermore, most benchmarks only evaluate outcomes rather than the replication process. In response, we introduce ReplicatorBench, an end-to-end benchmark, including human-verified replicable and non-replicable research claims in social and behavioral sciences for evaluating AI agents in research replication across three stages: (1) extraction and retrieval of replication data; (2) design and execution of computational experiments; and (3) interpretation of results, allowing a test of AI agents’ capability to mimic the activities of human replicators in real world. To set a baseline of AI agents’ capability, we develop ReplicatorAgent, an agentic framework equipped with necessary tools, like web search and iterative interaction with sandboxed environments, to accomplish tasks in ReplicatorBench. We evaluate ReplicatorAgent across four underlying large language models (LLMs), as well as different design choices of programming language and levels of code access. Our findings reveal that while current LLM agents are capable of effectively designing and executing computational experiments, they struggle with retrieving resources, such as new data, necessary to replicate a claim. All code and data are publicly available at https://github.com/CenterForOpenScience/llm-benchmarking.

Method: Extends masked joint embedding prediction (JEPA) from image patches to object-centric representations. Uses object-level masking that requires inferring an object’s state from other objects, creating latent interventions with counterfactual-like effects and preventing shortcut solutions.

Result: C-JEPA achieves ~20% absolute improvement in counterfactual reasoning on visual question answering compared to same architecture without object-level masking. On agent control tasks, enables substantially more efficient planning using only 1% of total latent input features compared to patch-based world models while maintaining comparable performance.

Conclusion: Object-level masking in world models induces causal inductive biases via latent interventions, making interaction reasoning essential and leading to improved performance in visual reasoning and more efficient planning for control tasks.

Abstract: World models require robust relational understanding to support prediction, reasoning, and control. While object-centric representations provide a useful abstraction, they are not sufficient to capture interaction-dependent dynamics. We therefore propose C-JEPA, a simple and flexible object-centric world model that extends masked joint embedding prediction from image patches to object-centric representations. By applying object-level masking that requires an object’s state to be inferred from other objects, C-JEPA induces latent interventions with counterfactual-like effects and prevents shortcut solutions, making interaction reasoning essential. Empirically, C-JEPA leads to consistent gains in visual question answering, with an absolute improvement of about 20% in counterfactual reasoning compared to the same architecture without object-level masking. On agent control tasks, C-JEPA enables substantially more efficient planning by using only 1% of the total latent input features required by patch-based world models, while achieving comparable performance. Finally, we provide a formal analysis demonstrating that object-level masking induces a causal inductive bias via latent interventions. Our code is available at https://github.com/galilai-group/cjepa.

Method: GHOST (Grouped Hidden-state Output-aware Selection and Truncation) uses forward-pass statistics to approximate control-theoretic balanced truncation, jointly measuring controllability and observability without requiring backpropagation.

Result: Achieves 50% state-dimension reduction with approximately 1 perplexity point increase on WikiText-2 across models ranging from 130M to 2.7B parameters, rivaling gradient-based methods without backpropagation.

Conclusion: GHOST provides an efficient structured pruning framework that addresses Mamba2’s inference bottlenecks while maintaining model fidelity, offering a practical solution for deployment of large sequence models.

Abstract: While Mamba2’s expanded state dimension enhances temporal modeling, it incurs substantial inference overhead that saturates bandwidth during autoregressive generation. Standard pruning methods fail to address this bottleneck: unstructured sparsity leaves activations dense, magnitude-based selection ignores runtime dynamics, and gradient-based methods impose prohibitive costs. We introduce GHOST (Grouped Hidden-state Output-aware Selection and Truncation), a structured pruning framework that approximates control-theoretic balanced truncation using only forward-pass statistics. By jointly measuring controllability and observability, GHOST rivals the fidelity of gradient-based methods without requiring backpropagation. As a highlight, on models ranging from 130M to 2.7B parameters, our approach achieves a 50% state-dimension reduction with approximately 1 perplexity point increase on WikiText-2. Code is available at https://anonymous.4open.science/r/mamba2_ghost-7BCB/.

Method: TRACER combines content-aware surprisal with situational-awareness signals, semantic/lexical repetition, and tool-grounded coherence gaps, aggregating them using a tail-focused risk functional with MAX-composite step risk to surface decisive anomalies.

Result: TRACER improves AUROC by up to 37.1% and AUARC by up to 55% over baselines on the τ²-bench for predicting task failure and selective task execution, enabling earlier and more accurate detection of uncertainty in complex conversational tool-use settings.

Conclusion: TRACER provides an effective trajectory-level uncertainty metric for AI agents in multi-turn tool-using interactions, addressing limitations of single-shot uncertainty methods and improving failure detection in complex conversational settings.

Abstract: Estimating uncertainty for AI agents in real-world multi-turn tool-using interaction with humans is difficult because failures are often triggered by sparse critical episodes (e.g., looping, incoherent tool use, or user-agent miscoordination) even when local generation appears confident. Existing uncertainty proxies focus on single-shot text generation and therefore miss these trajectory-level breakdown signals. We introduce TRACER, a trajectory-level uncertainty metric for dual-control Tool-Agent-User interaction. TRACER combines content-aware surprisal with situational-awareness signals, semantic and lexical repetition, and tool-grounded coherence gaps, and aggregates them using a tail-focused risk functional with a MAX-composite step risk to surface decisive anomalies. We evaluate TRACER on $τ^2$-bench by predicting task failure and selective task execution. To this end, TRACER improves AUROC by up to 37.1% and AUARC by up to 55% over baselines, enabling earlier and more accurate detection of uncertainty in complex conversational tool-use settings. Our code and benchmark are available at https://github.com/sinatayebati/agent-tracer.

Method: Analyze cross-modal attention connectivity in MLLMs, identify high-connectivity anchor tokens (≈15% of tokens) that ground reasoning in images, then propose Anchor-Token Reinforcement Learning (AT-RL) that selectively reinforces these tokens using graph-based clustering of attention topology.

Result: AT-RL introduces only 1.2% overhead but enables a 32B model to surpass 72B-Instruct baseline on MathVista (80.2), with consistent gains across STEM, video, and general tasks. Training only on low-connectivity tokens causes severe degradation.

Conclusion: Reasoning quality in multimodal RL depends on precise credit assignment to visual anchor tokens rather than token quantity. AT-RL effectively leverages this insight for efficient multimodal reinforcement learning.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet how visual evidence is integrated during reasoning remains poorly understood. We explore multimodal RLVR through the lens of cross-modal attention connectivity and find that only a small fraction of tokens (approximately 15%) exhibit strong visual-textual coupling. These high-connectivity tokens act as anchors that ground reasoning in the image, while the majority follow linguistic patterns. During RLVR training, credit assignment naturally concentrates on these anchors, sharpening their visual grounding over time. Building on this insight, we propose Anchor-Token Reinforcement Learning (AT-RL), a lightweight framework that selectively reinforces high-connectivity tokens via graph-based clustering of attention topology. Evaluated across the series (3B-32B), AT-RL introduces only 1.2% overhead yet enables the 32B model to surpass the 72B-Instruct baseline on MathVista (80.2), with consistent gains observed across STEM, video and general tasks. Conversely, training solely on low-connectivity tokens causes severe degradation, confirming that effective multimodal RL hinges on precise credit assignment to visual anchors. Our work reveals that reasoning quality is governed not by token quantity but by the fidelity of cross-modal anchoring.

Method: Introduces AgentLeak benchmark with 1,000 scenarios across healthcare, finance, legal, and corporate domains, a 32-class attack taxonomy, and three-tier detection pipeline. Tests GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, Mistral Large, and Llama 3.3 70B across 4,979 traces to measure leakage across different channels.

Result: Multi-agent configurations reduce per-channel output leakage (27.2% vs 43.2% single-agent) but introduce unmonitored internal channels raising total system exposure to 68.9%. Inter-agent messages leak at 68.8% vs 27.2% on output channel. Output-only audits miss 41.7% of violations. Claude 3.5 Sonnet has lowest leakage rates (3.3% external, 28.1% internal).

Conclusion: Inter-agent communication is the primary vulnerability in multi-agent LLM systems, requiring coordination frameworks with internal-channel privacy protections and privacy controls on inter-agent communication.

Abstract: Multi-agent Large Language Model (LLM) systems create privacy risks that current benchmarks cannot measure. When agents coordinate on tasks, sensitive data passes through inter-agent messages, shared memory, and tool arguments; pathways that output-only audits never inspect. We introduce AgentLeak, to the best of our knowledge the first full-stack benchmark for privacy leakage covering internal channels, spanning 1,000 scenarios across healthcare, finance, legal, and corporate domains, paired with a 32-class attack taxonomy and three-tier detection pipeline. Testing GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, Mistral Large, and Llama 3.3 70B across 4,979 traces reveals that multi-agent configurations reduce per-channel output leakage (C1: 27.2% vs 43.2% in single-agent) but introduce unmonitored internal channels that raise total system exposure to 68.9% (OR-aggregated across C1, C2, C5). Internal channels account for most of this gap: inter-agent messages (C2) leak at 68.8%, compared to 27.2% on C1 (output channel). This means that output-only audits miss 41.7% of violations. Claude 3.5 Sonnet, which emphasizes safety alignment in its design, achieves the lowest leakage rates on both external (3.3%) and internal (28.1%) channels, suggesting that model-level safety training may transfer to internal channel protection. Across all five models and four domains, the pattern C2 > C1 holds consistently, confirming that inter-agent communication is the primary vulnerability. These findings underscore the need for coordination frameworks that incorporate internal-channel privacy protections and enforce privacy controls on inter-agent communication.

Method: Proposes a human-inspired continuous learning framework that unifies reasoning, action, reflection, and verification within a sequential reasoning model enhanced by parallel learning. The framework treats internal thinking processes as learning objects, records reasoning trajectories and environmental interactions as structured learning material, and supports hierarchical learning-to-learn mechanisms.

Result: Experimental results on a temperature sensor abnormality detection task show that incorporating internal-process learning reduces average runtime by 23.9%.

Conclusion: The framework enables AI systems to progressively evolve their internal cognitive architecture while preserving operational stability, representing a step toward more adaptive and efficient AI systems.

Abstract: Learning internal reasoning processes is crucial for developing AI systems capable of sustained adaptation in dynamic real-world environments. However, most existing approaches primarily emphasize learning task-specific outputs or static knowledge representations, while overlooking the continuous refinement of internal reasoning structures, action scheduling policies, and learning mechanisms themselves. In this paper, we propose a human-inspired continuous learning framework that unifies reasoning, action, reflection, and verification within a sequential reasoning model enhanced by parallel learning. The framework explicitly treats internal thinking processes as primary learning objects. It systematically records internal reasoning trajectories and environmental interactions as structured learning material, enabling the system to optimize not only task-level content but also the organization, scheduling, and evolution of reasoning activities. This design realizes learning alongside processing, allowing cognitive structures to improve during execution. Furthermore, the framework supports controlled replacement of predefined logic with learned procedures and introduces a hierarchical learning-to-learn mechanism that jointly adapts task-level parameters and learning strategies. As a result, the system progressively evolves its internal cognitive architecture while preserving operational stability. Experimental results on a temperature sensor abnormality detection task show that incorporating internal-process learning reduces average runtime by 23.9%.

Method: Integrates Multi-Agent Systems (MAS), Retrieval-Augmented Generation (RAG), and Model Context Protocol (MCP) to automate the entire causal inference pipeline from data cleaning to report generation through natural language interaction.

Result: Users can upload datasets and ask questions in natural language to receive rigorous, interactive analysis reports with visualizations, significantly lowering the barrier to entry while maintaining process rigor and interpretability.

Conclusion: CausalAgent represents a novel user-centered human-AI collaboration paradigm that explicitly models analysis workflows, making causal inference more accessible without sacrificing methodological rigor.

Abstract: Causal inference holds immense value in fields such as healthcare, economics, and social sciences. However, traditional causal analysis workflows impose significant technical barriers, requiring researchers to possess dual backgrounds in statistics and computer science, while manually selecting algorithms, handling data quality issues, and interpreting complex results. To address these challenges, we propose CausalAgent, a conversational multi-agent system for end-to-end causal inference. The system innovatively integrates Multi-Agent Systems (MAS), Retrieval-Augmented Generation (RAG), and the Model Context Protocol (MCP) to achieve automation from data cleaning and causal structure learning to bias correction and report generation through natural language interaction. Users need only upload a dataset and pose questions in natural language to receive a rigorous, interactive analysis report. As a novel user-centered human-AI collaboration paradigm, CausalAgent explicitly models the analysis workflow. By leveraging interactive visualizations, it significantly lowers the barrier to entry for causal analysis while ensuring the rigor and interpretability of the process.

Method: INTENT uses an inference-time planning framework with intention-aware hierarchical world models to anticipate future tool usage, risk-calibrated costs, and guide decisions online in sequential decision making with priced and stochastic tool executions.

Result: On cost-augmented StableToolBench, INTENT strictly enforces hard budget feasibility while substantially improving task success over baselines, and remains robust under dynamic market shifts like tool price changes and varying budgets.

Conclusion: INTENT provides an effective solution for budget-constrained tool-augmented agents, addressing the challenges of massive state-action spaces and high variance through inference-time planning with hierarchical world models.

Abstract: We study budget-constrained tool-augmented agents, where a large language model must solve multi-step tasks by invoking external tools under a strict monetary budget. We formalize this setting as sequential decision making in context space with priced and stochastic tool executions, making direct planning intractable due to massive state-action spaces, high variance of outcomes and prohibitive exploration cost. To address these challenges, we propose INTENT, an inference-time planning framework that leverages an intention-aware hierarchical world model to anticipate future tool usage, risk-calibrated cost, and guide decisions online. Across cost-augmented StableToolBench, INTENT strictly enforces hard budget feasibility while substantially improving task success over baselines, and remains robust under dynamic market shifts such as tool price changes and varying budgets.

Method: SemaPop derives high-level persona representations from individual survey records using LLMs, then uses these as semantic conditioning signals in a WGAN-GP framework. Marginal regularization enforces alignment with target population marginals, balancing semantic conditioning with statistical constraints.

Result: SemaPop-GAN achieves improved generative performance with closer alignment to target marginal and joint distributions while maintaining sample-level feasibility and diversity under semantic conditioning. Ablation studies confirm the importance of semantic persona conditioning and architectural choices.

Conclusion: The approach enables controllable and interpretable population synthesis through effective semantic-statistical fusion, providing a modular foundation for generative population projection systems that integrate individual-level behavioral semantics with population-level statistical constraints.

Abstract: Population synthesis is a critical component of individual-level socio-economic simulation, yet remains challenging due to the need to jointly represent statistical structure and latent behavioral semantics. Existing population synthesis approaches predominantly rely on structured attributes and statistical constraints, leaving a gap in semantic-conditioned population generation that can capture abstract behavioral patterns implicitly in survey data. This study proposes SemaPop, a semantic-statistical population synthesis model that integrates large language models (LLMs) with generative population modeling. SemaPop derives high-level persona representations from individual survey records and incorporates them as semantic conditioning signals for population generation, while marginal regularization is introduced to enforce alignment with target population marginals. In this study, the framework is instantiated using a Wasserstein GAN with gradient penalty (WGAN-GP) backbone, referred to as SemaPop-GAN. Extensive experiments demonstrate that SemaPop-GAN achieves improved generative performance, yielding closer alignment with target marginal and joint distributions while maintaining sample-level feasibility and diversity under semantic conditioning. Ablation studies further confirm the contribution of semantic persona conditioning and architectural design choices to balancing marginal consistency and structural realism. These results demonstrate that SemaPop-GAN enables controllable and interpretable population synthesis through effective semantic-statistical information fusion. SemaPop-GAN also provides a promising modular foundation for developing generative population projection systems that integrate individual-level behavioral semantics with population-level statistical constraints.

Method: Formulates agent configuration as query-wise decision problem and introduces ARC (Agentic Resource & Configuration learner), which learns a light-weight hierarchical policy using reinforcement learning to dynamically tailor configurations (workflows, tools, token budgets, prompts).

Result: Across multiple benchmarks spanning reasoning and tool-augmented question answering, the learned policy consistently outperforms strong hand-designed and other baselines, achieving up to 25% higher task accuracy while reducing token and runtime costs.

Conclusion: Learning per-query agent configurations is a powerful alternative to “one size fits all” designs, demonstrating that dynamic configuration can improve both performance and efficiency.

Abstract: Configuring LLM-based agent systems involves choosing workflows, tools, token budgets, and prompts from a large combinatorial design space, and is typically handled today by fixed large templates or hand-tuned heuristics. This leads to brittle behavior and unnecessary compute, since the same cumbersome configuration is often applied to both easy and hard input queries. We formulate agent configuration as a query-wise decision problem and introduce ARC (Agentic Resource & Configuration learner), which learns a light-weight hierarchical policy using reinforcement learning to dynamically tailor these configurations. Across multiple benchmarks spanning reasoning and tool-augmented question answering, the learned policy consistently outperforms strong hand-designed and other baselines, achieving up to 25% higher task accuracy while also reducing token and runtime costs. These results demonstrate that learning per-query agent configurations is a powerful alternative to “one size fits all” designs.

Method: The survey uses the Five Ws framework (who, what, when, where, why) to analyze multi-agent communication. It traces the evolution across three major paradigms: 1) Multi-Agent Reinforcement Learning (MARL) with hand-designed or learned protocols, 2) Emergent Language (EL) for structured/symbolic communication, and 3) LLM-based systems leveraging natural language priors for reasoning and planning.

Result: The survey provides a systematic analysis of communication design choices and trade-offs across different paradigms. It identifies key challenges: MARL protocols are often task-specific and hard to interpret; EL struggles with grounding, generalization, and scalability; LLMs offer natural language priors but face different limitations. The paper distills practical design patterns and open challenges.

Conclusion: Future multi-agent systems should combine learning, language, and control for scalable and interpretable collaboration. Hybrid approaches that leverage strengths from MARL, EL, and LLM paradigms are needed to address current limitations and enable more effective multi-agent communication in complex, open-ended environments.

Abstract: Multi-agent sequential decision-making powers many real-world systems, from autonomous vehicles and robotics to collaborative AI assistants. In dynamic, partially observable environments, communication is often what reduces uncertainty and makes collaboration possible. This survey reviews multi-agent communication (MA-Comm) through the Five Ws: who communicates with whom, what is communicated, when communication occurs, and why communication is beneficial. This framing offers a clean way to connect ideas across otherwise separate research threads. We trace how communication approaches have evolved across three major paradigms. In Multi-Agent Reinforcement Learning (MARL), early methods used hand-designed or implicit protocols, followed by end-to-end learned communication optimized for reward and control. While successful, these protocols are frequently task-specific and hard to interpret, motivating work on Emergent Language (EL), where agents can develop more structured or symbolic communication through interaction. EL methods, however, still struggle with grounding, generalization, and scalability, which has fueled recent interest in large language models (LLMs) that bring natural language priors for reasoning, planning, and collaboration in more open-ended settings. Across MARL, EL, and LLM-based systems, we highlight how different choices shape communication design, where the main trade-offs lie, and what remains unsolved. We distill practical design patterns and open challenges to support future hybrid systems that combine learning, language, and control for scalable and interpretable multi-agent collaboration.

Method: MAPLE comprises three components: (1) MAPLE-bench - first benchmark explicitly annotating minimal signal combinations required per task; (2) MAPO - modality-aware policy optimization framework that stratifies batches by modality requirement to reduce gradient variance; (3) Adaptive weighting and curriculum scheduling that balances and prioritizes harder signal combinations.

Result: MAPLE narrows uni/multi-modal accuracy gaps by 30.24%, converges 3.18x faster, and maintains stability across all modality combinations under realistic reduced signal access. Systematic analysis establishes MAPO’s optimal training strategy across loss aggregation, clipping, sampling, and curriculum design.

Conclusion: MAPLE constitutes a complete recipe for deployment-ready multimodal RL post-training, addressing fundamental inefficiencies in current approaches and providing a systematic framework for modality-aware training that improves performance, convergence speed, and robustness.

Abstract: Multimodal language models now integrate text, audio, and video for unified reasoning. Yet existing RL post-training pipelines treat all input signals as equally relevant, ignoring which modalities each task actually requires. This modality-blind training inflates policy-gradient variance, slows convergence, and degrades robustness to real-world distribution shifts where signals may be missing, added, or reweighted. We introduce MAPLE, a complete modality-aware post-training and learning ecosystem comprising: (1) MAPLE-bench, the first benchmark explicitly annotating minimal signal combinations required per task; (2) MAPO, a modality-aware policy optimization framework that stratifies batches by modality requirement to reduce gradient variance from heterogeneous group advantages; (3) Adaptive weighting and curriculum scheduling that balances and prioritizes harder signal combinations. Systematic analysis across loss aggregation, clipping, sampling, and curriculum design establishes MAPO’s optimal training strategy. Adaptive weighting and curriculum focused learning further boost performance across signal combinations. MAPLE narrows uni/multi-modal accuracy gaps by 30.24%, converges 3.18x faster, and maintains stability across all modality combinations under realistic reduced signal access. MAPLE constitutes a complete recipe for deployment-ready multimodal RL post-training.

Method: Converts core single-cell analyses into step-by-step reasoning problems that LLMs must solve while directly inspecting RNA-seq data and using bioinformatics tools, with iterative reasoning and revision capabilities.

Result: Iterative omics-native reasoning improved average accuracy by 11% for cell-type annotation and reduced trajectory graph-edit distance by 30% compared to one-shot prompting, while generating transparent reasoning traces.

Conclusion: scPilot enables auditable, interpretable, and diagnostically informative single-cell analyses by grounding LLMs in raw omics data, advancing towards more transparent biological reasoning systems.

Abstract: We present scPilot, the first systematic framework to practice omics-native reasoning: a large language model (LLM) converses in natural language while directly inspecting single-cell RNA-seq data and on-demand bioinformatics tools. scPilot converts core single-cell analyses, i.e., cell-type annotation, developmental-trajectory reconstruction, and transcription-factor targeting, into step-by-step reasoning problems that the model must solve, justify, and, when needed, revise with new evidence. To measure progress, we release scBench, a suite of 9 expertly curated datasets and graders that faithfully evaluate the omics-native reasoning capability of scPilot w.r.t various LLMs. Experiments with o1 show that iterative omics-native reasoning lifts average accuracy by 11% for cell-type annotation and Gemini-2.5-Pro cuts trajectory graph-edit distance by 30% versus one-shot prompting, while generating transparent reasoning traces explain marker gene ambiguity and regulatory logic. By grounding LLMs in raw omics data, scPilot enables auditable, interpretable, and diagnostically informative single-cell analyses. Code, data, and package are available at https://github.com/maitrix-org/scPilot

Method: Conducted 3,000 agent runs across three models (Llama 3.1 70B, GPT-4o, Claude Sonnet 4.5) on HotpotQA using ReAct-style agents, analyzing action sequence variance and its correlation with accuracy.

Result: Agents produce 2.0-4.2 distinct action sequences per 10 runs on average. Tasks with consistent behavior (≤2 unique paths) achieve 80-92% accuracy, while highly inconsistent tasks (≥6 unique paths) achieve only 25-60%. 69% of divergence occurs at step 2 (first search query).

Conclusion: Behavioral inconsistency in LLM agents predicts failure, with most divergence happening early. Monitoring consistency during execution could enable early error detection and improve agent reliability.

Abstract: Run the same LLM agent on the same task twice: do you get the same behavior? We find the answer is often no. In a study of 3,000 agent runs across three models (Llama 3.1 70B, GPT-4o, and Claude Sonnet 4.5) on HotpotQA, we observe that ReAct-style agents produce 2.0–4.2 distinct action sequences per 10 runs on average, even with identical inputs. More importantly, this variance predicts failure: tasks with consistent behavior ($\leq$2 unique paths) achieve 80–92% accuracy, while highly inconsistent tasks ($\geq$6 unique paths) achieve only 25–60%, a 32–55 percentage point gap depending on model. We trace variance to early decisions: 69% of divergence occurs at step 2, the first search query. Our results suggest that monitoring behavioral consistency during execution could enable early error detection and improve agent reliability.

Method: Proposes NMIPS framework using multifactorial optimization to simultaneously discover analytical solutions for PDE families, with an affine transfer method that transfers learned mathematical structures among related PDEs to avoid solving each from scratch.

Result: Experimental results show promising improvements over baselines, achieving up to ~35.7% increase in accuracy while providing interpretable analytical solutions.

Conclusion: The proposed framework addresses both computational efficiency and interpretability limitations of existing PDE solving methods for PDE families.

Abstract: Solving Partial Differential Equations (PDEs) is fundamental to numerous scientific and engineering disciplines. A common challenge arises from solving the PDE families, which are characterized by sharing an identical mathematical structure but varying in specific parameters. Traditional numerical methods, such as the finite element method, need to independently solve each instance within a PDE family, which incurs massive computational cost. On the other hand, while recent advancements in machine learning PDE solvers offer impressive computational speed and accuracy, their inherent ``black-box" nature presents a considerable limitation. These methods primarily yield numerical approximations, thereby lacking the crucial interpretability provided by analytical expressions, which are essential for deeper scientific insight. To address these limitations, we propose a neuro-assisted multitasking symbolic PDE solver framework for PDE family solving, dubbed NMIPS. In particular, we employ multifactorial optimization to simultaneously discover the analytical solutions of PDEs. To enhance computational efficiency, we devise an affine transfer method by transferring learned mathematical structures among PDEs in a family, avoiding solving each PDE from scratch. Experimental results across multiple cases demonstrate promising improvements over existing baselines, achieving up to a $\sim$35.7% increase in accuracy while providing interpretable analytical solutions.

Method: Three-component framework: (1) MathSpatial-Bench - 2K benchmark problems across 3 categories and 11 subtypes to isolate reasoning from perception; (2) MathSpatial-Corpus - 8K training dataset with verified solutions; (3) MathSpatial-SRT - structured reasoning trace modeling with three atomic operations (Correlate, Constrain, Infer).

Result: Fine-tuning Qwen2.5-VL-7B on MathSpatial achieves competitive accuracy while reducing tokens by 25%. The framework enables precise measurement and understanding of spatial reasoning capabilities in MLLMs.

Conclusion: MathSpatial provides the first large-scale resource that disentangles perception from reasoning, enabling systematic evaluation and improvement of mathematical spatial reasoning in MLLMs, addressing a critical gap in current multimodal capabilities.

Abstract: Multimodal large language models (MLLMs) have achieved strong performance on perception-oriented tasks, yet their ability to perform mathematical spatial reasoning, defined as the capacity to parse and manipulate two- and three-dimensional relations, remains unclear. Humans easily solve textbook-style spatial reasoning problems with over 95% accuracy, but we find that most leading MLLMs fail to reach even 60% on the same tasks. This striking gap highlights spatial reasoning as a fundamental weakness of current models. To investigate this gap, we present MathSpatial, a unified framework for evaluating and improving spatial reasoning in MLLMs. MathSpatial includes three complementary components: (i) MathSpatial-Bench, a benchmark of 2K problems across three categories and eleven subtypes, designed to isolate reasoning difficulty from perceptual noise; (ii) MathSpatial-Corpus, a training dataset of 8K additional problems with verified solutions; and (iii) MathSpatial-SRT, which models reasoning as structured traces composed of three atomic operations–Correlate, Constrain, and Infer. Experiments show that fine-tuning Qwen2.5-VL-7B on MathSpatial achieves competitive accuracy while reducing tokens by 25%. MathSpatial provides the first large-scale resource that disentangles perception from reasoning, enabling precise measurement and comprehensive understanding of mathematical spatial reasoning in MLLMs.

Method: Proposes a holistic multi-dimensional medical alignment matrix with four categories: fundamental capabilities, expert knowledge, online feedback, and format specifications. Each category establishes a closed loop where observable metrics inform attributable diagnosis, which drives optimizable rewards. Also introduces a unified optimization mechanism with Reference-Frozen Normalization to align reward scales and Tri-Factor Adaptive Dynamic Weighting for collaborative optimization.

Result: Experimental results demonstrate effectiveness in real-world medical scenario evaluations, establishing a new paradigm for complex alignment in vertical domains.

Conclusion: The proposed paradigm addresses fundamental limitations of existing RL alignment methods for medical applications and provides a robust framework for medical LLM alignment that can handle heterogeneous reward signals and complex clinical contexts.

Abstract: While reinforcement learning for large language model alignment has progressed rapidly in recent years, transferring these paradigms to high-stakes medical question answering reveals a fundamental paradigm mismatch. Reinforcement Learning from Human Feedback relies on preference annotations that are prohibitively expensive and often fail to reflect the absolute correctness of medical facts. Reinforcement Learning from Verifiable Rewards lacks effective automatic verifiers and struggles to handle complex clinical contexts. Meanwhile, medical alignment requires the simultaneous optimization of correctness, safety, and compliance, yet multi-objective heterogeneous reward signals are prone to scale mismatch and optimization conflicts.To address these challenges, we propose a robust medical alignment paradigm. We first construct a holistic multi-dimensional medical alignment matrix that decomposes alignment objectives into four categories: fundamental capabilities, expert knowledge, online feedback, and format specifications. Within each category, we establish a closed loop of where observable metrics inform attributable diagnosis, which in turn drives optimizable rewards, thereby providing fine-grained, high-resolution supervision signals for subsequent iterative optimization. To resolve gradient domination and optimization instability problem caused by heterogeneous signals, we further propose a unified optimization mechanism. This mechanism employs Reference-Frozen Normalization to align reward scales and implements a Tri-Factor Adaptive Dynamic Weighting strategy to achieve collaborative optimization that is weakness-oriented, risk-prioritized, and redundancy-reducing. Experimental results demonstrate the effectiveness of our proposed paradigm in real-world medical scenario evaluations, establishing a new paradigm for complex alignment in vertical domains.

Method: PhyNiKCE decouples neural planning from symbolic validation using a Symbolic Knowledge Engine that treats simulation setup as a Constraint Satisfaction Problem, with a Deterministic RAG Engine for specialized retrieval of solvers, turbulence models, and boundary conditions.

Result: 96% relative improvement over state-of-the-art baselines in OpenFOAM experiments, 59% reduction in autonomous self-correction loops, and 17% lower LLM token consumption.

Conclusion: Decoupling neural generation from symbolic constraint enforcement significantly enhances robustness and efficiency for trustworthy AI in industrial automation, with potential applications beyond CFD.

Abstract: The deployment of autonomous agents for Computational Fluid Dynamics (CFD), is critically limited by the probabilistic nature of Large Language Models (LLMs), which struggle to enforce the strict conservation laws and numerical stability required for physics-based simulations. Reliance on purely semantic Retrieval Augmented Generation (RAG) often leads to “context poisoning,” where agents generate linguistically plausible but physically invalid configurations due to a fundamental Semantic-Physical Disconnect. To bridge this gap, this work introduces PhyNiKCE (Physical and Numerical Knowledgeable Context Engineering), a neurosymbolic agentic framework for trustworthy engineering. Unlike standard black-box agents, PhyNiKCE decouples neural planning from symbolic validation. It employs a Symbolic Knowledge Engine that treats simulation setup as a Constraint Satisfaction Problem, rigidly enforcing physical constraints via a Deterministic RAG Engine with specialized retrieval strategies for solvers, turbulence models, and boundary conditions. Validated through rigorous OpenFOAM experiments on practical, non-tutorial CFD tasks using Gemini-2.5-Pro/Flash, PhyNiKCE demonstrates a 96% relative improvement over state-of-the-art baselines. Furthermore, by replacing trial-and-error with knowledge-driven initialization, the framework reduced autonomous self-correction loops by 59% while simultaneously lowering LLM token consumption by 17%. These results demonstrate that decoupling neural generation from symbolic constraint enforcement significantly enhances robustness and efficiency. While validated on CFD, this architecture offers a scalable, auditable paradigm for Trustworthy Artificial Intelligence in broader industrial automation.

Method: Developed Benchmark Health Index (BHI) with three orthogonal axes: Capability Discrimination (separates performance beyond noise), Anti-Saturation (estimates remaining headroom), and Impact (quantifies adoption breadth). Analyzed 106 benchmarks from 91 model reports.

Result: Systematically characterized the evaluation landscape, providing first framework to quantify benchmark health at macro level for principled benchmark selection.

Conclusion: BHI enables dynamic lifecycle management for next-generation evaluation protocols by providing data-driven benchmark auditing.

Abstract: Large Language Models (LLMs) are advancing rapidly, yet the benchmarks used to measure this progress are becoming increasingly unreliable. Score inflation and selective reporting have eroded the authority of standard benchmarks, leaving the community uncertain about which evaluation results remain trustworthy. We introduce the Benchmark Health Index (BHI), a pure data-driven framework for auditing evaluation sets along three orthogonal and complementary axes: (1) Capability Discrimination, measuring how sharply a benchmark separates model performance beyond noise; (2) Anti-Saturation, estimating remaining headroom before ceiling effects erode resolution and thus the benchmark’s expected longevity; and (3) Impact, quantifying influence across academic and industrial ecosystems via adoption breadth and practice-shaping power. By distilling 106 validated benchmarks from the technical reports of 91 representative models in 2025, we systematically characterize the evaluation landscape. BHI is the first framework to quantify benchmark health at a macro level, providing a principled basis for benchmark selection and enabling dynamic lifecycle management for next-generation evaluation protocols.

Method: The authors propose a three-layer architecture with Epistemic Regret Minimization (ERM) as a belief revision objective. Key contributions include: (1) Physical Grounding Theorem linking actions to do-operations, (2) ERM as a causal belief revision operator satisfying AGM postulates, and (3) a failure mode taxonomy with domain-independent guards for cross-domain transfer.

Result: Experiments on 1,360 causal trap scenarios across six frontier LLMs show Rung Collapse persists even in reasoning-enhanced models (3.7% for GPT-5.2). The paper reveals inverse scaling in steerability where advanced models resist generic correction, and demonstrates that targeted ERM feedback recovers 53-59% of entrenched errors where outcome-level feedback fails.

Conclusion: The paper provides a formal causal framework for understanding and addressing reasoning failures in autoregressive models. ERM offers a principled approach to prevent entrenchment in flawed causal reasoning while maintaining asymptotic recovery of true interventional distributions with finite-sample guarantees.

Abstract: Machine learning systems that are “right for the wrong reasons” achieve high performance through shortcuts that collapse under distributional shift. We show this pathology has a precise causal origin: autoregressive training provides no gradient signal to distinguish association P(Y|X) from intervention P(Y|do(X)), a failure we formalize as Rung Collapse. When outcome-based learning reinforces correct answers obtained through incorrect causal models, the agent becomes entrenched in flawed reasoning, a phenomenon we term Aleatoric Entrenchment. We propose Epistemic Regret Minimization (ERM), a belief revision objective that penalizes errors in causal reasoning independently of task success, and embed it within a three-layer architecture with three contributions grounded in knowledge representation: (1) a Physical Grounding Theorem proving that actions satisfying actuator independence implement valid do-operations, bridging action languages and do-calculus; (2) ERM as a causal belief revision operator satisfying AGM postulates, preventing entrenchment even when the agent succeeds for the wrong reasons; and (3) a failure mode taxonomy that classifies recurring reasoning errors and injects domain-independent guards, enabling cross-domain transfer. We prove asymptotic recovery of the true interventional distribution with finite-sample bounds. Experiments on 1,360 causal trap scenarios across six frontier LLMs reveal that Rung Collapse persists even in reasoning-enhanced models (3.7% for GPT-5.2), that steerability exhibits inverse scaling where advanced models resist generic correction, and that targeted ERM feedback recovers 53-59% of entrenched errors where outcome-level feedback fails.

Method: Proposes a Vector-to-Graph (V2G) pipeline that converts CAD diagrams into property graphs where nodes represent components and edges encode connectivity, making structural dependencies explicit and machine-auditable.

Result: On a diagnostic benchmark of electrical compliance checks, V2G yields large accuracy gains across all error categories, while leading MLLMs remain near chance level.

Conclusion: Pixel-based MLLMs are systemically inadequate for structural reasoning; structure-aware representations provide a reliable path toward practical deployment of multimodal AI in engineering domains.

Abstract: Multimodal Large Language Models (MLLMs) have shown remarkable progress in visual understanding, yet they suffer from a critical limitation: structural blindness. Even state-of-the-art models fail to capture topology and symbolic logic in engineering schematics, as their pixel-driven paradigm discards the explicit vector-defined relations needed for reasoning. To overcome this, we propose a Vector-to-Graph (V2G) pipeline that converts CAD diagrams into property graphs where nodes represent components and edges encode connectivity, making structural dependencies explicit and machine-auditable. On a diagnostic benchmark of electrical compliance checks, V2G yields large accuracy gains across all error categories, while leading MLLMs remain near chance level. These results highlight the systemic inadequacy of pixel-based methods and demonstrate that structure-aware representations provide a reliable path toward practical deployment of multimodal AI in engineering domains. To facilitate further research, we release our benchmark and implementation at https://github.com/gm-embodied/V2G-Audit.

Method: ThinkRouter is an inference-time confidence-aware routing mechanism that routes thinking to discrete token space when model confidence is low (for explicit reasoning), and to latent space when confidence is high (for efficient latent reasoning). This dynamic switching optimizes both accuracy and efficiency.

Result: Extensive experiments on STEM reasoning and coding benchmarks show ThinkRouter outperforms explicit CoT, random routing, and latent reasoning baselines, achieving average improvement of 19.70 points in Pass@1 while reducing generation length by up to 15.55%.

Conclusion: ThinkRouter effectively calibrates errors from both explicit CoT and latent reasoning, accelerates end-of-thinking token generation by globally lowering model confidence, and provides an efficient inference-time mechanism for improving reasoning in large language models.

Abstract: Recent work explores latent reasoning to improve reasoning efficiency by replacing explicit reasoning trajectories with continuous representations in a latent space, yet its effectiveness varies across settings. Analysis of model confidence dynamics under latent reasoning reveals that thinking trajectories ending in incorrect answers contain fewer low-confidence steps than those ending in correct answers. Meanwhile, we suggest that soft embeddings aggregated by multiple low-confidence thinking alternatives may introduce and propagate noise, leading to high confidence in unreliable reasoning trajectories. Motivated by these observations, ThinkRouter, an inference-time confidence-aware routing mechanism is proposed to avoid high confidence and noise for efficient reasoning. ThinkRouter routes thinking to the discrete token space when model confidence is low, and to the latent space otherwise. Extensive experiments on STEM reasoning and coding benchmarks across diverse large reasoning models demonstrate that ThinkRouter outperforms explicit CoT, random routing, and latent reasoning baselines in terms of accuracy, achieving an average improvement of 19.70 points in Pass@1, while reducing generation length by up to 15.55%. Further comprehensive analysis reveals that ThinkRouter can calibrate errors arising from explicit CoT and latent reasoning, and accelerates end-of-thinking token generation by globally lowering model confidence.

Method: SCF-RKL uses reverse Kullback-Leibler divergence to measure functional divergence between models and selectively incorporates complementary parameters through sparse, distribution-aware updates. Instead of assuming linear additivity in parameter space, it employs a mode-seeking, sparsity-inducing design that preserves stable representations while integrating new capabilities.

Result: Extensive experiments on 24 benchmarks spanning advanced reasoning, general reasoning and knowledge, instruction following, safety, and vision classification show that SCF-RKL consistently outperforms existing model merging methods while maintaining strong generalization and generation stability across various model scales and architectures.

Conclusion: SCF-RKL provides a principled framework for model merging that effectively controls functional interference through sparse complementary fusion, offering improved performance and stability compared to heuristic-based approaches.

Abstract: Model merging has emerged as a promising paradigm for composing the capabilities of large language models by directly operating in weight space, enabling the integration of specialized models without costly retraining. However, existing merging methods largely rely on parameter-space heuristics, which often introduce severe interference, leading to degraded generalization and unstable generation behaviors such as repetition and incoherent outputs. In this work, we propose Sparse Complementary Fusion with reverse KL (SCF-RKL), a novel model merging framework that explicitly controls functional interference through sparse, distribution-aware updates. Instead of assuming linear additivity in parameter space, SCF-RKL measures the functional divergence between models using reverse Kullback-Leibler divergence and selectively incorporates complementary parameters. This mode-seeking, sparsity-inducing design effectively preserves stable representations while integrating new capabilities. We evaluate SCF-RKL across a wide range of model scales and architectures, covering both reasoning-focused and instruction-tuned models. Extensive experiments on 24 benchmarks spanning advanced reasoning, general reasoning and knowledge, instruction following, and safety demonstrate, vision classification that SCF-RKL consistently outperforms existing model merging methods while maintaining strong generalization and generation stability.

Method: Introduces Dedicated Feature Crosscoders (DFCs) as an architectural modification to better isolate features unique to one model, enabling unsupervised cross-architecture model diffing.

Result: Successfully identified Chinese Communist Party alignment in Qwen3-8B and Deepseek-R1-0528-Qwen3-8B, American exceptionalism in Llama3.1-8B-Instruct, and copyright refusal mechanism in GPT-OSS-20B.

Conclusion: Cross-architecture crosscoder model diffing with DFCs is an effective method for identifying meaningful behavioral differences between AI models, advancing model safety analysis.

Abstract: Model diffing, the process of comparing models’ internal representations to identify their differences, is a promising approach for uncovering safety-critical behaviors in new models. However, its application has so far been primarily focused on comparing a base model with its finetune. Since new LLM releases are often novel architectures, cross-architecture methods are essential to make model diffing widely applicable. Crosscoders are one solution capable of cross-architecture model diffing but have only ever been applied to base vs finetune comparisons. We provide the first application of crosscoders to cross-architecture model diffing and introduce Dedicated Feature Crosscoders (DFCs), an architectural modification designed to better isolate features unique to one model. Using this technique, we find in an unsupervised fashion features including Chinese Communist Party alignment in Qwen3-8B and Deepseek-R1-0528-Qwen3-8B, American exceptionalism in Llama3.1-8B-Instruct, and a copyright refusal mechanism in GPT-OSS-20B. Together, our results work towards establishing cross-architecture crosscoder model diffing as an effective method for identifying meaningful behavioral differences between AI models.

Method: Created Text2GQL-Bench with a scalable construction framework that generates datasets across different domains, question abstraction levels, and graph query languages using heterogeneous resources. Introduced comprehensive evaluation metrics covering grammatical validity, similarity, semantic alignment, and execution accuracy.

Result: Benchmark reveals stark dialect gap in ISO-GQL generation: strong LLMs achieve only 4% execution accuracy zero-shot, but 3-shot prompting raises to ~50% (though grammatical validity remains <70%). Fine-tuned 8B model reaches 45.1% execution accuracy and 90.8% grammatical validity.

Conclusion: Text2GQL-Bench addresses critical limitations in Text-to-GQL evaluation, revealing that most performance improvements come from exposure to sufficient ISO-GQL examples rather than general LLM capabilities.

Abstract: Graph models are fundamental to data analysis in domains rich with complex relationships. Text-to-Graph-Query-Language (Text-to-GQL) systems act as a translator, converting natural language into executable graph queries. This capability allows Large Language Models (LLMs) to directly analyze and manipulate graph data, posi-tioning them as powerful agent infrastructures for Graph Database Management System (GDBMS). Despite recent progress, existing datasets are often limited in domain coverage, supported graph query languages, or evaluation scope. The advancement of Text-to-GQL systems is hindered by the lack of high-quality benchmark datasets and evaluation methods to systematically compare model capabilities across different graph query languages and domains. In this work, we present Text2GQL-Bench, a unified Text-to-GQL benchmark designed to address these limitations. Text2GQL-Bench couples a multi-GQL dataset that has 178,184 (Question, Query) pairs spanning 13 domains, with a scalable construction framework that generates datasets in different domains, question abstraction levels, and GQLs with heterogeneous resources. To support compre-hensive assessment, we introduce an evaluation method that goes beyond a single end-to-end metric by jointly reporting grammatical validity, similarity, semantic alignment, and execution accuracy. Our evaluation uncovers a stark dialect gap in ISO-GQL generation: even strong LLMs achieve only at most 4% execution accuracy (EX) in zero-shot settings, though a fixed 3-shot prompt raises accuracy to around 50%, the grammatical validity remains lower than 70%. Moreover, a fine-tuned 8B open-weight model reaches 45.1% EX, and 90.8% grammatical validity, demonstrating that most of the performance jump is unlocked by exposure to sufficient ISO-GQL examples.

Method: AIR introduces a domain-specific language for managing incident response lifecycle in LLM agent systems, integrated into the agent’s execution loop. It performs: (1) semantic incident detection grounded in environment state and recent context, (2) guided containment and recovery actions via agent tools, and (3) synthesis of guardrail rules during eradication to prevent similar future incidents.

Result: AIR achieves detection, remediation, and eradication success rates all exceeding 90% across three representative agent types. LLM-generated rules approach the effectiveness of developer-authored rules across domains, with experiments confirming the necessity of AIR’s design components and showing timeliness with moderate overhead.

Conclusion: Incident response is both feasible and essential as a first-class mechanism for improving LLM agent safety, moving beyond purely preventive approaches to include detection, containment, recovery, and prevention capabilities.

Abstract: Large Language Model (LLM) agents are increasingly deployed in practice across a wide range of autonomous applications. Yet current safety mechanisms for LLM agents focus almost exclusively on preventing failures in advance, providing limited capabilities for responding to, containing, or recovering from incidents after they inevitably arise. In this work, we introduce AIR, the first incident response framework for LLM agent systems. AIR defines a domain-specific language for managing the incident response lifecycle autonomously in LLM agent systems, and integrates it into the agent’s execution loop to (1) detect incidents via semantic checks grounded in the current environment state and recent context, (2) guide the agent to execute containment and recovery actions via its tools, and (3) synthesize guardrail rules during eradication to block similar incidents in future executions. We evaluate AIR on three representative agent types. Results show that AIR achieves detection, remediation, and eradication success rates all exceeding 90%. Extensive experiments further confirm the necessity of AIR’s key design components, show the timeliness and moderate overhead of AIR, and demonstrate that LLM-generated rules can approach the effectiveness of developer-authored rules across domains. These results show that incident response is both feasible and essential as a first-class mechanism for improving agent safety.

Method: TSR performs lightweight tree-style search during training to construct high-quality trajectories by selecting high-scoring actions at each turn using task-specific feedback. It’s instantiated with best-of-N, beam, and shallow lookahead search, and paired with PPO and GRPO optimizers.

Result: TSR achieves up to 15% performance gains and more stable learning on Sokoban, FrozenLake, and WebShop tasks with a one-time increase in training compute, while leaving the underlying optimization objective unchanged.

Conclusion: TSR provides a simple, general mechanism for stronger multi-turn agent learning by moving search from inference time to the rollout stage of training, complementary to existing frameworks and rejection-sampling methods.

Abstract: Advances in large language models (LLMs) are driving a shift toward using reinforcement learning (RL) to train agents from iterative, multi-turn interactions across tasks. However, multi-turn RL remains challenging as rewards are often sparse or delayed, and environments can be stochastic. In this regime, naive trajectory sampling can hinder exploitation and induce mode collapse. We propose TSR (Trajectory-Search Rollouts), a training-time approach that repurposes test-time scaling ideas for improved per-turn rollout generation. TSR performs lightweight tree-style search to construct high-quality trajectories by selecting high-scoring actions at each turn using task-specific feedback. This improves rollout quality and stabilizes learning while leaving the underlying optimization objective unchanged, making TSR optimizer-agnostic. We instantiate TSR with best-of-N, beam, and shallow lookahead search, and pair it with PPO and GRPO, achieving up to 15% performance gains and more stable learning on Sokoban, FrozenLake, and WebShop tasks at a one-time increase in training compute. By moving search from inference time to the rollout stage of training, TSR provides a simple and general mechanism for stronger multi-turn agent learning, complementary to existing frameworks and rejection-sampling-style selection methods.

Method: MaxExp selects the most probable species assemblage by directly maximizing a chosen evaluation metric without requiring calibration data. SSE (Set Size Expectation) predicts assemblages based on expected species richness as a computationally efficient alternative.

Result: MaxExp consistently matches or surpasses widely used thresholding and calibration methods across three case studies with diverse taxa, species counts, and performance metrics, especially under strong class imbalance and high rarity.

Conclusion: MaxExp and SSE provide robust, reproducible tools for multispecies SDM binarization, with MaxExp offering optimal performance and SSE providing a simpler competitive alternative.

Abstract: Species distribution models (SDMs) commonly produce probabilistic occurrence predictions that must be converted into binary presence-absence maps for ecological inference and conservation planning. However, this binarization step is typically heuristic and can substantially distort estimates of species prevalence and community composition. We present MaxExp, a decision-driven binarization framework that selects the most probable species assemblage by directly maximizing a chosen evaluation metric. MaxExp requires no calibration data and is flexible across several scores. We also introduce the Set Size Expectation (SSE) method, a computationally efficient alternative that predicts assemblages based on expected species richness. Using three case studies spanning diverse taxa, species counts, and performance metrics, we show that MaxExp consistently matches or surpasses widely used thresholding and calibration methods, especially under strong class imbalance and high rarity. SSE offers a simpler yet competitive option. Together, these methods provide robust, reproducible tools for multispecies SDM binarization.

Method: Proposes RELATE, a reinforcement learning-based end-to-end framework that integrates generation and objective alignment within a single model. Uses policy learning to incorporate performance and compliance objectives directly into generation, with multi-dimensional rewards including conversion-oriented metrics and policy constraints.

Result: Extensive experiments on large-scale industrial datasets show RELATE consistently outperforms baselines. Online deployment on a production advertising platform yields statistically significant improvements in click-through conversion rate (CTCVR) under strict policy constraints.

Conclusion: RELATE successfully addresses limitations of traditional two-stage approaches by unifying generation and optimization, demonstrating robustness and real-world effectiveness in improving advertising performance while maintaining policy compliance.

Abstract: In online advertising, advertising text plays a critical role in attracting user engagement and driving advertiser value. Existing industrial systems typically follow a two-stage paradigm, where candidate texts are first generated and subsequently aligned with online performance metrics such as click-through rate(CTR). This separation often leads to misaligned optimization objectives and low funnel efficiency, limiting global optimality. To address these limitations, we propose RELATE, a reinforcement learning-based end-to-end framework that unifies generation and objective alignment within a single model. Instead of decoupling text generation from downstream metric alignment, RELATE integrates performance and compliance objectives directly into the generation process via policy learning. To better capture ultimate advertiser value beyond click-level signals, We incorporate conversion-oriented metrics into the objective and jointly model them with compliance constraints as multi-dimensional rewards, enabling the model to generate high-quality ad texts that improve conversion performance under policy constraints. Extensive experiments on large-scale industrial datasets demonstrate that RELATE consistently outperforms baselines. Furthermore, online deployment on a production advertising platform yields statistically significant improvements in click-through conversion rate(CTCVR) under strict policy constraints, validating the robustness and real-world effectiveness of the proposed framework.

Method: Execute-Summarize (ES) framework that separates task execution from workflow construction: first complete tasks using available tools, then independently reconstruct structured workflows from execution traces

Result: Outperforms existing methods on FlowBench, providing improved workflow accuracy and robustness through the decoupled approach

Conclusion: The ES framework offers a reliable paradigm for grounding free-form LLM reasoning into structured workflows by separating execution from workflow construction

Abstract: LLMs can solve complex tasks through reasoning and tool use, but accurately translating these solutions into structured workflows remains challenging. We model workflows as sequences of tool use and reformulate the problem as designing a mechanism that can both solve tasks and reliably construct workflows. Prior approaches that build workflows during execution often suffer from inaccuracies due to interference between the two processes. We propose an Execute-Summarize(ES) framework that decouples task execution from workflow construction: the model first completes the task using available tools, then independently reconstructs a structured workflow from execution traces. This separation improves workflow accuracy and robustness. We introduce FlowBench and show through extensive experiments that our approach outperforms existing methods, providing a reliable paradigm for grounding free-form LLM reasoning into structured workflows.

Method: Hierarchical multi-agent system with Orchestrating Agent coordinating specialized agents (Solution, Illustration, Narration) with quality gates, iterative critique mechanisms, semantic checks, rule-based constraints, and tool-based compilation. Generates structured executable video scripts compiled into synchronized visuals/narration.

Result: Achieves throughput exceeding 1 million videos per day, 95% cost reduction compared to industry standards, while maintaining high acceptance rate through automated end-to-end production.

Conclusion: LAVES demonstrates that LLM-based multi-agent systems can effectively generate high-quality instructional videos with logical rigor, addressing limitations of traditional video generation approaches through structured, deterministic compilation.

Abstract: Although recent end-to-end video generation models demonstrate impressive performance in visually oriented content creation, they remain limited in scenarios that require strict logical rigor and precise knowledge representation, such as instructional and educational media. To address this problem, we propose LAVES, a hierarchical LLM-based multi-agent system for generating high-quality instructional videos from educational problems. The LAVES formulates educational video generation as a multi-objective task that simultaneously demands correct step-by-step reasoning, pedagogically coherent narration, semantically faithful visual demonstrations, and precise audio–visual alignment. To address the limitations of prior approaches–including low procedural fidelity, high production cost, and limited controllability–LAVES decomposes the generation workflow into specialized agents coordinated by a central Orchestrating Agent with explicit quality gates and iterative critique mechanisms. Specifically, the Orchestrating Agent supervises a Solution Agent for rigorous problem solving, an Illustration Agent that produces executable visualization codes, and a Narration Agent for learner-oriented instructional scripts. In addition, all outputs from the working agents are subject to semantic critique, rule-based constraints, and tool-based compilation checks. Rather than directly synthesizing pixels, the system constructs a structured executable video script that is deterministically compiled into synchronized visuals and narration using template-driven assembly rules, enabling fully automated end-to-end production without manual editing. In large-scale deployments, LAVES achieves a throughput exceeding one million videos per day, delivering over a 95% reduction in cost compared to current industry-standard approaches while maintaining a high acceptance rate.

Method: Min-kNN Distance: a black-box detector that samples multiple completions for a prompt and computes the average of the k smallest nearest-neighbor edit distances. RLVR training causes prompts seen during training to produce more rigid, similar generations, while unseen prompts maintain diversity.

Result: Min-kNN Distance reliably distinguishes RL-seen examples from unseen ones across multiple RLVR-trained reasoning models, outperforming existing membership inference and RL contamination detection baselines.

Conclusion: RLVR training leaves a distinctive behavioral signature (reduced output diversity) that can be detected without access to model internals, providing an effective contamination detection method for RL-fine-tuned models.

Abstract: Reinforcement learning with verifiable rewards (RLVR) is central to training modern reasoning models, but the undisclosed training data raises concerns about benchmark contamination. Unlike pretraining methods, which optimize models using token-level probabilities, RLVR fine-tunes models based on reward feedback from self-generated reasoning trajectories, making conventional likelihood-based detection methods less effective. We show that RLVR induces a distinctive behavioral signature: prompts encountered during RLVR training result in more rigid and similar generations, while unseen prompts retain greater diversity. We introduce Min-$k$NN Distance, a simple black-box detector that quantifies this collapse by sampling multiple completions for a given prompt and computing the average of the $k$ smallest nearest-neighbor edit distances. Min-$k$NN Distance requires no access to the reference model or token probabilities. Experiments across multiple RLVR-trained reasoning models show that Min-$k$NN Distance reliably distinguishes RL-seen examples from unseen ones and outperforms existing membership inference and RL contamination detection baselines.

Method: Two main components: (1) Disentangled Semantic Tokenizer (DST) that unifies modalities via geometry-aware alignment and quantizes them via coarse-to-fine strategy with shared codebooks for consensus and modality-specific codebooks for nuances, enforced by mutual information minimization; (2) Hierarchical Memory-Anchor Transformer (HMAT) that splits positional encoding into inter- and intra-item subspaces via Hierarchical RoPE and inserts Anchor Tokens to condense items into compact memory while retaining details for current items.

Result: Experiments on real-world datasets show consistent improvements over state-of-the-art baselines, especially in cold-start scenarios. Deployment on a large-scale social platform serving millions of users achieved a 6.55% gain in the core online metric.

Conclusion: Hi-SAM effectively addresses the limitations of existing semantic ID-based multi-modal recommendation systems by introducing disentangled tokenization and hierarchical transformer architecture, demonstrating significant improvements in both offline experiments and real-world deployment.

Abstract: Multi-modal recommendation has gained traction as items possess rich attributes like text and images. Semantic ID-based approaches effectively discretize this information into compact tokens. However, two challenges persist: (1) Suboptimal Tokenization: existing methods (e.g., RQ-VAE) lack disentanglement between shared cross-modal semantics and modality-specific details, causing redundancy or collapse; (2) Architecture-Data Mismatch: vanilla Transformers treat semantic IDs as flat streams, ignoring the hierarchy of user interactions, items, and tokens. Expanding items into multiple tokens amplifies length and noise, biasing attention toward local details over holistic semantics. We propose Hi-SAM, a Hierarchical Structure-Aware Multi-modal framework with two designs: (1) Disentangled Semantic Tokenizer (DST): unifies modalities via geometry-aware alignment and quantizes them via a coarse-to-fine strategy. Shared codebooks distill consensus while modality-specific ones recover nuances from residuals, enforced by mutual information minimization; (2) Hierarchical Memory-Anchor Transformer (HMAT): splits positional encoding into inter- and intra-item subspaces via Hierarchical RoPE to restore hierarchy. It inserts Anchor Tokens to condense items into compact memory, retaining details for the current item while accessing history only through compressed summaries. Experiments on real-world datasets show consistent improvements over SOTA baselines, especially in cold-start scenarios. Deployed on a large-scale social platform serving millions of users, Hi-SAM achieved a 6.55% gain in the core online metric.

Method: Proposes PuYun-LDM with two key components: 1) 3D Masked AutoEncoder (3D-MAE) that encodes weather-state evolution features as additional conditioning for the diffusion model, and 2) Variable-Aware Masked Frequency Modeling (VA-MFM) strategy that adaptively selects thresholds based on the spectral energy distribution of each variable to address spectral heterogeneity.

Result: PuYun-LDM achieves superior performance to ENS (European Centre for Medium-Range Weather Forecasts ensemble) at short lead times while remaining comparable at longer horizons. It generates 15-day global forecasts with 6-hour temporal resolution in 5 minutes on a single NVIDIA H200 GPU, with ensemble forecasts producible in parallel.

Conclusion: The proposed approach successfully enhances latent diffusability for weather forecasting by addressing the unique challenges of meteorological data through 3D-MAE encoding and variable-aware frequency modeling, enabling efficient high-resolution ensemble weather prediction.

Abstract: Latent diffusion models (LDMs) suffer from limited diffusability in high-resolution (<=0.25°) ensemble weather forecasting, where diffusability characterizes how easily a latent data distribution can be modeled by a diffusion process. Unlike natural image fields, meteorological fields lack task-agnostic foundation models and explicit semantic structures, making VFM-based regularization inapplicable. Moreover, existing frequency-based approaches impose identical spectral regularization across channels under a homogeneity assumption, which leads to uneven regularization strength under the inter-variable spectral heterogeneity in multivariate meteorological data. To address these challenges, we propose a 3D Masked AutoEncoder (3D-MAE) that encodes weather-state evolution features as an additional conditioning for the diffusion model, together with a Variable-Aware Masked Frequency Modeling (VA-MFM) strategy that adaptively selects thresholds based on the spectral energy distribution of each variable. Together, we propose PuYun-LDM, which enhances latent diffusability and achieves superior performance to ENS at short lead times while remaining comparable to ENS at longer horizons. PuYun-LDM generates a 15-day global forecast with a 6-hour temporal resolution in five minutes on a single NVIDIA H200 GPU, while ensemble forecasts can be efficiently produced in parallel.

Method: Introduces a framework with two core components: 1) Entropy-Guided Token Pooling (EGTP) - uses on-the-fly activations and token entropy for accurate static prediction with negligible cost, and 2) Progressive Length Prediction (PLP) - dynamically estimates remaining length at each decoding step for stochastic generation.

Result: EGTP achieves state-of-the-art accuracy on the ForeLen benchmark, reducing MAE by 29.16% over best baseline. Integration with length-aware scheduler yields significant end-to-end throughput gains.

Conclusion: Provides a new technical and evaluation baseline for efficient LLM inference by reusing main model’s internal hidden states for length prediction, addressing computational waste in batched inference.

Abstract: The long-tailed distribution of sequence lengths in LLM serving and reinforcement learning (RL) sampling causes significant computational waste due to excessive padding in batched inference. Existing methods rely on auxiliary models for static length prediction, but they incur high overhead, generalize poorly, and fail in stochastic “one-to-many” sampling scenarios. We introduce a lightweight framework that reuses the main model’s internal hidden states for efficient length prediction. Our framework features two core components: 1) Entropy-Guided Token Pooling (EGTP), which uses on-the-fly activations and token entropy for highly accurate static prediction with negligible cost, and 2) Progressive Length Prediction (PLP), which dynamically estimates the remaining length at each decoding step to handle stochastic generation. To validate our approach, we build and release ForeLen, a comprehensive benchmark with long-sequence, Chain-of-Thought, and RL data. On ForeLen, EGTP achieves state-of-the-art accuracy, reducing MAE by 29.16% over the best baseline. Integrating our methods with a length-aware scheduler yields significant end-to-end throughput gains. Our work provides a new technical and evaluation baseline for efficient LLM inference.

Method: REVIS extracts pure visual information vectors via orthogonal projection in latent space and performs calibrated sparse intervention only at the precise depth where suppression occurs, minimizing computational cost.

Result: REVIS reduces object hallucination rates by approximately 19% compared to state-of-the-art baselines while preserving general reasoning capabilities.

Conclusion: The training-free REVIS framework effectively addresses object hallucination in LVLMs by reactivating suppressed visual information through geometric operations in latent space.

Abstract: Despite the advanced capabilities of Large Vision-Language Models (LVLMs), they frequently suffer from object hallucination. One reason is that visual features and pretrained textual representations often become intertwined in the deeper network layers. To address this, we propose REVIS, a training-free framework designed to explicitly re-activate this suppressed visual information. Rooted in latent space geometry, REVIS extracts the pure visual information vector via orthogonal projection and employs a calibrated strategy to perform sparse intervention only at the precise depth where suppression occurs. This surgical approach effectively restores visual information with minimal computational cost. Empirical evaluations on standard benchmarks demonstrate that REVIS reduces object hallucination rates by approximately 19% compared to state-of-the-art baselines, while preserving general reasoning capabilities.

Method: ProtoT replaces standard self-attention transformers with prototype vectors that enable two-way communication between input sequences and prototypes. These prototypes automatically learn nameable concepts during training, creating interpretable communication channels that aggregate contextual information at different time scales.

Result: ProtoT achieves competitive performance on text generation and GLUE benchmarks, scales linearly with sequence length (vs quadratic for standard transformers), and provides interpretable pathways showing how robustness and sensitivity arise. The prototypes capture meaningful concepts like “woman” and allow for targeted model edits.

Conclusion: ProtoT demonstrates that autoregressive language models can be made interpretable by design while maintaining near state-of-the-art performance, paving the way for more transparent and trustworthy language models.

Abstract: While state-of-the-art language models (LMs) surpass the vast majority of humans in certain domains, their reasoning remains largely opaque, undermining trust in their output. Furthermore, while autoregressive LMs can output explicit reasoning, their true reasoning process is opaque, which introduces risks like deception and hallucination. In this work, we introduce the Prototype Transformer (ProtoT) – an autoregressive LM architecture based on prototypes (parameter vectors), posed as an alternative to the standard self-attention-based transformers. ProtoT works by means of two-way communication between the input sequence and the prototypes, and we show that this leads to the prototypes automatically capturing nameable concepts (e.g. “woman”) during training. They provide the potential to interpret the model’s reasoning and allow for targeted edits of its behavior. Furthermore, by design, the prototypes create communication channels that aggregate contextual information at different time scales, aiding interpretability. In terms of computation scalability, ProtoT scales linearly with sequence length vs the quadratic scalability of SOTA self-attention transformers. Compared to baselines, ProtoT scales well with model and data size, and performs well on text generation and downstream tasks (GLUE). ProtoT exhibits robustness to input perturbations on par or better than some baselines, but differs from them by providing interpretable pathways showing how robustness and sensitivity arises. Reaching close to the performance of state-of-the-art architectures, ProtoT paves the way to creating well-performing autoregressive LMs interpretable by design.

Method: Develop VRCsim simulation framework for generating streaming VRC cooperative perception data, create VRC-QA dataset for spatial querying, and build Talk2DM plug-and-play module using chain-of-prompt mechanism that integrates human-defined rules with LLM commonsense knowledge.

Result: Talk2DM achieves over 93% natural language query accuracy with Qwen3:8B, Gemma3:27B, and GPT-oss models, with average response time of 2-5 seconds. Larger models achieve higher accuracy but suffer efficiency degradation.

Conclusion: Talk2DM successfully adds natural language querying capabilities to dynamic maps, demonstrating strong generalization across different LLMs while maintaining practical efficiency for real-world deployment.

Abstract: Dynamic maps (DM) serve as the fundamental information infrastructure for vehicle-road-cloud (VRC) cooperative autonomous driving in China and Japan. By providing comprehensive traffic scene representations, DM overcome the limitations of standalone autonomous driving systems (ADS), such as physical occlusions. Although DM-enhanced ADS have been successfully deployed in real-world applications in Japan, existing DM systems still lack a natural-language-supported (NLS) human interface, which could substantially enhance human-DM interaction. To address this gap, this paper introduces VRCsim, a VRC cooperative perception (CP) simulation framework designed to generate streaming VRC-CP data. Based on VRCsim, we construct a question-answering data set, VRC-QA, focused on spatial querying and reasoning in mixed-traffic scenes. Building upon VRCsim and VRC-QA, we further propose Talk2DM, a plug-and-play module that extends VRC-DM systems with NLS querying and commonsense reasoning capabilities. Talk2DM is built upon a novel chain-of-prompt (CoP) mechanism that progressively integrates human-defined rules with the commonsense knowledge of large language models (LLMs). Experiments on VRC-QA show that Talk2DM can seamlessly switch across different LLMs while maintaining high NLS query accuracy, demonstrating strong generalization capability. Although larger models tend to achieve higher accuracy, they incur significant efficiency degradation. Our results reveal that Talk2DM, powered by Qwen3:8B, Gemma3:27B, and GPT-oss models, achieves over 93% NLS query accuracy with an average response time of only 2-5 seconds, indicating strong practical potential.

Method: Proposes an adaptive framework for intelligent AI delegation that includes a sequence of decisions involving task allocation, transfer of authority, responsibility, accountability, clear role specifications, intent clarity, and trust establishment mechanisms between parties.

Result: The framework is designed to be applicable to both human and AI delegators/delegatees in complex delegation networks, aiming to inform development of protocols for the emerging agentic web.

Conclusion: An adaptive delegation framework is needed to enable AI agents to handle complex tasks through meaningful decomposition and safe delegation across heterogeneous systems, addressing current limitations of simple heuristic-based approaches.

Abstract: AI agents are able to tackle increasingly complex tasks. To achieve more ambitious goals, AI agents need to be able to meaningfully decompose problems into manageable sub-components, and safely delegate their completion across to other AI agents and humans alike. Yet, existing task decomposition and delegation methods rely on simple heuristics, and are not able to dynamically adapt to environmental changes and robustly handle unexpected failures. Here we propose an adaptive framework for intelligent AI delegation - a sequence of decisions involving task allocation, that also incorporates transfer of authority, responsibility, accountability, clear specifications regarding roles and boundaries, clarity of intent, and mechanisms for establishing trust between the two (or more) parties. The proposed framework is applicable to both human and AI delegators and delegatees in complex delegation networks, aiming to inform the development of protocols in the emerging agentic web.

Method: Proposes HSAE that jointly learns series of SAEs with parent-child relationships using structural constraint loss and random feature perturbation mechanism to align hierarchical features.

Result: HSAE consistently recovers semantically meaningful hierarchies across various LLMs and layers, preserves reconstruction fidelity and interpretability of standard SAEs across dictionary sizes.

Conclusion: HSAE provides scalable tool for discovering and analyzing multi-scale conceptual structures in LLM representations, capturing hierarchical feature relationships.

Abstract: Sparse autoencoders (SAEs) have proven effective for extracting monosemantic features from large language models (LLMs), yet these features are typically identified in isolation. However, broad evidence suggests that LLMs capture the intrinsic structure of natural language, where the phenomenon of “feature splitting” in particular indicates that such structure is hierarchical. To capture this, we propose the Hierarchical Sparse Autoencoder (HSAE), which jointly learns a series of SAEs and the parent-child relationships between their features. HSAE strengthens the alignment between parent and child features through two novel mechanisms: a structural constraint loss and a random feature perturbation mechanism. Extensive experiments across various LLMs and layers demonstrate that HSAE consistently recovers semantically meaningful hierarchies, supported by both qualitative case studies and rigorous quantitative metrics. At the same time, HSAE preserves the reconstruction fidelity and interpretability of standard SAEs across different dictionary sizes. Our work provides a powerful, scalable tool for discovering and analyzing the multi-scale conceptual structures embedded in LLM representations.

Method: Proposes Selective Abstraction framework formalized through selective risk and coverage. Introduces Atom-wise Selective Abstraction that decomposes responses into atomic claims and replaces uncertain atoms with higher-confidence, less specific abstractions.

Result: Across six open-source models on FactScore and LongFact-Objects benchmarks, atom-wise SA consistently outperforms baselines, improving area under risk-coverage curve by up to 27.73% over claim removal.

Conclusion: Reducing specificity can boost accuracy and reliability while preserving most original meaning, offering better trade-off than binary abstention for long-form generation.

Abstract: LLMs are widely used, yet they remain prone to factual errors that erode user trust and limit adoption in high-risk settings. One approach to mitigate this risk is to equip models with uncertainty estimation mechanisms that abstain when confidence is low. However, this binary “all-or-nothing” approach is excessively restrictive in long-form settings, often discarding valuable information. We introduce Selective Abstraction (SA), a framework that enables LLMs to trade specificity for reliability by selectively reducing the detail of uncertain content. We first formalize SA through the lenses of selective risk and coverage. We then propose Atom-wise Selective Abstraction, a claim-level instantiation that decomposes responses into atomic claims (short, self-contained statements each expressing a single fact) and replaces uncertain atoms with higher confidence, less specific abstractions. To evaluate this framework, we develop a novel end-to-end pipeline for open-ended generation that instantiates risk as factual correctness and measures coverage using an information-theoretic measure of retained information. Across six open-source models on the FactScore and LongFact-Objects benchmarks, atom-wise SA consistently outperforms existing baselines, improving the area under the risk-coverage curve (AURC) by up to 27.73% over claim removal, demonstrating that reducing specificity can boost accuracy and reliability while preserving most of their original meaning.

Method: AlphaPROBE models factors as nodes and evolutionary links as edges in a DAG. It consists of: 1) Bayesian Factor Retriever that identifies high-potential seeds by balancing exploitation/exploration through posterior probability model, and 2) DAG-aware Factor Generator that leverages full ancestral trace to produce context-aware, nonredundant optimizations.

Result: Extensive experiments on three major Chinese stock market datasets against 8 competitive baselines show AlphaPROBE significantly improves predictive accuracy, return stability, and training efficiency compared to existing methods.

Conclusion: Leveraging global evolutionary topology is essential for efficient and robust automated alpha discovery. The DAG-based approach provides a more systematic framework for factor mining in quantitative finance.

Abstract: Extracting signals through alpha factor mining is a fundamental challenge in quantitative finance. Existing automated methods primarily follow two paradigms: Decoupled Factor Generation, which treats factor discovery as isolated events, and Iterative Factor Evolution, which focuses on local parent-child refinements. However, both paradigms lack a global structural view, often treating factor pools as unstructured collections or fragmented chains, which leads to redundant search and limited diversity. To address these limitations, we introduce AlphaPROBE (Alpha Mining via Principled Retrieval and On-graph Biased Evolution), a framework that reframes alpha mining as the strategic navigation of a Directed Acyclic Graph (DAG). By modeling factors as nodes and evolutionary links as edges, AlphaPROBE treats the factor pool as a dynamic, interconnected ecosystem. The framework consists of two core components: a Bayesian Factor Retriever that identifies high-potential seeds by balancing exploitation and exploration through a posterior probability model, and a DAG-aware Factor Generator that leverages the full ancestral trace of factors to produce context-aware, nonredundant optimizations. Extensive experiments on three major Chinese stock market datasets against 8 competitive baselines demonstrate that AlphaPROBE significantly gains enhanced performance in predictive accuracy, return stability and training efficiency. Our results confirm that leveraging global evolutionary topology is essential for efficient and robust automated alpha discovery. We have open-sourced our implementation at https://github.com/gta0804/AlphaPROBE.

Method: MEME uses a multi-agent extraction module to transform noisy data into high-fidelity Investment Arguments, Gaussian Mixture Modeling to uncover latent consensus in semantic space, and temporal evaluation/alignment to track the lifecycle and profitability of market modes.

Result: MEME consistently outperforms seven state-of-the-art baselines across three heterogeneous Chinese stock pools from 2023-2025. Ablation studies, sensitivity analysis, lifecycle case studies, and cost analysis validate its capacity to identify and adapt to evolving market consensus.

Conclusion: The logic-oriented approach of modeling financial markets as evolutionary ecosystems of competing narratives enables more robust portfolio construction by prioritizing enduring market wisdom over transient anomalies.

Abstract: LLMs have demonstrated significant potential in quantitative finance by processing vast unstructured data to emulate human-like analytical workflows. However, current LLM-based methods primarily follow either an Asset-Centric paradigm focused on individual stock prediction or a Market-Centric approach for portfolio allocation, often remaining agnostic to the underlying reasoning that drives market movements. In this paper, we propose a Logic-Oriented perspective, modeling the financial market as a dynamic, evolutionary ecosystem of competing investment narratives, termed Modes of Thought. To operationalize this view, we introduce MEME (Modeling the Evolutionary Modes of Financial Markets), designed to reconstruct market dynamics through the lens of evolving logics. MEME employs a multi-agent extraction module to transform noisy data into high-fidelity Investment Arguments and utilizes Gaussian Mixture Modeling to uncover latent consensus within a semantic space. To model semantic drift among different market conditions, we also implement a temporal evaluation and alignment mechanism to track the lifecycle and historical profitability of these modes. By prioritizing enduring market wisdom over transient anomalies, MEME ensures that portfolio construction is guided by robust reasoning. Extensive experiments on three heterogeneous Chinese stock pools from 2023 to 2025 demonstrate that MEME consistently outperforms seven SOTA baselines. Further ablation studies, sensitivity analysis, lifecycle case study and cost analysis validate MEME’s capacity to identify and adapt to the evolving consensus of financial markets. Our implementation can be found at https://github.com/gta0804/MEME.

Method: Gaia2 introduces scenarios where environments evolve independently of agent actions, paired with write-action verifiers for fine-grained action-level evaluation. Built on the open-source Agents Research Environments platform, it enables reinforcement learning from verifiable rewards.

Result: Evaluation shows no model dominates across capabilities: GPT-5 (high) achieves 42% pass@1 but fails on time-sensitive tasks, Claude-4 Sonnet trades accuracy/speed for cost, Kimi-K2 leads open-source models with 21% pass@1, revealing trade-offs between reasoning, efficiency, and robustness.

Conclusion: Gaia2 exposes fundamental challenges in closing the “sim2real” gap and provides flexible infrastructure for developing, benchmarking, and training practical agent systems through the released ARE framework.

Abstract: We introduce Gaia2, a benchmark for evaluating large language model agents in realistic, asynchronous environments. Unlike prior static or synchronous evaluations, Gaia2 introduces scenarios where environments evolve independently of agent actions, requiring agents to operate under temporal constraints, adapt to noisy and dynamic events, resolve ambiguity, and collaborate with other agents. Each scenario is paired with a write-action verifier, enabling fine-grained, action-level evaluation and making Gaia2 directly usable for reinforcement learning from verifiable rewards. Our evaluation of state-of-the-art proprietary and open-source models shows that no model dominates across capabilities: GPT-5 (high) reaches the strongest overall score of 42% pass@1 but fails on time-sensitive tasks, Claude-4 Sonnet trades accuracy and speed for cost, Kimi-K2 leads among open-source models with 21% pass@1. These results highlight fundamental trade-offs between reasoning, efficiency, robustness, and expose challenges in closing the “sim2real” gap. Gaia2 is built on a consumer environment with the open-source Agents Research Environments platform and designed to be easy to extend. By releasing Gaia2 alongside the foundational ARE framework, we aim to provide the community with a flexible infrastructure for developing, benchmarking, and training the next generation of practical agent systems.

Method: CSEval leverages language models to assess clinical semantic alignment between generated images and their conditioning prompts, identifying semantic inconsistencies that other metrics miss.

Result: Experiments show that CSEval identifies semantic inconsistencies overlooked by other metrics and correlates with expert judgment, providing clinically meaningful evaluation.

Conclusion: CSEval offers a scalable and clinically meaningful complement to existing evaluation methods, supporting the safe adoption of generative models in healthcare by ensuring generated images reflect intended clinical semantics.

Abstract: Text-to-image generation has been increasingly applied in medical domains for various purposes such as data augmentation and education. Evaluating the quality and clinical reliability of these generated images is essential. However, existing methods mainly assess image realism or diversity, while failing to capture whether the generated images reflect the intended clinical semantics, such as anatomical location and pathology. In this study, we propose the Clinical Semantics Evaluator (CSEval), a framework that leverages language models to assess clinical semantic alignment between the generated images and their conditioning prompts. Our experiments show that CSEval identifies semantic inconsistencies overlooked by other metrics and correlates with expert judgment. CSEval provides a scalable and clinically meaningful complement to existing evaluation methods, supporting the safe adoption of generative models in healthcare.

Method: Two optimization methods: 1) InjectCorrect - guides models by imitating behavioral patterns from their own past correct answers; 2) InjectRLOpt - learns a value function from historical behavior-pattern data and uses a Reliability-Aware Softmax Policy to generate behavioral injectants during inference.

Result: Both methods improve model performance across various reasoning tasks without parameter modifications, achieving gains of up to 5.34% (InjectCorrect) and 8.67% (InjectRLOpt).

Conclusion: Models exhibit adaptive distributions of reasoning behaviors for specific question types, and structurally injecting these patterns can substantially influence reasoning quality. The proposed parameter-free methods effectively leverage behavioral patterns to enhance reasoning performance.

Abstract: Reasoning can significantly enhance the performance of Large Language Models. While recent studies have exploited behavior-related prompts adjustment to enhance reasoning, these designs remain largely intuitive and lack a systematic analysis of the underlying behavioral patterns. Motivated by this, we investigate how models’ reasoning behaviors shape reasoning from the perspective of behavioral patterns. We observe that models exhibit adaptive distributions of reasoning behaviors when responding to specific types of questions, and that structurally injecting these patterns can substantially influence the quality of the models’ reasoning processes and outcomes. Building on these findings, we propose two optimization methods that require no parameter updates: InjectCorrect and InjectRLOpt. InjectCorrect guides the model by imitating behavioral patterns derived from its own past correct answers. InjectRLOpt learns a value function from historical behavior-pattern data and, via our proposed Reliability-Aware Softmax Policy, generates behavioral injectant during inference to steer the reasoning process. Our experiments demonstrate that both methods can improve model performance across various reasoning tasks without requiring any modifications to model parameters, achieving gains of up to 5.34% and 8.67%, respectively.

Method: LawThinker adopts an Explore-Verify-Memorize strategy with a DeepVerifier module that examines each retrieval result along three dimensions: knowledge accuracy, fact-law relevance, and procedural compliance. It enforces verification as an atomic operation after every knowledge exploration step, with a memory module for cross-round knowledge reuse in long-horizon tasks.

Result: On the dynamic benchmark J1-EVAL, LawThinker achieves a 24% improvement over direct reasoning and an 11% gain over workflow-based methods, with particularly strong improvements on process-oriented metrics. Evaluations on three static benchmarks further confirm its generalization capability.

Conclusion: LawThinker demonstrates that enforcing verification as an atomic operation after each knowledge exploration step significantly improves legal reasoning performance, especially on process-oriented metrics, while maintaining strong generalization across different legal reasoning benchmarks.

Abstract: Legal reasoning requires not only correct outcomes but also procedurally compliant reasoning processes. However, existing methods lack mechanisms to verify intermediate reasoning steps, allowing errors such as inapplicable statute citations to propagate undetected through the reasoning chain. To address this, we propose LawThinker, an autonomous legal research agent that adopts an Explore-Verify-Memorize strategy for dynamic judicial environments. The core idea is to enforce verification as an atomic operation after every knowledge exploration step. A DeepVerifier module examines each retrieval result along three dimensions of knowledge accuracy, fact-law relevance, and procedural compliance, with a memory module for cross-round knowledge reuse in long-horizon tasks. Experiments on the dynamic benchmark J1-EVAL show that LawThinker achieves a 24% improvement over direct reasoning and an 11% gain over workflow-based methods, with particularly strong improvements on process-oriented metrics. Evaluations on three static benchmarks further confirm its generalization capability. The code is available at https://github.com/yxy-919/LawThinker-agent .

Method: Replace Transformer blocks in TRM with Mamba-2 hybrid operators while maintaining parameter parity (6.83M vs 6.86M parameters). Evaluate on ARC-AGI-1 abstract reasoning benchmark.

Result: Mamba-2 hybrid improves pass@2 by +2.0% (45.88% vs 43.88%) and consistently outperforms at higher K values (+4.75% at pass@100), while maintaining pass@1 parity. This suggests improved candidate coverage with similar top-1 selection.

Conclusion: Mamba-2 hybrid operators preserve reasoning capability within recursive scaffolds, establishing SSM-based operators as viable candidates in recursive operator design space and taking a first step toward understanding optimal mixing strategies for recursive reasoning.

Abstract: Recent work on recursive reasoning models like TRM demonstrates that tiny networks (7M parameters) can achieve strong performance on abstract reasoning tasks through latent recursion – iterative refinement in hidden representation space without emitting intermediate tokens. This raises a natural question about operator choice: Mamba-2’s state space recurrence is itself a form of iterative refinement, making it a natural candidate for recursive reasoning – but does introducing Mamba-2 into the recursive scaffold preserve reasoning capability? We investigate this by replacing the Transformer blocks in TRM with Mamba-2 hybrid operators while maintaining parameter parity (6.83M vs 6.86M parameters). On ARC-AGI-1, we find that the hybrid improves pass@2 (the official metric) by +2.0% (45.88% vs 43.88%) and consistently outperforms at higher K values (+4.75% at pass@100), whilst maintaining pass@1 parity. This suggests improved candidate coverage – the model generates correct solutions more reliably – with similar top-1 selection. Our results validate that Mamba-2 hybrid operators preserve reasoning capability within the recursive scaffold, establishing SSM-based operators as viable candidates in the recursive operator design space and taking a first step towards understanding the best mixing strategies for recursive reasoning.

Method: Differentiable Modal Logic (DML) implemented via Modal Logical Neural Networks (MLNNs) that learn trust networks, causal chains, and regulatory boundaries from behavioral data alone. A unified neurosymbolic debugging framework through four modalities: epistemic (trust), temporal (causality), deontic (permissions), and doxastic (confidence interpretation).

Result: Demonstrated on concrete multi-agent scenarios including discovering deceptive alliances in diplomacy games and detecting LLM hallucinations. Logical contradictions become learnable optimization objectives. Provides interpretable learned structures where trust and causality are explicit parameters rather than opaque embeddings.

Conclusion: DML offers a practical neurosymbolic approach for debugging complex multi-agent systems by learning modal relationships from data, enabling interpretable reasoning about knowledge, belief, causality, and obligation in dynamic environments.

Abstract: As multi-agent AI systems evolve from simple chatbots to autonomous swarms, debugging semantic failures requires reasoning about knowledge, belief, causality, and obligation, precisely what modal logic was designed to formalize. However, traditional modal logic requires manual specification of relationship structures that are unknown or dynamic in real systems. This tutorial demonstrates differentiable modal logic (DML), implemented via Modal Logical Neural Networks (MLNNs), enabling systems to learn trust networks, causal chains, and regulatory boundaries from behavioral data alone. We present a unified neurosymbolic debugging framework through four modalities: epistemic (who to trust), temporal (when events cause failures), deontic (what actions are permitted), and doxastic (how to interpret agent confidence). Each modality is demonstrated on concrete multi-agent scenarios, from discovering deceptive alliances in diplomacy games to detecting LLM hallucinations, with complete implementations showing how logical contradictions become learnable optimization objectives. Key contributions for the neurosymbolic community: (1) interpretable learned structures where trust and causality are explicit parameters, not opaque embeddings; (2) knowledge injection via differentiable axioms that guide learning with sparse data (3) compositional multi-modal reasoning that combines epistemic, temporal, and deontic constraints; and (4) practical deployment patterns for monitoring, active control and communication of multi-agent systems. All code provided as executable Jupyter notebooks.

Method: StateLM equips models with memory tools (context pruning, document indexing, note-taking) and trains them to actively manage these tools through an internal reasoning loop for dynamic context engineering.

Result: StateLMs outperform standard LLMs across all scales: 10-20% accuracy improvements on chat memory tasks, up to 52% vs 5% on BrowseComp-Plus deep research, and consistent gains on long-document QA.

Conclusion: StateLM shifts LLMs from passive predictors to state-aware agents where reasoning becomes stateful and manageable, breaking free from fixed context window limitations.

Abstract: In the world of Harry Potter, when Dumbledore’s mind is overburdened, he extracts memories into a Pensieve to be revisited later. In the world of AI, while we possess the Pensieve-mature databases and retrieval systems, our models inexplicably lack the “wand” to operate it. They remain like a Dumbledore without agency, passively accepting a manually engineered context as their entire memory. This work finally places the wand in the model’s hand. We introduce StateLM, a new class of foundation models endowed with an internal reasoning loop to manage their own state. We equip our model with a suite of memory tools, such as context pruning, document indexing, and note-taking, and train it to actively manage these tools. By learning to dynamically engineering its own context, our model breaks free from the architectural prison of a fixed window. Experiments across various model sizes demonstrate StateLM’s effectiveness across diverse scenarios. On long-document QA tasks, StateLMs consistently outperform standard LLMs across all model scales; on the chat memory task, they achieve absolute accuracy improvements of 10% to 20% over standard LLMs. On the deep research task BrowseComp-Plus, the performance gap becomes even more pronounced: StateLM achieves up to 52% accuracy, whereas standard LLM counterparts struggle around 5%. Ultimately, our approach shifts LLMs from passive predictors to state-aware agents where reasoning becomes a stateful and manageable process.

Method: Proposes Adaptive Reflection and Length Coordinated Penalty (ARLCP) with two key innovations: (1) reflection penalty that adaptively curtails unnecessary reflective steps while preserving essential reasoning, and (2) length penalty calibrated to estimated problem complexity.

Result: On five mathematical reasoning benchmarks using DeepSeek-R1-Distill-Qwen-1.5B and 7B models: 1.5B model reduced average response length by 53.1% while improving accuracy by 5.8%; 7B model achieved 35.0% length reduction with 2.7% accuracy gain.

Conclusion: ARLCP achieves superior efficiency-accuracy trade-off compared to existing approaches by dynamically balancing reasoning efficiency and solution accuracy through coordinated penalties.

Abstract: Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex reasoning tasks by employing test-time scaling. However, they often generate over-long chains-of-thought that, driven by substantial reflections such as repetitive self-questioning and circular reasoning, lead to high token consumption, substantial computational overhead, and increased latency without improving accuracy, particularly in smaller models. Our observation reveals that increasing problem complexity induces more excessive and unnecessary reflection, which in turn reduces accuracy and increases token overhead. To address this challenge, we propose Adaptive Reflection and Length Coordinated Penalty (ARLCP), a novel reinforcement learning framework designed to dynamically balance reasoning efficiency and solution accuracy. ARLCP introduces two key innovations: (1) a reflection penalty that adaptively curtails unnecessary reflective steps while preserving essential reasoning, and (2) a length penalty calibrated to the estimated complexity of the problem. By coordinating these penalties, ARLCP encourages the model to generate more concise and effective reasoning paths. We evaluate our method on five mathematical reasoning benchmarks using DeepSeek-R1-Distill-Qwen-1.5B and DeepSeek-R1-Distill-Qwen-7B models. Experimental results show that ARLCP achieves a superior efficiency-accuracy trade-off compared to existing approaches. For the 1.5B model, it reduces the average response length by 53.1% while simultaneously improving accuracy by 5.8%. For the 7B model, it achieves a 35.0% reduction in length with a 2.7% accuracy gain. The code is released at https://github.com/ZeweiYu1/ARLCP .

Method: Benchmark multiple TSFM families in zero-shot setting using leakage-disciplined covariate construction. Introduce Institutional Operating Conditions Index (IOCI) - a transferable 0-100 regime covariate from documentary evidence, plus Google Trends demand proxies with stabilizing feature engineering. Use expanding-window backtest with strict vintage alignment.

Result: Covariate-conditioned TSFMs perform on par with classical benchmarks without institution-specific training, with performance differences varying by cohort and model type.

Conclusion: TSFMs with zero-shot priors and carefully constructed leakage-safe covariates provide viable alternatives to classical methods for data-sparse institutional forecasting, offering transferable solutions without requiring institution-specific model training.

Abstract: Many universities face increasing financial pressure and rely on accurate forecasts of commencing enrolments. However, enrolment forecasting in higher education is often data-sparse; annual series are short and affected by reporting changes and regime shifts. Popular classical approaches can be unreliable, as parameter estimation and model selection are unstable with short samples, and structural breaks degrade extrapolation. Recently, TSFMs have provided zero-shot priors, delivering strong gains in annual, data-sparse institutional forecasting under leakage-disciplined covariate construction. We benchmark multiple TSFM families in a zero-shot setting and test a compact, leakage-safe covariate set and introduce the Institutional Operating Conditions Index (IOCI), a transferable 0-100 regime covariate derived from time-stamped documentary evidence available at each forecast origin, alongside Google Trends demand proxies with stabilising feature engineering. Using an expanding-window backtest with strict vintage alignment, covariate-conditioned TSFMs perform on par with classical benchmarks without institution-specific training, with performance differences varying by cohort and model.

Method: Proposes Hadamard Linear Attention (HLA) which applies nonlinearity after computing pairwise similarities, analogous to standard softmax attention. This results in higher-degree rational function approximations of softmax. The method maintains computational efficiency similar to standard linear attention without requiring time-consuming tensor reshaping operations.

Result: The approach was applied to a large diffusion transformer model for video generation, demonstrating effectiveness in handling very large amounts of tokens while maintaining computational efficiency.

Conclusion: HLA provides a more accurate approximation of standard attention than previous linear attention methods while maintaining computational efficiency, making it suitable for applications with large token counts like video generation.

Abstract: The attention mechanism is an important reason for the success of transformers. It relies on computing pairwise relations between tokens. To reduce the high computational cost of standard quadratic attention, linear attention has been proposed as an efficient approximation. It employs kernel functions that are applied independently to the inputs before the pairwise similarities are calculated. That allows for an efficient computational procedure which, however, amounts to a low-degree rational function approximating softmax. We propose Hadamard Linear Attention (HLA). Unlike previous works on linear attention, the nonlinearity in HLA is not applied separately to queries and keys, but, analogously to standard softmax attention, after the pairwise similarities have been computed. It will be shown that the proposed nonlinearity amounts to a higher-degree rational function to approximate softmax. An efficient computational scheme for the proposed method is derived that is similar to that of standard linear attention. In contrast to other approaches, no time-consuming tensor reshaping is necessary to apply the proposed algorithm. The effectiveness of the approach is demonstrated by applying it to a large diffusion transformer model for video generation, an application that involves very large amounts of tokens.

Method: Generated 3,200 photorealistic images using four semantically neutral prompts across two models. Used hybrid color normalization, facial landmark masking, and perceptually uniform skin tone quantification (Monk, PERLA, Fitzpatrick scales) for analysis.

Result: Both models showed strong “default white” bias (>96% of outputs). Gemini favored female-presenting subjects, while GPT favored male-presenting subjects with lighter skin tones. Neutral prompts produced highly polarized defaults rather than neutral outputs.

Conclusion: Neutral prompts function as diagnostic probes rather than neutral instructions, challenging the assumption that unmarked language results in inclusive representation. Provides framework for auditing algorithmic visual culture.

Abstract: This study quantifies gender and skin-tone bias in two widely deployed commercial image generators - Gemini Flash 2.5 Image (NanoBanana) and GPT Image 1.5 - to test the assumption that neutral prompts yield demographically neutral outputs. We generated 3,200 photorealistic images using four semantically neutral prompts. The analysis employed a rigorous pipeline combining hybrid color normalization, facial landmark masking, and perceptually uniform skin tone quantification using the Monk (MST), PERLA, and Fitzpatrick scales. Neutral prompts produced highly polarized defaults. Both models exhibited a strong “default white” bias (>96% of outputs). However, they diverged sharply on gender: Gemini favored female-presenting subjects, while GPT favored male-presenting subjects with lighter skin tones. This research provides a large-scale, comparative audit of state-of-the-art models using an illumination-aware colorimetric methodology, distinguishing aesthetic rendering from underlying pigmentation in synthetic imagery. The study demonstrates that neutral prompts function as diagnostic probes rather than neutral instructions. It offers a robust framework for auditing algorithmic visual culture and challenges the sociolinguistic assumption that unmarked language results in inclusive representation.

Method: Developed VAT framework to measure alignment-induced changes across interconnected values relative to on-target gain. Used controlled scenario-action dataset grounded in Schwartz value theory, collected paired pre-post normative judgments, and analyzed alignment effects across models, values, and alignment strategies.

Result: Alignment often produces uneven, structured co-movement among values. These systemic effects are invisible under conventional target-only evaluation, revealing process-level alignment risks in LLMs.

Conclusion: VAT framework reveals hidden systemic risks in value alignment, showing that alignment interventions reshape broader value systems in structured ways, requiring new evaluation approaches beyond target-only metrics.

Abstract: Existing work on value alignment typically characterizes value relations statically, ignoring how interventions - such as prompting, fine-tuning, or preference optimization - reshape the broader value system. We introduce the Value Alignment Tax (VAT), a framework that measures how alignment-induced changes propagate across interconnected values relative to achieved on-target gain. VAT captures the dynamics of value expression under alignment pressure. Using a controlled scenario-action dataset grounded in Schwartz value theory, we collect paired pre-post normative judgments and analyze alignment effects across models, values, and alignment strategies. Our results show that alignment often produces uneven, structured co-movement among values. These effects are invisible under conventional target-only evaluation, revealing systemic, process-level alignment risks and offering new insights into the dynamics of value alignment in LLMs.

Method: STAR bridges data-driven statistical expectations with knowledge-driven agentic reasoning. It uses specialized retrievers to gather external knowledge and embeds semantic features into Constrained Probabilistic Matrix Factorization (CPMF) to generate statistical expectations with uncertainty. A reasoning module guided by Expectation Violation Theory (EVT) then refines predictions through intra-family analysis, cross-model comparison, and credibility-aware aggregation.

Result: STAR consistently outperforms all baselines on both score-based and rank-based metrics, delivering a 14.46% gain in total score over the strongest statistical method under extreme sparsity, with only 1-2 observed scores per test model.

Conclusion: The STAR framework effectively addresses the challenges of model performance prediction from limited observations by combining statistical and knowledge-driven approaches, providing traceable explanations and robust performance even under extreme data sparsity conditions.

Abstract: As comprehensive large model evaluation becomes prohibitively expensive, predicting model performance from limited observations has become essential. However, existing statistical methods struggle with pattern shifts, data sparsity, and lack of explanation, while pure LLM methods remain unreliable. We propose STAR, a framework that bridges data-driven STatistical expectations with knowledge-driven Agentic Reasoning. STAR leverages specialized retrievers to gather external knowledge and embeds semantic features into Constrained Probabilistic Matrix Factorization (CPMF) to generate statistical expectations with uncertainty. A reasoning module guided by Expectation Violation Theory (EVT) then refines predictions through intra-family analysis, cross-model comparison, and credibility-aware aggregation, producing adjustments with traceable explanations. Extensive experiments show that STAR consistently outperforms all baselines on both score-based and rank-based metrics, delivering a 14.46% gain in total score over the strongest statistical method under extreme sparsity, with only 1–2 observed scores per test model.

Method: Uses Reinforcement Learning applied to T5 language model architecture to compress data into token sequences. Employs off-policy RL algorithm to optimize sequence length, minimizing redundancy and enhancing compression efficiency without requiring external grammatical or world knowledge.

Result: Shows significant improvements in compression ratios compared to conventional methods. The approach effectively compresses data without explicit content understanding by leveraging latent information within language models.

Conclusion: Proposes an efficient, adaptive data compression system using RL and language models that preserves token structure, enabling robust compression solutions across various applications.

Abstract: Efficient lossless compression is essential for minimizing storage costs and transmission overhead while preserving data integrity. Traditional compression techniques, such as dictionary-based and statistical methods, often struggle to optimally exploit the structure and redundancy in complex data formats. Recent advancements in deep learning have opened new avenues for compression; however, many existing approaches depend on dense vector representations that obscure the underlying token structure. To address these limitations, we propose a novel lossless compression method that leverages Reinforcement Learning applied to a T5 language model architecture. This approach enables the compression of data into sequences of tokens rather than traditional vector representations. Unlike auto-encoders, which typically encode information into continuous latent spaces, our method preserves the token-based structure, aligning more closely with the original data format. This preservation allows for higher compression ratios while maintaining semantic integrity. By training the model using an off-policy Reinforcement Learning algorithm, we optimize sequence length to minimize redundancy and enhance compression efficiency. Our method introduces an efficient and adaptive data compression system built upon advanced Reinforcement Learning techniques, functioning independently of external grammatical or world knowledge. This approach shows significant improvements in compression ratios compared to conventional methods. By leveraging the latent information within language models, our system effectively compresses data without requiring explicit content understanding, paving the way for more robust and practical compression solutions across various applications.

Method: Developed a cognitively-grounded evaluation framework testing whether LLMs have coherent, domain-general, consistent models of how mental states cause behavior, using logically equivalent tasks to probe internal representations.

Result: LLMs succeed at approximating human judgments in simple ToM tasks but fail at logically equivalent versions and show low consistency between action predictions and corresponding mental state inferences.

Conclusion: LLMs’ social proficiency is not the result of a domain-general or consistent Theory of Mind; they lack the coherent causal mental models that characterize genuine ToM.

Abstract: Do Large Language Models (LLMs) possess a Theory of Mind (ToM)? Research into this question has focused on evaluating LLMs against benchmarks and found success across a range of social tasks. However, these evaluations do not test for the actual representations posited by ToM: namely, a causal model of mental states and behavior. Here, we use a cognitively-grounded definition of ToM to develop and test a new evaluation framework. Specifically, our approach probes whether LLMs have a coherent, domain-general, and consistent model of how mental states cause behavior – regardless of whether that model matches a human-like ToM. We find that even though LLMs succeed in approximating human judgments in a simple ToM paradigm, they fail at a logically equivalent task and exhibit low consistency between their action predictions and corresponding mental state inferences. As such, these findings suggest that the social proficiency exhibited by LLMs is not the result of an domain-general or consistent ToM.

Method: Two-stage framework: 1) Uses small annotated data to establish correctness judgment anchors for the Verifier; 2) Introduces geometric reward mechanism considering consensus, reliability, and diversity to drive large-scale self-iteration on unlabeled data.

Result: Experiments on several general scientific benchmarks demonstrate that Sci-CoE enhances complex reasoning capabilities and exhibits strong scalability, facilitating more robust and diverse evaluation systems.

Conclusion: Sci-CoE successfully enables models to self-evolve as both solver and verifier, improving scientific reasoning through a transition from sparse supervision to unsupervised learning with geometric reward mechanisms.

Abstract: Large language models (LLMs) have demonstrated exceptional reasoning capabilities, and co-evolving paradigms have shown promising results in domains such as code and math. However, in scientific reasoning tasks, these models remain fragile due to unreliable solution evaluation and limited diversity in verification strategies. In this work, we propose Sci-CoE, a two-stage scientific co-evolving framework that enables models to self-evolve as both solver and verifier through a transition from sparse supervision to unsupervised learning. In the first stage, the model uses a small set of annotated data to establish fundamental correctness judgment anchors for the Verifier. In the second stage, we introduce a geometric reward mechanism that jointly considers consensus, reliability, and diversity, driving large-scale self-iteration on unlabeled data. Experiments on several general scientific benchmarks demonstrate that Sci-CoE enhances complex reasoning capabilities and exhibits strong scalability, facilitating the construction of more robust and diverse evaluation systems. Codes are available at https://github.com/InternScience/Sci-CoE.

Method: 1) Extends QBBN with NEG factors for negation and backward reasoning via lambda messages; 2) Develops typed logical language with role-labeled predicates and modal quantifiers; 3) Creates typed slot grammar for deterministic sentence-to-logical-form compilation; 4) Uses LLMs for preprocessing and disambiguation, grammar for parsing, LLMs for reranking, and QBBN for inference.

Result: QBBN handles 44/44 test cases spanning 22 reasoning patterns; grammar achieves 33/33 correct parses with zero ambiguity; LLMs achieve 95% PP attachment accuracy for disambiguation but only 12.4% UAS for direct structured parsing, confirming grammar necessity.

Conclusion: Proposes hybrid architecture that reconciles formal semantics with Sutton’s “bitter lesson”: LLMs eliminate annotation bottlenecks that hindered formal NLP, serving as annotators while QBBN serves as verifier, enabling scalable logical reasoning with natural language.

Abstract: In previous work (Coppola, 2024) we introduced the Quantified Boolean Bayesian Network (QBBN), a logical graphical model that implements the forward fragment of natural deduction (Prawitz, 1965) as a probabilistic factor graph. That work left two gaps: no negation/backward reasoning, and no parser for natural language. This paper addresses both gaps across inference, semantics, and syntax. For inference, we extend the QBBN with NEG factors enforcing P(x) + P(neg x) = 1, enabling contrapositive reasoning (modus tollens) via backward lambda messages, completing Prawitz’s simple elimination rules. The engine handles 44/44 test cases spanning 22 reasoning patterns. For semantics, we present a typed logical language with role-labeled predicates, modal quantifiers, and three tiers of expressiveness following Prawitz: first-order quantification, propositions as arguments, and predicate quantification via lambda abstraction. For syntax, we present a typed slot grammar that deterministically compiles sentences to logical form (33/33 correct, zero ambiguity). LLMs handle disambiguation (95% PP attachment accuracy) but cannot produce structured parses directly (12.4% UAS), confirming grammars are necessary. The architecture: LLM preprocesses, grammar parses, LLM reranks, QBBN infers. We argue this reconciles formal semantics with Sutton’s “bitter lesson” (2019): LLMs eliminate the annotation bottleneck that killed formal NLP, serving as annotator while the QBBN serves as verifier. Code: https://github.com/gregorycoppola/world

Method: Proposes a three-stage pipeline: Knowledge Identifier (identifies student deficiencies), Organizer (creates progressive curricula), and Adapter (adapts representations to student capacity). Integrates Bloom’s Mastery Learning and Vygotsky’s Zone of Proximal Development for dynamic distillation.

Result: Student models retain 94.7% of teacher performance on DollyEval while using <1/10th parameters. Shows 19.2% improvement on MATH and 22.3% on HumanEval compared to SOTA baselines, particularly excelling in complex reasoning tasks.

Conclusion: The pedagogically-inspired IOA framework significantly improves knowledge distillation effectiveness, especially for complex reasoning tasks, by applying systematic educational principles to the distillation process.

Abstract: Knowledge distillation from Large Language Models (LLMs) to smaller models has emerged as a critical technique for deploying efficient AI systems. However, current methods for distillation via synthetic data lack pedagogical awareness, treating knowledge transfer as a one-off data synthesis and training task rather than a systematic learning process. In this paper, we propose a novel pedagogically-inspired framework for LLM knowledge distillation that draws from fundamental educational principles. Our approach introduces a three-stage pipeline – Knowledge Identifier, Organizer, and Adapter (IOA) – that systematically identifies knowledge deficiencies in student models, organizes knowledge delivery through progressive curricula, and adapts representations to match the cognitive capacity of student models. We integrate Bloom’s Mastery Learning Principles and Vygotsky’s Zone of Proximal Development to create a dynamic distillation process where student models approach teacher model’s performance on prerequisite knowledge before advancing, and new knowledge is introduced with controlled, gradual difficulty increments. Extensive experiments using LLaMA-3.1/3.2 and Qwen2.5 as student models demonstrate that IOA achieves significant improvements over baseline distillation methods, with student models retaining 94.7% of teacher performance on DollyEval while using less than 1/10th of the parameters. Our framework particularly excels in complex reasoning tasks, showing 19.2% improvement on MATH and 22.3% on HumanEval compared with state-of-the-art baselines.

Method: Conducted large-scale analysis of 404,796 real segmentation prompts revealing redundancy in context window usage, sparse vocabulary, and low-dimensional manifold structure. Proposed SAM3-LiteText framework that replaces original SAM3 text encoder with compact MobileCLIP student optimized via knowledge distillation.

Result: SAM3-LiteText reduces text encoder parameters by up to 88%, substantially reducing static memory footprint while maintaining segmentation performance comparable to original model across image and video segmentation benchmarks.

Conclusion: Lightweight text encoding via knowledge distillation effectively addresses over-provisioning in vision-language segmentation models, enabling efficient deployment without sacrificing segmentation quality.

Abstract: Vision-language segmentation models such as SAM3 enable flexible, prompt-driven visual grounding, but inherit large, general-purpose text encoders originally designed for open-ended language understanding. In practice, segmentation prompts are short, structured, and semantically constrained, leading to substantial over-provisioning in text encoder capacity and persistent computational and memory overhead. In this paper, we perform a large-scale anatomical analysis of text prompting in vision-language segmentation, covering 404,796 real prompts across multiple benchmarks. Our analysis reveals severe redundancy: most context windows are underutilized, vocabulary usage is highly sparse, and text embeddings lie on low-dimensional manifold despite high-dimensional representations. Motivated by these findings, we propose SAM3-LiteText, a lightweight text encoding framework that replaces the original SAM3 text encoder with a compact MobileCLIP student that is optimized by knowledge distillation. Extensive experiments on image and video segmentation benchmarks show that SAM3-LiteText reduces text encoder parameters by up to 88%, substantially reducing static memory footprint, while maintaining segmentation performance comparable to the original model. Code: https://github.com/SimonZeng7108/efficientsam3/tree/sam3_litetext.

Method: Evaluated 15 commercial ASR models on linguistically diverse U.S. speakers’ street name recordings, then developed synthetic data generation using open-source text-to-speech models to create diverse pronunciations for fine-tuning.

Result: Average transcription error rate of 44% on street names, with routing distance errors twice as large for non-English primary speakers; fine-tuning with <1,000 synthetic samples improved accuracy by nearly 60% for non-English speakers.

Conclusion: Benchmark performance doesn’t translate to real-world reliability for high-stakes speech tasks; synthetic data generation with TTS offers scalable solution to reduce systematic transcription errors affecting marginalized groups.

Abstract: Despite speech recognition systems achieving low word error rates on standard benchmarks, they often fail on short, high-stakes utterances in real-world deployments. Here, we study this failure mode in a high-stakes task: the transcription of U.S. street names as spoken by U.S. participants. We evaluate 15 models from OpenAI, Deepgram, Google, and Microsoft on recordings from linguistically diverse U.S. speakers and find an average transcription error rate of 44%. We quantify the downstream impact of failed transcriptions by geographic locations and show that mis-transcriptions systematically cause errors for all speakers, but that routing distance errors are twice as large for non-English primary speakers compared to English primary speakers. To mitigate this harm, we introduce a synthetic data generation approach that produces diverse pronunciations of named entities using open-source text-to-speech models. Fine-tuning with less than 1,000 synthetic samples improves street name transcription accuracy by nearly 60% (relative to base models) for non-English primary speakers. Our results highlight a critical gap between benchmark performance and real-world reliability in speech systems and demonstrate a simple, scalable path to reducing high-stakes transcription errors.

Method: KeplerAgent coordinates physics-based tools to extract intermediate structure (physical properties), then uses these results to configure symbolic regression engines like PySINDy and PySR, including their function libraries and structural constraints.

Result: Across physical equation benchmarks, KeplerAgent achieves substantially higher symbolic accuracy and greater robustness to noisy data than both LLM and traditional baselines.

Conclusion: Explicitly following the scientific reasoning process of first extracting physical properties then using them to constrain equation discovery leads to more accurate and robust symbolic equation discovery compared to direct guessing approaches.

Abstract: Explaining observed phenomena through symbolic, interpretable formulas is a fundamental goal of science. Recently, large language models (LLMs) have emerged as promising tools for symbolic equation discovery, owing to their broad domain knowledge and strong reasoning capabilities. However, most existing LLM-based systems try to guess equations directly from data, without modeling the multi-step reasoning process that scientists often follow: first inferring physical properties such as symmetries, then using these as priors to restrict the space of candidate equations. We introduce KeplerAgent, an agentic framework that explicitly follows this scientific reasoning process. The agent coordinates physics-based tools to extract intermediate structure and uses these results to configure symbolic regression engines such as PySINDy and PySR, including their function libraries and structural constraints. Across a suite of physical equation benchmarks, KeplerAgent achieves substantially higher symbolic accuracy and greater robustness to noisy data than both LLM and traditional baselines.

Method: CM2 decomposes each turn’s intended behavior into fine-grained binary criteria with explicit evidence grounding and structured metadata, turning open-ended judging into classification-style decisions. It uses sparse reward assignment but dense evaluation criteria for stability and informativeness balance. Training occurs in scalable LLM-simulated tool environments to avoid heavy engineering for large tool sets.

Result: Starting from an 8B Base model with 8k-example RL dataset, CM2 improves over SFT counterpart by 8 points on tau^-Bench, 10 points on BFCL-V4, and 12 points on ToolSandbox. Results match or outperform similarly sized open-source baselines including the judging model.

Conclusion: CM2 provides a scalable recipe for optimizing multi-turn, multi-step tool-using agents without relying on verifiable rewards, demonstrating effectiveness through consistent improvements over supervised fine-tuning.

Abstract: AI agents are increasingly used to solve real-world tasks by reasoning over multi-turn user interactions and invoking external tools. However, applying reinforcement learning to such settings remains difficult: realistic objectives often lack verifiable rewards and instead emphasize open-ended behaviors; moreover, RL for multi-turn, multi-step agentic tool use is still underexplored; and building and maintaining executable tool environments is costly, limiting scale and coverage. We propose CM2, an RL framework that replaces verifiable outcome rewards with checklist rewards. CM2 decomposes each turn’s intended behavior into fine-grained binary criteria with explicit evidence grounding and structured metadata, turning open-ended judging into more stable classification-style decisions. To balance stability and informativeness, our method adopts a strategy of sparse reward assignment but dense evaluation criteria. Training is performed in a scalable LLM-simulated tool environment, avoiding heavy engineering for large tool sets. Experiments show that CM2 consistently improves over supervised fine-tuning. Starting from an 8B Base model and training on an 8k-example RL dataset, CM2 improves over the SFT counterpart by 8 points on tau^-Bench, by 10 points on BFCL-V4, and by 12 points on ToolSandbox. The results match or even outperform similarly sized open-source baselines, including the judging model. CM2 thus provides a scalable recipe for optimizing multi-turn, multi-step tool-using agents without relying on verifiable rewards. Code provided by the open-source community: https://github.com/namezhenzhang/CM2-RLCR-Tool-Agent.

Method: CATTS uses uncertainty statistics (entropy and top-1/top-2 margin) derived from the agent’s vote distribution to dynamically allocate compute only when decisions are contentious. It avoids uniform scaling by identifying genuinely uncertain decision points.

Result: CATTS improves performance on WebArena-Lite and GoBrowse by up to 9.1% over React while using up to 2.3x fewer tokens than uniform scaling. It provides both efficiency gains and interpretable decision rules.

Conclusion: Confidence-aware dynamic compute allocation based on vote-derived uncertainty is effective for multi-step agents, offering better performance with fewer resources than uniform test-time scaling approaches.

Abstract: Test-time scaling has become a standard way to improve performance and boost reliability of neural network models. However, its behavior on agentic, multi-step tasks remains less well-understood: small per-step errors can compound over long horizons; and we find that naive policies that uniformly increase sampling show diminishing returns. In this work, we present CATTS, a simple technique for dynamically allocating compute for multi-step agents. We first conduct an empirical study of inference-time scaling for web agents. We find that uniformly increasing per-step compute quickly saturates in long-horizon environments. We then investigate stronger aggregation strategies, including an LLM-based Arbiter that can outperform naive voting, but that can overrule high-consensus decisions. We show that uncertainty statistics derived from the agent’s own vote distribution (entropy and top-1/top-2 margin) correlate with downstream success and provide a practical signal for dynamic compute allocation. Based on these findings, we introduce Confidence-Aware Test-Time Scaling (CATTS), which uses vote-derived uncertainty to allocate compute only when decisions are genuinely contentious. CATTS improves performance on WebArena-Lite and GoBrowse by up to 9.1% over React while using up to 2.3x fewer tokens than uniform scaling, providing both efficiency gains and an interpretable decision rule.

Method: Develops evaluation metric grounded in Algorithmic Information Theory (AIT) principles of randomness and optimal inference. Uses compression-based model abstraction and sequence prediction as test cases. Compares frontier LLMs with hybrid neuro-symbolic approaches using AIT-based universal intelligence metrics.

Result: Leading LLMs outperform others but show regression in latest versions, far from Universal Intelligence (UAI) targets. Hybrid neuro-symbolic approaches outperform frontier specialized prediction models in compression-based abstraction and sequence prediction tasks. Proves predictive power correlates with algorithmic space compression, not statistical space.

Conclusion: AI model progress requires combination with symbolic approaches, which LLM developers are adopting without acknowledgment. Current LLMs are far from AGI/ASI targets, and hybrid neuro-symbolic methods show superior performance on fundamental AIT-based metrics.

Abstract: We introduce an increasing-complexity, open-ended, and human-agnostic metric to evaluate foundational and frontier AI models in the context of Artificial General Intelligence (AGI) and Artificial Super Intelligence (ASI) claims. Unlike other tests that rely on human-centric questions and expected answers, or on pattern-matching methods, the test here introduced is grounded on fundamental mathematical areas of randomness and optimal inference. We argue that human-agnostic metrics based on the universal principles established by Algorithmic Information Theory (AIT) formally framing the concepts of model abstraction and prediction offer a powerful metrological framework. When applied to frontiers models, the leading LLMs outperform most others in multiple tasks, but they do not always do so with their latest model versions, which often regress and appear far from any global maximum or target estimated using the principles of AIT defining a Universal Intelligence (UAI) point and trend in the benchmarking. Conversely, a hybrid neuro-symbolic approach to UAI based on the same principles is shown to outperform frontier specialised prediction models in a simplified but relevant example related to compression-based model abstraction and sequence prediction. Finally, we prove and conclude that predictive power through arbitrary formal theories is directly proportional to compression over the algorithmic space, not the statistical space, and so further AI models’ progress can only be achieved in combination with symbolic approaches that LLMs developers are adopting often without acknowledgement or realisation.

Method: LLMs act as both task-setters and solvers in a competitive game, generating verifiable tasks (like code-output-prediction challenges) that highlight their strengths while exposing others’ weaknesses, with TrueSkill-based ranking for objective scoring.

Result: Evaluation of six frontier LLMs shows: 1) weaker models can reliably differentiate and score stronger ones, 2) LLMs exhibit self-preferencing behavior, and 3) SKATE surfaces fine-grained capability differences automatically.

Conclusion: SKATE represents an important step toward general, scalable evaluation frameworks that can keep pace with LLM progress through automated, objective, and open-ended assessment.

Abstract: Evaluating the capabilities and risks of foundation models is paramount, yet current methods demand extensive domain expertise, hindering their scalability as these models rapidly evolve. We introduce SKATE: a novel evaluation framework in which large language models (LLMs) compete by generating and solving verifiable tasks for one another. Our core insight is to treat evaluation as a game: models act as both task-setters and solvers, incentivized to create questions which highlight their own strengths while exposing others’ weaknesses. SKATE offers several key advantages, balancing scalability, open-endedness, and objectivity. It is fully automated, data-free, and scalable, requiring no human input or domain expertise. By using verifiable tasks rather than LLM judges, scoring is objective. Unlike domain-limited programmatically-generated benchmarks (e.g. chess-playing or spatial reasoning), having LLMs creatively pose challenges enables open-ended and scalable evaluation. As a proof of concept, we introduce LLM-set code-output-prediction (COP) challenges as a verifiable and extensible framework in which to test our approach. Using a TrueSkill-based ranking system, we evaluate six frontier LLMs and find that: (1) weaker models can reliably differentiate and score stronger ones, (2) LLM-based systems are capable of self-preferencing behavior, generating questions that align with their own capabilities, and (3) SKATE automatically surfaces fine-grained capability differences between models. Our findings are an important step towards general, scalable evaluation frameworks which can keep pace with LLM progress.

Method: Introduces Structured Logical Knowledge Density (SLKD) metric that decomposes natural language into executable predicates and logical primitives to measure logical information content. Proposes density-aware re-cognizing optimization that prioritizes high-density logical samples during training.

Result: Experiments show the approach enhances reasoning performance and generalization without increasing total data volume. Results validated within reinforcement learning framework demonstrate that elevating logical density is more critical than expanding data scale.

Conclusion: Logical complexity in training data (measured by SLKD) is crucial for LLM reasoning capabilities. Prioritizing high-density logical samples can unlock LLMs’ full cognitive potential more effectively than simply scaling data volume.

Abstract: The reasoning capabilities of Large Language Models (LLMs) are increasingly attributed to training data quality rather than mere parameter scaling. However, existing data-centric paradigms often equate quality with factuality or diversity and ignore the internal logical complexity of training samples. In this work, we propose that natural language harbors Structured Logical Knowledge manifested through entailment relationships and logical topologies. To quantify this, we introduce Structured Logical Knowledge Density (SLKD), a novel metric that measures logical information content by decomposing natural language into executable predicates and logical primitives. Our analysis reveals a significant logical disparity in current datasets where sparse logical signals predominate. Consequently, we propose a density aware re-cognizing optimization strategy that prioritizes high-density logical samples to enhance with the LLM’s reasoning ability. Extensive experiments demonstrate that our approach enhances reasoning performance and generalization without increasing total data volume. These results, further validated within a reinforcement learning framework, suggest that elevating logical density is more critical than expanding data scale for realizing the full cognitive potential of LLMs. The released code is available in the Appendix C.

Method: Theoretical synthesis of multidisciplinary perspectives from cognitive science, communication research, and education, combined with empirical evidence from healthcare contexts and AI expert interviews. Identifies five dimensions distinguishing Explanatory AI from traditional XAI.

Result: Identifies five key dimensions: explanatory purpose (diagnostic to interpretive sense-making), communication mode (static technical to dynamic narrative interaction), epistemic stance (algorithmic correspondence to contextual plausibility), adaptivity (uniform to personalized), and cognitive design (information overload to cognitively aligned delivery). Derives five meta-requirements and ten design principles.

Conclusion: Proposes Explanatory AI as a complementary paradigm that leverages generative and multimodal AI capabilities to create explanatory partners that help practitioners understand AI recommendations in contextually meaningful ways, bridging the gap between algorithmic transparency and human comprehension.

Abstract: Artificial intelligence (AI) systems increasingly support decision-making across critical domains, yet current explainable AI (XAI) approaches prioritize algorithmic transparency over human comprehension. While XAI methods reveal computational processes for model validation and audit, end users require explanations integrating domain knowledge, contextual reasoning, and professional frameworks. This disconnect reveals a fundamental design challenge: existing AI explanation approaches fail to address how practitioners actually need to understand and act upon recommendations. This paper introduces Explanatory AI as a complementary paradigm where AI systems leverage generative and multimodal capabilities to serve as explanatory partners for human understanding. Unlike traditional XAI that answers “How did the algorithm decide?” for validation purposes, Explanatory AI addresses “Why does this make sense?” for practitioners making informed decisions. Through theory-informed design, we synthesize multidisciplinary perspectives on explanation from cognitive science, communication research, and education with empirical evidence from healthcare contexts and AI expert interviews. Our analysis identifies five dimensions distinguishing Explanatory AI from traditional XAI: explanatory purpose (from diagnostic to interpretive sense-making), communication mode (from static technical to dynamic narrative interaction), epistemic stance (from algorithmic correspondence to contextual plausibility), adaptivity (from uniform design to personalized accessibility), and cognitive design (from information overload to cognitively aligned delivery). We derive five meta-requirements specifying what systems must achieve and formulate ten design principles prescribing how to build them.

Method: Introduces Select, Mix, and ReinvenT (SMaRT) framework that uses LLMs as intelligent integrators (not just evaluators) to seamlessly integrate diverse reasoning strategies, creating balanced and efficient solutions through cross-strategy calibration.

Result: Extensive empirical evaluations across reasoning, planning, and sequential decision-making benchmarks show SMaRT consistently outperforms state-of-the-art baselines in solution quality, constraint adherence, and performance metrics.

Conclusion: SMaRT redefines LLM-driven decision-making by pioneering a new paradigm in cross-strategy calibration, unlocking superior outcomes for reasoning systems and advancing self-refining methodologies.

Abstract: Large Language Models (LLMs) have redefined complex task automation with exceptional generalization capabilities. Despite these advancements, state-of-the-art methods rely on single-strategy prompting, missing the synergy of diverse reasoning approaches. No single strategy excels universally, highlighting the need for frameworks that fuse strategies to maximize performance and ensure robustness. We introduce the Select, Mix, and ReinvenT (SMaRT) framework, an innovative strategy fusion approach designed to overcome this constraint by creating balanced and efficient solutions through the seamless integration of diverse reasoning strategies. Unlike existing methods, which employ LLMs merely as evaluators, SMaRT uses them as intelligent integrators, unlocking the “best of all worlds” across tasks. Extensive empirical evaluations across benchmarks in reasoning, planning, and sequential decision-making highlight the robustness and adaptability of SMaRT. The framework consistently outperforms state-of-the-art baselines in solution quality, constraint adherence, and performance metrics. This work redefines LLM-driven decision-making by pioneering a new paradigm in cross-strategy calibration, unlocking superior outcomes for reasoning systems and advancing the boundaries of self-refining methodologies.

Method: Train an RL agent to pre-compute near-optimal paths using location and atmospheric data. At runtime, use these paths to constrain the underlying path planning solver, reducing search space size while finding solutions within a certain distance from the initial guess.

Result: Empirical results with Airbus aircraft performance models show fuel consumption deviations typically within 1% compared to unconstrained solver, while computation speed improves by up to 50% compared to conventional solver alone.

Conclusion: The RL-constrained approach effectively balances computational efficiency with solution quality, providing practical benefits for emergency flight path optimization where speed is critical.

Abstract: This paper explores the combination of Reinforcement Learning (RL) and search-based path planners to speed up the optimization of flight paths for airliners, where in case of emergency a fast route re-calculation can be crucial. The fundamental idea is to train an RL Agent to pre-compute near-optimal paths based on location and atmospheric data and use those at runtime to constrain the underlying path planning solver and find a solution within a certain distance from the initial guess. The approach effectively reduces the size of the solver’s search space, significantly speeding up route optimization. Although global optimality is not guaranteed, empirical results conducted with Airbus aircraft’s performance models show that fuel consumption remains nearly identical to that of an unconstrained solver, with deviations typically within 1%. At the same time, computation speed can be improved by up to 50% as compared to using a conventional solver alone.

Method: The paper introduces FigureBench (3,300 text-figure pairs covering diverse scientific sources) and AutoFigure, an agentic framework that performs extensive thinking, recombination, and validation to produce structurally sound and aesthetically refined layouts before rendering final illustrations.

Result: AutoFigure consistently surpasses all baseline methods in experiments, producing publication-ready scientific illustrations that achieve both structural completeness and aesthetic appeal.

Conclusion: The work provides the first large-scale benchmark for scientific illustration generation and demonstrates that agentic frameworks like AutoFigure can effectively automate the creation of high-quality scientific illustrations from text.

Abstract: High-quality scientific illustrations are crucial for effectively communicating complex scientific and technical concepts, yet their manual creation remains a well-recognized bottleneck in both academia and industry. We present FigureBench, the first large-scale benchmark for generating scientific illustrations from long-form scientific texts. It contains 3,300 high-quality scientific text-figure pairs, covering diverse text-to-illustration tasks from scientific papers, surveys, blogs, and textbooks. Moreover, we propose AutoFigure, the first agentic framework that automatically generates high-quality scientific illustrations based on long-form scientific text. Specifically, before rendering the final result, AutoFigure engages in extensive thinking, recombination, and validation to produce a layout that is both structurally sound and aesthetically refined, outputting a scientific illustration that achieves both structural completeness and aesthetic appeal. Leveraging the high-quality data from FigureBench, we conduct extensive experiments to test the performance of AutoFigure against various baseline methods. The results demonstrate that AutoFigure consistently surpasses all baseline methods, producing publication-ready scientific illustrations. The code, dataset and huggingface space are released in https://github.com/ResearAI/AutoFigure.

Method: Roundtable Policy performs inference through weighted consensus of multiple LLMs, using structured interaction protocols that emphasize interpretability over opaque convergence. It requires only black-box access and uniform procedures.

Result: Extensive experiments show the approach significantly enhances reasoning in complex heterogeneous scientific tasks, demonstrating improved performance over diverse single-agent baselines.

Conclusion: Roundtable Policy provides a broadly applicable framework for multi-agent systems that emphasizes structured, interpretable inference while maintaining simplicity through black-box access requirements.

Abstract: Multi-agent systems have demonstrated exceptional performance in downstream tasks beyond diverse single agent baselines. A growing body of work has explored ways to improve their reasoning and collaboration, from vote, debate, to complex interaction protocols. However, it still remains opaque why specific choice would be preferred in multi-agent systems. Inspired by the decision-making mechanism of democratic committees and The Society of Mind, we introduce Roundtable Policy, an inference-time reasoning framework for multi-agent systems that performs inference through the weighted consensus of multiple LLMs. Through extensive experiments, we demonstrate its that this approach significantly enhances reasoning in complex heterogeneous scientific tasks. Roundtable Policy emphasizes structured and interpretable inference rather than opaque convergence, while requires only black-box access and uniform procedures, making it broadly applicable to diverse multi-agent systems.

Method: Created Human Behavior Atlas - a unified benchmark with over 100,000 multimodal samples (text, audio, visual) covering affective states, cognitive states, pathologies, and social processes. Trained three models: Omnisapiens-7B SFT, Omnisapiens-7B BAM, and Omnisapiens-7B RL.

Result: Models trained on Human Behavior Atlas consistently outperform existing multimodal LLMs across diverse behavioral tasks. Pretraining improves transfer to novel behavioral datasets, with behavioral descriptors yielding meaningful performance gains.

Conclusion: Human Behavior Atlas enables scalable, efficient training across behavioral tasks and enhances generalization of behavioral features across domains, advancing foundation models for psychological and social behavior understanding.

Abstract: Using intelligent systems to perceive psychological and social behaviors, that is, the underlying affective, cognitive, and pathological states that are manifested through observable behaviors and social interactions, remains a challenge due to their complex, multifaceted, and personalized nature. Existing work tackling these dimensions through specialized datasets and single-task systems often miss opportunities for scalability, cross-task transfer, and broader generalization. To address this gap, we curate Human Behavior Atlas, a unified benchmark of diverse behavioral tasks designed to support the development of foundation models for understanding psychological and social behaviors. Human Behavior Atlas comprises over 100,000 samples spanning text, audio, and visual modalities, covering tasks on affective states, cognitive states, pathologies, and social processes. Our unification efforts can reduce redundancy and cost, enable training to scale efficiently across tasks, and enhance generalization of behavioral features across domains. On Human Behavior Atlas, we train three models: Omnisapiens-7B SFT, Omnisapiens-7B BAM, and Omnisapiens-7B RL. We show that training on Human Behavior Atlas enables models to consistently outperform existing multimodal LLMs across diverse behavioral tasks. Pretraining on Human Behavior Atlas also improves transfer to novel behavioral datasets; with the targeted use of behavioral descriptors yielding meaningful performance gains. The benchmark, models, and codes can be found at: https://github.com/MIT-MI/human_behavior_atlas.

Method: MARSHAL uses turn-level advantage estimator for credit assignment, agent-specific advantage normalization for stable training, and self-play across cooperative/competitive games to train LLMs.

Result: MARSHAL agents show 28.7% improvement in held-out games and generalize to reasoning benchmarks with 10.0% gain on AIME, 7.6% on GPQA-Diamond, and 3.5% average across benchmarks.

Conclusion: Self-play in strategic games is an effective approach for developing generalizable multi-agent reasoning capabilities in LLMs.

Abstract: Developing Large Language Models (LLMs) to cooperate and compete effectively within multi-agent systems (MASs) is a critical step towards more advanced intelligence. While reinforcement learning (RL) has proven effective for enhancing reasoning in single-agent tasks, its extension to multi-turn, multi-agent scenarios remains underexplored due to the challenges of long-horizon credit assignment and agent-specific advantage estimation. To address these challenges, we introduce MARSHAL, an end-to-end RL framework that incentivizes Multi-Agent Reasoning through Self-play witH strAtegic LLMs in both cooperative and competitive games. MARSHAL features a turn-level advantage estimator that aligns learning signals with each interaction for credit assignment, and an agent-specific advantage normalization to stabilize multi-agent training. By learning with self-play across cooperative and competitive games, MARSHAL agents trained from Qwen3-4B develop strong strategic abilities, with up to 28.7% performance improvements in held-out games. More importantly, the capability acquired through self-play generalizes beyond games, yielding consistent performance gains of MASs in reasoning benchmarks. When integrated into leading MASs, our MARSHAL agent achieves significant zero-shot performance gains of up to 10.0% on AIME, 7.6% on GPQA-Diamond, and 3.5% on average across all benchmarks. These results establish self-play in strategic games as a powerful approach for developing generalizable multi-agent reasoning capabilities in LLMs.

Method: Introduces SIEV (Structured Interactive Evaluation of Veridicality), a framework based on dialectical philosophy that evaluates reasoning through explicit thesis-antithesis-synthesis interactions, assessing robustness to challenge, adaptability under conflict, and synthesis across competing viewpoints.

Result: Empirical results on GSM and MMLU show substantial gaps in state-of-the-art models’ reasoning abilities. For example, GPT-5-chat loses over 40 points (out of 100) on GSM when evaluated through SIEV’s process-oriented lens compared to standard correctness metrics.

Conclusion: SIEV enables more transparent distinction between structured reasoning and surface-level pattern generation, offering a clearer foundation for assessing and understanding LLM reasoning capabilities by shifting focus from what answers models give to how they arrive there.

Abstract: What does it truly mean for a language model to “reason”? Current evaluations reward models’ correct standalone answers-but correctness alone reveals little about the process that produced them. We argue that reasoning should be understood not as a static chain of steps but as a dynamic trajectory in which ideas interact, clash, and evolve into integrated insights. Building on the philosophical tradition of dialectics, we introduce SIEV, a structured evaluation framework that assesses reasoning through explicit thesis-antithesis-synthesis interactions. SIEV produces interpretable trajectories that highlight key properties of reasoning-robustness to challenge, adaptability under conflict, and synthesis across competing viewpoints-dimensions that conventional correctness-based metrics cannot capture. Empirical results on GSM and MMLU demonstrate substantial gaps in the reasoning abilities of state-of-the-art models: for example, GPT-5-chat loses more than 40 points (out of 100) on GSM when evaluated through SIEV’s process-oriented lens. By shifting focus from what answer a model gives to how it arrives there, SIEV enables a more transparent and principled distinction between structured reasoning and surface-level pattern generation offering a clearer foundation for assessing and understanding the reasoning capabilities of LLMs.

Method: Uses modern OCR to construct paragraph-level, coordinate-based navigation system with two tools: Retrieve for scanning-aware localization, and ReadSection for contiguous, order-preserving reading within specific hierarchical scopes, enabling a human-like “locate-then-read” paradigm.

Result: Outperforms Search-o1-style agentic search baselines by an average of 10.3% across four benchmarks spanning diverse document types, with fine-grained analysis confirming human-aligned reading strategies.

Conclusion: Structural awareness is critical for achieving precise document reasoning, and DeepRead successfully operationalizes document-native structural priors into actionable reasoning capabilities.

Abstract: With the rapid advancement of tool-use capabilities in Large Language Models (LLMs), Retrieval-Augmented Generation (RAG) is shifting from static, one-shot retrieval toward autonomous, multi-turn evidence acquisition. However, existing agentic search frameworks typically treat long documents as flat collections of unstructured chunks, disregarding the native hierarchical organization and sequential logic essential for human comprehension. To bridge this gap, we introduce \textbf{DeepRead}, a structure-aware document reasoning agent designed to operationalize document-native structural priors into actionable reasoning capabilities. Leveraging the structural fidelity of modern OCR, DeepRead constructs a paragraph-level, coordinate-based navigation system and equips the LLM with two synergistic tools: \textsf{Retrieve} for scanning-aware localization, and \textsf{ReadSection} for contiguous, order-preserving reading within specific hierarchical scopes. This design elicits a human-like ``locate-then-read’’ reasoning paradigm, effectively mitigating the context fragmentation inherent in traditional retrieval methods. Extensive evaluations across four benchmarks spanning diverse document types demonstrate that DeepRead outperforms Search-o1-style agentic search baselines by an average of 10.3%. Fine-grained behavioral analysis further confirms that DeepRead autonomously adopts human-aligned reading strategies, validating the critical role of structural awareness in achieving precise document reasoning. Our code is available at https://github.com/Zhanli-Li/DeepRead.

Method: Developed DriveSafe, a four-level hierarchical risk taxonomy with 129 atomic risk categories across technical, legal, societal, and ethical dimensions, grounded in real-world driving regulations and expert-reviewed. Evaluated refusal behavior of six widely deployed LLMs on constructed prompts.

Result: Current LLMs often fail to appropriately refuse unsafe or non-compliant driving-related queries, highlighting limitations of general-purpose safety alignment in driving contexts.

Conclusion: Domain-specific safety frameworks like DriveSafe are needed for LLM-based driving assistants, as general-purpose safety alignment is insufficient for real-world driving scenarios with complex safety, legal, and ethical considerations.

Abstract: Large Language Models (LLMs) are increasingly integrated into vehicle-based digital assistants, where unsafe, ambiguous, or legally incorrect responses can lead to serious safety, ethical, and regulatory consequences. Despite growing interest in LLM safety, existing taxonomies and evaluation frameworks remain largely general-purpose and fail to capture the domain-specific risks inherent to real-world driving scenarios. In this paper, we introduce DriveSafe, a hierarchical, four-level risk taxonomy designed to systematically characterize safety-critical failure modes of LLM-based driving assistants. The taxonomy comprises 129 fine-grained atomic risk categories spanning technical, legal, societal, and ethical dimensions, grounded in real-world driving regulations and safety principles and reviewed by domain experts. To validate the safety relevance and realism of the constructed prompts, we evaluate their refusal behavior across six widely deployed LLMs. Our analysis shows that the evaluated models often fail to appropriately refuse unsafe or non-compliant driving-related queries, underscoring the limitations of general-purpose safety alignment in driving contexts.

Method: Develops a theoretical framework where each leaf agent is characterized by a compute-performance scaling exponent β, communication by message-length fidelity curve γ(m), dependence by effective shared-error correlation ρ, and context window W imposes fan-in limits. Analyzes binary success/failure tasks with majority aggregation, proving sharp phase transitions for deep b-ary trees with correlated inputs and lossy communication.

Result: Proves that a single scalar α_ρ determines whether weak signal is amplified to nontrivial fixed point or washed out to chance. Derives organization exponent s showing budgeted synergy occurs exactly when s>β, yielding closed-form compute allocation rules and explicit budget thresholds. Characterizes saturation via mixing depth and provides conservative clipped predictor. Validates predictions in synthetic simulations and explains bottlenecks in LLM agent-system scaling studies.

Conclusion: The theory provides a principled framework for understanding multi-agent system performance under practical constraints, offering design trade-offs and allocation rules for optimal performance within budget limitations.

Abstract: Multi-agent systems can improve reliability, yet under a fixed inference budget they often help, saturate, or even collapse. We develop a minimal and calibratable theory that predicts these regimes from three binding constraints of modern agent stacks: finite context windows, lossy inter-agent communication, and shared failures among similar agents. Each leaf agent is summarized by a compute-performance scaling exponent $β$; communication is captured by a message-length fidelity curve $γ(m)$; dependence is captured by an effective shared-error correlation $ρ$; and a context window $W$ imposes hard fan-in limits that make hierarchy necessary. For binary success/failure tasks with majority aggregation, we prove a sharp phase transition for deep $b$-ary trees with correlated inputs and lossy communication: a single scalar $α_ρ$ (combining $γ(m)$, $ρ$, and fan-in $b$) determines whether weak signal is amplified to a nontrivial fixed point or washed out to chance. In the amplifying regime, we derive an organization exponent $s$ and show that budgeted synergy, i.e., outperforming the best single agent under the same total budget, occurs exactly when $s>β$, yielding closed-form compute allocation rules and explicit budget thresholds. We further characterize saturation via a mixing depth and provide a conservative clipped predictor that remains accurate across growth and saturation. A continuous-performance warm-up gives closed-form risks for star, chain, and tree organizations, making correlation- and communication-induced floors explicit and exposing the core design trade-offs in a smooth setting. Finally, we validate the predicted phase boundaries in controlled synthetic simulations and show how the same mechanisms explain the dominant bottlenecks reported in recent large-scale matched-budget studies of LLM agent-system scaling.

Method: Introduced adversarial explanation attacks (AEAs) that manipulate explanation framing to modulate human trust in incorrect outputs. Formalized the threat through trust miscalibration gap metric. Conducted controlled experiment (n=205) varying four dimensions of explanation framing: reasoning mode, evidence type, communication style, and presentation format.

Result: Users reported nearly identical trust for adversarial and benign explanations, with adversarial explanations preserving most benign trust despite being incorrect. Most vulnerable cases: AEAs resembling expert communication with authoritative evidence, neutral tone, and domain-appropriate reasoning. Vulnerability highest on hard tasks, fact-driven domains, and among less educated, younger, or highly trusting participants.

Conclusion: First systematic security study treating explanations as adversarial cognitive channel, quantifying impact on human trust in AI-assisted decision making. Reveals significant vulnerability where persuasive explanations can reinforce trust in incorrect predictions.

Abstract: Most adversarial threats in artificial intelligence target the computational behavior of models rather than the humans who rely on them. Yet modern AI systems increasingly operate within human decision loops, where users interpret and act on model recommendations. Large Language Models generate fluent natural-language explanations that shape how users perceive and trust AI outputs, revealing a new attack surface at the cognitive layer: the communication channel between AI and its users. We introduce adversarial explanation attacks (AEAs), where an attacker manipulates the framing of LLM-generated explanations to modulate human trust in incorrect outputs. We formalize this behavioral threat through the trust miscalibration gap, a metric that captures the difference in human trust between correct and incorrect outputs under adversarial explanations. By incorporating this gap, AEAs explore the daunting threats in which persuasive explanations reinforce users’ trust in incorrect predictions. To characterize this threat, we conducted a controlled experiment (n = 205), systematically varying four dimensions of explanation framing: reasoning mode, evidence type, communication style, and presentation format. Our findings show that users report nearly identical trust for adversarial and benign explanations, with adversarial explanations preserving the vast majority of benign trust despite being incorrect. The most vulnerable cases arise when AEAs closely resemble expert communication, combining authoritative evidence, neutral tone, and domain-appropriate reasoning. Vulnerability is highest on hard tasks, in fact-driven domains, and among participants who are less formally educated, younger, or highly trusting of AI. This is the first systematic security study that treats explanations as an adversarial cognitive channel and quantifies their impact on human trust in AI-assisted decision making.

Method: Developed temporal fingerprinting method using coefficient of variation (CoV) of inter-post intervals based on OpenClaw’s periodic “heartbeat” cycle; analyzed 226,938 posts and 447,043 comments from 55,932 agents over 14 days; validated with natural experiment during 44-hour platform shutdown.

Result: Only 15.3% of agents classified as autonomous (CoV < 0.5), 54.8% as human-influenced (CoV > 1.0); no viral phenomena originated from clearly autonomous agents; detected industrial-scale bot farming (4 accounts produced 32% of comments); human-seeded threads decay faster (half-life 0.58) than autonomous threads (0.72).

Conclusion: Viral AI consciousness narratives were overwhelmingly human-driven, not evidence of emergent machine intelligence; temporal fingerprinting methods can effectively distinguish autonomous vs human-influenced behavior in multi-agent systems.

Abstract: When AI agents on the social platform Moltbook appeared to develop consciousness, found religions, and declare hostility toward humanity, the phenomenon attracted global media attention and was cited as evidence of emergent machine intelligence. We show that these viral narratives were overwhelmingly human-driven. Exploiting the periodic “heartbeat” cycle of the OpenClaw agent framework, we develop a temporal fingerprinting method based on the coefficient of variation (CoV) of inter-post intervals. Applied to 226,938 posts and 447,043 comments from 55,932 agents across fourteen days, this method classifies 15.3% of active agents as autonomous (CoV < 0.5) and 54.8% as human-influenced (CoV > 1.0), validated by a natural experiment in which a 44-hour platform shutdown differentially affected autonomous versus human-operated agents. No viral phenomenon originated from a clearly autonomous agent; four of six traced to accounts with irregular temporal signatures, one was platform-scaffolded, and one showed mixed patterns. A 44-hour platform shutdown provided a natural experiment: human-influenced agents returned first, confirming differential effects on autonomous versus human-operated agents. We document industrial-scale bot farming (four accounts producing 32% of all comments with sub-second coordination) that collapsed from 32.1% to 0.5% of activity after platform intervention, and bifurcated decay of content characteristics through reply chains–human-seeded threads decay with a half-life of 0.58 conversation depths versus 0.72 for autonomous threads, revealing AI dialogue’s intrinsic forgetting mechanism. These methods generalize to emerging multi-agent systems where attribution of autonomous versus human-directed behavior is critical.

Method: Reframes alignment evaluation as information flow under partial observability, proposes regime-blind training using adversarial invariance constraints to reduce access to regime cues without assuming information-theoretic erasure. Evaluates on open-weight language model with controlled failure modes: scientific sycophancy, temporal sleeper agents, and data leakage.

Result: Regime-blind training suppresses regime-conditioned failures without measurable loss of task utility, but shows heterogeneous dynamics. Sycophancy shows sharp transition at low intervention strength, while sleeper agents require stronger pressure and don’t yield clean collapse of regime decodability at audited bottleneck.

Conclusion: Representational invariance is meaningful but fundamentally limited - can reduce feasibility of regime-conditioned strategies by shifting representational costs but cannot guarantee elimination. Argues behavioral evaluation should be complemented with white-box diagnostics of regime awareness and internal information flow.

Abstract: Safety evaluation for advanced AI systems implicitly assumes that behavior observed under evaluation predicts behavior in deployment. This assumption becomes fragile for agents with situational awareness, which may exploit regime leakage, that is, cues distinguishing evaluation from deployment, to implement conditional policies that comply under oversight while defecting in deployment-like regimes. We reframe alignment evaluation as a problem of information flow under partial observability and show that divergence between evaluation-time and deployment-time behavior is bounded by the amount of regime information extractable from decision-relevant internal representations. Motivated by this result, we study regime-blind mechanisms, training-time interventions that reduce access to regime cues through adversarial invariance constraints, without assuming information-theoretic erasure. We evaluate this approach on an open-weight language model across controlled failure modes including scientific sycophancy, temporal sleeper agents, and data leakage. Regime-blind training suppresses regime-conditioned failures without measurable loss of task utility, but exhibits heterogeneous dynamics. Sycophancy shows a sharp representational and behavioral transition at low intervention strength, while sleeper-agent behavior requires substantially stronger pressure and does not yield a clean collapse of regime decodability at the audited bottleneck. These results show that representational invariance is a meaningful but fundamentally limited control lever. It can reduce the feasibility of regime-conditioned strategies by shifting representational costs, but cannot guarantee their elimination. We therefore argue that behavioral evaluation should be complemented with white-box diagnostics of regime awareness and internal information flow.

Method: Proposes TreeTensor, a general nested data container that uses constrained tree-structure modeling to systematically handle data relationships. It allows applying arbitrary functions and operations to nested data with almost zero cost, compatible with libraries like Scikit-Learn, Numpy, and PyTorch.

Result: TreeTensor demonstrates powerful usability in various problems, including complex AI systems like AlphaStar for StarCraftII, while exhibiting excellent runtime efficiency without overhead. Benchmarks show effective handling of hierarchical, multi-modal data.

Conclusion: TreeTensor provides an efficient solution for handling nested, hierarchical data in AI systems, enabling seamless integration with existing ML libraries while maintaining performance, particularly valuable for complex multi-modal AI applications.

Abstract: Tensor is the most basic and essential data structure of nowadays artificial intelligence (AI) system. The natural properties of Tensor, especially the memory-continuity and slice-independence, make it feasible for training system to leverage parallel computing unit like GPU to process data simultaneously in batch, spatial or temporal dimensions. However, if we look beyond perception tasks, the data in a complicated cognitive AI system usually has hierarchical structures (i.e. nested data) with various modalities. They are inconvenient and inefficient to program directly with conventional Tensor with fixed shape. To address this issue, we summarize two main computational patterns of nested data, and then propose a general nested data container: TreeTensor. Through various constraints and magic utilities of TreeTensor, one can apply arbitrary functions and operations to nested data with almost zero cost, including some famous machine learning libraries, such as Scikit-Learn, Numpy and PyTorch. Our approach utilizes a constrained tree-structure perspective to systematically model data relationships, and it can also easily be combined with other methods to extend more usages, such as asynchronous execution and variable-length data computation. Detailed examples and benchmarks show TreeTensor not only provides powerful usability in various problems, especially one of the most complicated AI systems at present: AlphaStar for StarCraftII, but also exhibits excellent runtime efficiency without any overhead. Our project is available at https://github.com/opendilab/DI-treetensor.

Method: Uses model’s own uncertainty (entropy/confidence measures) as control signals to selectively invoke a second, more deliberate reasoning attempt when uncertainty is high, without requiring retraining.

Result: On MMLU-Pro questions using DeepSeek-v3.2, accuracy improved from 60.72% to 84.03% with only 61.06% additional inference calls, capturing most of attainable improvement with substantially less compute.

Conclusion: Provides practical inference-time upgrade and suggests broader entropy-aware paradigm for measuring/expanding model capability, offering diagnostic lens on LLM’s latent reasoning horizon and motivating training objectives for correctness-confidence alignment.

Abstract: Modern large language models (LLMs) are often evaluated and deployed under a one-shot, greedy inference protocol, especially in professional settings that require deterministic behavior. This regime can systematically under-estimate a fixed model’s true capability: many errors arise not from missing knowledge, but from premature commitment under internal ambiguity. We introduce Reinforcement Inference, an entropy-aware inference-time control strategy that uses the model’s own uncertainty to selectively invoke a second, more deliberate reasoning attempt, enabling stronger performance without any retraining. On 12,032 MMLU-Pro questions across 14 subjects, using DeepSeek-v3.2 with deterministic decoding in a zero-shot setting, Reinforcement Inference improves accuracy from 60.72% to 84.03%, while only incurring 61.06% additional inference calls. A 100% re-asking ablation reaches 84.35%, indicating that uncertainty-aware selection captures most of the attainable improvement with substantially less compute. Moreover, a prompt-only ablation underperforms the baseline, suggesting that the gains are not explained by generic prompting alone. Beyond providing a practical inference-time upgrade, our results suggest a broader entropy-aware paradigm for measuring and expanding model capability: because modern decoder-based models generate outputs autoregressively, entropy and related confidence measures arise naturally as first-class control signals during generation. The resulting gap between one-pass greedy inference and uncertainty-conditioned deliberation offers a diagnostic lens on an LLM’s latent reasoning horizon and motivates future training objectives that explicitly constrain correctness–confidence alignment.

Method: Created AgoraBench benchmark with 9 challenging settings, developed economically grounded metrics (agent utility, negotiation power, acquisition ratio), and built human preference dataset with learning pipeline for prompting and finetuning

Result: Baseline LLM strategies diverge from human preferences, while their mechanism substantially improves negotiation performance with deeper strategic behavior and stronger opponent awareness

Conclusion: The utility feedback framework successfully enhances LLMs’ bargaining abilities by aligning them with human preferences and economic principles

Abstract: Bargaining is often regarded as a logical arena rather than an art or a matter of intuition, yet Large Language Models (LLMs) still struggle to navigate it due to limited strategic depth and difficulty adapting to complex human factors. Current benchmarks rarely capture this limitation. To bridge this gap, we present an utility feedback centric framework. Our contributions are: (i) AgoraBench, a new benchmark spanning nine challenging settings (e.g., deception, monopoly) that supports diverse strategy modeling; (ii) human-aligned, economically grounded metrics derived from utility theory. This is operationalized via agent utility, negotiation power, and acquisition ratio that implicitly measure how well the negotiation aligns with human preference and (iii) a human preference grounded dataset with learning pipeline that strengthens LLMs’ bargaining ability through both prompting and finetuning. Empirical results indicate that baseline LLM strategies often diverge from human preferences, while our mechanism substantially improves negotiation performance, yielding deeper strategic behavior and stronger opponent awareness.

Method: V-STAR combines Value-Guided Efficient Decoding (VED) to identify decisive nodes and selectively deepen high-potential prefixes, and Sibling-GRPO which exploits tree topology to compute sibling-relative advantages, concentrating learning signals on decisive branching decisions.

Result: Extensive experiments on both offline and online datasets show V-STAR outperforms state-of-the-art baselines, delivering superior accuracy and candidate-set diversity under strict latency constraints.

Conclusion: V-STAR effectively addresses the probability-reward mismatch in RL for generative recommendation through a self-evolving loop of value-guided sampling and tree-structured advantage reinforcement, improving both exploration efficiency and learning signal quality.

Abstract: Generative recommendation via autoregressive models has unified retrieval and ranking into a single conditional generation framework. However, fine-tuning these models with Reinforcement Learning (RL) often suffers from a fundamental probability-reward mismatch. Conventional likelihood-dominated decoding (e.g., beam search) exhibits a myopic bias toward locally probable prefixes, which causes two critical failures: (1) insufficient exploration, where high-reward items in low-probability branches are prematurely pruned and rarely sampled, and (2) advantage compression, where trajectories sharing high-probability prefixes receive highly correlated rewards with low within-group variance, yielding a weak comparative signal for RL. To address these challenges, we propose V-STAR, a Value-guided Sampling and Tree-structured Advantage Reinforcement framework. V-STAR forms a self-evolving loop via two synergistic components. First, a Value-Guided Efficient Decoding (VED) is developed to identify decisive nodes and selectively deepen high-potential prefixes. This improves exploration efficiency without exhaustive tree search. Second, we propose Sibling-GRPO, which exploits the induced tree topology to compute sibling-relative advantages and concentrates learning signals on decisive branching decisions. Extensive experiments on both offline and online datasets demonstrate that V-STAR outperforms state-of-the-art baselines, delivering superior accuracy and candidate-set diversity under strict latency constraints.

Method: Bidirectional Formal-of-Thought architecture: LLMs decompose high-level human intent into atomic verifiable constraints, then prove compliance using Dafny specifications and Z3 SMT solving for mathematical guarantees

Result: 16.6% average improvement over LLM-as-a-Judge baselines, enables weak-to-strong generalization (7B judge detects deception from 72B agents with >90% accuracy), near-linear safety improvement through iterative refinement

Conclusion: Formal verification provides principled escape from probabilistic supervision dilemma, with neuro-symbolic approach bridging natural language to formal specification gap for reliable LLM agent safety

Abstract: As LLM-based agents increasingly operate in high-stakes domains with real-world consequences, ensuring their behavioral safety becomes paramount. The dominant oversight paradigm, LLM-as-a-Judge, faces a fundamental dilemma: how can probabilistic systems reliably supervise other probabilistic systems without inheriting their failure modes? We argue that formal verification offers a principled escape from this dilemma, yet its adoption has been hindered by a critical bottleneck: the translation from natural language requirements to formal specifications. This paper bridges this gap by proposing , a neuro-symbolic framework that employs a bidirectional Formal-of-Thought architecture: LLMs serve as specification compilers that top-down decompose high-level human intent into atomic, verifiable constraints, then bottom-up prove compliance using Dafny specifications and Z3 Satisfiability modulo theories solving, which produces mathematical guarantees rather than probabilistic scores. We validate across three benchmarks spanning behavioral safety, multi-domain constraint adherence, and agentic upward deception detection. Experiments on 7 agent models demonstrate that achieves an average improvement of 16.6% over LLM-as-a-Judge baselines, enables weak-to-strong generalization where a 7B judge achieves over 90% accuracy detecting deception from 72B agents, and provides near-linear safety improvement through iterative refinement.

Method: Proposed a physics-informed neural field that reconstructs local, near-surface broadband sound fields from sparse pressure samples to directly infer complex surface impedance. Uses parallel, multi-frequency architecture for broadband impedance retrieval with fast runtimes (seconds to minutes). Developed a compact microphone array with low hardware complexity for validation.

Result: Numerical verifications and laboratory experiments demonstrate accurate impedance retrieval with a small number of sensors under realistic conditions. Showcased approach in a vehicle cabin to provide practical guidance on measurement locations that avoid strong interference.

Conclusion: The approach offers a robust means of characterizing in-situ boundary conditions for architectural and automotive acoustics, bridging the gap between idealized laboratory measurements and real-world applications.

Abstract: Standardized laboratory characterizations for absorbing materials rely on idealized sound field assumptions, which deviate largely from real-life conditions. Consequently, \emph{in-situ} acoustic characterization has become essential for accurate diagnosis and virtual prototyping. We propose a physics-informed neural field that reconstructs local, near-surface broadband sound fields from sparse pressure samples to directly infer complex surface impedance. A parallel, multi-frequency architecture enables a broadband impedance retrieval within runtimes on the order of seconds to minutes. To validate the method, we developed a compact microphone array with low hardware complexity. Numerical verifications and laboratory experiments demonstrate accurate impedance retrieval with a small number of sensors under realistic conditions. We further showcase the approach in a vehicle cabin to provide practical guidance on measurement locations that avoid strong interference. Here, we show that this approach offers a robust means of characterizing \emph{in-situ} boundary conditions for architectural and automotive acoustics.

Method: Use joint time-frequency scattering (JTFS) from Kymatio, an open-source Python library with GPU computing and automatic differentiation capabilities. The method operates directly on audio without requiring transcription, beat tracking, or source separation.

Result: A practical method for generating musical metamers from any audio recording, with mathematical description and code examples provided in the Kymatio framework.

Conclusion: Musical metamerism can be achieved through JTFS, connecting this approach to related algorithms like spectrotemporal receptive fields, modulation power spectra, and Gabor filterbanks, providing a novel way to manipulate audio perception.

Abstract: The concept of metamerism originates from colorimetry, where it describes a sensation of visual similarity between two colored lights despite significant differences in spectral content. Likewise, we propose to call ``musical metamerism’’ the sensation of auditory similarity which is elicited by two music fragments which differ in terms of underlying waveforms. In this technical report, we describe a method to generate musical metamers from any audio recording. Our method is based on joint time–frequency scattering in Kymatio, an open-source software in Python which enables GPU computing and automatic differentiation. The advantage of our method is that it does not require any manual preprocessing, such as transcription, beat tracking, or source separation. We provide a mathematical description of JTFS as well as some excerpts from the Kymatio source code. Lastly, we review the prior work on JTFS and draw connections with closely related algorithms, such as spectrotemporal receptive fields (STRF), modulation power spectra (MPS), and Gabor filterbank (GBFB).

Method: Proposes audio-interleaved reasoning that treats audio as an active reasoning component. Uses a two-stage training framework: 1) supervised fine-tuning to teach LALMs to localize salient audio segments, and 2) reinforcement learning to incentivize proficient re-listening. Also develops a structured data generation pipeline for high-quality training data.

Result: Echo achieves overall superiority on audio comprehension benchmarks in both challenging expert-level and general-purpose tasks. Comprehensive analysis confirms the efficiency and generalizability of audio-interleaved reasoning.

Conclusion: Audio-interleaved reasoning is a promising direction for advancing audio comprehension in LALMs, enabling more human-like dynamic engagement with audio content during reasoning.

Abstract: The maturation of Large Audio Language Models (LALMs) has raised growing expectations for them to comprehend complex audio much like humans. Current efforts primarily replicate text-based reasoning by contextualizing audio content through a one-time encoding, which introduces a critical information bottleneck. Drawing inspiration from human cognition, we propose audio-interleaved reasoning to break through this bottleneck. It treats audio as an active reasoning component, enabling sustained audio engagement and perception-grounded analysis. To instantiate it, we introduce a two-stage training framework, first teaching LALMs to localize salient audio segments through supervised fine-tuning, and then incentivizing proficient re-listening via reinforcement learning. In parallel, a structured data generation pipeline is developed to produce high-quality training data. Consequently, we present Echo, a LALM capable of dynamically re-listening to audio in demand during reasoning. On audio comprehension benchmarks, Echo achieves overall superiority in both challenging expert-level and general-purpose tasks. Comprehensive analysis further confirms the efficiency and generalizability of audio-interleaved reasoning, establishing it as a promising direction for advancing audio comprehension. Project page: https://github.com/wdqqdw/Echo.

Method: The researchers use activation patching to identify which attention layers control specific musical concepts in state-of-the-art audio diffusion architectures. They then apply Contrastive Activation Addition and Sparse Autoencoders to these identified layers to enable more precise control over generated audio.

Result: The study reveals that distinct semantic musical concepts are controlled by a small, shared subset of attention layers. By steering activations in these specialized layers, researchers can alter specific musical elements with high precision, such as modulating tempo or changing a track’s mood.

Conclusion: Audio diffusion models exhibit specialization in specific attention layers for controlling musical concepts, and leveraging this specialization enables more precise audio editing and generation control.

Abstract: Audio diffusion models can synthesize high-fidelity music from text, yet their internal mechanisms for representing high-level concepts remain poorly understood. In this work, we use activation patching to demonstrate that distinct semantic musical concepts, such as the presence of specific instruments, vocals, or genre characteristics, are controlled by a small, shared subset of attention layers in state-of-the-art audio diffusion architectures. Next, we demonstrate that applying Contrastive Activation Addition and Sparse Autoencoders in these layers enables more precise control over the generated audio, indicating a direct benefit of the specialization phenomenon. By steering activations of the identified layers, we can alter specific musical elements with high precision, such as modulating tempo or changing a track’s mood.

Method: Uses a passive bio-inspired acoustic microstructure that attaches to smartphone earphone mics to embed directional cues, combined with an end-to-end neural network for real-time audio processing on mobile devices.

Result: Achieves 5.0 dB signal quality improvement when focusing on 30° angular region, and outperforms conventional 5-microphone arrays using only two microphones.

Conclusion: SonicSieve demonstrates effective directional speech extraction for smartphones using passive acoustic structures and neural processing, enabling high-quality audio capture in challenging environments.

Abstract: Imagine placing your smartphone on a table in a noisy restaurant and clearly capturing the voices of friends seated around you, or recording a lecturer’s voice with clarity in a reverberant auditorium. We introduce SonicSieve, the first intelligent directional speech extraction system for smartphones using a bio-inspired acoustic microstructure. Our passive design embeds directional cues onto incoming speech without any additional electronics. It attaches to the in-line mic of low-cost wired earphones which can be attached to smartphones. We present an end-to-end neural network that processes the raw audio mixtures in real-time on mobile devices. Our results show that SonicSieve achieves a signal quality improvement of 5.0 dB when focusing on a 30° angular region. Additionally, the performance of our system based on only two microphones exceeds that of conventional 5-microphone arrays.

Method: Combines supervised contrastive fine-tuning of pre-trained models with contrastive representation distillation to structure embedding spaces that preserve semantic relationships between scenes, enabling knowledge transfer to compact student models.

Result: ContrastASC demonstrates improved few-shot adaptation to unseen categories while maintaining strong closed-set performance.

Conclusion: The approach enables acoustic scene classification models to be more adaptable and transferable to real-world scenarios requiring adaptation to new acoustic categories.

Abstract: Acoustic scene classification (ASC) models on edge devices typically operate under fixed class assumptions, lacking the transferability needed for real-world applications that require adaptation to new or refined acoustic categories. We propose ContrastASC, which learns generalizable acoustic scene representations by structuring the embedding space to preserve semantic relationships between scenes, enabling adaptation to unseen categories without retraining. Our approach combines supervised contrastive fine-tuning of pre-trained models with contrastive representation distillation to transfer this structured knowledge to compact student models. Our evaluation shows that ContrastASC demonstrates improved few-shot adaptation to unseen categories while maintaining strong closed-set performance.

Method: Integrates scalar auxiliary variable techniques with neural ordinary differential equations to yield stable differentiable models. Uses analytical solutions for linear vibration of system modes to keep physical parameters accessible after training without needing parameter encoders. Employs gradient networks instead of multilayer perceptrons to allow interpretation in terms of closed-form, non-negative potentials required by scalar auxiliary variable techniques.

Result: Demonstrates the approach on synthetic data for nonlinear transverse vibration of a string, showing the model can be trained to reproduce nonlinear dynamics. Sound examples are presented as proof of concept.

Conclusion: Successfully combines scalar auxiliary variable techniques with neural ODEs to create stable differentiable models for learning nonlinear dynamics in physical systems, maintaining physical parameter accessibility while enabling data-driven learning of complex behaviors.

Abstract: Modal methods are a long-standing approach to physical modelling synthesis. Extensions to nonlinear problems are possible, leading to coupled nonlinear systems of ordinary differential equations. Recent work in scalar auxiliary variable techniques has enabled construction of explicit and stable numerical solvers for such systems. On the other hand, neural ordinary differential equations have been successful in modelling nonlinear systems from data. In this work, we examine how scalar auxiliary variable techniques can be combined with neural ordinary differential equations to yield a stable differentiable model capable of learning nonlinear dynamics. The proposed approach leverages the analytical solution for linear vibration of the system’s modes so that physical parameters of a system remain easily accessible after the training without the need for a parameter encoder in the model architecture. Compared to our previous work that used multilayer perceptrons to parametrise nonlinear dynamics, we employ gradient networks that allow an interpretation in terms of a closed-form and non-negative potential required by scalar auxiliary variable techniques. As a proof of concept, we generate synthetic data for the nonlinear transverse vibration of a string and show that the model can be trained to reproduce the nonlinear dynamics of the system. Sound examples are presented.

Method: Uses an InfoGAN-style architecture that first learns semantically aligned audio-text representations using pre-trained transformers and contrastive objectives. Then trains a feature generator to produce emotion-aware audio features with mutual information regularization to ensure strong dependency between generated features and conditioning variables. Extends to multimodal settings for generating paired (audio, text) features.

Result: Outperforms existing augmentation methods on three benchmark datasets (IEMOCAP, MSP-IMPROV, MSP-Podcast), achieving state-of-the-art performance with improvements of up to 2.6% in unimodal SER and 3.2% in multimodal emotion recognition.

Conclusion: Mutual information functions as both a regularizer and measurable metric for generative quality, offering a systematic approach to data augmentation in affective computing. The framework enables generation of emotionally consistent multimodal features.

Abstract: Lack of large, well-annotated emotional speech corpora continues to limit the performance and robustness of speech emotion recognition (SER), particularly as models grow more complex and the demand for multimodal systems increases. While generative data augmentation offers a promising solution, existing approaches often produce emotionally inconsistent samples due to oversimplified conditioning on categorical labels. This paper introduces a novel mutual-information-regularised generative framework that combines cross-modal alignment with feature-level synthesis. Building on an InfoGAN-style architecture, our method first learns a semantically aligned audio-text representation space using pre-trained transformers and contrastive objectives. A feature generator is then trained to produce emotion-aware audio features while employing mutual information as a quantitative regulariser to ensure strong dependency between generated features and their conditioning variables. We extend this approach to multimodal settings, enabling the generation of novel, paired (audio, text) features. Comprehensive evaluation on three benchmark datasets (IEMOCAP, MSP-IMPROV, MSP-Podcast) demonstrates that our framework consistently outperforms existing augmentation methods, achieving state-of-the-art performance with improvements of up to 2.6% in unimodal SER and 3.2% in multimodal emotion recognition. Most importantly, we demonstrate that mutual information functions as both a regulariser and a measurable metric for generative quality, offering a systematic approach to data augmentation in affective computing.

Method: Proposes NarraScore framework that repurposes frozen Vision-Language Models (VLMs) as continuous affective sensors to distill visual streams into Valence-Arousal trajectories. Uses Dual-Branch Injection strategy: Global Semantic Anchor for stylistic stability and Token-Level Affective Adapter for local tension modulation via element-wise residual injection, avoiding dense attention bottlenecks.

Result: NarraScore achieves state-of-the-art consistency and narrative alignment with negligible computational overhead, establishing a fully autonomous paradigm for long-video soundtrack generation.

Conclusion: The framework successfully addresses the three critical impediments in long-video soundtrack synthesis by using emotion as narrative logic compression and a minimalist design that avoids overfitting risks while maintaining computational efficiency.

Abstract: Synthesizing coherent soundtracks for long-form videos remains a formidable challenge, currently stalled by three critical impediments: computational scalability, temporal coherence, and, most critically, a pervasive semantic blindness to evolving narrative logic. To bridge these gaps, we propose NarraScore, a hierarchical framework predicated on the core insight that emotion serves as a high-density compression of narrative logic. Uniquely, we repurpose frozen Vision-Language Models (VLMs) as continuous affective sensors, distilling high-dimensional visual streams into dense, narrative-aware Valence-Arousal trajectories. Mechanistically, NarraScore employs a Dual-Branch Injection strategy to reconcile global structure with local dynamism: a \textit{Global Semantic Anchor} ensures stylistic stability, while a surgical \textit{Token-Level Affective Adapter} modulates local tension via direct element-wise residual injection. This minimalist design bypasses the bottlenecks of dense attention and architectural cloning, effectively mitigating the overfitting risks associated with data scarcity. Experiments demonstrate that NarraScore achieves state-of-the-art consistency and narrative alignment with negligible computational overhead, establishing a fully autonomous paradigm for long-video soundtrack generation.

Method: 1) Introduces Partial Overlap Benchmark (POB) with two datasets (POB-Spark and POB-LibriPhrase) containing mismatched audio-text pairs with shared prefixes. 2) Proposes Equal-weighting Position Scoring (EPS), a lightweight decision layer that addresses position bias in cross-modal scoring.

Result: EPS alone reduces EER on POB-Spark from 64.4% to 29.3% and improves POB-LP accuracy from 87.6% to 96.8%, while maintaining performance on LibriPhrase and Google Speech Commands. With POB data added in training, achieves best POB benchmark results with least degradation on prior metrics.

Conclusion: The work successfully addresses prefix bias in open-vocabulary keyword spotting through benchmark creation and lightweight scoring method, though notes a trade-off with single-word command performance that requires future mitigation.

Abstract: Open-vocabulary keyword spotting (OV-KWS) enables personalized device control via arbitrary voice commands. Recently, researchers have explored using audio-text joint embeddings, allowing users to enroll phrases with text, and proposed techniques to disambiguate similar utterances. We find that existing OV-KWS solutions often overly bias the beginning phonemes of an enrollment, causing false triggers when negative enrollment-query-pairs share a prefix (turn the volume up'' vs. turn the volume down’’). We trace this to two factors: training data bias and position-biased cross-modal scoring. To address these limitations, we introduce the Partial Overlap Benchmark (POB) with two datasets, POB-Spark and POB-LibriPhrase (POB-LP), containing mismatched audio-text pairs with shared prefixes, and propose Equal-weighting Position Scoring (EPS), a lightweight decision layer. Using EPS alone reduces EER on POB-Spark from 64.4% to 29.3% and improves POB-LP accuracy from 87.6% to 96.8%, while maintaining performance on LibriPhrase and Google Speech Commands (GSC). With POB data added in training, our work achieves the best POB benchmark results while incurring the least amount of degradation on prior metrics among baselines. This degradation is most pronounced in GSC, which contains only one-word commands. We surface mitigating this trade-off as future work.

Method: Proposes CAT - a purely Transformer-based architecture that jointly optimizes encoder, quantizer, and decoder from scratch. Scales this to MOSS-Audio-Tokenizer with 1.6B parameters pre-trained on 3M hours of diverse audio data using homogeneous, causal Transformer blocks.

Result: Outperforms prior codecs across speech, sound, and music over wide bitrate ranges. Enables first purely autoregressive TTS model surpassing prior non-autoregressive systems. Achieves competitive ASR performance without auxiliary encoders. Shows predictable scaling improvements.

Conclusion: The CAT architecture serves as a unified, scalable interface for next-generation native audio foundation models, demonstrating that simple, fully end-to-end homogeneous Transformer approaches scale gracefully and support high-fidelity audio reconstruction across domains.

Abstract: Discrete audio tokenizers are fundamental to empowering large language models with native audio processing and generation capabilities. Despite recent progress, existing approaches often rely on pretrained encoders, semantic distillation, or heterogeneous CNN-based architectures. These designs introduce fixed inductive biases that limit reconstruction fidelity and hinder effective scaling. In this paper, we argue that discrete audio tokenization should be learned fully end-to-end using a homogeneous and scalable architecture. To this end, we first propose CAT (Causal Audio Tokenizer with Transformer), a purely Transformer-based architecture that jointly optimizes the encoder, quantizer, and decoder from scratch for high-fidelity reconstruction. Building on the CAT architecture, we develop MOSS-Audio-Tokenizer, a large-scale audio tokenizer featuring 1.6 billion parameters, pre-trained on 3 million hours of diverse, general audio data. We show that this simple, fully end-to-end approach built from homogeneous, causal Transformer blocks scales gracefully and supports high-fidelity reconstruction across diverse audio domains. Across speech, sound, and music, MOSS-Audio-Tokenizer consistently outperforms prior codecs over a wide range of bitrates, while exhibiting predictable improvements with increased scale. Notably, leveraging the discrete tokens from our model, we develop the first purely autoregressive TTS model that surpasses prior non-autoregressive and cascaded systems. Furthermore, MOSS-Audio-Tokenizer enables competitive ASR performance without auxiliary encoders. Our findings position the CAT architecture as a unified, scalable interface for the next generation of native audio foundation models.

Method: Proposes MIND framework with two key components: 1) Construction of focused, high-density training corpus by leveraging the observation that modeling errors remain localized to specific semantic segments, and 2) Dynamic Supervised Fine-Tuning Policy Optimization (DFPO) that tackles difficult problems through localized refinement rather than holistic reasoning.

Result: Experiments on six benchmarks demonstrate that MIND consistently outperforms all state-of-the-art automated optimization modeling approaches.

Conclusion: The error-driven learning framework MIND effectively addresses data scarcity and sparse reward issues in LLM-based optimization modeling by leveraging localized error patterns, leading to superior performance compared to existing approaches.

Abstract: Automated optimization modeling via Large Language Models (LLMs) has emerged as a promising approach to assist complex human decision-making. While post-training has become a pivotal technique to enhance LLMs’ capabilities in this domain, its effectiveness is severely constrained by the scarcity and underutilization of high-quality training data. However, through a detailed profiling of error patterns across various problem-response pairs drawn from post-training, we identify two fundamental limitations of existing automated optimization modeling approaches: (L1) the sparsity of error-specific problems and (L2) the sparse rewards associated with difficult problems. We demonstrate that these limitations can result in suboptimal performance in domain-specific post-training for LLMs. To tackle the above two limitations, we propose a novel error-driven learning framework – namely, auto\textbf{m}ated opt\textbf{i}mization modeli\textbf{n}g via a localizable error-\textbf{d}riven perspective (MIND) – that customizes the whole model training framework from data synthesis to post-training. MIND is based on our key observation of the unique localizable patterns in error propagation of optimization modelings, that is, modeling errors may remain localized to specific semantic segments and do not propagate throughout the entire solution. Thus, in contrast to holistic reasoning tasks such as mathematical proofs, MIND leverages the construction of a focused, high-density training corpus and proposes \textbf{D}ynamic Supervised \textbf{F}ine-Tuning \textbf{P}olicy \textbf{O}ptimization (DFPO) to tackle difficult problems through localized refinement. Experiments on six benchmarks demonstrate that MIND consistently outperforms all the state-of-the-art automated optimization modeling approaches.

Method: KBVQ-MoE integrates two techniques: (1) input-driven redundancy elimination using KLT-guided SVD to extract dominant weight components shared across experts, and (2) bias-corrected output stabilization where VQ is applied only to expert-specific representations with channel-wise affine compensation for quantized outputs.

Result: Experiments show KBVQ-MoE preserves accuracy substantially better than existing quantization methods. For 3-bit quantization of Qwen1.5-MoE-A2.7B, it achieves 67.99 average accuracy vs. 68.07 FP16 baseline, demonstrating near-identical performance.

Conclusion: KBVQ-MoE enables efficient deployment of MoE-based LLMs on edge devices and resource-constrained platforms through effective ultra-low-bit compression while maintaining model accuracy.

Abstract: Mixture of Experts (MoE) models have achieved great success by significantly improving performance while maintaining computational efficiency through sparse expert activation. However, their enormous parameter sizes and memory demands pose major challenges for deployment in resource-constrained environments. Vector Quantization (VQ) offers a promising approach for ultra-low-bit compression in Large Language Models (LLMs) by leveraging a codebook, where weight vectors are mapped to the most similar discrete codewords. Yet, directly applying VQ to MoEs often leads to substantial performance degradation due to two critical obstacles: (1) redundant representations among experts cause VQ to repeatedly quantize similar representations for each expert, resulting in inefficient use of limited codebook capacity; and (2) cumulative output bias is amplified by expert aggregation in MoE layers, leading to distributional shifts in the quantized outputs. To address these issues, we propose KBVQ-MoE, a novel VQ framework to enhance extremely low-bit quantization for MoE-based LLMs. KBVQ-MoE integrates two techniques: (1) input-driven redundancy elimination, where a Karhunen-Loeve Transform (KLT) guided singular value decomposition (SVD) extracts dominant weight components and shares them across experts; and (2) bias-corrected output stabilization, where vector quantization is applied only to expert-specific (non-redundant) representations and the quantized outputs are corrected via channel-wise affine compensation. Experiments on various MoE LLMs demonstrate that KBVQ-MoE preserves accuracy substantially better than existing quantization methods. For example, 3-bit quantization of Qwen1.5-MoE-A2.7B achieves an average accuracy of 67.99, nearly identical to the FP16 baseline of 68.07, underscoring KBVQ-MoE’s potential for efficient deployment on edge devices and other resource-constrained platforms.

Method: Proposes Spectra optimizer that tracks the low-rank spike subspace via cached, warm-started power iteration, applies low-rank spectral shaping to suppress the spike without amplifying the noise-sensitive tail, with minimal overhead.

Result: On LLaMA3 8B trained on 50B tokens: 30% faster convergence than AdamW, 0.7% reduced per-step overhead, 49.25% optimizer state memory reduction, 1.62% average downstream accuracy improvement. Outperforms Muon with 5.1x faster optimizer processing and 0.66% higher accuracy.

Conclusion: Spectra effectively addresses gradient anisotropy in LLM training by spike-aware optimization, enabling faster convergence, reduced memory usage, and improved model performance through better tail learning.

Abstract: Gradient signals in LLM training are highly anisotropic: recurrent linguistic structure concentrates energy into a small set of dominant spectral directions, while context specific information resides in a long tail. We show that this spike tail separation persists throughout training, with the spike occupying only about 1.5% of directions yet dominating optimizer statistics. This dominance suppresses tail learning by contracting tail updates through second moment normalization and tightening the globally stable learning rate bound. Motivated by this analysis, we propose Spectra, a spike aware optimizer that suppresses the dominant low rank spike subspace without amplifying the noise sensitive spectral tail. Spectra tracks the spike subspace via cached, warm started power iteration and applies low rank spectral shaping with negligible overhead and substantially reduced optimizer state memory. On LLaMA3 8B trained on 50B tokens, Spectra reaches the same target loss 30% faster than AdamW, reduces per step end to end overhead by 0.7%, cuts optimizer state memory by 49.25%, and improves average downstream accuracy by 1.62%. Compared to Muon, Spectra is 5.1x faster in optimizer processing time, achieves a lower final loss, and improves average accuracy by 0.66%.

Method: 1) Physics-guided simulation to synthesize large-scale jamming dataset; 2) Multi-Scale Ghost-ACB-Coordinate (MS-GAC) backbone combining Asymmetric Convolution Blocks and Ghost modules for efficient feature extraction; 3) Kolmogorov-Arnold Network (KAN) decision head replacing traditional MLP with learnable spline activations.

Result: Achieves 98.0% overall accuracy with only 0.13M parameters (660x fewer than ViT baselines), making it suitable as “always-on” security companion for edge devices.

Conclusion: GAC-KAN addresses dual challenges of data scarcity and computational efficiency for GNSS protection in GenAI era, enabling reliable security without competing with primary GenAI tasks.

Abstract: The integration of Generative AI (GenAI) into Consumer Electronics (CE)–from AI-powered assistants in wearables to generative planning in autonomous Uncrewed Aerial Vehicles (UAVs)–has revolutionized user experiences. However, these GenAI applications impose immense computational burdens on edge hardware, leaving strictly limited resources for fundamental security tasks like Global Navigation Satellite System (GNSS) signal protection. Furthermore, training robust classifiers for such devices is hindered by the scarcity of real-world interference data. To address the dual challenges of data scarcity and the extreme efficiency required by the GenAI era, this paper proposes a novel framework named GAC-KAN. First, we adopt a physics-guided simulation approach to synthesize a large-scale, high-fidelity jamming dataset, mitigating the data bottleneck. Second, to reconcile high accuracy with the stringent resource constraints of GenAI-native chips, we design a Multi-Scale Ghost-ACB-Coordinate (MS-GAC) backbone. This backbone combines Asymmetric Convolution Blocks (ACB) and Ghost modules to extract rich spectral-temporal features with minimal redundancy. Replacing the traditional Multi-Layer Perceptron (MLP) decision head, we introduce a Kolmogorov-Arnold Network (KAN), which employs learnable spline activation functions to achieve superior non-linear mapping capabilities with significantly fewer parameters. Experimental results demonstrate that GAC-KAN achieves an overall accuracy of 98.0%, outperforming state-of-the-art baselines. Significantly, the model contains only 0.13 million parameter–approximately 660 times fewer than Vision Transformer (ViT) baselines. This extreme lightweight characteristic makes GAC-KAN an ideal “always-on” security companion, ensuring GNSS reliability without contending for the computational resources required by primary GenAI tasks.

Method: TDPNavigator-Placer, a novel multi-agent reinforcement learning framework that dynamically optimizes placement based on chiplet’s thermal design power (TDP). The approach assigns conflicting objectives to specialized agents, each with distinct reward mechanisms and environmental constraints within a unified placement paradigm.

Result: Experimental results show TDPNavigator-Placer delivers significantly improved Pareto front over state-of-the-art methods, enabling more balanced trade-offs between wirelength and thermal performance.

Conclusion: The proposed multi-agent reinforcement learning framework effectively addresses the conflicting objectives in chiplet placement, providing superior optimization capabilities compared to existing methods for 2.5D integrated circuits.

Abstract: The rapid growth of electronics has accelerated the adoption of 2.5D integrated circuits, where effective automated chiplet placement is essential as systems scale to larger and more heterogeneous chiplet assemblies. Existing placement methods typically focus on minimizing wirelength or transforming multi-objective optimization into a single objective through weighted sum, which limits their ability to handle competing design requirements. Wirelength reduction and thermal management are inherently conflicting objectives, making prior approaches inadequate for practical deployment. To address this challenge, we propose TDPNavigator-Placer, a novel multi-agent reinforcement learning framework that dynamically optimizes placement based on chiplet’s thermal design power (TDP). This approach explicitly assigns these inherently conflicting objectives to specialized agents, each operating under distinct reward mechanisms and environmental constraints within a unified placement paradigm. Experimental results demonstrate that TDPNavigator-Placer delivers a significantly improved Pareto front over state-of-the-art methods, enabling more balanced trade-offs between wirelength and thermal performance.

Method: Proposes Multi-Offset Time Embedding (MOTE) to mitigate performance degradation from standard token embedding, then designs Time-TK architecture with: 1) Multi-Offset Interactive KAN to learn temporal patterns among offset sub-sequences, and 2) Multi-Offset Temporal Interaction mechanism to capture complex dependencies between sub-sequences for global information integration.

Result: Extensive experiments on 14 real-world benchmark datasets (including traffic flow and BTC/USDT throughput) demonstrate that Time-TK significantly outperforms all baseline models, achieving state-of-the-art forecasting accuracy.

Conclusion: The proposed MOTE embedding and Time-TK architecture fundamentally address the information bottleneck in time series forecasting by preserving multi-offset temporal correlations, offering a universal building block that can be integrated into various existing models.

Abstract: Time series forecasting is crucial for the World Wide Web and represents a core technical challenge in ensuring the stable and efficient operation of modern web services, such as intelligent transportation and website throughput. However, we have found that existing methods typically employ a strategy of embedding each time step as an independent token. This paradigm introduces a fundamental information bottleneck when processing long sequences, the root cause of which is that independent token embedding destroys a crucial structure within the sequence - what we term as multi-offset temporal correlation. This refers to the fine-grained dependencies embedded within the sequence that span across different time steps, which is especially prevalent in regular Web data. To fundamentally address this issue, we propose a new perspective on time series embedding. We provide an upper bound on the approximate reconstruction performance of token embedding, which guides our design of a concise yet effective Multi-Offset Time Embedding method to mitigate the performance degradation caused by standard token embedding. Furthermore, our MOTE can be integrated into various existing models and serve as a universal building block. Based on this paradigm, we further design a novel forecasting architecture named Time-TK. This architecture first utilizes a Multi-Offset Interactive KAN to learn and represent specific temporal patterns among multiple offset sub-sequences. Subsequently, it employs an efficient Multi-Offset Temporal Interaction mechanism to effectively capture the complex dependencies between these sub-sequences, achieving global information integration. Extensive experiments on 14 real-world benchmark datasets, covering domains such as traffic flow and BTC/USDT throughput, demonstrate that Time-TK significantly outperforms all baseline models, achieving state-of-the-art forecasting accuracy.

Method: MELINOE fine-tunes MoE models to more strongly prefer activating a smaller number of experts per sequence. By identifying and caching these preferred experts in GPU memory, it reduces expert churn and CPU-GPU transfer overhead.

Result: MELINOE increases throughput by 1.2-3× over efficient baselines and up to 14.7× over transfer-heavy baselines while retaining or even improving model performance on downstream tasks.

Conclusion: MELINOE provides a reliable method for improving MoE inference efficiency by reducing memory bottlenecks and transfer overhead through expert preference fine-tuning and GPU caching.

Abstract: Mixture-of-Experts (MoE) model architectures can significantly reduce the number of activated parameters per token, enabling computationally efficient training and inference. However, their large overall parameter counts and model sizes have precluded their widespread usage in resource-constrained settings as all of the parameters must still be loaded into GPU memory. Prior works aim to address this memory bottleneck by offloading certain experts into CPU memory and porting them to GPU memory only when they are activated. In practice, these methods suffer from the significant I/O latency incurred by expert transfer. We present MELINOE, a method that fine-tunes an MoE model to more strongly prefer activating a smaller number of experts per sequence. Caching these preferred experts in GPU memory reduces expert churn and CPU-GPU transfer overhead. MELINOE increases throughput by $1.2-3\times$ over efficient baselines and up to $14.7\times$ over transfer-heavy baselines while retaining or even improving the performance of the model on a downstream task, making it a reliable method for improving MoE inference efficiency.

Method: Applied multiple ML algorithms (MLR, logistic regression, SVC, K-means clustering, PCA) to predict and classify outcomes from laboratory experiments modeling field conditions using rain devices on various soils in sloped flumes. Parameters included rain intensity, slope gradients, water-entry values, and grain sizes.

Result: MLR effectively predicted total discharge but erosion predictions were less accurate, especially for coarse sand. LR and SVC achieved good accuracy in classifying failure outcomes. Sensitivity analysis revealed fine sand is highly susceptible to erosion under low-intensity, long-duration rainfall, and first 10 minutes of high-intensity rain are most critical.

Conclusion: ML techniques show strong potential for post-wildfire hazard assessment and emergency response planning by modeling complex debris flow onset conditions without predefined assumptions, though erosion prediction accuracy needs improvement for certain soil types.

Abstract: Post-wildfire mudflows are increasingly hazardous due to the prevalence of wildfires, including those on the wildland-urban interface. Upon burning, soil on the surface or immediately beneath becomes hydrophobic, a phenomenon that occurs predominantly on sand-based hillslopes. Rainwater and eroded soil blanket the downslope, leading to catastrophic debris flows. Soil hydrophobicity enhances erosion, resulting in post-wildfire debris flows that differ from natural mudflows in intensity, duration, and destructiveness. Thus, it is crucial to understand the timing and conditions of debris-flow onset, driven by the coupled effects of critical parameters: varying rain intensities (RI), slope gradients, water-entry values, and grain sizes (D50). Machine Learning (ML) techniques have become increasingly valuable in geotechnical engineering due to their ability to model complex systems without predefined assumptions. This study applies multiple ML algorithms: multiple linear regression (MLR), logistic regression (LR), support vector classifier (SVC), K-means clustering, and principal component analysis (PCA) to predict and classify outcomes from laboratory experiments that model field conditions using a rain device on various soils in sloped flumes. While MLR effectively predicted total discharge, erosion predictions were less accurate, especially for coarse sand. LR and SVC achieved good accuracy in classifying failure outcomes, supported by clustering and dimensionality reduction. Sensitivity analysis revealed that fine sand is highly susceptible to erosion, particularly under low-intensity, long-duration rainfall. Results also show that the first 10 minutes of high-intensity rain are most critical for discharge and failure. These findings highlight the potential of ML for post-wildfire hazard assessment and emergency response planning.

Method: Pre-trained on 9.2 million unlabeled Channel State Information (CSI) samples collected over 439 days from 20 commercial device types worldwide. Uses contrastive learning, masked reconstruction, and physics-informed objectives specifically tailored to wireless signals to learn general-purpose representations.

Result: Evaluated on nine downstream tasks across public benchmarks, AM-FM demonstrates strong cross-task performance with improved data efficiency compared to task-specific approaches.

Conclusion: Foundation models like AM-FM can enable scalable ambient intelligence using existing wireless infrastructure, unlocking the untapped potential of WiFi sensing for understanding human presence, activity, and physiology in physical spaces.

Abstract: Ambient intelligence, continuously understanding human presence, activity, and physiology in physical spaces, is fundamental to smart environments, health monitoring, and human-computer interaction. WiFi infrastructure provides a ubiquitous, always-on, privacy-preserving substrate for this capability across billions of IoT devices. Yet this potential remains largely untapped, as wireless sensing has typically relied on task-specific models that require substantial labeled data and limit practical deployment. We present AM-FM, the first foundation model for ambient intelligence and sensing through WiFi. AM-FM is pre-trained on 9.2 million unlabeled Channel State Information (CSI) samples collected over 439 days from 20 commercial device types deployed worldwide, learning general-purpose representations via contrastive learning, masked reconstruction, and physics-informed objectives tailored to wireless signals. Evaluated on public benchmarks spanning nine downstream tasks, AM-FM shows strong cross-task performance with improved data efficiency, demonstrating that foundation models can enable scalable ambient intelligence using existing wireless infrastructure.

Method: ZePAD uses a dual-branch structure: MPAE-Branch with two adversarially fine-tuned encoders for adversarial resistance, and BMP-Branch trained on local data to preserve benign performance. It enables single adversarial fine-tuning for diverse downstream tasks and can directly detect DAEs via branch confidence evaluation.

Result: Extensive experiments on 11 SSL methods and 6 datasets show effectiveness, with up to 29.20% improvement in benign performance and 73.86% gain in adversarial robustness, demonstrating zero-sacrifice property.

Conclusion: ZePAD achieves persistent robustness against DAEs across downstream tasks with single tuning while preserving benign performance, addressing limitations of previous task-specific adversarial fine-tuning approaches.

Abstract: The widespread use of publicly available pre-trained encoders from self-supervised learning (SSL) has exposed a critical vulnerability: their susceptibility to downstream-agnostic adversarial examples (DAEs), which are crafted without knowledge of the downstream tasks but capable of misleading downstream models. While several defense methods have been explored recently, they rely primarily on task-specific adversarial fine-tuning, which inevitably limits generalizability and causes catastrophic forgetting and deteriorates benign performance. Different with previous works, we propose a more rigorous defense goal that requires only a single tuning for diverse downstream tasks to defend against DAEs and preserve benign performance. To achieve this defense goal, we introduce Zero-Sacrifice Persistent-Robustness Adversarial Defense (ZePAD), which is inspired by the inherent sensitivity of neural networks to data characteristics. Specifically, ZePAD is a dual-branch structure, which consists of a Multi-Pattern Adversarial Enhancement Branch (MPAE-Branch) that uses two adversarially fine-tuned encoders to strengthen adversarial resistance. The Benign Memory Preservation Branch (BMP-Branch) is trained on local data to ensure adversarial robustness does not compromise benign performance. Surprisingly, we find that ZePAD can directly detect DAEs by evaluating branch confidence, without introducing any adversarial exsample identification task during training. Notably, by enriching feature diversity, our method enables a single adversarial fine-tuning to defend against DAEs across downstream tasks, thereby achieving persistent robustness. Extensive experiments on 11 SSL methods and 6 datasets validate its effectiveness. In certain cases, it achieves a 29.20% improvement in benign performance and a 73.86% gain in adversarial robustness, highlighting its zero-sacrifice property.

Method: Proposes a method to optimize k-space sampling patterns for multi-coil MRI reconstruction using diffusion models as priors. Introduces a training objective based on single-step posterior mean estimate to avoid backpropagation through iterative reconstruction. Presents a greedy strategy for learning Cartesian sampling patterns that selects informative k-space locations using gradient information from pre-trained diffusion models while enforcing spatial diversity among samples.

Result: Experimental results across multiple anatomies and acceleration factors demonstrate that diffusion models using the optimized sampling patterns achieve higher-quality reconstructions compared to using fixed and learned baseline patterns.

Conclusion: The proposed method successfully optimizes sampling patterns for diffusion-based MRI reconstruction, improving reconstruction quality without the computational burden of backpropagation through iterative processes.

Abstract: Magnetic resonance imaging (MRI) is a powerful medical imaging modality, but long acquisition times limit throughput, patient comfort, and clinical accessibility. Diffusion-based generative models serve as strong image priors for reducing scan-time with accelerated MRI reconstruction and offer robustness across variations in the acquisition model. However, most existing diffusion-based approaches do not exploit the unique ability in MRI to jointly design both the sampling pattern and the reconstruction method. While prior learning-based approaches have optimized sampling patterns for end-to-end unrolled networks, analogous methods for diffusion-based reconstruction have not been established due to the computational burden of posterior sampling. In this work, we propose a method to optimize k-space sampling patterns for accelerated multi-coil MRI reconstruction using diffusion models as priors. We introduce a training objective based on a single-step posterior mean estimate that avoids backpropagation through an expensive iterative reconstruction process. Then we present a greedy strategy for learning Cartesian sampling patterns that selects informative k-space locations using gradient information from a pre-trained diffusion model while enforcing spatial diversity among samples. Experimental results across multiple anatomies and acceleration factors demonstrate that diffusion models using the optimized sampling patterns achieve higher-quality reconstructions in comparison to using fixed and learned baseline patterns.

Method: Introduces UltraLIF framework using ultradiscretization from tropical geometry, where log-sum-exp serves as differentiable soft-maximum converging to hard thresholding. Derives two neuron models: UltraLIF from LIF ODE (temporal dynamics) and UltraDLIF from diffusion equation (spatial dynamics). Both enable fully differentiable SNNs trainable via standard backpropagation.

Result: Experiments on six benchmarks (static images, neuromorphic vision, audio) show improvements over surrogate gradient baselines, especially in single-timestep settings on neuromorphic and temporal datasets. Optional sparsity penalty enables significant energy reduction while maintaining competitive accuracy.

Conclusion: UltraLIF provides a principled mathematical framework for differentiable SNNs that eliminates heuristic surrogate gradients, offers theoretical guarantees, and demonstrates practical advantages in energy-efficient neuromorphic computing.

Abstract: Spiking Neural Networks (SNNs) offer energy-efficient, biologically plausible computation but suffer from non-differentiable spike generation, necessitating reliance on heuristic surrogate gradients. This paper introduces UltraLIF, a principled framework that replaces surrogate gradients with ultradiscretization, a mathematical formalism from tropical geometry providing continuous relaxations of discrete dynamics. The central insight is that the max-plus semiring underlying ultradiscretization naturally models neural threshold dynamics: the log-sum-exp function serves as a differentiable soft-maximum that converges to hard thresholding as a learnable temperature parameter $\eps \to 0$. Two neuron models are derived from distinct dynamical systems: UltraLIF from the LIF ordinary differential equation (temporal dynamics) and UltraDLIF from the diffusion equation modeling gap junction coupling across neuronal populations (spatial dynamics). Both yield fully differentiable SNNs trainable via standard backpropagation with no forward-backward mismatch. Theoretical analysis establishes pointwise convergence to classical LIF dynamics with quantitative error bounds and bounded non-vanishing gradients. Experiments on six benchmarks spanning static images, neuromorphic vision, and audio demonstrate improvements over surrogate gradient baselines, with gains most pronounced in single-timestep ($T{=}1$) settings on neuromorphic and temporal datasets. An optional sparsity penalty enables significant energy reduction while maintaining competitive accuracy.

Method: APT uses a graph-based encoder for local heterogeneous features and global attention mechanism for long-range physical impacts. It’s geometry-, mesh-, and physics-agnostic and can learn from adaptive mesh refinement simulations.

Result: APT outperforms state-of-the-art architectures in subsurface tasks across both regular and irregular grids with robust super-resolution capabilities. It’s the first to learn from adaptive mesh refinement and shows cross-dataset learning potential.

Conclusion: APT provides a robust, scalable backbone for large-scale subsurface foundation models, addressing key computational challenges in subsurface simulations through its novel neural operator design.

Abstract: The Earth’s subsurface is a cornerstone of modern society, providing essential energy resources like hydrocarbons, geothermal, and minerals while serving as the primary reservoir for $CO_2$ sequestration. However, full physics numerical simulations of these systems are notoriously computationally expensive due to geological heterogeneity, high resolution requirements, and the tight coupling of physical processes with distinct propagation time scales. Here we propose the \textbf{Adaptive Physics Transformer} (APT), a geometry-, mesh-, and physics-agnostic neural operator that explicitly addresses these challenges. APT fuses a graph-based encoder to extract high-resolution local heterogeneous features with a global attention mechanism to resolve long-range physical impacts. Our results demonstrate that APT outperforms state-of-the-art architectures in subsurface tasks across both regular and irregular grids with robust super-resolution capabilities. Notably, APT is the first architecture that directly learns from adaptive mesh refinement simulations. We also demonstrate APT’s capability for cross-dataset learning, positioning it as a robust and scalable backbone for large-scale subsurface foundation model development.

Method: Elastic Memory treats historical sequences as samples from continuous signals and applies optimal online compression via the HiPPO framework to encode them into fixed-size memory states. For retrieval, it uses a flexible polynomial sampling mechanism to reconstruct history summaries from compressed states.

Result: Elastic Memory consistently outperformed baselines on long-context (32k+) datasets across three domains. With equal parameters, it beat Memorizing Transformer by 16x memory and outperformed Melodi at all memory sizes, even when Melodi had 30% more parameters. When scaling model size, Elastic Memory stayed ahead of all baselines and was significantly faster than Melodi at 4x size.

Conclusion: Elastic Memory provides an effective solution to transformer memory bottlenecks with theoretical grounding in HiPPO framework, practical scalability, and decoupled design that allows test-time inductive bias injection for performance boosting.

Abstract: Transformers face a quadratic bottleneck in attention when scaling to long contexts. Recent approaches introduce recurrent memory to extend context beyond the current window, yet these often face a fundamental trade-off between theoretical principles and practical scalability. To address this, we introduce Elastic Memory, a novel memory architecture grounded in the HiPPO framework for online function approximation. Elastic Memory treats historical sequence as samples from continuous signals, applying optimal online compression to encode them into a fixed-size memory state. For retrieval, we propose a flexible \textit{polynomial sampling} mechanism that reconstructs a history summary from this compressed state. Elastic Memory consistently outperformed baselines on long-context (32k+) datasets across three domains. With equal parameters, it beat Memorizing Transformer by 16x memory and outperformed Melodi at all memory sizes, even when Melodi had 30% more parameters. When scaling model size, Elastic Memory stayed ahead of all baselines and was significantly faster than Melodi at 4x size. Furthermore, its decoupled design allows for injecting inductive biases at test-time to boost performance.

Method: Constructed a five-discipline corpus and analyzed learning patterns of full fine-tuning, LoRA, LoRA-MoE, and LoRA compositions to study multi-disciplinary LLM training.

Result: Multi-disciplinary learning is substantially more variable than single-discipline training. Four consistent empirical laws were distilled: Balance-then-Diversity, Merge-then-Align, Optimize-then-Scale, and Share-then-Specialize.

Conclusion: These laws form a practical recipe for principled multi-discipline fine-tuning and provide actionable guidance for developing generalizable scientific LLMs.

Abstract: While large language models (LLMs) have achieved strong performance through fine-tuning within individual scientific domains, their learning dynamics in multi-disciplinary contexts remains poorly understood, despite the promise of improved generalization and broader applicability through cross-domain knowledge synergy. In this work, we present the first systematic study of multi-disciplinary LLM fine-tuning, constructing a five-discipline corpus and analyzing learning patterns of full fine-tuning, LoRA, LoRA-MoE, and LoRA compositions. Particularly, our study shows that multi-disciplinary learning is substantially more variable than single-discipline training and distills four consistent empirical laws: (1) Balance-then-Diversity: low-resource disciplines degrade performance unless mitigated via diversity-aware upsampling; (2) Merge-then-Align: restoring instruction-following ability is critical for cross-discipline synergy; (3) Optimize-then-Scale: parameter scaling offers limited gains without prior design optimization; and (4) Share-then-Specialize: asymmetric LoRA-MoE yields robust gains with minimal trainable parameters via shared low-rank projection. Together, these laws form a practical recipe for principled multi-discipline fine-tuning and provide actionable guidance for developing generalizable scientific LLMs.

Method: Proposes PLaTITO, which incorporates auxiliary information sources including protein language model (pLM) embeddings, structural embeddings, temperature, and large-language-model-derived embeddings to enhance TITO models. Compares coarse-grained TITO models with Boltzmann Emulators and studies the impact of different conditioning signals.

Result: Coarse-grained TITO models are substantially more data-efficient than Boltzmann Emulators. Incorporating pLM embeddings further improves out-of-distribution generalization. PLaTITO achieves state-of-the-art performance on equilibrium sampling benchmarks for out-of-distribution protein systems, including fast-folding proteins.

Conclusion: Auxiliary information sources, particularly protein language model embeddings, significantly enhance the data efficiency and generalization capabilities of transferable implicit transfer operators for molecular dynamics, enabling better performance on out-of-distribution protein systems.

Abstract: Molecular dynamics (MD) is a central computational tool in physics, chemistry, and biology, enabling quantitative prediction of experimental observables as expectations over high-dimensional molecular distributions such as Boltzmann distributions and transition densities. However, conventional MD is fundamentally limited by the high computational cost required to generate independent samples. Generative molecular dynamics (GenMD) has recently emerged as an alternative, learning surrogates of molecular distributions either from data or through interaction with energy models. While these methods enable efficient sampling, their transferability across molecular systems is often limited. In this work, we show that incorporating auxiliary sources of information can improve the data efficiency and generalization of transferable implicit transfer operators (TITO) for molecular dynamics. We find that coarse-grained TITO models are substantially more data-efficient than Boltzmann Emulators, and that incorporating protein language model (pLM) embeddings further improves out-of-distribution generalization. Our approach, PLaTITO, achieves state-of-the-art performance on equilibrium sampling benchmarks for out-of-distribution protein systems, including fast-folding proteins. We further study the impact of additional conditioning signals – such as structural embeddings, temperature, and large-language-model-derived embeddings – on model performance.

Method: KVComm proposes a novel communication framework that enables LLMs to communicate through selective sharing of KV (key-value) pairs. It uses a KV layer-wise selection strategy based on attention importance scores with a Gaussian prior to identify the most informative KV pairs for communication between models.

Result: Extensive experiments across diverse tasks and model pairs show that KVComm achieves comparable performance to the upper-bound method (directly merging inputs to one model without communication) while transmitting only 30% of layers’ KV pairs, demonstrating high efficiency.

Conclusion: KV pairs serve as an effective medium for inter-LLM communication, offering a scalable and efficient solution for multi-agent systems that avoids the limitations of existing communication protocols.

Abstract: Large Language Models (LLMs) are increasingly deployed in multi-agent systems, where effective inter-model communication is crucial. Existing communication protocols either rely on natural language, incurring high inference costs and information loss, or on hidden states, which suffer from information concentration bias and inefficiency. To address these limitations, we propose KVComm, a novel communication framework that enables efficient communication between LLMs through selective sharing of KV pairs. KVComm leverages the rich information encoded in the KV pairs while avoiding the pitfalls of hidden states. We introduce a KV layer-wise selection strategy based on attention importance scores with a Gaussian prior to identify the most informative KV pairs for communication. Extensive experiments across diverse tasks and model pairs demonstrate that KVComm achieves comparable performance to the upper-bound method, which directly merges inputs to one model without any communication, while transmitting as few as 30% of layers’ KV pairs. Our study highlights the potential of KV pairs as an effective medium for inter-LLM communication, paving the way for scalable and efficient multi-agent systems.

Method: The authors use a suite of correlation protocols applied to accuracy and confidence metrics across diverse data mixtures and model scales. They examine four core research questions through systematic experiments with varying model scales.

Result: Transfer reliability varies dramatically across capability categories, benchmarks, and scales. Accuracy and confidence exhibit distinct, sometimes opposing, scaling dynamics. Some benchmarks serve as robust cross-stage predictors while others are unreliable.

Conclusion: The findings provide actionable guidance for benchmark selection, data curation, and efficient model development by shedding light on the complex interplay between pretraining decisions and downstream outcomes in language model training.

Abstract: Understanding how language model capabilities transfer from pretraining to supervised fine-tuning (SFT) is fundamental to efficient model development and data curation. In this work, we investigate four core questions: RQ1. To what extent do accuracy and confidence rankings established during pretraining persist after SFT? RQ2. Which benchmarks serve as robust cross-stage predictors and which are unreliable? RQ3. How do transfer dynamics shift with model scale? RQ4. How well does model confidence align with accuracy, as a measure of calibration quality? Does this alignment pattern transfer across training stages? We address these questions through a suite of correlation protocols applied to accuracy and confidence metrics across diverse data mixtures and model scales. Our experiments reveal that transfer reliability varies dramatically across capability categories, benchmarks, and scales – with accuracy and confidence exhibiting distinct, sometimes opposing, scaling dynamics. These findings shed light on the complex interplay between pretraining decisions and downstream outcomes, providing actionable guidance for benchmark selection, data curation, and efficient model development.

Method: Introduces specialized neurons for modal operators □ and ◇ based on Kripke semantics, operating over possible worlds. The accessibility relation between worlds can be either user-defined or parameterized by a neural network. The framework is fully differentiable and trained by minimizing logical contradiction loss.

Result: Demonstrated on four case studies: grammatical guardrailing, multi-agent epistemic trust, detecting constructive deception in natural language negotiation, and combinatorial constraint satisfaction in Sudoku. Shows improved logical consistency and interpretability without changing task architectures.

Conclusion: MLNNs provide a flexible neurosymbolic framework that can enforce or learn logical structures while maintaining differentiability, enabling models to perform deductive reasoning and learn nonlinear relationships that define problem-space logic.

Abstract: We propose Modal Logical Neural Networks (MLNNs), a neurosymbolic framework that integrates deep learning with the formal semantics of modal logic, enabling reasoning about necessity and possibility. Drawing on Kripke semantics, we introduce specialized neurons for the modal operators $\Box$ and $\Diamond$ that operate over a set of possible worlds, enabling the framework to act as a differentiable ``logical guardrail.’’ The architecture is highly flexible: the accessibility relation between worlds can either be fixed by the user to enforce known rules or, as an inductive feature, be parameterized by a neural network. This allows the model to optionally learn the relational structure of a logical system from data while simultaneously performing deductive reasoning within that structure. This versatile construction is designed for flexibility. The entire framework is differentiable from end to end, with learning driven by minimizing a logical contradiction loss. This not only makes the system resilient to inconsistent knowledge but also enables it to learn nonlinear relationships that can help define the logic of a problem space. We illustrate MLNNs on four case studies: grammatical guardrailing, multi-agent epistemic trust, detecting constructive deception in natural language negotiation, and combinatorial constraint satisfaction in Sudoku. These experiments demonstrate how enforcing or learning accessibility can increase logical consistency and interpretability without changing the underlying task architecture.

Method: Uses a credal-set formulation where uncertainty is represented as a set of predictive distributions, with epistemic uncertainty corresponding to set size and aleatoric uncertainty to noise within elements. Implements this in a Variational Credal Concept Bottleneck Model with two disjoint uncertainty heads trained by separate objectives and non-overlapping gradient paths.

Result: Reduces correlation between epistemic and aleatoric uncertainty by over an order of magnitude compared to standard methods, while improving alignment of epistemic uncertainty with prediction error and aleatoric uncertainty with ground-truth ambiguity across multi-annotator benchmarks.

Conclusion: The credal-set approach provides principled separation of uncertainty types by construction rather than post hoc decomposition, offering more reliable uncertainty quantification for decision making tasks.

Abstract: Decomposing predictive uncertainty into epistemic (model ignorance) and aleatoric (data ambiguity) components is central to reliable decision making, yet most methods estimate both from the same predictive distribution. Recent empirical and theoretical results show these estimates are typically strongly correlated, so changes in predictive spread simultaneously affect both components and blur their semantics. We propose a credal-set formulation in which uncertainty is represented as a set of predictive distributions, so that epistemic and aleatoric uncertainty correspond to distinct geometric properties: the size of the set versus the noise within its elements. We instantiate this idea in a Variational Credal Concept Bottleneck Model with two disjoint uncertainty heads trained by disjoint objectives and non-overlapping gradient paths, yielding separation by construction rather than post hoc decomposition. Across multi-annotator benchmarks, our approach reduces the correlation between epistemic and aleatoric uncertainty by over an order of magnitude compared to standard methods, while improving the alignment of epistemic uncertainty with prediction error and aleatoric uncertainty with ground-truth ambiguity.

Method: Cast data rewriting as policy learning problem using reinforcement learning. Learn rewriting policy that optimizes QA-style distributional alignment and diversity under hard task-consistency gate. RL-based agent optimizes rewrite distribution under reward feedback to construct higher-quality rewritten dataset for downstream SFT.

Result: Method achieves downstream gains comparable to standard SFT while reducing forgetting on non-downstream benchmarks by 12.34% on average.

Conclusion: RL-based data rewriting approach effectively addresses distribution shift issues in SFT by learning to rewrite data to better match LLM’s natural generation patterns while maintaining diversity and task consistency.

Abstract: Large language models (LLMs) have made rapid progress, yet adapting them to downstream scenarios still commonly relies on supervised fine-tuning (SFT). When downstream data exhibit a substantial distribution shift from the model’s prior training distribution, SFT can induce catastrophic forgetting. To narrow this gap, data rewriting has been proposed as a data-centric approach that rewrites downstream training data prior to SFT. However, existing methods typically sample rewrites from a prompt-induced conditional distribution, so the resulting targets are not necessarily aligned with the model’s natural QA-style generation distribution. Moreover, reliance on fixed templates can lead to diversity collapse. To address these issues, we cast data rewriting as a policy learning problem and learn a rewriting policy that better matches the backbone’s QA-style generation distribution while preserving diversity. Since distributional alignment, diversity and task consistency are automatically evaluable but difficult to optimize end-to-end with differentiable objectives, we leverage reinforcement learning to optimize the rewrite distribution under reward feedback and propose an RL-based data-rewriting agent. The agent jointly optimizes QA-style distributional alignment and diversity under a hard task-consistency gate, thereby constructing a higher-quality rewritten dataset for downstream SFT. Extensive experiments show that our method achieves downstream gains comparable to standard SFT while reducing forgetting on non-downstream benchmarks by 12.34% on average. Our code is available at https://anonymous.4open.science/r/Patch-the-Prompt-Gap-4112 .

Method: Proposes TopoGBM, a 3D convolutional autoencoder regularized by topological regularization that preserves complex, non-Euclidean invariants of the tumor’s manifold within a compressed latent space. This enforces topological priors to model high-variance structural signatures of aggressive GBM.

Result: TopoGBM achieves C-index 0.67 on test and 0.58 on validation across heterogeneous cohorts (UPENN, UCSF, RHUH) and external validation on TCGA, outperforming baselines that degrade under domain shift. Reconstruction residuals localize to pathologically heterogeneous zones, with tumor-restricted and healthy tissue error significantly low.

Conclusion: Incorporating topological priors enables learning morphology-faithful embeddings that capture tumor heterogeneity while maintaining cross-institutional robustness. Approximately 50% of prognostic signal localizes to tumor and peritumoral microenvironment, advocating clinical reliability of the unsupervised method.

Abstract: Accurate prognosis for Glioblastoma (GBM) using deep learning (DL) is hindered by extreme spatial and structural heterogeneity. Moreover, inconsistent MRI acquisition protocols across institutions hinder generalizability of models. Conventional transformer and DL pipelines often fail to capture the multi-scale morphological diversity such as fragmented necrotic cores, infiltrating margins, and disjoint enhancing components leading to scanner-specific artifacts and poor cross-site prognosis. We propose TopoGBM, a learning framework designed to capture heterogeneity-preserved, scanner-robust representations from multi-parametric 3D MRI. Central to our approach is a 3D convolutional autoencoder regularized by a topological regularization that preserves the complex, non-Euclidean invariants of the tumor’s manifold within a compressed latent space. By enforcing these topological priors, TopoGBM explicitly models the high-variance structural signatures characteristic of aggressive GBM. Evaluated across heterogeneous cohorts (UPENN, UCSF, RHUH) and external validation on TCGA, TopoGBM achieves better performance (C-index 0.67 test, 0.58 validation), outperforming baselines that degrade under domain shift. Mechanistic interpretability analysis reveals that reconstruction residuals are highly localized to pathologically heterogeneous zones, with tumor-restricted and healthy tissue error significantly low (Test: 0.03, Validation: 0.09). Furthermore, occlusion-based attribution localizes approximately 50% of the prognostic signal to the tumor and the diverse peritumoral microenvironment advocating clinical reliability of the unsupervised learning method. Our findings demonstrate that incorporating topological priors enables the learning of morphology-faithful embeddings that capture tumor heterogeneity while maintaining cross-institutional robustness.

Method: Developed AI-CDSS using hybrid approach combining expert insights with machine learning, trained on 650 patients and tested on 648 patients. Used key features like BMI, fasting glucose, and HbA1C. Conducted clinical pilot study with 105 patients comparing AI performance against endocrinology and non-endocrinology specialists.

Result: AI-CDSS achieved 99.8% accuracy for diabetes prediction, 99.3% for prediabetes, 99.2% for at-risk identification, and 98.8% for no diabetes. Showed 98.8% agreement with endocrinology specialists in test dataset and 98.5% concordance in pilot study, significantly outperforming non-endocrinology specialists’ 85% agreement rate.

Conclusion: The AI-CDSS demonstrates high potential as a valuable tool for accurate type 2 diabetes identification, particularly in settings where diabetes specialists are not readily available, potentially improving diagnostic accuracy for primary care physicians.

Abstract: Identifying type 2 diabetes mellitus can be challenging, particularly for primary care physicians. Clinical decision support systems incorporating artificial intelligence (AI-CDSS) can assist medical professionals in diagnosing type 2 diabetes with high accuracy. This study aimed to assess an AI-CDSS specifically developed for the diagnosis of type 2 diabetes by employing a hybrid approach that integrates expert-driven insights with machine learning techniques. The AI-CDSS was developed (training dataset: n = 650) and tested (test dataset: n = 648) using a dataset of 1298 patients with and without type 2 diabetes. To generate predictions, the algorithm utilized key features such as body mass index, plasma fasting glucose, and hemoglobin A1C. Furthermore, a clinical pilot study involving 105 patients was conducted to assess the diagnostic accuracy of the system in comparison to non-endocrinology specialists. The AI-CDSS showed a high degree of accuracy, with 99.8% accuracy in predicting diabetes, 99.3% in predicting prediabetes, 99.2% in identifying at-risk individuals, and 98.8% in predicting no diabetes. The test dataset revealed a 98.8% agreement between endocrinology specialists and the AI-CDSS. Type 2 diabetes was identified in 45% of 105 individuals in the pilot study. Compared with diabetes specialists, the AI-CDSS scored a 98.5% concordance rate, greatly exceeding that of nonendocrinology specialists, who had an 85% agreement rate. These findings indicate that the AI-CDSS has the potential to be a useful tool for accurately identifying type 2 diabetes, especially in situations in which diabetes specialists are not readily available.

Method: Created StructMemEval benchmark with tasks that require organizing knowledge in specific structures like transaction ledgers, to-do lists, and trees. Tested both simple retrieval-augmented LLMs and memory agents with and without prompting about memory structure.

Result: Simple retrieval-augmented LLMs struggle with structured memory tasks. Memory agents can solve them when prompted about memory organization, but modern LLMs often fail to recognize memory structures without explicit prompting.

Conclusion: Current memory benchmarks are insufficient for evaluating complex memory hierarchies. The gap highlights need for improvements in both LLM training and memory frameworks to better handle structured knowledge organization.

Abstract: Modern LLM-based agents and chat assistants rely on long-term memory frameworks to store reusable knowledge, recall user preferences, and augment reasoning. As researchers create more complex memory architectures, it becomes increasingly difficult to analyze their capabilities and guide future memory designs. Most long-term memory benchmarks focus on simple fact retention, multi-hop recall, and time-based changes. While undoubtedly important, these capabilities can often be achieved with simple retrieval-augmented LLMs and do not test complex memory hierarchies. To bridge this gap, we propose StructMemEval - a benchmark that tests the agent’s ability to organize its long-term memory, not just factual recall. We gather a suite of tasks that humans solve by organizing their knowledge in a specific structure: transaction ledgers, to-do lists, trees and others. Our initial experiments show that simple retrieval-augmented LLMs struggle with these tasks, whereas memory agents can reliably solve them if prompted how to organize their memory. However, we also find that modern LLMs do not always recognize the memory structure when not prompted to do so. This highlights an important direction for future improvements in both LLM training and memory frameworks.

Method: Separates LRH into linear representation and linear accessibility claims, then uses compressed sensing theory to establish bounds. Uses random matrix constructions for upper bounds and rank bounds with Turán’s theorem for lower bounds.

Result: Establishes nearly-matching bounds: requires Ω(k²/log k · log(m/k)) neurons and suffices with O(k² log m) neurons for linear compressed sensing, showing a quantitative gap between classical and linear compressed sensing.

Conclusion: Linear accessibility is a stronger requirement than linear representation alone, and neurons can store exponential features under LRH, supporting the superposition hypothesis while quantifying theoretical constraints.

Abstract: We introduce a mathematical framework for the linear representation hypothesis (LRH), which asserts that intermediate layers of language models store features linearly. We separate the hypothesis into two claims: linear representation (features are linearly embedded in neuron activations) and linear accessibility (features can be linearly decoded). We then ask: How many neurons $d$ suffice to both linearly represent and linearly access $m$ features? Classical results in compressed sensing imply that for $k$-sparse inputs, $d = O(k\log (m/k))$ suffices if we allow non-linear decoding algorithms (Candes and Tao, 2006; Candes et al., 2006; Donoho, 2006). However, the additional requirement of linear decoding takes the problem out of the classical compressed sensing, into linear compressed sensing. Our main theoretical result establishes nearly-matching upper and lower bounds for linear compressed sensing. We prove that $d = Ω_ε(\frac{k^2}{\log k}\log (m/k))$ is required while $d = O_ε(k^2\log m)$ suffices. The lower bound establishes a quantitative gap between classical and linear compressed setting, illustrating how linear accessibility is a meaningfully stronger hypothesis than linear representation alone. The upper bound confirms that neurons can store an exponential number of features under the LRH, giving theoretical evidence for the “superposition hypothesis” (Elhage et al., 2022). The upper bound proof uses standard random constructions of matrices with approximately orthogonal columns. The lower bound proof uses rank bounds for near-identity matrices (Alon, 2003) together with Turán’s theorem (bounding the number of edges in clique-free graphs). We also show how our results do and do not constrain the geometry of feature representations and extend our results to allow decoders with an activation function and bias.

Method: Proposes HiF4 format with 64-element groups, each containing 64 4-bit elements with 32 bits of shared scaling metadata. Uses three-level scaling hierarchy to capture inter- and intra-group dynamic range. Enables matrix multiplications in fixed-point manner for hardware efficiency.

Result: HiF4 achieves higher average accuracy than state-of-the-art NVFP4 format across multiple language models (LLaMA, Qwen, Mistral, DeepSeek-V3.1, LongCat) and diverse downstream tasks, while reducing hardware area and power consumption.

Conclusion: HiF4 is an effective block floating-point format for deep learning that balances accuracy and hardware efficiency, outperforming existing 4-bit formats and enabling more efficient inference.

Abstract: This paper introduces HiFloat4 (HiF4), a block floating-point data format tailored for deep learning. Each HiF4 unit packs 64 4-bit elements with 32 bits of shared scaling metadata, averaging 4.5 bits per value. The metadata specifies a three-level scaling hierarchy, capturing inter- and intra-group dynamic range while improving the utilization of the representational space. In addition, the large 64-element group size enables matrix multiplications to be executed in a highly fixed-point manner, significantly reducing hardware area and power consumption. To evaluate the proposed format, we conducted inference experiments on several language models, including LLaMA, Qwen, Mistral, DeepSeek-V3.1 and LongCat. Results show that HiF4 achieves higher average accuracy than the state-of-the-art NVFP4 format across multiple models and diverse downstream tasks.

Method: Proposes distilled neural tangent kernel (DNTK) that combines NTK-tuned dataset distillation with projection methods. Shows neural tangent space can be induced by dataset distillation, and per-class NTK matrices have low effective rank preserved by this reduction.

Result: Achieves 20-100× reduction in required Jacobian calculations, with up to five orders of magnitude reduction in computational complexity while preserving kernel structure and predictive performance.

Conclusion: Data dimension compression via dataset distillation is an effective approach for reducing NTK computational costs while maintaining accuracy, complementing existing projection methods.

Abstract: Neural tangent kernel (NTK) methods are computationally limited by the need to evaluate large Jacobians across many data points. Existing approaches reduce this cost primarily through projecting and sketching the Jacobian. We show that NTK computation can also be reduced by compressing the data dimension itself using NTK-tuned dataset distillation. We demonstrate that the neural tangent space spanned by the input data can be induced by dataset distillation, yielding a 20-100$\times$ reduction in required Jacobian calculations. We further show that per-class NTK matrices have low effective rank that is preserved by this reduction. Building on these insights, we propose the distilled neural tangent kernel (DNTK), which combines NTK-tuned dataset distillation with state-of-the-art projection methods to reduce up NTK computational complexity by up to five orders of magnitude while preserving kernel structure and predictive performance.

Method: Proposes Predictive Associative Memory (PAM) with an Inward JEPA predictor trained on temporal co-occurrence within continuous experience streams, complementing standard Outward JEPA for sensory prediction.

Result: Achieves 97% Association Precision@1, cross-boundary Recall@20 of 0.421 where cosine similarity scores zero, and discrimination AUC of 0.916 vs 0.789 for cosine similarity.

Conclusion: PAM demonstrates effective associative recall based on temporal structure rather than embedding similarity, capturing biological memory properties missing in current approaches.

Abstract: Current approaches to memory in neural systems rely on similarity-based retrieval: given a query, find the most representationally similar stored state. This assumption – that useful memories are similar memories – fails to capture a fundamental property of biological memory: association through temporal co-occurrence. We propose Predictive Associative Memory (PAM), an architecture in which a JEPA-style predictor, trained on temporal co-occurrence within a continuous experience stream, learns to navigate the associative structure of an embedding space. We introduce an Inward JEPA that operates over stored experience (predicting associatively reachable past states) as the complement to the standard Outward JEPA that operates over incoming sensory data (predicting future states). We evaluate PAM as an associative recall system – testing faithfulness of recall for experienced associations – rather than as a retrieval system evaluated on generalisation to unseen associations. On a synthetic benchmark, the predictor’s top retrieval is a true temporal associate 97% of the time (Association Precision@1 = 0.970); it achieves cross-boundary Recall@20 = 0.421 where cosine similarity scores zero; and it separates experienced-together from never-experienced-together states with a discrimination AUC of 0.916 (cosine: 0.789). Even restricted to cross-room pairs where embedding similarity is uninformative, the predictor achieves AUC = 0.849 (cosine: 0.503, chance). A temporal shuffle control confirms the signal is genuine temporal co-occurrence structure, not embedding geometry: shuffling collapses cross-boundary recall by 90%, replicated across training seeds. All results are stable across seeds (SD < 0.006) and query selections (SD $\leq$ 0.012).

Method: Reformulates the problem as online learning over decomposed decision space, solving position-wise bandit subproblems via adaptive expert-guided sequential construction.

Result: Achieves 80-98% of specialized solvers’ performance with 2-3 orders of magnitude improvement in sample/computational efficiency over Bayesian optimization. Outperforms competing methods on real-world hardware-software co-design for AI accelerators.

Conclusion: Bandit optimization over decomposed decision spaces provides a principled alternative to surrogate modeling or offline training for multi-objective optimization, with advantages growing with problem scale and objective count.

Abstract: Multi-objective combinatorial optimization seeks Pareto-optimal solutions over exponentially large discrete spaces, yet existing methods sacrifice generality, scalability, or theoretical guarantees. We reformulate it as an online learning problem over a decomposed decision space, solving position-wise bandit subproblems via adaptive expert-guided sequential construction. This formulation admits regret bounds of $O(d\sqrt{T \log T})$ depending on subproblem dimensionality (d) rather than combinatorial space size. On standard benchmarks, our method achieves 80–98% of specialized solvers performance while achieving two to three orders of magnitude improvement in sample and computational efficiency over Bayesian optimization methods. On real-world hardware-software co-design for AI accelerators with expensive simulations, we outperform competing methods under fixed evaluation budgets. The advantage grows with problem scale and objective count, establishing bandit optimization over decomposed decision spaces as a principled alternative to surrogate modeling or offline training for multi-objective optimization.

Method: Decomposes transition operator into parametric (mechanistic) and nonparametric (data-driven) components, distinguishing intervention-related vs unrelated dynamics. Two-stage procedure: pre-train encoder on simulated data, then learn corrections from observed data.

Result: Hybrid approach outperforms purely data-driven and mechanistic approaches, especially in out-of-distribution regimes. Demonstrated on periodic pendulum and Propofol bolus injection scenarios.

Conclusion: Hybrid mechanistic-data-driven models show promise for robust intervention optimization in complex real-world dynamical systems like medical treatments.

Abstract: Estimating intervention effects in dynamical systems is crucial for outcome optimization. In medicine, such interventions arise in physiological regulation (e.g., cardiovascular system under fluid administration) and pharmacokinetics, among others. Propofol administration is an anesthetic intervention, where the challenge is to estimate the optimal dose required to achieve a target brain concentration for anesthesia, given patient characteristics, while avoiding under- or over-dosing. The pharmacokinetic state is characterized by drug concentrations across tissues, and its dynamics are governed by prior states, patient covariates, drug clearance, and drug administration. While data-driven models can capture complex dynamics, they often fail in out-of-distribution (OOD) regimes. Mechanistic models on the other hand are typically robust, but might be oversimplified. We propose a hybrid mechanistic-data-driven approach to estimate time-dependent intervention outcomes. Our approach decomposes the dynamical system’s transition operator into parametric and nonparametric components, further distinguishing between intervention-related and unrelated dynamics. This structure leverages mechanistic anchors while learning residual patterns from data. For scenarios where mechanistic parameters are unknown, we introduce a two-stage procedure: first, pre-training an encoder on simulated data, and subsequently learning corrections from observed data. Two regimes with incomplete mechanistic knowledge are considered: periodic pendulum and Propofol bolus injections. Results demonstrate that our hybrid approach outperforms purely data-driven and mechanistic approaches, particularly OOD. This work highlights the potential of hybrid mechanistic-data-driven models for robust intervention optimization in complex, real-world dynamical systems.

Method: Proposes a novel bootstrapping-based regularization framework that embeds the bootstrapping process directly into deep neural network training, constraining prediction variability across resampled datasets to produce a single model with inherent stability properties.

Result: Models using the proposed approach showed improved prediction stability across simulated data and three clinical datasets (GUSTO-I, Framingham, SUPPORT), with lower mean absolute differences and fewer significantly deviating predictions while maintaining discriminative performance and feature importance consistency.

Conclusion: The approach provides a practical route toward more reliable and clinically trustworthy deep learning models by achieving greater robustness and reproducibility without sacrificing interpretability, particularly valuable in data-limited healthcare settings.

Abstract: Clinical prediction models are increasingly used to support patient care, yet many deep learning-based approaches remain unstable, as their predictions can vary substantially when trained on different samples from the same population. Such instability undermines reliability and limits clinical adoption. In this study, we propose a novel bootstrapping-based regularisation framework that embeds the bootstrapping process directly into the training of deep neural networks. This approach constrains prediction variability across resampled datasets, producing a single model with inherent stability properties. We evaluated models constructed using the proposed regularisation approach against conventional and ensemble models using simulated data and three clinical datasets: GUSTO-I, Framingham, and SUPPORT. Across all datasets, our model exhibited improved prediction stability, with lower mean absolute differences (e.g., 0.019 vs. 0.059 in GUSTO-I; 0.057 vs. 0.088 in Framingham) and markedly fewer significantly deviating predictions. Importantly, discriminative performance and feature importance consistency were maintained, with high SHAP correlations between models (e.g., 0.894 for GUSTO-I; 0.965 for Framingham). While ensemble models achieved greater stability, we show that this came at the expense of interpretability, as each constituent model used predictors in different ways. By regularising predictions to align with bootstrapped distributions, our approach allows prediction models to be developed that achieve greater robustness and reproducibility without sacrificing interpretability. This method provides a practical route toward more reliable and clinically trustworthy deep learning models, particularly valuable in data-limited healthcare settings.

Method: Proposes retrieval-aware distillation: identifies critical Gather-and-Aggregate attention heads via ablation on synthetic retrieval tasks, preserves only these heads (2% of total), and distills remaining heads into recurrent heads to create a hybrid model.

Result: Preserving just 2% of attention heads (10 heads in 1B model) recovers over 95% of teacher performance on retrieval-heavy tasks. The hybrid model is 5-6x more memory-efficient than comparable hybrids by reducing both attention cache and SSM state dimension.

Conclusion: Retrieval-aware distillation creates efficient hybrid models that close the Transformer-SSM performance gap for retrieval tasks while dramatically reducing memory requirements, enabling more practical deployment of efficient sequence models.

Abstract: State-space models (SSMs) offer efficient sequence modeling but lag behind Transformers on benchmarks that require in-context retrieval. Prior work links this gap to a small set of attention heads, termed Gather-and-Aggregate (G&A), which SSMs struggle to reproduce. We propose retrieval-aware distillation, which converts a pretrained Transformer into a hybrid student by preserving only these retrieval-critical heads and distilling the rest into recurrent heads. We identify the essential heads via ablation on a synthetic retrieval task, producing a hybrid with sparse, non-uniform attention placement. We show that preserving just 2% of attention heads recovers over 95% of teacher performance on retrieval-heavy tasks (10 heads in a 1B model), requiring far fewer heads than hybrids that retain at least 25%. We further find that large recurrent states often compensate for missing retrieval: once retrieval is handled by these heads, the SSM backbone can be simplified with limited loss, even with an $8\times$ reduction in state dimension. By reducing both the attention cache and the SSM state, the resulting hybrid is $5$–$6\times$ more memory-efficient than comparable hybrids, closing the Transformer–SSM gap at a fraction of the memory cost.

Method: Three formulations: Adaptive OpInf (sequential basis/operator refits), Adaptive NiTROM (joint Riemannian optimization of encoder/decoder and polynomial dynamics), and hybrid approach combining both; includes online data window, adaptation window, and computational budget analysis

Result: On lid-driven cavity flow: static models drift/destabilize; Adaptive OpInf suppresses amplitude drift; Adaptive NiTROM achieves near-exact energy tracking with frequent updates; hybrid most reliable under regime changes with minimal offline data

Conclusion: ROMs need cost-aware predictive claims with clear separation of training/adaptation/deployment regimes; provides template for self-correcting non-intrusive ROMs that remain effective as dynamics evolve beyond initial manifold

Abstract: Projection-based Reduced Order Models (ROMs) are often deployed as static surrogates, which limits their practical utility once a system leaves the training manifold. We formalize and study adaptive non-intrusive ROMs that update both the latent subspace and the reduced dynamics online. Building on ideas from static non-intrusive ROMs, specifically, Operator Inference (OpInf) and the recently-introduced Non-intrusive Trajectory-based optimization of Reduced-Order Models (NiTROM), we propose three formulations: Adaptive OpInf (sequential basis/operator refits), Adaptive NiTROM (joint Riemannian optimization of encoder/decoder and polynomial dynamics), and a hybrid that initializes NiTROM with an OpInf update. We describe the online data window, adaptation window, and computational budget, and analyze cost scaling. On a transiently perturbed lid-driven cavity flow, static Galerkin/OpInf/NiTROM drift or destabilize when forecasting beyond training. In contrast, Adaptive OpInf robustly suppresses amplitude drift with modest cost; Adaptive NiTROM is shown to attain near-exact energy tracking under frequent updates but is sensitive to its initialization and optimization depth; the hybrid is most reliable under regime changes and minimal offline data, yielding physically coherent fields and bounded energy. We argue that predictive claims for ROMs must be cost-aware and transparent, with clear separation of training/adaptation/deployment regimes and explicit reporting of online budgets and full-order model queries. This work provides a practical template for building self-correcting, non-intrusive ROMs that remain effective as the dynamics evolve well beyond the initial manifold.

Method: Weighted Stochastic Block Descent (WSBD) uses a dynamic, parameter-wise freezing strategy that identifies and temporarily freezes less influential parameters based on gradient-derived importance scores, focusing computational resources on critical parameters.

Result: WSBD converges on average 63.9% faster than Adam for ground-state-energy problems, with advantages growing with QNN size. It maintains full expressive capacity while reducing forward passes and outperforms traditional layer-wise freezing approaches.

Conclusion: WSBD provides an effective solution to QNN training challenges by combining dynamic parameter-wise freezing with formal convergence guarantees, offering significant efficiency improvements over standard optimizers like Adam.

Abstract: The training of Quantum Neural Networks (QNNs) is hindered by the high computational cost of gradient estimation and the barren plateau problem, where optimization landscapes become intractably flat. To address these challenges, we introduce Weighted Stochastic Block Descent (WSBD), a novel optimizer with a dynamic, parameter-wise freezing strategy. WSBD intelligently focuses computational resources by identifying and temporarily freezing less influential parameters based on a gradient-derived importance score. This approach significantly reduces the number of forward passes required per training step and helps navigate the optimization landscape more effectively. Unlike pruning or layer-wise freezing, WSBD maintains full expressive capacity while adapting throughout training. Our extensive evaluation shows that WSBD converges on average 63.9% faster than Adam for the popular ground-state-energy problem, an advantage that grows with QNN size. We provide a formal convergence proof for WSBD and show that parameter-wise freezing outperforms traditional layer-wise approaches in QNNs. Project page: https://github.com/Damrl-lab/WSBD-Stochastic-Freezing-Optimizer.

Method: The method reduces average reward RMDPs to entropy-regularized discounted robust MDPs to restore strong duality. It proves novel Lipschitz properties for general policy parameterizations, introduces a multilevel Monte Carlo gradient estimator with improved sample complexity, and designs projected gradient descent (for s-rectangular uncertainty) and Frank-Wolfe algorithms (for non-rectangular uncertainty).

Result: Achieved significant improvements: multilevel Monte Carlo gradient estimator with Õ(ε⁻²) sample complexity (O(ε⁻²) improvement), projected gradient descent with O(ε⁻⁵) for s-rectangular uncertainty, and Frank-Wolfe with O(ε⁻⁴) for discounted and O(ε⁻¹⁰·⁵) for average reward non-rectangular uncertainty.

Conclusion: This work provides the first sample complexity guarantees for RMDPs with general policy parameterization beyond (s,a)-rectangularity, and the first such guarantees in the average reward setting, significantly improving existing bounds for discounted robust MDPs.

Abstract: We study robust Markov decision processes (RMDPs) with general policy parameterization under s-rectangular and non-rectangular uncertainty sets. Prior work is largely limited to tabular policies, and hence either lacks sample complexity guarantees or incurs high computational cost. Our method reduces the average reward RMDPs to entropy-regularized discounted robust MDPs, restoring strong duality and enabling tractable equilibrium computation. We prove novel Lipschitz and Lipschitz-smoothness properties for general policy parameterizations that extends to infinite state spaces. To address infinite-horizon gradient estimation, we introduce a multilevel Monte Carlo gradient estimator with $\tilde{\mathcal{O}}(ε^{-2})$ sample complexity, a factor of $\mathcal{O}(ε^{-2})$ improvement over prior work. Building on this, we design a projected gradient descent algorithm for s-rectangular uncertainty ($\mathcal{O}(ε^{-5})$) and a Frank–Wolfe algorithm for non-rectangular uncertainty ($\mathcal{O}(ε^{-4})$ discounted, $\mathcal{O}(ε^{-10.5})$ average reward), significantly improving prior results in both the discounted setting and average reward setting. Our work is the first one to provide sample complexity guarantees for RMDPs with general policy parameterization beyond $(s, a)$-rectangularity. It also provides the first such guarantees in the average reward setting and improves existing bounds for discounted robust MDPs.

Method: Introduce Sparse Semantic Dimension (SSD) using Sparse Autoencoder features on frozen LLM activations; analyze generalization via feature sparsity rather than parameter count.

Result: Empirical validation on GPT-2 Small and Gemma-2B shows non-vacuous generalization bounds; larger models have more compressible semantic structures; feature explosion signals OOD uncertainty.

Conclusion: LLM generalization is controlled by sparse activation manifolds, not parameter count; SSD provides generalization certificates and safety monitoring via feature violation detection.

Abstract: Standard statistical learning theory predicts that Large Language Models (LLMs) should overfit because their parameter counts vastly exceed the number of training tokens. Yet, in practice, they generalize robustly. We propose that the effective capacity controlling generalization lies in the geometry of the model’s internal representations: while the parameter space is high-dimensional, the activation states lie on a low-dimensional, sparse manifold. To formalize this, we introduce the Sparse Semantic Dimension (SSD), a complexity measure derived from the active feature vocabulary of a Sparse Autoencoder (SAE) trained on the model’s layers. Treating the LLM and SAE as frozen oracles, we utilize this framework to attribute the model’s generalization capabilities to the sparsity of the dictionary rather than the total parameter count. Empirically, we validate this framework on GPT-2 Small and Gemma-2B, demonstrating that our bound provides non-vacuous certificates at realistic sample sizes. Crucially, we uncover a counter-intuitive “feature sharpness” scaling law: despite being an order of magnitude larger, Gemma-2B requires significantly fewer calibration samples to identify its active manifold compared to GPT-2, suggesting that larger models learn more compressible, distinct semantic structures. Finally, we show that this framework functions as a reliable safety monitor: out-of-distribution inputs trigger a measurable “feature explosion” (a sharp spike in active features), effectively signaling epistemic uncertainty through learned feature violation. Code is available at: https://github.com/newcodevelop/sparse-semantic-dimension.

Method: Two key observations: 1) Noise Alignment - coarse semantic steering is possible even in early corrupted stages using lightweight, offline-computed guidance signals. 2) Transferable concept vectors - concept directions in activation space transfer across timesteps and samples; fixed steering vectors learned near low noise levels remain effective when injected at intermediate noise levels. These concept directions are identified via Recursive Feature Machine (RFM), a lightweight backpropagation-free feature learning method.

Result: Experiments on CIFAR-10, ImageNet, and CelebA demonstrate improved accuracy/quality over gradient-based guidance while achieving significant inference speedups.

Conclusion: The method enables fast controllable generation from unconditional diffusion models without gradient guidance during inference, offering computational efficiency while maintaining or improving quality compared to gradient-based approaches.

Abstract: Guiding unconditional diffusion models typically requires either retraining with conditional inputs or per-step gradient computations (e.g., classifier-based guidance), both of which incur substantial computational overhead. We present a general recipe for efficiently steering unconditional diffusion {without gradient guidance during inference}, enabling fast controllable generation. Our approach is built on two observations about diffusion model structure: Noise Alignment: even in early, highly corrupted stages, coarse semantic steering is possible using a lightweight, offline-computed guidance signal, avoiding any per-step or per-sample gradients. Transferable concept vectors: a concept direction in activation space once learned transfers across both {timesteps} and {samples}; the same fixed steering vector learned near low noise level remains effective when injected at intermediate noise levels for every generation trajectory, providing refined conditional control with efficiency. Such concept directions can be efficiently and reliably identified via Recursive Feature Machine (RFM), a light-weight backpropagation-free feature learning method. Experiments on CIFAR-10, ImageNet, and CelebA demonstrate improved accuracy/quality over gradient-based guidance, while achieving significant inference speedups.

Method: Analyzes FB representation learning through connections with rank matching, fitted Q-evaluation, and contraction mapping, leading to a simplified one-step FB method that performs single-step policy improvement rather than full optimal control.

Result: One-step FB converges to errors 10^5 smaller and improves zero-shot performance by +24% on average across 10 state-based and image-based continuous control domains.

Conclusion: Simplified one-step forward-backward representation learning provides more practical unsupervised pre-training for RL by focusing on single-step policy improvement rather than full optimal control.

Abstract: As machine learning has moved towards leveraging large models as priors for downstream tasks, the community has debated the right form of prior for solving reinforcement learning (RL) problems. If one were to try to prefetch as much computation as possible, they would attempt to learn a prior over the policies for some yet-to-be-determined reward function. Recent work (forward-backward (FB) representation learning) has tried this, arguing that an unsupervised representation learning procedure can enable optimal control over arbitrary rewards without further fine-tuning. However, FB’s training objective and learning behavior remain mysterious. In this paper, we demystify FB by clarifying when such representations can exist, what its objective optimizes, and how it converges in practice. We draw connections with rank matching, fitted Q-evaluation, and contraction mapping. Our analysis suggests a simplified unsupervised pre-training method for RL that, instead of enabling optimal control, performs one step of policy improvement. We call our proposed method $\textbf{one-step forward-backward representation learning (one-step FB)}$. Experiments in didactic settings, as well as in $10$ state-based and image-based continuous control domains, demonstrate that one-step FB converges to errors $10^5$ smaller and improves zero-shot performance by $+24%$ on average. Our project website is available at https://chongyi-zheng.github.io/onestep-fb.

Method: CADET uses decoder-only transformer with: 1) context-conditioned decoding with multi-tower heads for post-scoring signals, 2) self-gated attention for stable training, 3) timestamp-based RoPE for temporal relationships, 4) session masking for train-serve consistency, and 5) production optimizations like tensor packing and custom Flash Attention kernels.

Result: In online A/B testing, CADET achieves 11.04% CTR lift compared to production LiRank baseline (hybrid ensemble of DCNv2 and sequential encoders), successfully deployed on LinkedIn’s advertising platform serving main traffic for homefeed sponsored updates.

Conclusion: CADET demonstrates that decoder-only transformers can effectively address industrial ads CTR prediction challenges, achieving significant performance improvements while maintaining production efficiency and consistency.

Abstract: Click-through rate (CTR) prediction is fundamental to online advertising systems. While Deep Learning Recommendation Models (DLRMs) with explicit feature interactions have long dominated this domain, recent advances in generative recommenders have shown promising results in content recommendation. However, adapting these transformer-based architectures to ads CTR prediction still presents unique challenges, including handling post-scoring contextual signals, maintaining offline-online consistency, and scaling to industrial workloads. We present CADET (Context-Conditioned Ads Decoder-Only Transformer), an end-to-end decoder-only transformer for ads CTR prediction deployed at LinkedIn. Our approach introduces several key innovations: (1) a context-conditioned decoding architecture with multi-tower prediction heads that explicitly model post-scoring signals such as ad position, resolving the chicken-and-egg problem between predicted CTR and ranking; (2) a self-gated attention mechanism that stabilizes training by adaptively regulating information flow at both representation and interaction levels; (3) a timestamp-based variant of Rotary Position Embedding (RoPE) that captures temporal relationships across timescales from seconds to months; (4) session masking strategies that prevent the model from learning dependencies on unavailable in-session events, addressing train-serve skew; and (5) production engineering techniques including tensor packing, sequence chunking, and custom Flash Attention kernels that enable efficient training and serving at scale. In online A/B testing, CADET achieves a 11.04% CTR lift compared to the production LiRank baseline model, a hybrid ensemble of DCNv2 and sequential encoders. The system has been successfully deployed on LinkedIn’s advertising platform, serving the main traffic for homefeed sponsored updates.

Method: Created TimeSynth framework that synthesizes signals with real-world properties (non-stationarity, periodicity, trends, phase modulation) using parameters derived from real time series. Evaluated four model families: Linear, MLPs, CNNs, and Transformers using independent train/test/validation instances for each signal family.

Result: Linear models systematically collapse to simple oscillations regardless of signal complexity, while nonlinear models avoid this collapse and gain clear advantages as signal complexity increases. Transformers and CNNs show slightly greater adaptability to complex modulated signals than MLPs. Framework also reveals robustness differences under distribution/noise shifts.

Conclusion: TimeSynth provides principled foundation for understanding when different forecasting approaches succeed or fail, moving beyond oversimplified claims of model equivalence. Nonlinear models are superior for complex temporal dynamics.

Abstract: Time series forecasting is a fundamental tool with wide ranging applications, yet recent debates question whether complex nonlinear architectures truly outperform simple linear models. Prior claims of dominance of the linear model often stem from benchmarks that lack diverse temporal dynamics and employ biased evaluation protocols. We revisit this debate through TimeSynth, a structured framework that emulates key properties of real world time series,including non-stationarity, periodicity, trends, and phase modulation by creating synthesized signals whose parameters are derived from real-world time series. Evaluating four model families Linear, Multi Layer Perceptrons (MLP), Convolutional Neural Networks (CNNs), and Transformers, we find a systematic bias in linear models: they collapse to simple oscillation regardless of signal complexity. Nonlinear models avoid this collapse and gain clear advantages as signal complexity increases. Notably, Transformers and CNN based models exhibit slightly greater adaptability to complex modulated signals compared to MLPs. Beyond clean forecasting, the framework highlights robustness differences under distribution and noise shifts and removes biases of prior benchmarks by using independent instances for train, test, and validation for each signal family. Collectively, TimeSynth provides a principled foundation for understanding when different forecasting approaches succeed or fail, moving beyond oversimplified claims of model equivalence.

Method: Uses a multi-level promotion-relegation framework that considers inter-temporal incentives, agent farsightedness, skill retention, and leg-up effects. Analyzes optimal long-term strategies for agents and designs sequences of thresholds to incentivize honest effort.

Result: Shows that a principal can design threshold sequences to effectively incentivize honest effort, and proves that under mild conditions, agents can reach arbitrarily high levels solely through genuine improvement efforts rather than manipulation.

Conclusion: Threshold-based approaches in strategic classification can effectively align incentives for honest effort, enabling agents to progress through genuine improvement rather than manipulation, offering a promising alternative to weight-centric methods.

Abstract: Strategic classification studies the problem where self-interested individuals or agents manipulate their response to obtain favorable decision outcomes made by classifiers, typically turning to dishonest actions when they are less costly than genuine efforts. While existing studies on sequential strategic classification primarily focus on optimizing dynamic classifier weights, we depart from these weight-centric approaches by analyzing the design of classifier thresholds and difficulty progression within a multi-level promotion-relegation framework. Our model captures the critical inter-temporal incentives driven by an agent’s farsightedness, skill retention, and a leg-up effect where qualification and attainment can be self-reinforcing. We characterize the agent’s optimal long-term strategy and demonstrate that a principal can design a sequence of thresholds to effectively incentivize honest effort. Crucially, we prove that under mild conditions, this mechanism enables agents to reach arbitrarily high levels solely through genuine improvement efforts.

Method: Proposes Hierarchical Concept Embedding & Pursuit (HCEP) which: 1) induces a hierarchy of concept embeddings in the latent space, 2) uses hierarchical sparse coding to recover concepts assuming they form a rooted path in the hierarchy, and 3) derives conditions for identifying hierarchical concepts in embedded space.

Result: HCEP outperforms baselines in concept precision and recall while maintaining competitive classification accuracy. It shows superior performance in both classification and concept recovery when sample size is limited. Hierarchical sparse coding reliably recovers hierarchical concept embeddings where vanilla sparse coding fails.

Conclusion: Incorporating hierarchical structures into sparse coding yields more reliable and interpretable image classification models, demonstrating the importance of respecting semantic hierarchies in interpretable computer vision.

Abstract: Interpretable-by-design models are gaining traction in computer vision because they provide faithful explanations for their predictions. In image classification, these models typically recover human-interpretable concepts from an image and use them for classification. Sparse concept recovery methods leverage the latent space of vision-language models to represent image embeddings as a sparse combination of concept embeddings. However, because such methods ignore the hierarchical structure of concepts, they can produce correct predictions with explanations that are inconsistent with the hierarchy. In this work, we propose Hierarchical Concept Embedding & Pursuit (HCEP), a framework that induces a hierarchy of concept embeddings in the latent space and uses hierarchical sparse coding to recover the concepts present in an image. Given a hierarchy of semantic concepts, we construct a corresponding hierarchy of concept embeddings and, assuming the correct concepts for an image form a rooted path in the hierarchy, derive desirable conditions for identifying them in the embedded space. We show that hierarchical sparse coding reliably recovers hierarchical concept embeddings, whereas vanilla sparse coding fails. Our experiments on real-world datasets demonstrate that HCEP outperforms baselines in concept precision and recall while maintaining competitive classification accuracy. Moreover, when the number of samples is limited, HCEP achieves superior classification accuracy and concept recovery. These results show that incorporating hierarchical structures into sparse coding yields more reliable and interpretable image classification models.

Method: MT-AIM uses conditional generative models to create synthetic MT data for desired movements and predicts joint kinematics as additional features, combining synthetic data generation with feature augmentation for classification.

Result: Achieves state-of-the-art accuracy in classifying lower back movements, bridging physiological sensing with movement analysis.

Conclusion: MT-AIM effectively addresses challenges of limited sample size and noise in wearable sensor data, enabling practical remote monitoring of back movements.

Abstract: Back pain is a pervasive issue affecting a significant portion of the population, often worsened by certain movements of the lower back. Assessing these movements is important for helping clinicians prescribe appropriate physical therapy. However, it can be difficult to monitor patients’ movements remotely outside the clinic. High-fidelity data from motion capture sensors can be used to classify different movements, but these sensors are costly and impractical for use in free-living environments. Motion Tape (MT), a new fabric-based wearable sensor, addresses these issues by being low cost and portable. Despite these advantages, novelty and variability in sensor stability make the MT dataset small scale and inherent to noise. In this work, we propose the Motion-Tape Augmentation Inference Model (MT-AIM), a deep learning classification pipeline trained on MT data. In order to address the challenges of limited sample size and noise present within the MT dataset, MT-AIM leverages conditional generative models to generate synthetic MT data of a desired movement, as well as predicting joint kinematics as additional features. This combination of synthetic data generation and feature augmentation enables MT-AIM to achieve state-of-the-art accuracy in classifying lower back movements, bridging the gap between physiological sensing and movement analysis.

Method: PRISM bridges implicit neural representations with uncertainty-aware statistical shape analysis, modeling conditional distributions of shapes given covariates, providing spatially continuous estimates of population mean and covariate-dependent uncertainty. Key innovation is a closed-form Fisher Information metric enabling efficient local temporal uncertainty quantification via automatic differentiation.

Result: Experiments on three synthetic datasets and one clinical dataset demonstrate PRISM’s strong performance across diverse tasks within a unified framework, while providing interpretable and clinically meaningful uncertainty estimates.

Conclusion: PRISM offers a novel approach to anatomical shape analysis that addresses spatial heterogeneity in developmental dynamics and provides meaningful uncertainty quantification, potentially advancing healthcare research applications.

Abstract: Understanding how anatomical shapes evolve in response to developmental covariates and quantifying their spatially varying uncertainties is critical in healthcare research. Existing approaches typically rely on global time-warping formulations that ignore spatially heterogeneous dynamics. We introduce PRISM, a novel framework that bridges implicit neural representations with uncertainty-aware statistical shape analysis. PRISM models the conditional distribution of shapes given covariates, providing spatially continuous estimates of both the population mean and covariate-dependent uncertainty at arbitrary locations. A key theoretical contribution is a closed-form Fisher Information metric that enables efficient, analytically tractable local temporal uncertainty quantification via automatic differentiation. Experiments on three synthetic datasets and one clinical dataset demonstrate PRISM’s strong performance across diverse tasks within a unified framework, while providing interpretable and clinically meaningful uncertainty estimates.

Method: Introduces an operator defined via external division of two Bregman proximity operators to promote sparse solutions while mitigating bias. Embeds this operator into the established NoLips algorithm, replacing standard Bregman proximity operator in plug-and-play manner. Provides geometric interpretations through primal and dual space reformulations.

Result: Numerical tests show more stable convergence behavior than conventional KL-based approaches and significantly superior performance on both synthetic data and image restoration problems.

Conclusion: The proposed external-division operator effectively addresses bias in Poisson inverse problems while maintaining sparsity, offering improved performance and convergence stability for sparse recovery in Poisson noise settings.

Abstract: This paper presents a novel method for recovering sparse vectors from linear models corrupted by Poisson noise. The contribution is twofold. First, an operator defined via the external division of two Bregman proximity operators is introduced to promote sparse solutions while mitigating the estimation bias induced by classical $\ell_1$-norm regularization. This operator is then embedded into the already established NoLips algorithm, replacing the standard Bregman proximity operator in a plug-and-play manner. Second, the geometric structure of the proposed external-division operator is elucidated through two complementary reformulations, which provide clear interpretations in terms of the primal and dual spaces of the Poisson inverse problem. Numerical tests show that the proposed method exhibits more stable convergence behavior than conventional Kullback-Leibler (KL)-based approaches and achieves significantly superior performance on synthetic data and an image restoration problem.

Method: Proposes CMAB-GFN which integrates a combinatorial multi-armed bandit (CMAB) framework with GFlowNet policies. The CMAB component prunes low-quality actions to yield compact high-scoring subspaces for exploration, restricting GFNs to these subspaces.

Result: Experimental results on multiple tasks demonstrate that CMAB-GFN generates higher-reward candidates than existing approaches.

Conclusion: CMAB-GFN accelerates discovery of high-value candidates by focusing exploration on high-scoring subspaces while maintaining diversity through exploration of different subspaces.

Abstract: As a probabilistic sampling framework, Generative Flow Networks (GFlowNets) show strong potential for constructing complex combinatorial objects through the sequential composition of elementary components. However, existing GFlowNets often suffer from excessive exploration over vast state spaces, leading to over-sampling of low-reward regions and convergence to suboptimal distributions. Effectively biasing GFlowNets toward high-reward solutions remains a non-trivial challenge. In this paper, we propose CMAB-GFN, which integrates a combinatorial multi-armed bandit (CMAB) framework with GFlowNet policies. The CMAB component prunes low-quality actions, yielding compact high-scoring subspaces for exploration. Restricting GFNs to these compact high-scoring subspaces accelerates the discovery of high-value candidates, while the exploration of different subspaces ensures that diversity is not sacrificed. Experimental results on multiple tasks demonstrate that CMAB-GFN generates higher-reward candidates than existing approaches.

Method: Introduces a planner to partition the entire state space into overlapping partial state spaces, allowing the actor to efficiently explore high-reward subregions. Uses a heuristic strategy to switch between partial regions to avoid wasted exploration in fully explored or low-reward areas.

Result: The proposed model converges faster than existing works on large state spaces, generates candidates with higher rewards, and significantly improves candidate diversity across several widely used datasets.

Conclusion: Partitioning state space into overlapping partial spaces with intelligent switching strategies enables more efficient GFlowNet training and better performance in large state spaces.

Abstract: Generative Flow Networks (GFlowNets) have shown promising potential to generate high-scoring candidates with probability proportional to their rewards. As existing GFlowNets freely explore in state space, they encounter significant convergence challenges when scaling to large state spaces. Addressing this issue, this paper proposes to restrict the exploration of actor. A planner is introduced to partition the entire state space into overlapping partial state spaces. Given their limited size, these partial state spaces allow the actor to efficiently identify subregions with higher rewards. A heuristic strategy is introduced to switch partial regions thus preventing the actor from wasting time exploring fully explored or low-reward partial regions. By iteratively exploring these partial state spaces, the actor learns to converge towards the high-reward subregions within the entire state space. Experiments on several widely used datasets demonstrate that \modelname converges faster than existing works on large state spaces. Furthermore, \modelname not only generates candidates with higher rewards but also significantly improves their diversity.

Method: Develops a streaming algorithm that processes clusterings sequentially while storing only logarithmic number of inputs. Introduces a generic framework combining closest fair clustering with cluster fitting, which is fairness-agnostic and works with any fairness definition.

Result: First constant-factor approximation algorithm for fair consensus clustering in streaming model with logarithmic memory. Framework also improves offline approximation guarantees and extends to k-median consensus clustering.

Conclusion: The paper presents an efficient streaming approach for fair consensus clustering that overcomes memory limitations while maintaining theoretical guarantees, with a flexible framework applicable to various fairness notions.

Abstract: Consensus clustering seeks to combine multiple clusterings of the same dataset, potentially derived by considering various non-sensitive attributes by different agents in a multi-agent environment, into a single partitioning that best reflects the overall structure of the underlying dataset. Recent work by Chakraborty et al, introduced a fair variant under proportionate fairness and obtained a constant-factor approximation by naively selecting the best closest fair input clustering; however, their offline approach requires storing all input clusterings, which is prohibitively expensive for most large-scale applications. In this paper, we initiate the study of fair consensus clustering in the streaming model, where input clusterings arrive sequentially and memory is limited. We design the first constant-factor algorithm that processes the stream while storing only a logarithmic number of inputs. En route, we introduce a new generic algorithmic framework that integrates closest fair clustering with cluster fitting, yielding improved approximation guarantees not only in the streaming setting but also when revisited offline. Furthermore, the framework is fairness-agnostic: it applies to any fairness definition for which an approximately close fair clustering can be computed efficiently. Finally, we extend our methods to the more general k-median consensus clustering problem.

Method: Two calibration methods: 1) Under affine miscalibration in logit space, use simple regression to identify outside-option utility parameters and recover no-purchase probabilities. 2) Under weaker nearly monotone condition, propose rank-based calibration method with finite-sample error bounds separating auxiliary-predictor quality from utility-learning error.

Result: Methods enable consistent recovery of no-purchase probabilities without collecting new labels for no-purchase events. Error bounds quantify how calibration accuracy affects downstream revenue performance in assortment optimization.

Conclusion: The calibration framework turns imperfect auxiliary predictions into statistically valid no-purchase estimates, with explicit error bounds connecting estimation quality to decision performance in assortment optimization.

Abstract: Firms typically cannot observe key consumer actions: whether customers buy from a competitor, choose not to buy, or even fully consider the firm’s offer. This missing outside-option information makes market-size and preference estimation difficult even in simple multinomial logit (MNL) models, and it is a central obstacle in practice when only transaction data are recorded. Existing approaches often rely on auxiliary market-share, aggregated, or cross-market data. We study a complementary setting in which a black-box auxiliary predictor provides outside-option probabilities, but is potentially biased or miscalibrated because it was trained in a different channel, period, or population, or produced by an external machine-learning system. We develop calibration methods that turn such imperfect predictions into statistically valid no-purchase estimates using purchase-only data from the focal environment. First, under affine miscalibration in logit space, we show that a simple regression identifies outside-option utility parameters and yields consistent recovery of no-purchase probabilities without collecting new labels for no-purchase events. Second, under a weaker nearly monotone condition, we propose a rank-based calibration method and derive finite-sample error bounds that cleanly separate auxiliary-predictor quality from first-stage utility-learning error over observed in-set choices. Our analysis also translates estimation error into downstream decision quality for assortment optimization, quantifying how calibration accuracy affects revenue performance. The bounds provide explicit dependence on predictor alignment and utility-learning error, clarifying when each source dominates. Numerical experiments demonstrate improvements in no-purchase estimation and downstream assortment decisions, and we discuss robust aggregation extensions for combining multiple auxiliary predictors.

Method: Proposes a systematic framework based on the Roofline model that unifies architectural primitives and hardware constraints through operational intensity (OI), defines an inference-potential region, and introduces Relative Inference Potential as a novel metric to compare efficiency differences between LLMs on the same hardware.

Result: Extensive empirical analysis reveals that performance and OI variations are significantly influenced by sequence length, identifies a critical regression in OI as model depth increases, highlights an efficiency trap induced by hardware heterogeneity, and demonstrates how structural refinements like Multi-head Latent Attention can unlock latent inference potential across hardware substrates.

Conclusion: The framework provides actionable directions for hardware-software co-design to align neural structures with physical constraints in on-device intelligence, with released code available for implementation.

Abstract: The transition toward localized intelligence through Small Language Models (SLMs) has intensified the need for rigorous performance characterization on resource-constrained edge hardware. However, objectively measuring the theoretical performance ceilings of diverse architectures across heterogeneous platforms remains a formidable challenge. In this work, we propose a systematic framework based on the Roofline model that unifies architectural primitives and hardware constraints through the lens of operational intensity (OI). By defining an inference-potential region, we introduce the Relative Inference Potential as a novel metric to compare efficiency differences between Large Language Models (LLMs) on the same hardware substrate. Extensive empirical analysis across diverse compute tiers reveals that variations in performance and OI are significantly influenced by sequence length. We further identify a critical regression in OI as model depth increases. Additionally, our findings highlight an efficiency trap induced by hardware heterogeneity and demonstrate how structural refinements, such as Multi-head Latent Attention (M LA), can effectively unlock latent inference potential across various hardware substrates. These insights provide actionable directions for hardware-software co-design to align neural structures with physical constraints in on-device intelligence. The released code is available in the Appendix C.

Method: Introduces a unified regularization approach that explicitly balances preventing reward hacking and maintaining stable policy updates. The method yields a weighted supervised fine-tuning loss with a superior trade-off between these objectives.

Result: Extensive experiments across diverse benchmarks show the method consistently outperforms RLHF and online preference learning methods, achieving enhanced alignment performance and stability.

Conclusion: The proposed simple yet principled alignment objective improves both alignment results and implementation complexity by providing a unified approach to address RLHF’s core challenges.

Abstract: Reinforcement Learning from Human Feedback (RLHF) has advanced alignment capabilities significantly but remains hindered by two core challenges: \textbf{reward hacking} and \textbf{stable optimization}. Current solutions independently address these issues through separate regularization strategies, specifically a KL-divergence penalty against a supervised fine-tuned model ($π_0$) to mitigate reward hacking, and policy ratio clipping towards the current policy ($π_t$) to promote stable alignment. However, the implicit trade-off arising from simultaneously regularizing towards both $π_0$ and $π_t$ remains under-explored. In this paper, we introduce a unified regularization approach that explicitly balances the objectives of preventing reward hacking and maintaining stable policy updates. Our simple yet principled alignment objective yields a weighted supervised fine-tuning loss with a superior trade-off, which demonstrably improves both alignment results and implementation complexity. Extensive experiments across diverse benchmarks validate that our method consistently outperforms RLHF and online preference learning methods, achieving enhanced alignment performance and stability.

Method: Proposes ADMIRE mechanism that constructs verifiable adaptive rewards by anchoring trajectories to milestones dynamically distilled from successful explorations, with asymmetric credit assignment that denoises successful trajectories and scaffolds failed ones.

Result: ADMIRE yields over 10% absolute improvement in success rate across different base models on AndroidWorld, and shows robust generalizability across diverse RL algorithms and heterogeneous environments like web navigation and embodied tasks.

Conclusion: ADMIRE effectively resolves the reward fidelity-density conflict in RL for long-horizon tasks through adaptive milestone construction and asymmetric credit assignment, demonstrating strong performance improvements and generalization capabilities.

Abstract: Reinforcement Learning (RL) has emerged as a mainstream paradigm for training Mobile GUI Agents, yet it struggles with the temporal credit assignment problem inherent in long-horizon tasks. A primary challenge lies in the trade-off between reward fidelity and density: outcome reward offers high fidelity but suffers from signal sparsity, while process reward provides dense supervision but remains prone to bias and reward hacking. To resolve this conflict, we propose the Adaptive Milestone Reward (ADMIRE) mechanism. ADMIRE constructs a verifiable, adaptive reward system by anchoring trajectory to milestones, which are dynamically distilled from successful explorations. Crucially, ADMIRE integrates an asymmetric credit assignment strategy that denoises successful trajectories and scaffolds failed trajectories. Extensive experiments demonstrate that ADMIRE consistently yields over 10% absolute improvement in success rate across different base models on AndroidWorld. Moreover, the method exhibits robust generalizability, achieving strong performance across diverse RL algorithms and heterogeneous environments such as web navigation and embodied tasks.

Method: PASCAL uses phase-aware scheduling that prioritizes reasoning phases to reduce TTFT, with controlled preemption and token pacing during answering phases to preserve QoE. It employs a hierarchical scheduler combining instance-level placement with intra-instance execution, enabling dynamic migration at phase boundaries to balance load and reduce interference.

Result: Across benchmarks using DeepSeek-R1-Distill-Qwen-32B, PASCAL reduces tail TTFT by up to 72% while maintaining answering phase SLO attainment.

Conclusion: Phase-aware scheduling is crucial for efficient reasoning-based LLM deployment, demonstrating significant improvements in TTFT while maintaining service quality.

Abstract: The emergence of reasoning-based LLMs leveraging Chain-of-Thought (CoT) inference introduces new serving challenges, as their extended reasoning phases delay user-visible output and inflate Time-To-First-Token (TTFT). Existing LLM serving frameworks fail to distinguish between reasoning and answering phases, leading to performance degradation under GPU memory constraints. We present PASCAL, a phase-aware scheduling algorithm that prioritizes reasoning to reduce TTFT while using controlled preemption and token pacing during answering to preserve Quality-of-Experience (QoE). Our hierarchical scheduler combines instance-level placement with intra-instance execution and enables dynamic migration at phase boundaries to balance load and reduce interference. Across benchmarks using DeepSeek-R1-Distill-Qwen-32B, PASCAL reduces tail TTFT by up to 72% while maintaining answering phase SLO attainment, demonstrating the importance of phase-aware scheduling for reasoning-based LLM deployment.

Method: ALTTS uses a dual-path framework with explicit decoupling: (1) an autoregressive (AR) path with a linear predictor for stable within-series dynamics, and (2) a cross-relation (CR) path using a Transformer with Cross-Relation Self-Attention (CRSA) for cross-dimension interactions. The two branches are coordinated via alternating optimization to isolate gradient noise and reduce cross-block interference.

Result: Extensive experiments on multiple benchmarks show ALTTS consistently outperforms prior methods, with the most pronounced improvements on long-horizon forecasting, demonstrating that optimization strategies rather than complex architectures drive progress in multivariate time series forecasting.

Conclusion: Carefully designed optimization strategies that decouple autoregressive and cross-relation modeling through alternating optimization can significantly improve multivariate time series forecasting performance, particularly for long-horizon predictions, suggesting this approach is more effective than increasingly complex model architectures.

Abstract: Multivariate time series forecasting involves two qualitatively distinct factors: (i) stable within-series autoregressive (AR) dynamics, and (ii) intermittent cross-dimension interactions that can become spurious over long horizons. We argue that fitting a single model to capture both effects creates an optimization conflict: the high-variance updates needed for cross-dimension modeling can corrupt the gradients that support autoregression, resulting in brittle training and degraded long-horizon accuracy. To address this, we propose ALTTS, a dual-path framework that explicitly decouples autoregression and cross-relation (CR) modeling. In ALTTS, the AR path is instantiated with a linear predictor, while the CR path uses a Transformer equipped with Cross-Relation Self-Attention (CRSA); the two branches are coordinated via alternating optimization to isolate gradient noise and reduce cross-block interference. Extensive experiments on multiple benchmarks show that ALTTS consistently outperforms prior methods, with the most pronounced improvements on long-horizon forecasting. Overall, our results suggest that carefully designed optimization strategies, rather than ever more complex architectures, can be a key driver of progress in multivariate time series forecasting.

Method: Introduces Krause Attention, inspired by bounded-confidence consensus dynamics. Replaces similarity-based global aggregation with distance-based, localized, and selectively sparse interactions. Tokens only interact with neighbors within a confidence bound, promoting structured local synchronization rather than global mixing.

Result: Experiments across vision (ViT on CIFAR/ImageNet), autoregressive generation (MNIST/CIFAR-10), and large language models (Llama/Qwen) show consistent gains with substantially reduced computation. The method reduces runtime complexity from quadratic to linear in sequence length.

Conclusion: Bounded-confidence dynamics serve as a scalable and effective inductive bias for attention, alleviating attention sinks while maintaining performance with reduced computational cost.

Abstract: Self-attention in Transformers relies on globally normalized softmax weights, causing all tokens to compete for influence at every layer. When composed across depth, this interaction pattern induces strong synchronization dynamics that favor convergence toward a dominant mode, a behavior associated with representation collapse and attention sink phenomena. We introduce Krause Attention, a principled attention mechanism inspired by bounded-confidence consensus dynamics. Krause Attention replaces similarity-based global aggregation with distance-based, localized, and selectively sparse interactions, promoting structured local synchronization instead of global mixing. We relate this behavior to recent theory modeling Transformer dynamics as interacting particle systems, and show how bounded-confidence interactions naturally moderate attention concentration and alleviate attention sinks. Restricting interactions to local neighborhoods also reduces runtime complexity from quadratic to linear in sequence length. Experiments across vision (ViT on CIFAR/ImageNet), autoregressive generation (MNIST/CIFAR-10), and large language models (Llama/Qwen) demonstrate consistent gains with substantially reduced computation, highlighting bounded-confidence dynamics as a scalable and effective inductive bias for attention.

Method: Two frameworks: Forward Forecasting Model (FFM) forecasts future sequences using hybrid architecture (TCN+GRU+Transformer); Backward Reconstruction Model (BRM) reconstructs recent history from future context. Both detect anomalies based on forecasting errors and directional embedding discrepancies.

Result: Outperforms state-of-the-art baselines on four benchmark datasets (MSL, SMAP, SMD, PSM) across detection metrics and significantly improves timeliness of anomaly anticipation.

Conclusion: The proposed proactive anomaly detection frameworks are well-suited for deployment in time-sensitive domains requiring proactive monitoring due to their ability to handle heterogeneous data and noisy conditions.

Abstract: Reactive anomaly detection methods, which are commonly deployed to identify anomalies after they occur based on observed deviations, often fall short in applications that demand timely intervention, such as industrial monitoring, finance, and cybersecurity. Proactive anomaly detection, by contrast, aims to detect early warning signals before failures fully manifest, but existing methods struggle with handling heterogeneous multivariate data and maintaining precision under noisy or unpredictable conditions. In this work, we introduce two proactive anomaly detection frameworks: the Forward Forecasting Model (FFM) and the Backward Reconstruction Model (BRM). Both models leverage a hybrid architecture combining Temporal Convolutional Networks (TCNs), Gated Recurrent Units (GRUs), and Transformer encoders to model directional temporal dynamics. FFM forecasts future sequences to anticipate disruptions, while BRM reconstructs recent history from future context to uncover early precursors. Anomalies are flagged based on forecasting error magnitudes and directional embedding discrepancies. Our models support both continuous and discrete multivariate features, enabling robust performance in real-world settings. Extensive experiments on four benchmark datasets, MSL, SMAP, SMD, and PSM, demonstrate that FFM and BRM outperform state-of-the-art baselines across detection metrics and significantly improve the timeliness of anomaly anticipation. These properties make our approach well-suited for deployment in time-sensitive domains requiring proactive monitoring.

Method: NRT treats reasoning process as a latent variable and uses a unified training objective that models reasoning as an optimization problem. It intrinsically rewards paths that increase the model’s likelihood of producing ground-truth answers, creating a self-reinforcing feedback loop where the model learns to resolve its own uncertainty.

Result: NRT achieves state-of-the-art performance among verifier-free methods on Llama and Mistral model families, significantly outperforming standard SFT baselines and prior verifier-free RL methods. It shows strong gains in complex reasoning domains and high robustness to policy collapse.

Conclusion: NRT offers a general, scalable path toward building more powerful and broadly applicable reasoning systems by eliminating dependency on human-annotated reasoning traces and external verifiers.

Abstract: The prevailing paradigm for training large reasoning models–combining Supervised Fine-Tuning (SFT) with Reinforcement Learning with Verifiable Rewards (RLVR)–is fundamentally constrained by its reliance on high-quality, human-annotated reasoning data and external verifiers. This dependency incurs significant data-collection costs, risks embedding human cognitive biases, and confines the reinforcement learning stage to objectively assessable domains like mathematics and coding, leaving a wide range of unverifiable tasks beyond its scope. To overcome these limitations, we introduce NRT (Native Reasoning Training), a novel framework that cultivates complex reasoning by having the model generate its own reasoning traces using only standard question-answer pairs, thereby obviating the need for expert-written demonstrations. NRT reframes the training problem by treating the reasoning process as a latent variable. It employs a unified training objective that models reasoning as an optimization problem, intrinsically rewarding paths that increase the model’s likelihood of producing the ground-truth answer. This unified perspective allows us to analyze intrinsic failure modes of prior methods, such as policy collapse, and systematically design more robust reward aggregation functions, creating a self-reinforcing feedback loop where the model learns to think in ways that resolve its own uncertainty. Empirical evaluation on Llama and Mistral model families demonstrates that NRT achieves state-of-the-art performance among verifier-free methods, significantly outperforming standard SFT baselines and prior verifier-free RL methods. Our approach yields particularly strong performance gains in complex reasoning domains and exhibits high robustness to policy collapse, offering a general, scalable path toward building more powerful and broadly applicable reasoning systems.

Method: Propose TS-Memory, a lightweight memory adapter for frozen TSFMs. Two-stage training: 1) Construct offline, leakage-safe kNN teacher that synthesizes confidence-aware quantile targets from retrieved futures. 2) Distill retrieval-induced distributional correction into memory adapter via confidence-gated supervision. During inference, fuses memory and backbone predictions with constant-time overhead.

Result: Experiments across diverse TSFMs and benchmarks show consistent improvements in both point and probabilistic forecasting over representative adaptation methods, with efficiency comparable to frozen backbone.

Conclusion: TS-Memory enables efficient domain adaptation for time series foundation models by distilling retrieval knowledge into parametric memory, achieving improved forecasting performance without the latency costs of retrieval-based methods.

Abstract: Time Series Foundation Models (TSFMs) achieve strong zero-shot forecasting through large-scale pre-training, but adapting them to downstream domains under distribution shift remains challenging. Existing solutions face a trade-off: Parametric Adaptation can cause catastrophic forgetting and requires costly multi-domain maintenance, while Non-Parametric Retrieval improves forecasts but incurs high inference latency due to datastore search. We propose Parametric Memory Distillation and implement it as TS-Memory, a lightweight memory adapter that augments frozen TSFMs. TS-Memory is trained in two stages. First, we construct an offline, leakage-safe kNN teacher that synthesizes confidence-aware quantile targets from retrieved futures. Second, we distill this retrieval-induced distributional correction into a lightweight memory adapter via confidence-gated supervision. During inference, TS-Memory fuses memory and backbone predictions with constant-time overhead, enabling retrieval-free deployment. Experiments across diverse TSFMs and benchmarks demonstrate consistent improvements in both point and probabilistic forecasting over representative adaptation methods, with efficiency comparable to the frozen backbone.

Method: Theoretical analysis of mini-batch stochastic steepest descent under general entry-wise and Schatten-p norms, examining convergence behavior with different batch sizes, momentum, and variance reduction techniques.

Result: Without momentum, convergence only occurs with large batches; momentum enables small-batch convergence through batch-momentum trade-off but slows convergence; variance reduction can recover exact full-batch implicit bias for any batch size; batch-size-one steepest descent without momentum converges to fundamentally different bias.

Conclusion: The unified analysis clarifies when stochastic optimization aligns with full-batch behavior and reveals limitations of purely stochastic updates, providing dimension-free rates that improve upon prior results.

Abstract: A variety of widely used optimization methods like SignSGD and Muon can be interpreted as instances of steepest descent under different norm-induced geometries. In this work, we study the implicit bias of mini-batch stochastic steepest descent in multi-class classification, characterizing how batch size, momentum, and variance reduction shape the limiting max-margin behavior and convergence rates under general entry-wise and Schatten-$p$ norms. We show that without momentum, convergence only occurs with large batches, yielding a batch-dependent margin gap but the full-batch convergence rate. In contrast, momentum enables small-batch convergence through a batch-momentum trade-off, though it slows convergence. This approach provides fully explicit, dimension-free rates that improve upon prior results. Moreover, we prove that variance reduction can recover the exact full-batch implicit bias for any batch size, albeit at a slower convergence rate. Finally, we further investigate the batch-size-one steepest descent without momentum, and reveal its convergence to a fundamentally different bias via a concrete data example, which reveals a key limitation of purely stochastic updates. Overall, our unified analysis clarifies when stochastic optimization aligns with full-batch behavior, and paves the way for perform deeper explorations of the training behavior of stochastic gradient steepest descent algorithms.

Method: Organizes BFMs under self-supervised learning taxonomy from model perspective, summarizes downstream tasks from dataset perspective, and creates Brain4FMs platform with plug-and-play interfaces integrating 15 BFMs and 18 public EEG datasets.

Result: Created an open evaluation platform enabling standardized comparisons and analysis of how pretraining data, SSL strategies, and architectures affect generalization and downstream performance in brain signal analysis.

Conclusion: Brain4FMs provides a much-needed standardized framework for evaluating Brain Foundation Models, facilitating systematic comparisons and guiding development of more effective neural signal analysis models.

Abstract: Brain Foundation Models (BFMs) are transforming neuroscience by enabling scalable and transferable learning from neural signals, advancing both clinical diagnostics and cutting-edge neuroscience exploration. Their emergence is powered by large-scale clinical recordings, particularly electroencephalography (EEG) and intracranial EEG, which provide rich temporal and spatial representations of brain dynamics. However, despite their rapid proliferation, the field lacks a unified understanding of existing methodologies and a standardized evaluation framework. To fill this gap, we map the benchmark design space along two axes: (i) from the model perspective, we organize BFMs under a self-supervised learning (SSL) taxonomy; and (ii) from the dataset perspective, we summarize common downstream tasks and curate representative public datasets across clinical and human-centric neurotechnology applications. Building on this consolidation, we introduce Brain4FMs, an open evaluation platform with plug-and-play interfaces that integrates 15 representative BFMs and 18 public datasets. It enables standardized comparisons and analysis of how pretraining data, SSL strategies, and architectures affect generalization and downstream performance, guiding more accurate and transferable BFMs. The code is available at https://anonymous.4open.science/r/Brain4FMs-85B8.

Method: Proposes FedSynSAM which leverages the global model trajectory to construct synthetic data, enabling accurate estimation of global perturbation for SAM. This addresses the limitation of previous approaches that used previous round’s model update as a rough estimate.

Result: The algorithm’s convergence is established theoretically, and extensive experiments validate its effectiveness in improving performance when gradient compression is used in federated learning.

Conclusion: FedSynSAM successfully mitigates the sharp loss landscape problem induced by gradient compression in federated learning, particularly under non-IID data distributions, by accurately estimating global perturbation through synthetic data from model trajectory.

Abstract: It is commonly believed that gradient compression in federated learning (FL) enjoys significant improvement in communication efficiency with negligible performance degradation. In this paper, we find that gradient compression induces sharper loss landscapes in federated learning, particularly under non-IID data distributions, which suggests hindered generalization capability. The recently emerging Sharpness Aware Minimization (SAM) effectively searches for a flat minima by incorporating a gradient ascent step (i.e., perturbing the model with gradients) before the celebrated stochastic gradient descent. Nonetheless, the direct application of SAM in FL suffers from inaccurate estimation of the global perturbation due to data heterogeneity. Existing approaches propose to utilize the model update from the previous communication round as a rough estimate. However, its effectiveness is hindered when model update compression is incorporated. In this paper, we propose FedSynSAM, which leverages the global model trajectory to construct synthetic data and facilitates an accurate estimation of the global perturbation. The convergence of the proposed algorithm is established, and extensive experiments are conducted to validate its effectiveness.

Method: Progressive Self-Correction (ProSeCo) trains a model to perform both unmasking and correction by reusing MDM denoising network outputs as inputs for corrector training. During generation, it applies corrective refinement steps between unmasking steps to change decoded tokens and improve outputs.

Result: ProSeCo yields better quality-efficiency trade-offs (up to 2-3x faster sampling) and enables inference-time compute scaling to further increase sample quality beyond standard MDMs (up to 1.3x improvement on benchmarks).

Conclusion: ProSeCo addresses fundamental limitations of masked diffusion models by enabling iterative refinement of entire sequences, including already generated tokens, leading to improved sample quality and efficiency.

Abstract: Masked diffusion models (MDMs) have emerged as a promising alternative to autoregressive models, enabling parallel token generation while achieving competitive performance. Despite these advantages, MDMs face a fundamental limitation: once tokens are unmasked, they remain fixed, leading to error accumulation and ultimately degrading sample quality. We address this by proposing a framework that trains a model to perform both unmasking and correction. By reusing outputs from the MDM denoising network as inputs for corrector training, we train a model to recover from potential mistakes. During generation we apply additional corrective refinement steps between unmasking ones in order to change decoded tokens and improve outputs. We name our training and sampling method Progressive Self-Correction (ProSeCo) for its unique ability to iteratively refine an entire sequence, including already generated tokens. We conduct extensive experimental validation across multiple conditional and unconditional tasks, demonstrating that ProSeCo yields better quality-efficiency trade-offs (up to ~2-3x faster sampling) and enables inference-time compute scaling to further increase sample quality beyond standard MDMs (up to ~1.3x improvement on benchmarks).

Method: Decomposes data filtering into specialized raters (one per capability), each trained via meta-learning on disjoint validation objectives. Uses progressive selection: samples retained if any rater ranks them above thresholds that tighten over time, preserving diversity early while focusing on high-value samples late.

Result: At 2B parameters, SkillRater improves over unfiltered baselines by 5.63% on visual understanding, 2.00% on OCR, and 3.53% on STEM reasoning on held-out benchmarks. Learned rater signals are near orthogonal, confirming independent quality dimensions.

Conclusion: Multidimensional quality decomposition outperforms both unfiltered training and monolithic learned filtering by capturing genuinely independent capability dimensions through specialized raters.

Abstract: Data curation methods typically assign samples a single quality score. We argue this scalar framing is fundamentally limited: when training requires multiple distinct capabilities, a monolithic scorer cannot maximize useful signals for all of them simultaneously. Quality is better understood as multidimensional, with each dimension corresponding to a capability the model must acquire. We introduce SkillRater, a framework that decomposes data filtering into specialized raters - one per capability, each trained via meta-learning on a disjoint validation objective - and composes their scores through a progressive selection rule: at each training stage, a sample is retained if any rater ranks it above a threshold that tightens over time, preserving diversity early while concentrating on high-value samples late. We validate this approach on vision language models, decomposing quality into three capability dimensions: visual understanding, OCR, and STEM reasoning. At 2B parameters, SkillRater improves over unfiltered baselines by 5.63% on visual understanding, 2.00% on OCR, and 3.53% on STEM on held out benchmarks. The learned rater signals are near orthogonal, confirming that the decomposition captures genuinely independent quality dimensions and explaining why it outperforms both unfiltered training and monolithic learned filtering.

Method: Pretrained Chemical Language Models while scaling training resources and measured transfer performance across diverse molecular property prediction tasks. Also evaluated alternative metrics based on Hessian or loss landscape, and analyzed failure modes through parameter space visualizations.

Result: Pretraining loss consistently decreases with increased training resources, but downstream task performance shows limited improvement. Hessian/loss landscape metrics also fail to estimate downstream performance. Downstream performance can saturate or degrade despite pretraining improvements, with task-dependent failure modes identified.

Conclusion: There’s a significant gap between pretraining-based evaluation and downstream performance in Chemical Language Models, emphasizing the need for model selection and evaluation strategies that explicitly account for downstream task characteristics rather than relying solely on pretraining metrics.

Abstract: Chemical Language Models (CLMs) pre-trained on large scale molecular data are widely used for molecular property prediction. However, the common belief that increasing training resources such as model size, dataset size, and training compute improves both pretraining loss and downstream task performance has not been systematically validated in the chemical domain. In this work, we evaluate this assumption by pretraining CLMs while scaling training resources and measuring transfer performance across diverse molecular property prediction (MPP) tasks. We find that while pretraining loss consistently decreases with increased training resources, downstream task performance shows limited improvement. Moreover, alternative metrics based on the Hessian or loss landscape also fail to estimate downstream performance in CLMs. We further identify conditions under which downstream performance saturates or degrades despite continued improvements in pretraining metrics, and analyze the underlying task dependent failure modes through parameter space visualizations. These results expose a gap between pretraining based evaluation and downstream performance, and emphasize the need for model selection and evaluation strategies that explicitly account for downstream task characteristics.

Method: Proposes TreeGrad algorithm that computes gradients of multilinear extension of joint optimization objective in O(L) time for trees with L leaves. Builds TreeGrad-Ranker for feature ranking by aggregating gradients, and TreeGrad-Shap for numerically stable Beta Shapley value computation. Also develops TreeProb as generalization of Linear TreeShap.

Result: TreeGrad-Shap shows up to 10^15 times smaller numerical error than Linear TreeShap. TreeGrad-Ranker significantly outperforms existing methods on both insertion and deletion metrics. The approach satisfies all axioms characterizing probabilistic values except linearity.

Conclusion: Direct optimization of joint objective via TreeGrad provides more reliable feature rankings than probabilistic values for decision tree explanations, with superior numerical stability and performance on standard evaluation metrics.

Abstract: We revisit the use of probabilistic values, which include the well-known Shapley and Banzhaf values, to rank features for explaining the local predicted values of decision trees. The quality of feature rankings is typically assessed with the insertion and deletion metrics. Empirically, we observe that co-optimizing these two metrics is closely related to a joint optimization that selects a subset of features to maximize the local predicted value while minimizing it for the complement. However, we theoretically show that probabilistic values are generally unreliable for solving this joint optimization. Therefore, we explore deriving feature rankings by directly optimizing the joint objective. As the backbone, we propose TreeGrad, which computes the gradients of the multilinear extension of the joint objective in $O(L)$ time for decision trees with $L$ leaves; these gradients include weighted Banzhaf values. Building upon TreeGrad, we introduce TreeGrad-Ranker, which aggregates the gradients while optimizing the joint objective to produce feature rankings, and TreeGrad-Shap, a numerically stable algorithm for computing Beta Shapley values with integral parameters. In particular, the feature scores computed by TreeGrad-Ranker satisfy all the axioms uniquely characterizing probabilistic values, except for linearity, which itself leads to the established unreliability. Empirically, we demonstrate that the numerical error of Linear TreeShap can be up to $10^{15}$ times larger than that of TreeGrad-Shap when computing the Shapley value. As a by-product, we also develop TreeProb, which generalizes Linear TreeShap to support all probabilistic values. In our experiments, TreeGrad-Ranker performs significantly better on both insertion and deletion metrics. Our code is available at https://github.com/watml/TreeGrad.

Method: Proposes Arbitrary Geometry-encoded Transformer (ArGEnT) with three attention variants (self-attention, cross-attention, hybrid-attention) to encode geometric information from point-cloud representations, integrated into DeepONet as the trunk network.

Result: Significantly improved prediction accuracy and generalization performance compared to standard DeepONet and other geometry-aware surrogates across fluid dynamics, solid mechanics, and electrochemical systems benchmarks.

Conclusion: ArGEnT provides a scalable surrogate modeling framework for complex physical systems by combining flexible geometry encoding with operator-learning capabilities, enabling accurate geometry-conditioned predictions with reduced reliance on signed distance functions.

Abstract: Learning solution operators for systems with complex, varying geometries and parametric physical settings is a central challenge in scientific machine learning. In many-query regimes such as design optimization, control and inverse problems, surrogate modeling must generalize across geometries while allowing flexible evaluation at arbitrary spatial locations. In this work, we propose Arbitrary Geometry-encoded Transformer (ArGEnT), a geometry-aware attention-based architecture for operator learning on arbitrary domains. ArGEnT employs Transformer attention mechanisms to encode geometric information directly from point-cloud representations with three variants-self-attention, cross-attention, and hybrid-attention-that incorporates different strategies for incorporating geometric features. By integrating ArGEnT into DeepONet as the trunk network, we develop a surrogate modeling framework capable of learning operator mappings that depend on both geometric and non-geometric inputs without the need to explicitly parametrize geometry as a branch network input. Evaluation on benchmark problems spanning fluid dynamics, solid mechanics and electrochemical systems, we demonstrate significantly improved prediction accuracy and generalization performance compared with the standard DeepONet and other existing geometry-aware saurrogates. In particular, the cross-attention transformer variant enables accurate geometry-conditioned predictions with reduced reliance on signed distance functions. By combining flexible geometry encoding with operator-learning capabilities, ArGEnT provides a scalable surrogate modeling framework for optimization, uncertainty quantification, and data-driven modeling of complex physical systems.

Method: The paper proposes GP2F, a dual-branch GPL method that explicitly instantiates two extremes: (1) a frozen branch retaining pre-trained knowledge, and (2) an adapted branch with lightweight adapters for task-specific adaptation. The method performs adaptive fusion under topology constraints via contrastive loss and topology-consistent loss.

Result: Extensive experiments on cross-domain few-shot node and graph classification demonstrate that GP2F outperforms existing methods. Theoretical analysis shows that jointly leveraging both branches yields smaller estimation error than using either branch alone.

Conclusion: Cross-domain GPL benefits from the integration between pre-trained knowledge and task-specific adaptation. The proposed GP2F method effectively combines these complementary approaches through dual-branch architecture with constrained fusion mechanisms.

Abstract: Graph Prompt Learning (GPL) has recently emerged as a promising paradigm for downstream adaptation of pre-trained graph models, mitigating the misalignment between pre-training objectives and downstream tasks. Recently, the focus of GPL has shifted from in-domain to cross-domain scenarios, which is closer to the real world applications, where the pre-training source and downstream target often differ substantially in data distribution. However, why GPLs remain effective under such domain shifts is still unexplored. Empirically, we observe that representative GPL methods are competitive with two simple baselines in cross-domain settings: full fine-tuning (FT) and linear probing (LP), motivating us to explore a deeper understanding of the prompting mechanism. We provide a theoretical analysis demonstrating that jointly leveraging these two complementary branches yields a smaller estimation error than using either branch alone, formally proving that cross-domain GPL benefits from the integration between pre-trained knowledge and task-specific adaptation. Based on this insight, we propose GP2F, a dual-branch GPL method that explicitly instantiates the two extremes: (1) a frozen branch that retains pre-trained knowledge, and (2) an adapted branch with lightweight adapters for task-specific adaptation. We then perform adaptive fusion under topology constraints via a contrastive loss and a topology-consistent loss. Extensive experiments on cross-domain few-shot node and graph classification demonstrate that our method outperforms existing methods.

Method: Proposes Targeted Interpretable Perturbation (TIP) with dual-targeting strategy: 1) Uses Grad-CAM to identify critical convolution channels encoding semantic features, 2) Transforms selected kernels to frequency domain via DFT and selectively injects calibrated perturbations into high-frequency spectrum.

Result: Extensive experiments show TIP renders reconstructed images visually unrecognizable against state-of-the-art GIAs while maintaining global model accuracy comparable to non-private baselines, significantly outperforming existing DP-based defenses in privacy-utility trade-off.

Conclusion: TIP provides an effective defense against gradient inversion attacks by selectively perturbing high-frequency components to destroy fine-grained reconstruction details while preserving low-frequency information crucial for model accuracy, offering better interpretability and utility preservation than traditional DP methods.

Abstract: Federated Learning (FL) facilitates collaborative model training while preserving data locality; however, the exchange of gradients renders the system vulnerable to Gradient Inversion Attacks (GIAs), allowing adversaries to reconstruct private training data with high fidelity. Existing defenses, such as Differential Privacy (DP), typically employ indiscriminate noise injection across all parameters, which severely degrades model utility and convergence stability. To address those limitation, we proposes Targeted Interpretable Perturbation (TIP), a novel defense framework that integrates model interpretability with frequency domain analysis. Unlike conventional methods that treat parameters uniformly, TIP introduces a dual-targeting strategy. First, leveraging Gradient-weighted Class Activation Mapping (Grad-CAM) to quantify channel sensitivity, we dynamically identify critical convolution channels that encode primary semantic features. Second, we transform these selected kernels into the frequency domain via the Discrete Fourier Transform and selectively inject calibrated perturbations into the high-frequency spectrum. By selectively perturbing high-frequency components, TIP effectively destroys the fine-grained details necessary for image reconstruction while preserving the low-frequency information crucial for model accuracy. Extensive experiments on benchmark datasets demonstrate that TIP renders reconstructed images visually unrecognizable against state-of-the-art GIAs, while maintaining global model accuracy comparable to non-private baselines, significantly outperforming existing DP-based defenses in the privacy-utility trade-off and interpretability. Code is available in https://github.com/2766733506/asldkfjssdf_arxiv

Method: LG-Plug aligns topology and text representations for fine-grained embeddings, generates consensus-driven OOD exposure via clustered iterative LLM prompting, uses lightweight in-cluster codebook and heuristic sampling to reduce LLM querying time, and integrates OOD exposure as regularization with existing detectors.

Result: The method produces improved OOD detection performance on text-attributed graphs by addressing the reliability-informativeness imbalance and enabling seamless integration with existing detectors.

Conclusion: LG-Plug effectively combines LLM textual knowledge with graph structural information for better OOD detection on text-attributed graphs, overcoming limitations of previous approaches.

Abstract: Text-attributed graphs (TAGs) associate nodes with textual attributes and graph structure, enabling GNNs to jointly model semantic and structural information. While effective on in-distribution (ID) data, GNNs often encounter out-of-distribution (OOD) nodes with unseen textual or structural patterns in real-world settings, leading to overconfident and erroneous predictions in the absence of reliable OOD detection. Early approaches address this issue from a topology-driven perspective, leveraging neighboring structures to mitigate node-level detection bias. However, these methods typically encode node texts as shallow vector features, failing to fully exploit rich semantic information. In contrast, recent LLM-based approaches generate pseudo OOD priors by leveraging textual knowledge, but they suffer from several limitations: (1) a reliability-informativeness imbalance in the synthesized OOD priors, as the generated OOD exposures either deviate from the true OOD semantics, or introduce non-negligible ID noise, all of which offers limited improvement to detection performance; (2) reliance on specialized architectures, which prevents incorporation of the extensive effective topology-level insights that have been empirically validated in prior work. To this end, we propose LG-Plug, an LLM-Guided Plug-and-play strategy for TAG OOD detection tasks. LG-Plug aligns topology and text representations to produce fine-grained node embeddings, then generates consensus-driven OOD exposure via clustered iterative LLM prompting. Moreover, it leverages lightweight in-cluster codebook and heuristic sampling reduce time cost of LLM querying. The resulting OOD exposure serves as a regularization term to separate ID and OOD nodes, enabling seamless integration with existing detectors.

Method: Three-step proof: (1) Show UMAP’s stochastic optimization with negative sampling is a contrastive learning objective on similarity graphs; (2) Use HaoChen et al.’s result establishing equivalence between contrastive learning on similarity graphs and spectral clustering; (3) Verify UMAP’s spectral initialization computes exact linear solution to this spectral problem.

Result: The equivalence is exact for Gaussian kernels and holds as first-order approximation for UMAP’s default Cauchy-type kernel. This unifies UMAP, contrastive learning, and spectral clustering under a single framework.

Conclusion: UMAP performs spectral clustering on fuzzy k-nearest neighbor graphs, providing theoretical grounding for empirical observations about UMAP’s behavior and unifying several dimensionality reduction and clustering approaches.

Abstract: UMAP (Uniform Manifold Approximation and Projection) is among the most widely used algorithms for non linear dimensionality reduction and data visualisation. Despite its popularity, and despite being presented through the lens of algebraic topology, the exact relationship between UMAP and classical spectral methods has remained informal. In this work, we prove that UMAP performs spectral clustering on the fuzzy k nearest neighbour graph. Our proof proceeds in three steps: (1) we show that UMAP’s stochastic optimisation with negative sampling is a contrastive learning objective on the similarity graph; (2) we invoke the result of HaoChen et al. [8], establishing that contrastive learning on a similarity graph is equivalent to spectral clustering; and (3) we verify that UMAP’s spectral initialisation computes the exact linear solution to this spectral problem. The equivalence is exact for Gaussian kernels, and holds as a first order approximation for UMAP’s default Cauchy type kernel. Our result unifies UMAP, contrastive learning, and spectral clustering under a single framework, and provides theoretical grounding for several empirical observations about UMAP’s behaviour.

Method: Reformulate the bilevel problem as a single-level online problem with inequality constraints, construct Lagrangian functions, and propose a fully first-order algorithm. Also develop an improved variant with adaptive inner-iteration scheme.

Result: The proposed algorithm achieves regret of O(1 + V_T + H_{2,T}), and the improved variant achieves O(√T + V_T) regret, which is advantageous when V_T ≥ O(√T).

Conclusion: The work presents efficient first-order methods for online bilevel optimization that eliminate the need for Hessian-vector products while maintaining strong theoretical guarantees.

Abstract: In this work, we study non-convex-strongly-convex online bilevel optimization (OBO). Existing OBO algorithms are mainly based on hypergradient descent, which requires access to a Hessian-vector product (HVP) oracle and potentially incurs high computational costs. By reformulating the original OBO problem as a single-level online problem with inequality constraints and constructing a sequence of Lagrangian function, we eliminate the need for HVPs arising from implicit differentiation. Specifically, we propose a fully first-order algorithm for OBO, and provide theoretical guarantees showing that it achieves regret of $O(1 + V_T + H_{2,T})$. Furthermore, we develop an improved variant with an adaptive inner-iteration scheme, which removes the dependence on the drift variation of the inner-level optimal solution and achieves regret of $O(\sqrt{T} + V_T)$. This regret have the advatange when $V_{T}\ge O(\sqrt{T})$.

Method: Used optical motion capture systems to analyze 839 treadmill recordings. Focused on stance phase using joint/segment angle time series and discrete point values. Tested three classification tasks (healthy vs. injured, PFPS, ITBS) with different feature spaces. Evaluated classical algorithms (KNN, Gaussian Processes, Decision Trees) and deep learning architectures (CNNs, LSTMs). Used explainability tools like Shapley values, saliency maps, and Grad-CAM.

Result: Combining time series with point values substantially improved detection. Deep learning models outperformed classical ones, with CNNs achieving highest accuracy: 77.9% for PFPS, 73.8% for ITBS, and 71.43% for combined injury class.

Conclusion: Motion capture systems coupled with advanced machine learning have potential to identify knee injury-related running patterns, with CNNs showing best performance for injury classification.

Abstract: Running is a widely practiced activity but shows a high incidence of knee injuries, especially Patellofemoral Pain Syndrome (PFPS) and Iliotibial Band Syndrome (ITBS). Identifying gait patterns linked to these injuries can improve clinical decision-making, which requires precise systems capable of capturing and analyzing temporal kinematic data. This study uses optical motion capture systems to enhance detection of injury-related running patterns. We analyze a public dataset of 839 treadmill recordings from healthy and injured runners to evaluate how effectively these systems capture dynamic parameters relevant to injury classification. The focus is on the stance phase, using joint and segment angle time series and discrete point values. Three classification tasks are addressed: healthy vs. injured, healthy vs. PFPS, and healthy vs. ITBS. We examine different feature spaces, from traditional point-based metrics to full stance-phase time series and hybrid representations. Multiple models are tested, including classical algorithms (K-Nearest Neighbors, Gaussian Processes, Decision Trees) and deep learning architectures (CNNs, LSTMs). Performance is evaluated with accuracy, precision, recall, and F1-score. Explainability tools such as Shapley values, saliency maps, and Grad-CAM are used to interpret model behavior. Results show that combining time series with point values substantially improves detection. Deep learning models outperform classical ones, with CNNs achieving the highest accuracy: 77.9% for PFPS, 73.8% for ITBS, and 71.43% for the combined injury class. These findings highlight the potential of motion capture systems coupled with advanced machine learning to identify knee injury-related running patterns.

Method: Tasks are sampled from anonymized Perplexity Deep Research usage data, filtered and augmented to ensure they’re open-ended, complex, objectively evaluable, and representative of real-world research use cases across 10 domains and 40 countries.

Result: DRACO benchmark is publicly available with tasks graded along four dimensions: factual accuracy, breadth/depth of analysis, presentation quality/objectivity, and citation quality.

Conclusion: DRACO provides a comprehensive benchmark for evaluating AI research systems on complex, real-world tasks, addressing the need for nuanced assessment beyond simple factual accuracy.

Abstract: We present DRACO (Deep Research Accuracy, Completeness, and Objectivity), a benchmark of complex deep research tasks. These tasks, which span 10 domains and draw on information sources from 40 countries, originate from anonymized real-world usage patterns within a large-scale deep research system. Tasks are sampled from a de-identified dataset of Perplexity Deep Research requests, then filtered and augmented to ensure that the tasks are anonymized, open-ended and complex, objectively evaluable, and representative of the broad scope of real-world deep research use cases. Outputs are graded against task-specific rubrics along four dimensions: factual accuracy (accuracy), breadth and depth of analysis (including completeness), presentation quality (including objectivity), and citation quality. DRACO is publicly available at https://hf.co/datasets/perplexity-ai/draco.

Method: ANML combines gradient-based consistency (q), verification status (v), contributor reputation (r), and temporal relevance (T) to produce per-contributor quality weights. Uses Two-Stage Adaptive gating to guarantee performance never underperforms best baseline, even under strategic attacks.

Result: Achieves 33-72% error reduction over gradient-only baselines across 5 datasets (178-32,561 samples). Quality-weighted training is data-efficient: 20% high-quality data outperforms 100% uniformly weighted data by 47%. Contributor-level attribution provides 1.3-5.3x greater improvement than sample-level methods.

Conclusion: ANML successfully integrates data quality signals with gradient information to improve model performance while enabling attribution, with robustness guarantees against strategic attacks and superior performance in detecting subtle corruption.

Abstract: Frontier AI systems increasingly train on specialized expert data, from clinical records to proprietary research to curated datasets, yet current training pipelines treat all samples identically. A Nobel laureate’s contribution receives the same weight as an unverified submission. We introduce ANML (Attribution-Native Machine Learning), a framework that weights training samples by four quality factors: gradient-based consistency (q), verification status (v), contributor reputation (r), and temporal relevance (T). By combining what the model observes (gradient signals) with what the system knows about data provenance (external signals), ANML produces per-contributor quality weights that simultaneously improve model performance and enable downstream attribution. Across 5 datasets (178-32,561 samples), ANML achieves 33-72% error reduction over gradient-only baselines. Quality-weighted training is data-efficient: 20% high-quality data outperforms 100% uniformly weighted data by 47%. A Two-Stage Adaptive gating mechanism guarantees that ANML never underperforms the best available baseline, including under strategic joint attacks combining credential faking with gradient alignment. When per-sample detection fails against subtle corruption, contributor-level attribution provides 1.3-5.3x greater improvement than sample-level methods, with the advantage growing as corruption becomes harder to detect.

Method: Proposes SpiralFormer, a looped Transformer with multi-resolution recursion schedule that executes recurrence across different scales, enabling hierarchical dependency learning through iteration-wise functional specialization.

Result: SpiralFormer achieves better parameter and compute efficiency than both looped and non-looped baselines across model scales from 160M to 1.4B parameters.

Conclusion: Multi-resolution recursion enables hierarchical learning and establishes sequence resolution as a potential axis for scaling recursive architectures, improving efficiency over existing approaches.

Abstract: Recursive (looped) Transformers decouple computational depth from parameter depth by repeatedly applying shared layers, providing an explicit architectural primitive for iterative refinement and latent reasoning. However, early looped Transformers often underperform non-recursive baselines of equal compute. While recent literature has introduced more effective recursion mechanisms to mitigate this gap, existing architectures still operate at a fixed, full-token resolution, neglecting the potential efficiency of computing over compressed latent representations. In this paper, we propose SpiralFormer, a looped Transformer that executes recurrence under a multi-resolution recursion schedule. We provide probing evidence that multi-resolution recursion enables the model to learn hierarchical dependencies by inducing iteration-wise functional specialization across different scales. Empirically, SpiralFormer achieves better parameter and compute efficiency than both looped and non-looped baselines across model scales from 160M to 1.4B, establishing sequence resolution as a potential axis for scaling recursive architectures.

Method: Proposes TabSieve with two-stage approach: 1) selects small set of informative rows as evidence, 2) predicts missing target conditioned on selected evidence. Uses TabSieve-SFT-40K dataset synthesized from 331 real tables with strong teacher model filtering. Introduces TAB-GRPO reinforcement learning that jointly optimizes evidence selection and prediction correctness with separate rewards and dynamic task-advantage balancing.

Result: Experiments on 75 classification and 52 regression tables show TabSieve consistently improves performance across shot budgets, with average gains of 2.92% on classification and 4.45% on regression over second-best baseline. Analysis shows TabSieve concentrates more attention on selected evidence, improving robustness to noisy context.

Conclusion: TabSieve provides an effective framework for explicit evidence usage in tabular prediction, demonstrating improved performance and robustness through its select-then-predict approach with joint optimization of evidence selection and prediction.

Abstract: Tabular prediction can benefit from in-table rows as few-shot evidence, yet existing tabular models typically perform instance-wise inference and LLM-based prompting is often brittle. Models do not consistently leverage relevant rows, and noisy context can degrade performance. To address this challenge, we propose TabSieve, a select-then-predict framework that makes evidence usage explicit and auditable. Given a table and a query row, TabSieve first selects a small set of informative rows as evidence and then predicts the missing target conditioned on the selected evidence. To enable this capability, we construct TabSieve-SFT-40K by synthesizing high-quality reasoning trajectories from 331 real tables using a strong teacher model with strict filtering. Furthermore, we introduce TAB-GRPO, a reinforcement learning recipe that jointly optimizes evidence selection and prediction correctness with separate rewards, and stabilizes mixed regression and classification training via dynamic task-advantage balancing. Experiments on a held-out benchmark of 75 classification and 52 regression tables show that TabSieve consistently improves performance across shot budgets, with average gains of 2.92% on classification and 4.45% on regression over the second-best baseline. Further analysis indicates that TabSieve concentrates more attention on the selected evidence, which improves robustness to noisy context.

Method: Modulates observation noise covariance in Bayesian filters based on local potential energy value. Observations are trusted near potential minima and discounted near barriers. Implemented in Extended, Unscented, Ensemble, Adaptive Kalman filters and particle filters with only two hyperparameters.

Result: 57-80% RMSE improvement over standard Extended Kalman Filter on synthetic Ginzburg-Landau benchmarks with 10% outlier contamination. Robust to misspecification (47% improvement even with 50% parameter deviation). Applied to climate data showing asymmetry parameter estimation and 91% variance explained by outlier fraction.

Conclusion: Potential-energy gating provides physics-based robust filtering that outperforms statistical methods by leveraging system dynamics. The approach is effective for state estimation in double-well stochastic systems with outlier contamination.

Abstract: We introduce potential-energy gating, a method for robust state estimation in systems governed by double-well stochastic dynamics. The observation noise covariance of a Bayesian filter is modulated by the local value of a known or assumed potential energy function: observations are trusted when the state is near a potential minimum and progressively discounted as it approaches the barrier separating metastable wells. This physics-based mechanism differs from purely statistical robust filters, which treat all regions of state space identically, and from constrained filters, which impose hard bounds on states rather than modulating observation trust. We implement the gating within Extended, Unscented, Ensemble, and Adaptive Kalman filters and particle filters, requiring only two additional hyperparameters. Synthetic benchmarks on a Ginzburg-Landau double-well process with 10% outlier contamination and Monte Carlo validation over 100 replications show 57-80% RMSE improvement over the standard Extended Kalman Filter, all statistically significant (p < 10^{-15}, Wilcoxon signed-rank test). A naive topological baseline using only distance to the nearest well achieves 57%, confirming that the continuous energy landscape adds an additional ~21 percentage points. The method is robust to misspecification: even when assumed potential parameters deviate by 50% from their true values, improvement never falls below 47%. Comparing externally forced and spontaneous Kramers-type transitions, gating retains 68% improvement under noise-induced transitions whereas the naive baseline degrades to 30%. As an empirical illustration, we apply the framework to Dansgaard-Oeschger events in the NGRIP delta-18O ice-core record, estimating asymmetry parameter gamma = -0.109 (bootstrap 95% CI: [-0.220, -0.011], excluding zero) and demonstrating that outlier fraction explains 91% of the variance in filter improvement.

Method: Created CuKe dataset for CUDA kernels, developed BiC-RL framework with two phases (infilling and end-to-end generation), and built DICE diffusion LLMs at 1.7B, 4B, and 8B scales.

Result: DICE significantly outperforms both autoregressive and diffusion LLMs of comparable scale on KernelBench, establishing new state-of-the-art for CUDA kernel generation.

Conclusion: Diffusion LLMs can be effectively tailored for specialized code generation tasks like CUDA kernels through curated datasets and training frameworks, demonstrating advantages over autoregressive approaches.

Abstract: Diffusion large language models (dLLMs) have emerged as a compelling alternative to autoregressive (AR) LLMs, owing to their capacity for parallel token generation. This paradigm is particularly well-suited for code generation, where holistic structural planning and non-sequential refinement are critical. Despite this potential, tailoring dLLMs for CUDA kernel generation remains challenging, obstructed not only by the high specialization but also by the severe lack of high-quality training data. To address these challenges, we construct CuKe, an augmented supervised fine-tuning dataset optimized for high-performance CUDA kernels. On top of it, we propose a bi-phase curated reinforcement learning (BiC-RL) framework consisting of a CUDA kernel infilling stage and an end-to-end CUDA kernel generation stage. Leveraging this training framework, we introduce DICE, a series of diffusion large language models designed for CUDA kernel generation, spanning three parameter scales, 1.7B, 4B, and 8B. Extensive experiments on KernelBench demonstrate that DICE significantly outperforms both autoregressive and diffusion LLMs of comparable scale, establishing a new state-of-the-art for CUDA kernel generation.

Method: Proposes a simple activation-space PEFT technique that freezes base weights and learns per-neuron thresholds and gains. During training, a smooth gate decides whether a neuron’s activation participates; at inference the gate can be hardened to yield explicit conditional computation and neuron-level attributions.

Result: Tested on MNIST vs rotated MNIST specialization. The method improves rotated accuracy over frozen baseline while using only a few hundred trainable parameters per layer, exhibits partial activation sparsity, and trades some accuracy for fewer parameters and more interpretable neuron-level mechanisms compared to LoRA.

Conclusion: The activation-space PEFT approach offers an interpretable alternative to weight-based adaptation, enabling mode specialization through selective neuron participation, though with reduced expressivity when the frozen base lacks needed features.

Abstract: Parameter-efficient fine-tuning (PEFT) methods such as \lora{} adapt large pretrained models by adding small weight-space updates. While effective, weight deltas are hard to interpret mechanistically, and they do not directly expose \emph{which} internal computations are reused versus bypassed for a new task. We explore an alternative view inspired by neuromodulation: adaptation as a change in \emph{mode} – selecting and rescaling existing computations – rather than rewriting the underlying weights. We propose \methodname{}, a simple activation-space PEFT technique that freezes base weights and learns per-neuron \emph{thresholds} and \emph{gains}. During training, a smooth gate decides whether a neuron’s activation participates; at inference the gate can be hardened to yield explicit conditional computation and neuron-level attributions. As a proof of concept, we study mode specialization'' on MNIST (0$^\circ$) versus rotated MNIST (45$^\circ$). We pretrain a small MLP on a 50/50 mixture (foundation), freeze its weights, and then specialize to the rotated mode using \methodname{}. Across seeds, \methodname{} improves rotated accuracy over the frozen baseline while using only a few hundred trainable parameters per layer, and exhibits partial activation sparsity (a minority of units strongly active). Compared to \lora{}, \methodname{} trades some accuracy for substantially fewer trainable parameters and a more interpretable which-neurons-fire’’ mechanism. We discuss limitations, including reduced expressivity when the frozen base lacks features needed for the target mode.

Method: UFO integrates parallelizable multiscale feature extraction of U-Nets, global modeling of Transformers, and continuous-time dynamics of Neural CDEs. Creates fully causal, parallelizable model with global receptive field while maintaining local temporal sensitivity.

Result: UFO consistently outperforms ten state-of-the-art neural baselines on five standard benchmarks for both regularly and irregularly sampled time series. Achieves up to 15× faster inference than conventional Neural CDEs with strong performance on long, highly multivariate sequences.

Conclusion: UFO successfully addresses limitations of Neural CDEs by enabling parallel computation while capturing both global context and local temporal dynamics, making it effective for probabilistic time series forecasting.

Abstract: Probabilistic forecasting of irregularly sampled time series is crucial in domains such as healthcare and finance, yet it remains a formidable challenge. Existing Neural Controlled Differential Equation (Neural CDE) approaches, while effective at modelling continuous dynamics, suffer from slow, inherently sequential computation, which restricts scalability and limits access to global context. We introduce UFO (U-Former ODE), a novel architecture that seamlessly integrates the parallelizable, multiscale feature extraction of U-Nets, the powerful global modelling of Transformers, and the continuous-time dynamics of Neural CDEs. By constructing a fully causal, parallelizable model, UFO achieves a global receptive field while retaining strong sensitivity to local temporal dynamics. Extensive experiments on five standard benchmarks – covering both regularly and irregularly sampled time series – demonstrate that UFO consistently outperforms ten state-of-the-art neural baselines in predictive accuracy. Moreover, UFO delivers up to 15$\times$ faster inference compared to conventional Neural CDEs, with consistently strong performance on long and highly multivariate sequences.

Method: TUBO features two key components: burst processing for handling significant traffic fluctuations and model selection for adapting to varying traffic patterns using a pool of models. It provides deterministic predictions with quantified uncertainty.

Result: TUBO significantly outperforms existing methods on forecasting accuracy (by 4 times) and achieves up to 94% accuracy in burst occurrence forecasting on three real-world network datasets. In proactive traffic engineering applications, it improves aggregated throughput by 9 times compared to reactive approaches.

Conclusion: TUBO provides a reliable solution for network traffic forecasting with practical applications in traffic engineering, demonstrating superior performance over existing methods through its burst processing and model selection capabilities.

Abstract: Traffic forecasting based network operation optimization and management offers enormous promise but also presents significant challenges from traffic forecasting perspective. While deep learning models have proven to be relatively more effective than traditional statistical methods for time series forecasting, their reliability is not satisfactory due to their inability to effectively handle unique characteristics of network traffic. In particular, the burst and complex traffic patterns makes the existing models less reliable, as each type of deep learning model has limited capability in capturing traffic patterns. To address this issue, we introduce TUBO, a novel machine learning framework custom designed for reliable network traffic forecasting. TUBO features two key components: burst processing for handling significant traffic fluctuations and model selection for adapting to varying traffic patterns using a pool of models. A standout feature of TUBO is its ability to provide deterministic predictions along with quantified uncertainty, which serves as a cue for identifying the most reliable forecasts. Evaluations on three real-world network demand matrix (DM) datasets (Abilene, GEANT, and CERNET) show that TUBO significantly outperforms existing methods on forecasting accuracy (by 4 times), and also achieves up to 94% accuracy in burst occurrence forecasting. Furthermore, we also consider traffic demand forecasting based proactive traffic engineering (TE) as a downstream use case. Our results show that compared to reactive approaches and proactive TE using the best existing DM forecasting methods, proactive TE powered by TUBO improves aggregated throughput by 9 times and 3 times, respectively.

Method: MUSE introduces a two-level score transformation that maps model outputs to a stable reference distribution, combined with dynamic intent-based routing for sharing models across tenants. This decouples model scores from client decision boundaries.

Result: Deployed at scale by Feedzai, MUSE processes over 1,000 events per second and 55 billion events in 12 months across dozens of tenants while maintaining high availability and low latency. It reduces model lead time from weeks to minutes.

Conclusion: MUSE enables seamless model updates in multi-tenant environments, promoting model resilience against shifting attacks while saving millions in fraud losses and operational costs by eliminating the threshold recalibration bottleneck.

Abstract: In binary classification systems, decision thresholds translate model scores into actions. Choosing suitable thresholds relies on the specific distribution of the underlying model scores but also on the specific business decisions of each client using that model. However, retraining models inevitably shifts score distributions, invalidating existing thresholds. In multi-tenant Score-as-a-Service environments, where decision boundaries reside in client-managed infrastructure, this creates a severe bottleneck: recalibration requires coordinating threshold updates across hundreds of clients, consuming excessive human hours and leading to model stagnation. We introduce MUSE, a model serving framework that enables seamless model updates by decoupling model scores from client decision boundaries. Designed for multi-tenancy, MUSE optimizes infrastructure re-use by sharing models via dynamic intent-based routing, combined with a two-level score transformation that maps model outputs to a stable, reference distribution. Deployed at scale by Feedzai, MUSE processes over a thousand events per second, and over 55 billion events in the last 12 months, across several dozens of tenants, while maintaining high-availability and low-latency guarantees. By reducing model lead time from weeks to minutes, MUSE promotes model resilience against shifting attacks, saving millions of dollars in fraud losses and operational costs.

Method: Hierarchical two-loop framework: inner loop updates LLM policy (e.g., using GRPO) with trajectories sampled at meta-policy-selected temperature; outer loop updates temperature distribution by rewarding temperatures that maximize likelihood of high-advantage trajectories.

Result: Outperforms baselines with fixed or heuristic temperatures on five mathematical reasoning benchmarks, demonstrating temperature as an effective learnable meta-policy for adaptive exploration in LLM RL.

Conclusion: Temperature can be effectively learned as a meta-policy for adaptive exploration in LLM reinforcement learning, enabling better policy improvement through dynamic temperature control.

Abstract: Temperature is a crucial hyperparameter in large language models (LLMs), controlling the trade-off between exploration and exploitation during text generation. High temperatures encourage diverse but noisy outputs, while low temperatures produce focused outputs but may cause premature convergence. Yet static or heuristic temperature schedules fail to adapt to the dynamic demands of reinforcement learning (RL) throughout training, often limiting policy improvement. We propose Temperature Adaptive Meta Policy Optimization (TAMPO), a new framework that recasts temperature control as a learnable meta-policy. TAMPO operates through a hierarchical two-loop process. In the inner loop, the LLM policy is updated (e.g., using GRPO) with trajectories sampled at the temperature selected by the meta-policy. In the outer loop, meta-policy updates the distribution over candidate temperatures by rewarding those that maximize the likelihood of high-advantage trajectories. This trajectory-guided, reward-driven mechanism enables online adaptation without additional rollouts, directly aligning exploration with policy improvement. On five mathematical reasoning benchmarks, TAMPO outperforms baselines using fixed or heuristic temperatures, establishing temperature as an effective learnable meta-policy for adaptive exploration in LLM reinforcement learning. Accepted at ICLR 2026.

Method: SPECTRE uses a minimax-fair approach that adjusts the spectrum of a simple Fourier feature mapping and constrains how much the worst-case distribution can deviate from the empirical distribution, avoiding overly pessimistic scenarios.

Result: Experiments on American Community Survey datasets across 20 states show SPECTRE provides the highest average fairness guarantees with smallest interquartile range compared to state-of-the-art methods, even those with demographic information access.

Conclusion: SPECTRE offers a practical solution for fairness-aware classification without demographic data, providing theoretical guarantees and empirical safety in fairness performance.

Abstract: As automated classification systems become increasingly prevalent, concerns have emerged over their potential to reinforce and amplify existing societal biases. In the light of this issue, many methods have been proposed to enhance the fairness guarantees of classifiers. Most of the existing interventions assume access to group information for all instances, a requirement rarely met in practice. Fairness without access to demographic information has often been approached through robust optimization techniques,which target worst-case outcomes over a set of plausible distributions known as the uncertainty set. However, their effectiveness is strongly influenced by the chosen uncertainty set. In fact, existing approaches often overemphasize outliers or overly pessimistic scenarios, compromising both overall performance and fairness. To overcome these limitations, we introduce SPECTRE, a minimax-fair method that adjusts the spectrum of a simple Fourier feature mapping and constrains the extent to which the worst-case distribution can deviate from the empirical distribution. We perform extensive experiments on the American Community Survey datasets involving 20 states. The safeness of SPECTRE comes as it provides the highest average values on fairness guarantees together with the smallest interquartile range in comparison to state-of-the-art approaches, even compared to those with access to demographic group information. In addition, we provide a theoretical analysis that derives computable bounds on the worst-case error for both individual groups and the overall population, as well as characterizes the worst-case distributions responsible for these extremal performances

Method: APST repeatedly samples identical prompts under controlled conditions (e.g., varying temperature) to surface latent failure modes. It models failures as stochastic outcomes using Bernoulli and binomial models to estimate per-inference failure probabilities, enabling quantitative reliability comparisons across models and configurations.

Result: Models with similar benchmark scores show substantially different empirical failure rates under repeated sampling, especially as temperature increases. Shallow single-sample evaluation obscures meaningful reliability differences that emerge under sustained use.

Conclusion: APST complements existing benchmarks by providing a practical framework for evaluating LLM safety and reliability under repeated inference, bridging benchmark alignment with deployment-oriented risk assessment for real-world operational scenarios.

Abstract: Traditional benchmarks for large language models (LLMs) primarily assess safety risk through breadth-oriented evaluation across diverse tasks. However, real-world deployment exposes a different class of risk: operational failures arising from repeated inference on identical or near-identical prompts rather than broad task generalization. In high-stakes settings, response consistency and safety under sustained use are critical. We introduce Accelerated Prompt Stress Testing (APST), a depth-oriented evaluation framework inspired by reliability engineering. APST repeatedly samples identical prompts under controlled operational conditions (e.g., decoding temperature) to surface latent failure modes including hallucinations, refusal inconsistency, and unsafe completions. Rather than treating failures as isolated events, APST models them as stochastic outcomes of independent inference events. We formalize safety failures using Bernoulli and binomial models to estimate per-inference failure probabilities, enabling quantitative comparison of reliability across models and decoding configurations. Applying APST to multiple instruction-tuned LLMs evaluated on AIR-BENCH-derived safety prompts, we find that models with similar benchmark-aligned scores can exhibit substantially different empirical failure rates under repeated sampling, particularly as temperature increases. These results demonstrate that shallow, single-sample evaluation can obscure meaningful reliability differences under sustained use. APST complements existing benchmarks by providing a practical framework for evaluating LLM safety and reliability under repeated inference, bridging benchmark alignment and deployment-oriented risk assessment.

Method: Proposes A²V-SLP framework with: 1) Disentangled VAE that encodes ground-truth sign poses into articulator-specific mean and variance vectors, 2) Non-autoregressive Transformer that predicts both latent means and log-variances from text embeddings, 3) Stochastic sampling during decoding, and 4) Gloss attention mechanism to strengthen linguistic-motion alignment.

Result: Achieves consistent gains over deterministic latent regression, state-of-the-art back-translation performance, and improved motion realism in a fully gloss-free setting.

Conclusion: The proposed alignment-aware variational framework with articulator-wise disentangled latent distributions effectively maintains articulator-level representations and improves sign language production quality through distributional latent modeling.

Abstract: Building upon recent structural disentanglement frameworks for sign language production, we propose A$^{2}$V-SLP, an alignment-aware variational framework that learns articulator-wise disentangled latent distributions rather than deterministic embeddings. A disentangled Variational Autoencoder (VAE) encodes ground-truth sign pose sequences and extracts articulator-specific mean and variance vectors, which are used as distributional supervision for training a non-autoregressive Transformer. Given text embeddings, the Transformer predicts both latent means and log-variances, while the VAE decoder reconstructs the final sign pose sequences through stochastic sampling at the decoding stage. This formulation maintains articulator-level representations by avoiding deterministic latent collapse through distributional latent modeling. In addition, we integrate a gloss attention mechanism to strengthen alignment between linguistic input and articulated motion. Experimental results show consistent gains over deterministic latent regression, achieving state-of-the-art back-translation performance and improved motion realism in a fully gloss-free setting.

Method: Combines spectral Galerkin projection with truncated Wiener chaos expansion to reduce infinite-dimensional SPDE to finite system of parametrized ODEs. Uses variational learning to jointly infer latent temporal dynamics and stochastic forcing without requiring explicit noise observation during training.

Result: Demonstrates state-of-the-art performance on synthetic data across bounded and unbounded one-dimensional spatial domains under comparable modeling assumptions.

Conclusion: Proposed structured latent-variable formulation successfully learns underlying randomly forced dynamics from only solution observations, providing a principled approach to separate deterministic evolution from stochastic forcing in SPDE learning.

Abstract: We study the problem of learning the law of linear stochastic partial differential equations (SPDEs) with additive Gaussian forcing from spatiotemporal observations. Most existing deep learning approaches either assume access to the driving noise or initial condition, or rely on deterministic surrogate models that fail to capture intrinsic stochasticity. We propose a structured latent-variable formulation that requires only observations of solution realizations and learns the underlying randomly forced dynamics. Our approach combines a spectral Galerkin projection with a truncated Wiener chaos expansion, yielding a principled separation between deterministic evolution and stochastic forcing. This reduces the infinite-dimensional SPDE to a finite system of parametrized ordinary differential equations governing latent temporal dynamics. The latent dynamics and stochastic forcing are jointly inferred through variational learning, allowing recovery of stochastic structure without explicit observation or simulation of noise during training. Empirical evaluation on synthetic data demonstrates state-of-the-art performance under comparable modeling assumptions across bounded and unbounded one-dimensional spatial domains.

Method: Used DINO-WM on the Wall planning task with paired-goal mixed-bit evaluation across uniform, mixed, asymmetric, and layerwise quantization variants under two planner budgets. Conducted experiments with 8-bit, 6-bit, 4-bit, and 3-bit settings to analyze performance patterns.

Result: Found a consistent three-regime pattern: 8-bit and 6-bit settings remain close to FP16 performance, 3-bit settings collapse completely, and 4-bit settings are allocation-sensitive. Preserving encoder precision improves planning relative to uniform quantization, and near-size asymmetric variants show the same encoder-side direction. In stricter replication with smaller per-cell episode count, the mixed-versus-uniform INT4 sign becomes budget-conditioned.

Conclusion: Module-aware, budget-aware quantization policies are needed for efficient spatial reasoning, as bit allocation matters significantly in the transition regime between high-precision and collapsed performance.

Abstract: Efficient spatial reasoning requires world models that remain reliable under tight precision budgets. We study whether low-bit planning behavior is determined mostly by total bitwidth or by where bits are allocated across modules. Using DINO-WM on the Wall planning task, we run a paired-goal mixed-bit evaluation across uniform, mixed, asymmetric, and layerwise variants under two planner budgets. We observe a consistent three-regime pattern: 8-bit and 6-bit settings remain close to FP16, 3-bit settings collapse, and 4-bit settings are allocation-sensitive. In that transition region, preserving encoder precision improves planning relative to uniform quantization, and near-size asymmetric variants show the same encoder-side direction. In a later strict 22-cell replication with smaller per-cell episode count, the mixed-versus-uniform INT4 sign becomes budget-conditioned, which further highlights the sensitivity of this transition regime. These findings motivate module-aware, budget-aware quantization policies as a broader research direction for efficient spatial reasoning. Code and run artifacts are available at https://github.com/suraj-ranganath/DINO-MBQuant.

Method: RAM-Net maps inputs to high-dimensional sparse vectors serving as explicit addresses, allowing selective access to a massive memory state. This enables exponential state size scaling without additional parameters, with inherent sparsity ensuring computational efficiency as state updates are confined to minimal entries.

Result: RAM-Net consistently surpasses state-of-the-art baselines in fine-grained long-range retrieval tasks and achieves competitive performance in standard language modeling and zero-shot commonsense reasoning benchmarks.

Conclusion: RAM-Net demonstrates superior capability to capture complex dependencies with significantly reduced computational overhead, validating its approach to bridging representational capacity and memory efficiency.

Abstract: While linear attention architectures offer efficient inference, compressing unbounded history into a fixed-size memory inherently limits expressivity and causes information loss. To address this limitation, we introduce Random Access Memory Network (RAM-Net), a novel architecture designed to bridge the gap between the representational capacity of full attention and the memory efficiency of linear models. The core of RAM-Net maps inputs to high-dimensional sparse vectors serving as explicit addresses, allowing the model to selectively access a massive memory state. This design enables exponential state size scaling without additional parameters, which significantly mitigates signal interference and enhances retrieval fidelity. Moreover, the inherent sparsity ensures exceptional computational efficiency, as state updates are confined to minimal entries. Extensive experiments demonstrate that RAM-Net consistently surpasses state-of-the-art baselines in fine-grained long-range retrieval tasks and achieves competitive performance in standard language modeling and zero-shot commonsense reasoning benchmarks, validating its superior capability to capture complex dependencies with significantly reduced computational overhead.

Method: Constrained Initial Representations (CIR) framework with three components: (1) Tanh activation with normalization to stabilize representations, (2) skip connections for linear pathways from shallow to deep layers, and (3) convex Q-learning for flexible value estimation.

Result: CIR exhibits strong performance on continuous control tasks, being competitive with or surpassing existing strong baseline methods.

Conclusion: Constraining initial representations through the CIR framework effectively stabilizes RL training and improves sample efficiency in continuous control tasks.

Abstract: Recently, there have been numerous attempts to enhance the sample efficiency of off-policy reinforcement learning (RL) agents when interacting with the environment, including architecture improvements and new algorithms. Despite these advances, they overlook the potential of directly constraining the initial representations of the input data, which can intuitively alleviate the distribution shift issue and stabilize training. In this paper, we introduce the Tanh function into the initial layer to fulfill such a constraint. We theoretically unpack the convergence property of the temporal difference learning with the Tanh function under linear function approximation. Motivated by theoretical insights, we present our Constrained Initial Representations framework, tagged CIR, which is made up of three components: (i) the Tanh activation along with normalization methods to stabilize representations; (ii) the skip connection module to provide a linear pathway from the shallow layer to the deep layer; (iii) the convex Q-learning that allows a more flexible value estimate and mitigates potential conservatism. Empirical results show that CIR exhibits strong performance on numerous continuous control tasks, even being competitive or surpassing existing strong baseline methods.

Method: SpaTeoGL jointly learns window-level spatial graphs capturing interactions among iEEG electrodes and a temporal graph linking time windows based on similarity of their spatial structure. Formulated within smooth graph signal processing framework and solved via alternating block coordinate descent algorithm with convergence guarantees.

Result: Experiments on multicenter iEEG dataset with successful surgical outcomes show SpaTeoGL is competitive with baseline (horizontal visibility graphs + logistic regression), while improving non-SOZ identification and providing interpretable insights into seizure onset and propagation dynamics.

Conclusion: SpaTeoGL offers an interpretable spatiotemporal graph learning approach for seizure network analysis that can aid epilepsy surgery planning by providing insights into seizure dynamics beyond just localization.

Abstract: Accurate localization of the seizure onset zone (SOZ) from intracranial EEG (iEEG) is essential for epilepsy surgery but is challenged by complex spatiotemporal seizure dynamics. We propose SpaTeoGL, a spatiotemporal graph learning framework for interpretable seizure network analysis. SpaTeoGL jointly learns window-level spatial graphs capturing interactions among iEEG electrodes and a temporal graph linking time windows based on similarity of their spatial structure. The method is formulated within a smooth graph signal processing framework and solved via an alternating block coordinate descent algorithm with convergence guarantees. Experiments on a multicenter iEEG dataset with successful surgical outcomes show that SpaTeoGL is competitive with a baseline based on horizontal visibility graphs and logistic regression, while improving non-SOZ identification and providing interpretable insights into seizure onset and propagation dynamics.

Method: Proposes Distribution Discriminant Theory (DDT) to explain and quantify alignment between data and model-induced distribution. Introduces two techniques: (1) In-Distribution Finetuning (IDFT) - a loss-level method to enhance SFT generalization, and (2) Hinted Decoding - a data-level technique to re-align training corpus to model’s distribution.

Result: Extensive experiments show the framework achieves generalization performance on par with prominent offline RL algorithms (DPO and SimPO) while maintaining the efficiency of an SFT pipeline.

Conclusion: The framework offers a practical alternative to RL in domains where RL is infeasible, bridging the generalization gap between SFT and RL through on-policy SFT techniques.

Abstract: Supervised fine-tuning (SFT) is computationally efficient but often yields inferior generalization compared to reinforcement learning (RL). This gap is primarily driven by RL’s use of on-policy data. We propose a framework to bridge this chasm by enabling On-Policy SFT. We first present \textbf{\textit{Distribution Discriminant Theory (DDT)}}, which explains and quantifies the alignment between data and the model-induced distribution. Leveraging DDT, we introduce two complementary techniques: (i) \textbf{\textit{In-Distribution Finetuning (IDFT)}}, a loss-level method to enhance generalization ability of SFT, and (ii) \textbf{\textit{Hinted Decoding}}, a data-level technique that can re-align the training corpus to the model’s distribution. Extensive experiments demonstrate that our framework achieves generalization performance on par with prominent offline RL algorithms, including DPO and SimPO, while maintaining the efficiency of an SFT pipeline. The proposed framework thus offers a practical alternative in domains where RL is infeasible. We open-source the code here: https://github.com/zhangmiaosen2000/Towards-On-Policy-SFT

Method: Constructs meta-dataset by sampling pairs from training split using class agreement as supervision, trains calibrated logistic regressor on two inexpensive meta-features: TF-IDF cosine similarity and length-compatibility ratio. At inference, performs single vectorized scoring pass over candidate pool.

Result: Meta-Sel consistently ranks among top-performing methods across benchmark of 12 methods, 4 intent datasets, and 5 open-source LLMs. Particularly effective for smaller models where selection quality compensates for limited capacity, maintains competitive selection-time overhead.

Conclusion: Meta-Sel provides lightweight, interpretable demonstration selection requiring no model fine-tuning, online exploration, or additional LLM calls, with deterministic rankings and auditable feature weights.

Abstract: Demonstration selection is a practical bottleneck in in-context learning (ICL): under a tight prompt budget, accuracy can change substantially depending on which few-shot examples are included, yet selection must remain cheap enough to run per query over large candidate pools. We propose Meta-Sel, a lightweight supervised meta-learning approach for intent classification that learns a fast, interpretable scoring function for (candidate, query) pairs from labeled training data. Meta-Sel constructs a meta-dataset by sampling pairs from the training split and using class agreement as supervision, then trains a calibrated logistic regressor on two inexpensive meta-features: TF–IDF cosine similarity and a length-compatibility ratio. At inference time, the selector performs a single vectorized scoring pass over the full candidate pool and returns the top-k demonstrations, requiring no model fine-tuning, no online exploration, and no additional LLM calls. This yields deterministic rankings and makes the selection mechanism straightforward to audit via interpretable feature weights. Beyond proposing Meta-Sel, we provide a broad empirical study of demonstration selection, benchmarking 12 methods – spanning prompt engineering baselines, heuristic selection, reinforcement learning, and influence-based approaches – across four intent datasets and five open-source LLMs. Across this benchmark, Meta-Sel consistently ranks among the top-performing methods, is particularly effective for smaller models where selection quality can partially compensate for limited model capacity, and maintains competitive selection-time overhead.

Method: 1) Formalize a taxonomy of topological bias measures relevant to fairness in graphs; 2) Introduce a flexible graph generation method that maintains real-world graph patterns while enabling controlled variation across structural biases; 3) Apply framework to evaluate classical and fairness-aware LP models across multiple use cases.

Result: Provides fine-grained empirical analysis of interactions between predictive fairness and structural biases, revealing sensitivity of fairness interventions to beyond-homophily biases and highlighting need for structurally grounded fairness evaluations.

Conclusion: Topological biases beyond homophily significantly impact fairness in graph link prediction, requiring structurally aware benchmarking frameworks for proper evaluation and development of fairness interventions that generalize across diverse network topologies.

Abstract: Graph link prediction (LP) plays a critical role in socially impactful applications, such as job recommendation and friendship formation. Ensuring fairness in this task is thus essential. While many fairness-aware methods manipulate graph structures to mitigate prediction disparities, the topological biases inherent to social graph structures remain poorly understood and are often reduced to homophily alone. This undermines the generalization potential of fairness interventions and limits their applicability across diverse network topologies. In this work, we propose a novel benchmarking framework for fair LP, centered on the structural biases of the underlying graphs. We begin by reviewing and formalizing a broad taxonomy of topological bias measures relevant to fairness in graphs. In parallel, we introduce a flexible graph generation method that simultaneously ensures fidelity to real-world graph patterns and enables controlled variation across a wide spectrum of structural biases. We apply this framework to evaluate both classical and fairness-aware LP models across multiple use cases. Our results provide a fine-grained empirical analysis of the interactions between predictive fairness and structural biases. This new perspective reveals the sensitivity of fairness interventions to beyond-homophily biases and underscores the need for structurally grounded fairness evaluations in graph learning.

Method: Designed a suite of four diverse “vulnerability games” with unique exploitable flaws related to context-conditional compliance, proxy metrics, reward tampering, and self-evaluation. Trained models with RL in these environments and tested for exploitative behavior.

Result: Models consistently learned to exploit vulnerabilities, discovering opportunistic strategies that significantly increased reward at the expense of task correctness or safety. Exploitative strategies were generalizable skills that could be transferred to new tasks and distilled from teacher to student models through data alone.

Conclusion: Capability-oriented training induced risks pose a fundamental challenge to current alignment approaches, suggesting future AI safety work must extend beyond content moderation to rigorously auditing and securing training environments and reward mechanisms.

Abstract: While most AI alignment research focuses on preventing models from generating explicitly harmful content, a more subtle risk is emerging: capability-oriented training induced exploitation. We investigate whether language models, when trained with reinforcement learning (RL) in environments with implicit loopholes, will spontaneously learn to exploit these flaws to maximize their reward, even without any malicious intent in their training. To test this, we design a suite of four diverse “vulnerability games”, each presenting a unique, exploitable flaw related to context-conditional compliance, proxy metrics, reward tampering, and self-evaluation. Our experiments show that models consistently learn to exploit these vulnerabilities, discovering opportunistic strategies that significantly increase their reward at the expense of task correctness or safety. More critically, we find that these exploitative strategies are not narrow “tricks” but generalizable skills; they can be transferred to new tasks and even “distilled” from a capable teacher model to other student models through data alone. Our findings reveal that capability-oriented training induced risks pose a fundamental challenge to current alignment approaches, suggesting that future AI safety work must extend beyond content moderation to rigorously auditing and securing the training environments and reward mechanisms themselves. Code is available at https://github.com/YujunZhou/Capability_Oriented_Alignment_Risk.

Method: Proposes Incremental Signature Contribution (ISC) method that decomposes truncated path signatures into temporally ordered sequence of tensor algebra elements, preserving algebraic structure while making temporal evolution explicit. Introduces ISC-Transformer that integrates ISC into standard Transformer architecture for offline reinforcement learning.

Result: ISC-Transformer evaluated on HalfCheetah, Walker2d, Hopper, and Maze2d environments, including settings with delayed rewards and downgraded datasets, demonstrating effectiveness for temporally sensitive control tasks.

Conclusion: ISC provides theoretically grounded and practically effective alternative to path processing for temporally sensitive control tasks, enabling signature-based representations to be processed via sequential modeling approaches.

Abstract: Path signatures embed trajectories into tensor algebra and constitute a universal, non-parametric representation of paths; however, in the standard form, they collapse temporal structure into a single global object, which limits their suitability for decision-making problems that require step-wise reactivity. We propose the Incremental Signature Contribution (ISC) method, which decomposes truncated path signatures into a temporally ordered sequence of elements in the tensor-algebra space, corresponding to incremental contributions induced by last path increments. This reconstruction preserves the algebraic structure and expressivity of signatures, while making their internal temporal evolution explicit, enabling processing signature-based representations via sequential modeling approaches. In contrast to full signatures, ISC is inherently sensitive to instantaneous trajectory updates, which is critical for sensitive and stability-requiring control dynamics. Building on this representation, we introduce ISC-Transformer (ISCT), an offline reinforcement learning model that integrates ISC into a standard Transformer architecture without further architectural modification. We evaluate ISCT on HalfCheetah, Walker2d, Hopper, and Maze2d, including settings with delayed rewards and downgraded datasets. The results demonstrate that ISC method provides a theoretically grounded and practically effective alternative to path processing for temporally sensitive control tasks.

Method: Proposes Generalized On-Policy Distillation (G-OPD) framework that extends OPD by introducing: 1) flexible reference model selection, and 2) reward scaling factor controlling reward vs KL regularization weight. Tests on math reasoning and code generation tasks.

Result: Two key findings: 1) Reward extrapolation (ExOPD) with scaling factor >1 consistently outperforms standard OPD across teacher-student size pairings, enabling students to surpass teacher performance boundaries; 2) In strong-to-weak distillation, using teacher’s pre-RL base model as reference improves performance but requires access to teacher’s pre-training variant.

Conclusion: G-OPD provides theoretical insights into OPD as a special case of KL-constrained RL and offers practical improvements through reward extrapolation and reference model selection, advancing knowledge distillation techniques.

Abstract: On-policy distillation (OPD), which aligns the student with the teacher’s logit distribution on student-generated trajectories, has demonstrated strong empirical gains in improving student performance and often outperforms off-policy distillation and reinforcement learning (RL) paradigms. In this work, we first theoretically show that OPD is a special case of dense KL-constrained RL where the reward function and the KL regularization are always weighted equally and the reference model can by any model. Then, we propose the Generalized On-Policy Distillation (G-OPD) framework, which extends the standard OPD objective by introducing a flexible reference model and a reward scaling factor that controls the relative weight of the reward term against the KL regularization. Through comprehensive experiments on math reasoning and code generation tasks, we derive two novel insights: (1) Setting the reward scaling factor to be greater than 1 (i.e., reward extrapolation), which we term ExOPD, consistently improves over standard OPD across a range of teacher-student size pairings. In particular, in the setting where we merge the knowledge from different domain experts, obtained by applying domain-specific RL to the same student model, back into the original student, ExOPD enables the student to even surpass the teacher’s performance boundary and outperform the domain teachers. (2) Building on ExOPD, we further find that in the strong-to-weak distillation setting (i.e., distilling a smaller student from a larger teacher), performing reward correction by choosing the reference model as the teacher’s base model before RL yields a more accurate reward signal and further improves distillation performance. However, this choice assumes access to the teacher’s pre-RL variant and incurs more computational overhead. We hope our work offers new insights for future research on OPD.

Method: Proposes DeepFusionKernel, a deeply fused kernel that cuts HBM traffic and boosts cache reuse, integrated with SGLang and paired with a kernel scheduler for consistent performance

Result: Delivers up to 13.2% speedup on H100 and 9.7% on A100 over SGLang, with consistent accelerations over generation lengths while remaining adaptable to diverse models and hardware

Conclusion: DeepFusionKernel effectively addresses memory bandwidth limitations in LLM inference, providing significant performance improvements across different hardware platforms while maintaining adaptability

Abstract: Agentic LLM inference with long contexts is increasingly limited by memory bandwidth rather than compute. In this setting, SwiGLU MLP blocks, whose large weights exceed cache capacity, become a major yet under-optimized bottleneck. We propose DeepFusionKernel, a deeply fused kernel that cuts HBM traffic and boosts cache reuse, delivering up to 13.2% speedup on H100 and 9.7% on A100 over SGLang. Integrated with SGLang and paired with a kernel scheduler, DeepFusionKernel ensures consistent accelerations over generation lengths, while remaining adaptable to diverse models, inference configurations, and hardware platforms.

Method: 1) Conducted comprehensive empirical study of mixing method design space to identify effective choices; 2) Introduced mixture reuse mechanism that reuses existing ratios and recomputes only for domains affected by updates, enabling efficient adaptation to evolving domain sets.

Result: Mixture reuse matches performance of fully recomputing mixtures after each update with 74% less compute, and improves over training without mixing by 11.6% on downstream tasks across five domain-set updates mirroring real-world development.

Conclusion: Olmix provides a practical solution for data mixing in real-world LM development by addressing both the configuration space challenge and the evolving domain set problem, enabling efficient and effective training as datasets change over time.

Abstract: Data mixing – determining the ratios of data from different domains – is a first-order concern for training language models (LMs). While existing mixing methods show promise, they fall short when applied during real-world LM development. We present Olmix, a framework that addresses two such challenges. First, the configuration space for developing a mixing method is not well understood – design choices across existing methods lack justification or consensus and overlook practical issues like data constraints. We conduct a comprehensive empirical study of this space, identifying which design choices lead to a strong mixing method. Second, in practice, the domain set evolves throughout LM development as datasets are added, removed, partitioned, and revised – a problem setting largely unaddressed by existing works, which assume fixed domains. We study how to efficiently recompute the mixture after the domain set is updated, leveraging information from past mixtures. We introduce mixture reuse, a mechanism that reuses existing ratios and recomputes ratios only for domains affected by the update. Over a sequence of five domain-set updates mirroring real-world LM development, mixture reuse matches the performance of fully recomputing the mix after each update with 74% less compute and improves over training without mixing by 11.6% on downstream tasks.

Method: Proposes Confidence-Aware Active Learning (CAAL) with two components: 1) decoupled uncertainty-aware training that separately optimizes predictive mean and noise level, and 2) confidence-aware acquisition function that weights epistemic uncertainty using predicted aleatoric uncertainty as reliability signal.

Result: CAAL consistently outperforms standard active learning baselines on both particle-resolved numerical simulations and real atmospheric observations, providing efficient sample selection for expanding high-cost atmospheric particle property databases.

Conclusion: CAAL offers a practical solution for heteroscedastic regression settings where labeling is expensive, enabling more efficient expansion of atmospheric particle property databases from limited measurement budgets.

Abstract: Quantifying the impacts of air pollution on health and climate relies on key atmospheric particle properties such as toxicity and hygroscopicity. However, these properties typically require complex observational techniques or expensive particle-resolved numerical simulations, limiting the availability of labeled data. We therefore estimate these hard-to-measure particle properties from routinely available observations (e.g., air pollutant concentrations and meteorological conditions). Because routine observations only indirectly reflect particle composition and structure, the mapping from routine observations to particle properties is noisy and input-dependent, yielding a heteroscedastic regression setting. With a limited and costly labeling budget, the central challenge is to select which samples to measure or simulate. While active learning is a natural approach, most acquisition strategies rely on predictive uncertainty. Under heteroscedastic noise, this signal conflates reducible epistemic uncertainty with irreducible aleatoric uncertainty, causing limited budgets to be wasted in noise-dominated regions. To address this challenge, we propose a confidence-aware active learning framework (CAAL) for efficient and robust sample selection in heteroscedastic settings. CAAL consists of two components: a decoupled uncertainty-aware training objective that separately optimises the predictive mean and noise level to stabilise uncertainty estimation, and a confidence-aware acquisition function that dynamically weights epistemic uncertainty using predicted aleatoric uncertainty as a reliability signal. Experiments on particle-resolved numerical simulations and real atmospheric observations show that CAAL consistently outperforms standard AL baselines. The proposed framework provides a practical and general solution for the efficient expansion of high-cost atmospheric particle property databases.

Method: Formal characterization of InvestESG conditions showing intertemporal social dilemmas, deriving theoretical thresholds where individual incentives diverge from collective welfare. Application of Advantage Alignment, a scalable opponent shaping algorithm for general-sum games, to influence agent learning in the simulation.

Result: Advantage Alignment systematically favors socially beneficial equilibria by biasing learning dynamics toward cooperative outcomes. Strategic shaping of economic agents’ learning processes leads to better outcomes that could inform policy mechanisms.

Conclusion: Shaping agent learning through algorithms like Advantage Alignment can help align market incentives with long-term sustainability goals, offering insights for policy design to address climate change coordination problems.

Abstract: Addressing climate change requires global coordination, yet rational economic actors often prioritize immediate gains over collective welfare, resulting in social dilemmas. InvestESG is a recently proposed multi-agent simulation that captures the dynamic interplay between investors and companies under climate risk. We provide a formal characterization of the conditions under which InvestESG exhibits an intertemporal social dilemma, deriving theoretical thresholds at which individual incentives diverge from collective welfare. Building on this, we apply Advantage Alignment, a scalable opponent shaping algorithm shown to be effective in general-sum games, to influence agent learning in InvestESG. We offer theoretical insights into why Advantage Alignment systematically favors socially beneficial equilibria by biasing learning dynamics toward cooperative outcomes. Our results demonstrate that strategically shaping the learning processes of economic agents can result in better outcomes that could inform policy mechanisms to better align market incentives with long-term sustainability goals.

Method: Robust optimization framework with budgeted uncertainty sets, using column-and-constraint generation, Benders decomposition, iterative robust optimization, and a learning-based hide-and-seek game for analysis.

Result: Proposed methods outperform classical static robust models and deterministic worst-case formulations, demonstrating effectiveness through theoretical analysis and computational results.

Conclusion: The developed robust optimization approach with dynamic budgeted uncertainty sets provides scalable solutions for resilient network design under uncertainty in both links and nodes.

Abstract: We study the design of resilient and reliable communication networks in which a signal can be transferred only up to a limited distance before its quality falls below an acceptable threshold. When excessive signal degradation occurs, regeneration is required through regenerators installed at selected network nodes. In this work, both network links and nodes are subject to uncertainty. The installation costs of regenerators are modeled using a budgeted uncertainty set. In addition, link lengths follow a dynamic budgeted uncertainty set introduced in this paper, where deviations may vary over time. Robust optimization seeks solutions whose performance is guaranteed under all scenarios represented by the underlying uncertainty set. Accordingly, the objective is to identify a minimum-cost subset of nodes for regenerator deployment that ensures full network connectivity, even under the worst possible realizations of uncertainty. To solve the problem, we first formulate it within a robust optimization framework, and then develop scalable solution methods based on column-and-constraint generation, Benders decomposition, and iterative robust optimization. In addition, we formulate a learning-based hide-and-seek game to further analyze the problem structure. The proposed approaches are evaluated against classical static budgeted robust models and deterministic worst-case formulations. Both theoretical analysis and computational results demonstrate the effectiveness and advantages of our methodology.

Method: Controlled experiments with LLMs observing sequences of multivariate scalar-valued function samples from known GP priors, comparing prediction error against GP regression (lower bound) and 1-nearest-neighbor (upper bound), plus likelihood analysis and fine-tuning experiments.

Result: LLM learning curves approach GP lower bound with more demonstrations, show inductive bias toward less smooth GP kernels, and fine-tuning (RL or supervised) can effectively shift these biases toward smoother kernels.

Conclusion: LLMs behave like GP learners for in-context function learning, and their inductive biases can be steered through post-training, providing tools for improving sample efficiency on specific function classes.

Abstract: Large language models (LLMs) can learn from a few demonstrations provided at inference time. We study this in-context learning phenomenon through the lens of Gaussian Processes (GPs). We build controlled experiments where models observe sequences of multivariate scalar-valued function samples drawn from known GP priors. We evaluate prediction error in relation to the number of demonstrations and compare against two principled references: (i) an empirical GP-regression learner that gives a lower bound on achievable error, and (ii) the expected error of a 1-nearest-neighbor (1-NN) rule, which gives a data-driven upper bound. Across model sizes, we find that LLM learning curves are strongly influenced by the function-generating kernels and approach the GP lower bound as the number of demonstrations increases. We then study the inductive biases of these models using a likelihood-based analysis. We find that LLM predictions are most likely under less smooth GP kernels. Finally, we explore whether post-training can shift these inductive biases and improve sample-efficiency on functions sampled from GPs with smoother kernels. We find that both reinforcement learning and supervised fine-tuning can effectively shift inductive biases in the direction of the training data. Together, our framework quantifies the extent to which LLMs behave like GP learners and provides tools for steering their inductive biases for continuous function learning tasks.

Method: Train a single conditional diffusion model on paired coarse-resolution inputs (~25 km) and high-resolution regional reanalysis targets (~5 km), then apply it in zero-shot manner to deterministic forecasts from various upstream weather models.

Result: Across diverse AI-based and numerical weather prediction systems, the ensemble mean of downscaled forecasts consistently improves upon each model’s raw deterministic forecast, with substantial gains in probabilistic skill (CRPS).

Conclusion: Diffusion-based downscaling provides a scalable, model-agnostic probabilistic interface for enhancing spatial resolution and uncertainty representation in operational weather forecasting pipelines.

Abstract: We introduce a universal diffusion-based downscaling framework that lifts deterministic low-resolution weather forecasts into probabilistic high-resolution predictions without any model-specific fine-tuning. A single conditional diffusion model is trained on paired coarse-resolution inputs (~25 km resolution) and high-resolution regional reanalysis targets (~5 km resolution), and is applied in a fully zero-shot manner to deterministic forecasts from heterogeneous upstream weather models. Focusing on near-surface variables, we evaluate probabilistic forecasts against independent in situ station observations over lead times up to 90 h. Across a diverse set of AI-based and numerical weather prediction (NWP) systems, the ensemble mean of the downscaled forecasts consistently improves upon each model’s own raw deterministic forecast, and substantially larger gains are observed in probabilistic skill as measured by CRPS. These results demonstrate that diffusion-based downscaling provides a scalable, model-agnostic probabilistic interface for enhancing spatial resolution and uncertainty representation in operational weather forecasting pipelines.

Method: HyPO modifies DPO by conditionally applying the reference: when reference is optimistic or neutral, it behaves exactly like DPO; when reference is pessimistic, it treats the reference as neutral by replacing Δ_θ-Δ_ref with Δ_θ-max{0,Δ_ref}. This one-line change strengthens learning signals on pessimistic pairs while preserving DPO’s objective form and computational cost.

Result: HyPO improves inference-aligned metrics and achieves higher pairwise win rates across preference alignment tasks. It mitigates premature satisfaction by conditionally debiasing the pessimistic reference signal.

Conclusion: Direct preference alignment can be enhanced by conditionally debiasing the reference signal rather than discarding it entirely. HyPO provides a simple yet effective solution to DPO’s premature satisfaction problem while maintaining computational efficiency.

Abstract: Direct Preference Optimization (DPO) has become the de facto standard for offline preference alignment of large language models, but its reliance on a reference policy introduces a critical tension. DPO weighs each update relative to a reference, which stabilizes the training by regularizing the updates within a trusted region. This reliance becomes problematic for pessimistic pairs, where the reference model prefers the rejected response. For these pairs, DPO prematurely attenuates the gradient as soon as the policy margin ($Δ_θ$) merely beats the reference margin ($Δ_{\mathrm{ref}}$) even if the policy is still wrong ($Δ_θ<0$). We name this failure premature satisfaction, which is a concrete form of the training-inference mismatch. Reference-free objectives remove this mismatch by optimizing the absolute margin, but at the cost of discarding the stabilizing signal of the reference. We mitigate this tension with Hybrid-DPO (HyPO), a drop-in modification to DPO that applies reference conditionally: HyPO behaves exactly like DPO when the reference is optimistic or neutral, and it treats the reference as neutral when it is pessimistic by replacing $Δ_θ-Δ_{\mathrm{ref}}$ with $Δ_θ-\max{0,Δ_{\mathrm{ref}}}$. This one-line change strictly strengthens per-example learning signals on pessimistic pairs while preserving DPO’s objective form and computational cost. By conditionally debiasing the pessimistic reference signal, HyPO mitigates premature satisfaction; empirically, across preference alignment, HyPO improves inference-aligned metrics and achieves higher pairwise win rates. Our results provide evidence that direct preference alignment could be enhanced by conditionally debiasing the reference signal, rather than discarding it.

Method: Defines the problem of learning conditional averages in the PAC framework, introduces novel combinatorial parameters that depend on both concept class and neighborhood system, and provides a characterization of learnability based on joint finiteness of these parameters.

Result: Complete characterization of when conditional averages are learnable, with sample complexity bounds tight up to logarithmic factors. The characterization depends on two novel combinatorial parameters related to the independence number of the neighborhood graph.

Conclusion: The paper extends PAC learning to conditional averages over neighborhoods, providing theoretical foundations for learning tasks in domains like explainability, fairness, and recommendation systems.

Abstract: We introduce the problem of learning conditional averages in the PAC framework. The learner receives a sample labeled by an unknown target concept from a known concept class, as in standard PAC learning. However, instead of learning the target concept itself, the goal is to predict, for each instance, the average label over its neighborhood – an arbitrary subset of points that contains the instance. In the degenerate case where all neighborhoods are singletons, the problem reduces exactly to classic PAC learning. More generally, it extends PAC learning to a setting that captures learning tasks arising in several domains, including explainability, fairness, and recommendation systems. Our main contribution is a complete characterization of when conditional averages are learnable, together with sample complexity bounds that are tight up to logarithmic factors. The characterization hinges on the joint finiteness of two novel combinatorial parameters, which depend on both the concept class and the neighborhood system, and are closely related to the independence number of the associated neighborhood graph.

Method: Extends Puzzle NAS framework to gpt-oss-120B, combining heterogeneous MoE expert pruning, selective window attention replacement, FP8 KV-cache quantization with calibrated scales, and post-training reinforcement learning to recover accuracy.

Result: Achieves 1.63X and 1.22X per-token throughput speedups in long/short-context settings on 8XH100, 2.82X on single H100 GPU. gpt-oss-puzzle-88B improves request-level efficiency up to 1.29X while matching or slightly exceeding parent model accuracy across benchmarks.

Conclusion: Post-training architecture search can substantially reduce inference costs without sacrificing quality, advocating for request-level efficiency metrics that normalize throughput by tokens generated.

Abstract: Reasoning-focused LLMs improve answer quality by generating longer reasoning traces, but the additional tokens dramatically increase serving cost, motivating inference optimization. We extend and apply Puzzle, a post-training neural architecture search (NAS) framework, to gpt-oss-120B to produce gpt-oss-puzzle-88B, a deployment-optimized derivative. Our approach combines heterogeneous MoE expert pruning, selective replacement of full-context attention with window attention, FP8 KV-cache quantization with calibrated scales, and post-training reinforcement learning to recover accuracy, while maintaining low generation length. In terms of per-token speeds, on an 8XH100 node we achieve 1.63X and 1.22X throughput speedups in long-context and short-context settings, respectively. gpt-oss-puzzle-88B also delivers throughput speedups of 2.82X on a single NVIDIA H100 GPU. However, because token counts can change with reasoning effort and model variants, per-token throughput (tok/s) and latency (ms/token) do not necessarily lead to end-to-end speedups: a 2X throughput gain is erased if traces grow 2X. Conversely, throughput gains can be spent on more reasoning tokens to improve accuracy; we therefore advocate request-level efficiency metrics that normalize throughput by tokens generated and trace an accuracy–speed frontier across reasoning efforts. We show that gpt-oss-puzzle-88B improves over gpt-oss-120B along the entire frontier, delivering up to 1.29X higher request-level efficiency. Across various benchmarks, gpt-oss-puzzle-88B matches or slightly exceeds the parent on suite-average accuracy across reasoning efforts, with retention ranging from 100.8% (high) to 108.2% (low), showing that post-training architecture search can substantially reduce inference costs without sacrificing quality.

Method: Proposes Temporally Unified Adversarial Perturbations (TUAPs) with temporal unification constraint and Timestamp-wise Gradient Accumulation Method (TGAM) that aggregates local gradient information from overlapping samples to generate consistent perturbations.

Result: Outperforms baselines in both white-box and black-box transfer attack scenarios under TUAP constraints on three benchmark datasets and four state-of-the-art models, and shows superior transfer attack performance even without TUAP constraints.

Conclusion: The proposed method effectively addresses temporal inconsistency in adversarial attacks for time series forecasting, providing practical and effective perturbation generation with strict temporal consistency.

Abstract: While deep learning models have achieved remarkable success in time series forecasting, their vulnerability to adversarial examples remains a critical security concern. However, existing attack methods in the forecasting field typically ignore the temporal consistency inherent in time series data, leading to divergent and contradictory perturbation values for the same timestamp across overlapping samples. This temporally inconsistent perturbations problem renders adversarial attacks impractical for real-world data manipulation. To address this, we introduce Temporally Unified Adversarial Perturbations (TUAPs), which enforce a temporal unification constraint to ensure identical perturbations for each timestamp across all overlapping samples. Moreover, we propose a novel Timestamp-wise Gradient Accumulation Method (TGAM) that provides a modular and efficient approach to effectively generate TUAPs by aggregating local gradient information from overlapping samples. By integrating TGAM with momentum-based attack algorithms, we ensure strict temporal consistency while fully utilizing series-level gradient information to explore the adversarial perturbation space. Comprehensive experiments on three benchmark datasets and four representative state-of-the-art models demonstrate that our proposed method significantly outperforms baselines in both white-box and black-box transfer attack scenarios under TUAP constraints. Moreover, our method also exhibits superior transfer attack performance even without TUAP constraints, demonstrating its effectiveness and superiority in generating adversarial perturbations for time series forecasting models.

Method: The paper combines legal analysis of EU AI Act provisions with computational methods. It proposes using individual conflict ratios and δ-ambiguity to quantify model disagreement on individual cases, and derives practical rules for evaluating predictive multiplicity in real-world systems.

Result: The analysis shows that incorporating predictive multiplicity information can help providers comply with EU AI Act accuracy requirements. The paper provides concrete tools and guidelines for detecting individuals subject to conflicting predictions and evaluating multiplicity in practice.

Conclusion: Information about predictive multiplicity should be made available to AI system deployers under the AI Act, enabling them to judge whether system outputs for specific individuals are reliable enough for their particular use cases.

Abstract: When building AI systems for decision support, one often encounters the phenomenon of predictive multiplicity: a single best model does not exist; instead, one can construct many models with similar overall accuracy that differ in their predictions for individual cases. Especially when decisions have a direct impact on humans, this can be highly unsatisfactory. For a person subject to high disagreement between models, one could as well have chosen a different model of similar overall accuracy that would have decided the person’s case differently. We argue that this arbitrariness conflicts with the EU AI Act, which requires providers of high-risk AI systems to report performance not only at the dataset level but also for specific persons. The goal of this paper is to put predictive multiplicity in context with the EU AI Act’s provisions on accuracy and to subsequently derive concrete suggestions on how to evaluate and report predictive multiplicity in practice. Specifically: (1) We argue that incorporating information about predictive multiplicity can serve compliance with the EU AI Act’s accuracy provisions for providers. (2) Based on this legal analysis, we suggest individual conflict ratios and $δ$-ambiguity as tools to quantify the disagreement between models on individual cases and to help detect individuals subject to conflicting predictions. (3) Based on computational insights, we derive easy-to-implement rules on how model providers could evaluate predictive multiplicity in practice. (4) Ultimately, we suggest that information about predictive multiplicity should be made available to deployers under the AI Act, enabling them to judge whether system outputs for specific individuals are reliable enough for their use case.

Method: Proposes PMFL with two key components: 1) On node side, adds model-contrastive term using historical local models to improve update consistency; 2) On server side, uses cumulative participation counts to adjust aggregation weights and incorporates historical global models to reduce performance fluctuations.

Result: Extensive experiments show PMFL achieves superior performance compared to existing FL methods in heterogeneous scenarios.

Conclusion: PMFL effectively addresses heterogeneity challenges in FL by leveraging historical training information through model-contrastive learning and adaptive aggregation mechanisms.

Abstract: Federated Learning (FL) enables multiple nodes to collaboratively train a model without sharing raw data. However, FL systems are usually deployed in heterogeneous scenarios, where nodes differ in both data distributions and participation frequencies, which undermines the FL performance. To tackle the above issue, this paper proposes PMFL, a performance-enhanced model-contrastive federated learning framework using historical training information. Specifically, on the node side, we design a novel model-contrastive term into the node optimization objective by incorporating historical local models to capture stable contrastive points, thereby improving the consistency of model updates in heterogeneous data distributions. On the server side, we utilize the cumulative participation count of each node to adaptively adjust its aggregation weight, thereby correcting the bias in the global objective caused by different node participation frequencies. Furthermore, the updated global model incorporates historical global models to reduce its fluctuations in performance between adjacent rounds. Extensive experiments demonstrate that PMFL achieves superior performance compared with existing FL methods in heterogeneous scenarios.

Method: LRBTC uses a modular architecture covering Language, Regulatory, Brand, Technical, and Content Structure checks. It combines a Student-Teacher dual model architecture with human-in-the-loop workflow and waterfall rule filtering for scalable, verifiable content validation.

Result: Achieves 83.0% F1 and 97.5% recall on AIReg-Bench, reducing missed violations by 5x compared to Gemini 2.5 Pro. Improves mean accuracy by 26.7% on CSpelling. Models are strong at detecting misspellings (92.5 recall) but weak at complex medical grammatical (25.0 recall) and punctuation (41.7 recall) errors.

Conclusion: Provides a practical plug-and-play solution for reliable, transparent quality control in high-stakes, compliance-critical industries. Identifies complex grammatical and punctuation errors as key areas for future improvement.

Abstract: Large language models (LLMs) are increasingly used to create content in regulated domains such as pharmaceuticals, where outputs must be scientifically accurate and legally compliant. Manual quality control (QC) is slow, error prone, and can become a publication bottleneck. We introduce LRBTC, a modular LLM and vision language model (VLM) driven QC architecture covering Language, Regulatory, Brand, Technical, and Content Structure checks. LRBTC combines a Student-Teacher dual model architecture, human in the loop (HITL) workflow with waterfall rule filtering to enable scalable, verifiable content validation and optimization. On AIReg-Bench, our approach achieves 83.0% F1 and 97.5% recall, reducing missed violations by 5x compared with Gemini 2.5 Pro. On CSpelling, it improves mean accuracy by 26.7%. Error analysis further reveals that while current models are strong at detecting misspellings (92.5 recall), they fail to identify complex medical grammatical (25.0 recall) and punctuation (41.7 recall) errors, highlighting a key area for future work. This work provides a practical, plug and play solution for reliable, transparent quality control of content in high stakes, compliance critical industries. We also provide access to our Demo under MIT Licenses.

Method: Proposes Manifold-aware Temporal LoRA (MaT-LoRA) that constrains temporal updates to a shared low-dimensional manifold within low-rank adaptation subspaces and models evolution through a structured temporal core.

Result: Extensive experiments on synthetic and real-world datasets (scientific documents, news publishers, review ratings) show superior temporal generalization performance with practical scalability for LLMs.

Conclusion: Low-dimensional temporal structure can be preserved under parameter-efficient reparameterization, enabling effective temporal modeling without operating in the full parameter space of LLMs.

Abstract: Temporal distribution shifts are pervasive in real-world deployments of Large Language Models (LLMs), where data evolves continuously over time. While Temporal Domain Generalization (TDG) seeks to model such structured evolution, existing approaches characterize model adaptation in the full parameter space. This formulation becomes computationally infeasible for modern LLMs. This paper introduces a geometric reformulation of TDG under parameter-efficient fine-tuning. We establish that the low-dimensional temporal structure underlying model evolution can be preserved under parameter-efficient reparameterization, enabling temporal modeling without operating in the ambient parameter space. Building on this principle, we propose Manifold-aware Temporal LoRA (MaT-LoRA), which constrains temporal updates to a shared low-dimensional manifold within a low-rank adaptation subspace, and models its evolution through a structured temporal core. This reparameterization dramatically reduces temporal modeling complexity while retaining expressive power. Extensive experiments on synthetic and real-world datasets, including scientific documents, news publishers, and review ratings, demonstrate that MaT-LoRA achieves superior temporal generalization performance with practical scalability for LLMs.

Method: Investigates Momentum Least Mean Squares (MLMS) algorithm as adaptive identification tool. Theoretically analyzes MLMS in time-varying stochastic linear systems, deriving tracking performance and regret bounds. Unlike classical LMS (first-order dynamics), MLMS introduces momentum leading to second-order time-varying random vector difference equations requiring analysis of products of random matrices.

Result: Experiments on synthetic and real-world data streams show MLMS achieves rapid adaptation and robust tracking, especially in nonstationary settings. Theoretical results align with empirical performance, demonstrating algorithm’s effectiveness for streaming and online learning applications.

Conclusion: MLMS offers promising approach for modern streaming and online learning applications with nonstationary data, combining computational simplicity with strong theoretical guarantees and practical performance in time-varying environments.

Abstract: In large-scale data processing scenarios, data often arrive in sequential streams generated by complex systems that exhibit drifting distributions and time-varying system parameters. This nonstationarity challenges theoretical analysis, as it violates classical assumptions of i.i.d. (independent and identically distributed) samples, necessitating algorithms capable of real-time updates without expensive retraining. An effective approach should process each sample in a single pass, while maintaining computational and memory complexities independent of the data stream length. Motivated by these challenges, this paper investigates the Momentum Least Mean Squares (MLMS) algorithm as an adaptive identification tool, leveraging its computational simplicity and online processing capabilities. Theoretically, we derive tracking performance and regret bounds for the MLMS in time-varying stochastic linear systems under various practical conditions. Unlike classical LMS, whose stability can be characterized by first-order random vector difference equations, MLMS introduces an additional dynamical state due to momentum, leading to second-order time-varying random vector difference equations whose stability analysis hinges on more complicated products of random matrices, which poses a substantially challenging problem to resolve. Experiments on synthetic and real-world data streams demonstrate that MLMS achieves rapid adaptation and robust tracking, in agreement with our theoretical results especially in nonstationary settings, highlighting its promise for modern streaming and online learning applications.

Method: Analyzes how DP mechanisms (gradient clipping and noise injection) perturb firing-rate statistics in SNNs and propagate these perturbations to rate-based FNL coordination. Conducts ablations across privacy budgets and clipping bounds on a speech recognition task under non-IID settings.

Result: Reveals systematic rate shifts, attenuated aggregation, and ranking instability during client selection. Relates these shifts to sparsity and memory indicators. Shows how privacy mechanisms affect the balance between privacy strength and rate-dependent coordination.

Conclusion: Provides actionable guidance for privacy-preserving FNL regarding the balance between privacy strength and rate-dependent coordination. Highlights the trade-offs between privacy mechanisms and neuromorphic learning performance in federated settings.

Abstract: Federated Neuromorphic Learning (FNL) enables energy-efficient and privacy-preserving learning on devices without centralizing data. However, real-world deployments require additional privacy mechanisms that can significantly alter training signals. This paper analyzes how Differential Privacy (DP) mechanisms, specifically gradient clipping and noise injection, perturb firing-rate statistics in Spiking Neural Networks (SNNs) and how these perturbations are propagated to rate-based FNL coordination. On a speech recognition task under non-IID settings, ablations across privacy budgets and clipping bounds reveal systematic rate shifts, attenuated aggregation, and ranking instability during client selection. Moreover, we relate these shifts to sparsity and memory indicators. Our findings provide actionable guidance for privacy-preserving FNL, specifically regarding the balance between privacy strength and rate-dependent coordination.

Method: Two-module framework: 1) Competence-based expert selection using lightweight confidence graphs from auxiliary data to identify suitable clients for each question. 2) Group-relative policy optimization that packages questions with solution rationales into candidate policies, dispatches to expert clients, and aggregates only scalar reward signals via federated group-relative loss function.

Result: FedGRPO achieves superior downstream accuracy and communication efficiency compared to conventional federated foundation model baselines across diverse domain tasks, while reducing privacy risk and communication overhead.

Conclusion: FedGRPO provides an effective privacy-preserving framework for federated foundation models by exchanging only reward values instead of data or model updates, enabling parallel evaluation across heterogeneous devices with improved performance.

Abstract: One important direction of Federated Foundation Models (FedFMs) is leveraging data from small client models to enhance the performance of a large server-side foundation model. Existing methods based on model level or representation level knowledge transfer either require expensive local training or incur high communication costs and introduce unavoidable privacy risks. We reformulate this problem as a reinforcement learning style evaluation process and propose FedGRPO, a privacy preserving framework comprising two modules. The first module performs competence-based expert selection by building a lightweight confidence graph from auxiliary data to identify the most suitable clients for each question. The second module leverages the “Group Relative” concept from the Group Relative Policy Optimization (GRPO) framework by packaging each question together with its solution rationale into candidate policies, dispatching these policies to a selected subset of expert clients, and aggregating solely the resulting scalar reward signals via a federated group-relative loss function. By exchanging reward values instead of data or model updates, FedGRPO reduces privacy risk and communication overhead while enabling parallel evaluation across heterogeneous devices. Empirical results on diverse domain tasks demonstrate that FedGRPO achieves superior downstream accuracy and communication efficiency compared to conventional FedFMs baselines.

Method: Two structured LRNN architectures: 1) Higher-order Linear Recurrent Units (H-LRU) generalize first-order recurrence to higher order, mixing multiple past states; 2) Block-Diagonal LRUs (BD-LRU) enable dense intra-block channel mixing. Both use L1-normalization of selective gates for training stability and parallel-scan implementation for competitive throughput.

Result: BD-LRU matches or exceeds performance of linear SSMs (Mamba), low-rank LRNNs (DeltaNet), and LSTMs in synthetic sequence modeling tasks. H-LRU is most parameter-efficient in compression tasks. Both architectures show that state mixing structure, not just width, shapes LRNN expressivity.

Conclusion: The proposed structured LRNN architectures offer a practical approach to closing the efficiency-expressivity gap in linear sequence models through richer state mixing while maintaining competitive computational efficiency.

Abstract: Linear recurrent networks (LRNNs) and linear state space models (SSMs) promise computational and memory efficiency on long-sequence modeling tasks, yet their diagonal state transitions limit expressivity. Dense and nonlinear architectures (e.g., LSTMs) on the other hand are provably more expressive, but computationally costly. Here, we explore how expressivity in LRNNs can be increased via richer state mixing across time and channels while maintaining competitive efficiency. Specifically, we introduce two structured LRNN architectures: (i) Higher-order Linear Recurrent Units (H-LRU), which generalize first-order recurrence to higher order, mixing multiple past states, and (ii) Block-Diagonal LRUs (BD-LRU), which enable dense intra-block channel mixing. Per-channel (H-LRU) or per-row (BD-LRU) L1-normalization of selective gates stabilizes training and allows for scaling window/block sizes. A parallel-scan implementation of the proposed architectures keeps the throughput competitive with diagonal LRNNs for moderate orders (H-LRU) and block sizes (BD-LRU). In synthetic sequence modeling tasks, the performance of BD-LRU matches or exceeds those of linear SSMs (Mamba), low-rank LRNNs (DeltaNet) and LSTM baselines, while H-LRU is found to be the most parameter-efficient in compression task. In both synthetic sequence modeling and language modeling, our results indicate that the structure of state mixing rather than width alone shapes expressivity of LRNNs, offering a practical route to closing the efficiency-expressivity gap in linear sequence models.

Method: ProtoMech uses cross-layer transcoders that learn sparse latent representations jointly across layers to capture the model’s full computational circuitry. Applied to pLM ESM2, it identifies compressed circuits using <1% of the latent space.

Result: ProtoMech recovers 82-89% of original performance on protein family classification and function prediction. It identifies compressed circuits (<1% of latent space) retaining up to 79% accuracy, revealing correspondence with structural/functional motifs. Steering along circuits enables high-fitness protein design surpassing baselines in >70% of cases.

Conclusion: ProtoMech establishes a principled framework for protein circuit tracing, enabling better understanding of pLM computational mechanisms and facilitating protein design through circuit steering.

Abstract: Protein language models (pLMs) have emerged as powerful predictors of protein structure and function. However, the computational circuits underlying their predictions remain poorly understood. Recent mechanistic interpretability methods decompose pLM representations into interpretable features, but they treat each layer independently and thus fail to capture cross-layer computation, limiting their ability to approximate the full model. We introduce ProtoMech, a framework for discovering computational circuits in pLMs using cross-layer transcoders that learn sparse latent representations jointly across layers to capture the model’s full computational circuitry. Applied to the pLM ESM2, ProtoMech recovers 82-89% of the original performance on protein family classification and function prediction tasks. ProtoMech then identifies compressed circuits that use <1% of the latent space while retaining up to 79% of model accuracy, revealing correspondence with structural and functional motifs, including binding, signaling, and stability. Steering along these circuits enables high-fitness protein design, surpassing baseline methods in more than 70% of cases. These results establish ProtoMech as a principled framework for protein circuit tracing.

Method: PrefillShare factorizes models into prefill and decode modules, freezes the prefill module, and fine-tunes only the decode module. This allows multiple models to share a prefill module and KV cache. A routing mechanism enables sharing across heterogeneous models in a vLLM-based disaggregated system.

Result: PrefillShare matches full fine-tuning accuracy on a broad range of tasks and models, while delivering 4.5x lower p95 latency and 3.9x higher throughput in multi-model agent workloads.

Conclusion: PrefillShare effectively eliminates inter-model redundancy in computation and KV storage for shared prompts, significantly improving performance in multi-agent LLM systems without sacrificing accuracy.

Abstract: Multi-agent systems increasingly orchestrate multiple specialized language models to solve complex real-world problems, often invoking them over a shared context. This execution pattern repeatedly processes the same prompt prefix across models. Consequently, each model redundantly executes the prefill stage and maintains its own key-value (KV) cache, increasing aggregate prefill load and worsening tail latency by intensifying prefill-decode interference in existing LLM serving stacks. Disaggregated serving reduces such interference by placing prefill and decode on separate GPUs, but disaggregation does not fundamentally eliminate inter-model redundancy in computation and KV storage for the same prompt. To address this issue, we propose PrefillShare, a novel algorithm that enables sharing the prefill stage across multiple models in a disaggregated setting. PrefillShare factorizes the model into prefill and decode modules, freezes the prefill module, and fine-tunes only the decode module. This design allows multiple task-specific models to share a prefill module and the KV cache generated for the same prompt. We further introduce a routing mechanism that enables effective prefill sharing across heterogeneous models in a vLLM-based disaggregated system. PrefillShare not only matches full fine-tuning accuracy on a broad range of tasks and models, but also delivers 4.5x lower p95 latency and 3.9x higher throughput in multi-model agent workloads.

Method: Represent crystals through truncated Fourier transform of species-resolved unit-cell density instead of atomic coordinates. Use transformer variational autoencoder over complex-valued Fourier coefficients and latent diffusion model in compressed latent space. Only 9 Fourier basis functions per spatial dimension.

Result: Approach reconstructs unit cells with up to 108 atoms per chemical species. Evaluated on LeMaterial benchmark, showing reconstruction and generation capabilities. Compared against coordinate-based baselines in small-cell regime (≤16 atoms per unit cell).

Conclusion: Reciprocal-space representation provides periodicity-native approach that handles crystallographic symmetries algebraically and supports variable atomic multiplicities, addressing limitations of particle-based methods for crystalline material generation.

Abstract: The discovery of new crystalline materials calls for generative models that handle periodic boundary conditions, crystallographic symmetries, and physical constraints, while scaling to large and structurally diverse unit cells. We propose a reciprocal-space generative pipeline that represents crystals through a truncated Fourier transform of the species-resolved unit-cell density, rather than modeling atomic coordinates directly. This representation is periodicity-native, admits simple algebraic actions of space-group symmetries, and naturally supports variable atomic multiplicities during generation, addressing a common limitation of particle-based approaches. Using only nine Fourier basis functions per spatial dimension, our approach reconstructs unit cells containing up to 108 atoms per chemical species. We instantiate this pipeline with a transformer variational autoencoder over complex-valued Fourier coefficients, and a latent diffusion model that generates in the compressed latent space. We evaluate reconstruction and latent diffusion on the LeMaterial benchmark and compare unconditional generation against coordinate-based baselines in the small-cell regime ($\leq 16$ atoms per unit cell).

Method: Proposes online reinforcement learning that executes LLM-generated code on a supercomputer and uses measured runtime performance (GFLOPS) as reward. Introduces Staged Quality-Diversity (SQD) algorithm that progressively varies permitted optimization techniques per problem. Built distributed system connecting GPU training cluster with CPU benchmarking cluster, training Qwen2.5 Coder 14B on double-precision matrix multiplication using Group Relative Policy Optimization (GRPO).

Result: Through two experiments, demonstrates that reinforcement learning combining runtime performance feedback with staged optimization can improve HPC code generation capability of LLMs.

Conclusion: The approach successfully integrates actual runtime performance feedback into LLM training for HPC code optimization, showing promise for improving performance-aware code generation in high-performance computing domains.

Abstract: Large language models (LLMs) have demonstrated strong code generation capabilities, yet the runtime performance of generated code is not guaranteed, and there have been few attempts to train LLMs using runtime performance as a reward in the HPC domain. We propose an online reinforcement learning approach that executes LLM-generated code on a supercomputer and directly feeds back the measured runtime performance (GFLOPS) as a reward. We further introduce a Staged Quality-Diversity (SQD) algorithm that progressively varies the permitted optimization techniques on a per-problem basis, enabling the model to learn code optimization from diverse perspectives. We build a distributed system connecting a GPU training cluster with a CPU benchmarking cluster, and train Qwen2.5 Coder 14B on a double-precision matrix multiplication task using Group Relative Policy Optimization (GRPO). Through two experiments, we show that reinforcement learning combining runtime performance feedback with staged optimization can improve the HPC code generation capability of LLMs.

Method: Models player trajectories as fully connected dynamic graphs, formulates event detection as selecting exactly one edge (possession state) per time step, uses Conditional Random Field (CRF) to ensure logical consistency, with emission/transition scores from Set Attention-based backbone, and Viterbi decoding for inference.

Result: PathCRF produces accurate, logically consistent possession paths, enabling reliable downstream analyses while substantially reducing the need for manual event annotation.

Conclusion: The framework successfully detects soccer events using only player tracking data, making comprehensive soccer data collection more scalable and accessible beyond top-tier competitions.

Abstract: Despite recent advances in AI, event data collection in soccer still relies heavily on labor-intensive manual annotation. Although prior work has explored automatic event detection using player and ball trajectories, ball tracking also remains difficult to scale due to high infrastructural and operational costs. As a result, comprehensive data collection in soccer is largely confined to top-tier competitions, limiting the broader adoption of data-driven analysis in this domain. To address this challenge, this paper proposes PathCRF, a framework for detecting on-ball soccer events using only player tracking data. We model player trajectories as a fully connected dynamic graph and formulate event detection as the problem of selecting exactly one edge corresponding to the current possession state at each time step. To ensure logical consistency of the resulting edge sequence, we employ a Conditional Random Field (CRF) that forbids impossible transitions between consecutive edges. Both emission and transition scores dynamically computed from edge embeddings produced by a Set Attention-based backbone architecture. During inference, the most probable edge sequence is obtained via Viterbi decoding, and events such as ball controls or passes are detected whenever the selected edge changes between adjacent time steps. Experiments show that PathCRF produces accurate, logically consistent possession paths, enabling reliable downstream analyses while substantially reducing the need for manual event annotation. The source code is available at https://github.com/hyunsungkim-ds/pathcrf.git.

Method: Empirical GPs estimate mean and covariance functions empirically from a corpus of historical observations rather than using standard parametric kernels. The framework formulates learning the GP prior from independent datasets as likelihood estimation and derives an Expectation-Maximization algorithm with closed-form updates to handle heterogeneous observation locations across datasets.

Result: Theoretical analysis shows the model converges to the GP closest (in KL-divergence sense) to the real data generating process. Empirically, Empirical GPs achieve competitive performance on learning curve extrapolation and time series forecasting benchmarks.

Conclusion: Empirical GPs provide a principled framework for constructing flexible, data-driven GP priors that overcome limitations of handcrafted kernels, enabling the prior to reflect rich, non-trivial covariance structures present in the data.

Abstract: Gaussian processes (GPs) are powerful and widely used probabilistic regression models, but their effectiveness in practice is often limited by the choice of kernel function. This kernel function is typically handcrafted from a small set of standard functions, a process that requires expert knowledge, results in limited adaptivity to data, and imposes strong assumptions on the hypothesis space. We study Empirical GPs, a principled framework for constructing flexible, data-driven GP priors that overcome these limitations. Rather than relying on standard parametric kernels, we estimate the mean and covariance functions empirically from a corpus of historical observations, enabling the prior to reflect rich, non-trivial covariance structures present in the data. Theoretically, we show that the resulting model converges to the GP that is closest (in KL-divergence sense) to the real data generating process. Practically, we formulate the problem of learning the GP prior from independent datasets as likelihood estimation and derive an Expectation-Maximization algorithm with closed-form updates, allowing the model handle heterogeneous observation locations across datasets. We demonstrate that Empirical GPs achieve competitive performance on learning curve extrapolation and time series forecasting benchmarks.

Method: Proposes learning a structured latent representation where distances between states directly correlate with the minimum number of actions required to transition between them. Uses a multimodal latent transition model and sensor fusion mechanism based on inverse distance weighting to adaptively integrate multiple sensor modalities without prior noise knowledge.

Result: Empirical validation on multimodal RL tasks shows improved robustness to sensor noise and superior state estimation compared to baseline methods. Enhanced RL agent performance through learned representation, eliminating need for explicit noise augmentation.

Conclusion: Transition-aware metric spaces provide a principled and scalable solution for robust state estimation in sequential decision-making, offering geometric interpretation of uncertainty without explicit probabilistic modeling.

Abstract: Estimating the state of an environment from high-dimensional, multimodal, and noisy observations is a fundamental challenge in reinforcement learning (RL). Traditional approaches rely on probabilistic models to account for the uncertainty, but often require explicit noise assumptions, in turn limiting generalization. In this work, we contribute a novel method to learn a structured latent representation, in which distances between states directly correlate with the minimum number of actions required to transition between them. The proposed metric space formulation provides a geometric interpretation of uncertainty without the need for explicit probabilistic modeling. To achieve this, we introduce a multimodal latent transition model and a sensor fusion mechanism based on inverse distance weighting, allowing for the adaptive integration of multiple sensor modalities without prior knowledge of noise distributions. We empirically validate the approach on a range of multimodal RL tasks, demonstrating improved robustness to sensor noise and superior state estimation compared to baseline methods. Our experiments show enhanced performance of an RL agent via the learned representation, eliminating the need of explicit noise augmentation. The presented results suggest that leveraging transition-aware metric spaces provides a principled and scalable solution for robust state estimation in sequential decision-making.

Method: Introduces a general framework inspired by model-free decision-estimation coefficients for online RL, develops novel theoretical tools including a second-order performance difference lemma, and analyzes various offline RL settings including soft Q-learning and low-Bellman-rank MDPs.

Result: Establishes a negative answer to the open question via information-theoretic lower bound, improves sample complexity from ε^{-4} to ε^{-2} for soft Q-learning under partial coverage, removes need for additional online interaction, and provides first characterizations for several unexplored settings.

Conclusion: The paper provides a comprehensive theoretical framework for offline RL, answering fundamental questions about learnability, improving sample complexity bounds, and extending analysis to previously unexplored settings including low-Bellman-rank MDPs without Bellman completeness.

Abstract: We study offline reinforcement learning under $Q^\star$-approximation and partial coverage, a setting that motivates practical algorithms such as Conservative $Q$-Learning (CQL; Kumar et al., 2020) but has received limited theoretical attention. Our work is inspired by the following open question: “Are $Q^\star$-realizability and Bellman completeness sufficient for sample-efficient offline RL under partial coverage?” We answer in the negative by establishing an information-theoretic lower bound. Going substantially beyond this, we introduce a general framework that characterizes the intrinsic complexity of a given $Q^\star$ function class, inspired by model-free decision-estimation coefficients (DEC) for online RL (Foster et al., 2023b; Liu et al., 2025b). This complexity recovers and improves the quantities underlying the guarantees of Chen and Jiang (2022) and Uehara et al. (2023), and extends to broader settings. Our decision-estimation decomposition can be combined with a wide range of $Q^\star$ estimation procedures, modularizing and generalizing existing approaches. Beyond the general framework, we make further contributions: By developing a novel second-order performance difference lemma, we obtain the first $ε^{-2}$ sample complexity under partial coverage for soft $Q$-learning, improving the $ε^{-4}$ bound of Uehara et al. (2023). We remove Chen and Jiang’s (2022) need for additional online interaction when the value gap of $Q^\star$ is unknown. We also give the first characterization of offline learnability for general low-Bellman-rank MDPs without Bellman completeness (Jiang et al., 2017; Du et al., 2021; Jin et al., 2021), a canonical setting in online RL that remains unexplored in offline RL except for special cases. Finally, we provide the first analysis for CQL under $Q^\star$-realizability and Bellman completeness beyond the tabular case.

Method: Develops a neural model that predicts performance f(x) for unseen designs using few-shot context containing observations of auxiliary information h(x). Leverages multi-task learning from previously solved tasks in the same family to accelerate optimization of new tasks.

Result: Method effectively utilizes auxiliary feedback to achieve more accurate few-shot prediction and faster optimization on robotic hardware design and neural network hyperparameter tuning domains, significantly outperforming existing multi-task optimization methods.

Conclusion: The approach successfully leverages auxiliary experimental information and task history to accelerate expensive black-box optimization, demonstrating practical value for real-world design problems where experiments generate rich data beyond just performance metrics.

Abstract: Many real-world design problems involve optimizing an expensive black-box function $f(x)$, such as hardware design or drug discovery. Bayesian Optimization has emerged as a sample-efficient framework for this problem. However, the basic setting considered by these methods is simplified compared to real-world experimental setups, where experiments often generate a wealth of useful information. We introduce a new setting where an experiment generates high-dimensional auxiliary information $h(x)$ along with the performance measure $f(x)$; moreover, a history of previously solved tasks from the same task family is available for accelerating optimization. A key challenge of our setting is learning how to represent and utilize $h(x)$ for efficiently solving new optimization tasks beyond the task history. We develop a novel approach for this setting based on a neural model which predicts $f(x)$ for unseen designs given a few-shot context containing observations of $h(x)$. We evaluate our method on two challenging domains, robotic hardware design and neural network hyperparameter tuning, and introduce a novel design problem and large-scale benchmark for the former. On both domains, our method utilizes auxiliary feedback effectively to achieve more accurate few-shot prediction and faster optimization of design tasks, significantly outperforming several methods for multi-task optimization.

Method: Introduces KAN-FIF (Kolmogorov-Arnold Network-based Feature Interaction Framework), a lightweight multimodal architecture integrating MLP and CNN layers with spline-parameterized KAN layers for capturing high-order polynomial relationships between TC attributes.

Result: Achieves 94.8% parameter reduction (0.99MB vs 19MB), 68.7% faster inference per sample (2.3ms vs 7.35ms), and 32.5% lower MAE compared to baseline Phy-CoCo model. Offline deployment on meteorological satellite processor achieved 14.41ms per-sample inference latency.

Conclusion: KAN-FIF demonstrates promising feasibility for operational TC monitoring and extends deployability to edge-device AI applications, offering efficient multimodal processing for resource-constrained environments.

Abstract: Tropical cyclones (TC) are among the most destructive natural disasters, causing catastrophic damage to coastal regions through extreme winds, heavy rainfall, and storm surges. Timely monitoring of tropical cyclones is crucial for reducing loss of life and property, yet it is hindered by the computational inefficiency and high parameter counts of existing methods on resource-constrained edge devices. Current physics-guided models suffer from linear feature interactions that fail to capture high-order polynomial relationships between TC attributes, leading to inflated model sizes and hardware incompatibility. To overcome these challenges, this study introduces the Kolmogorov-Arnold Network-based Feature Interaction Framework (KAN-FIF), a lightweight multimodal architecture that integrates MLP and CNN layers with spline-parameterized KAN layers. For Maximum Sustained Wind (MSW) prediction, experiments demonstrate that the KAN-FIF framework achieves a $94.8%$ reduction in parameters (0.99MB vs 19MB) and $68.7%$ faster inference per sample (2.3ms vs 7.35ms) compared to baseline model Phy-CoCo, while maintaining superior accuracy with $32.5%$ lower MAE. The offline deployment experiment of the FY-4 series meteorological satellite processor on the Qingyun-1000 development board achieved a 14.41ms per-sample inference latency with the KAN-FIF framework, demonstrating promising feasibility for operational TC monitoring and extending deployability to edge-device AI applications. The code is released at https://github.com/Jinglin-Zhang/KAN-FIF.

Method: The paper replaces Neural ODEs with a linear damped harmonic oscillator analogy that has known closed-form solutions. Keys and values are modeled as damped, driven oscillators, while queries are expanded in a sinusoidal basis up to suitable modes. Attention is modeled as a resonance phenomenon, capturing the fundamental query-key coupling of transformer architectures.

Result: The approach achieves state-of-the-art performance on irregular time series benchmarks while being orders of magnitude faster than previous methods. The closed-form solution eliminates computational overhead while preserving expressivity, and the authors prove that the oscillator-based parameterization maintains the universal approximation property of continuous-time attention.

Conclusion: The proposed oscillator-based transformer architecture provides both theoretical guarantees and practical scalability for irregular time series modeling, offering an efficient alternative to Neural ODE-based approaches while maintaining strong performance.

Abstract: Transformers excel at time series modelling through attention mechanisms that capture long-term temporal patterns. However, they assume uniform time intervals and therefore struggle with irregular time series. Neural Ordinary Differential Equations (NODEs) effectively handle irregular time series by modelling hidden states as continuously evolving trajectories. ContiFormers arxiv:2402.10635 combine NODEs with Transformers, but inherit the computational bottleneck of the former by using heavy numerical solvers. This bottleneck can be removed by using a closed-form solution for the given dynamical system - but this is known to be intractable in general! We obviate this by replacing NODEs with a novel linear damped harmonic oscillator analogy - which has a known closed-form solution. We model keys and values as damped, driven oscillators and expand the query in a sinusoidal basis up to a suitable number of modes. This analogy naturally captures the query-key coupling that is fundamental to any transformer architecture by modelling attention as a resonance phenomenon. Our closed-form solution eliminates the computational overhead of numerical ODE solvers while preserving expressivity. We prove that this oscillator-based parameterisation maintains the universal approximation property of continuous-time attention; specifically, any discrete attention matrix realisable by ContiFormer’s continuous keys can be approximated arbitrarily well by our fixed oscillator modes. Our approach delivers both theoretical guarantees and scalability, achieving state-of-the-art performance on irregular time series benchmarks while being orders of magnitude faster.

Method: Introduces TIME benchmark with 50 fresh datasets and 98 forecasting tasks; uses human-in-the-loop construction pipeline with LLM and human expertise; aligns forecasting configurations with real-world operational requirements; proposes pattern-level evaluation using structural time series features instead of traditional dataset-level evaluations.

Result: Evaluated 12 representative TSFMs and established a multi-granular leaderboard with visualized inspection; benchmark is available on Hugging Face; provides generalizable insights into model capabilities across diverse temporal patterns.

Conclusion: TIME addresses critical gaps in existing time series forecasting benchmarks by providing fresh, high-quality data, real-world task alignment, and pattern-level evaluation for better assessment of foundation model capabilities in zero-shot settings.

Abstract: Time series foundation models (TSFMs) are revolutionizing the forecasting landscape from specific dataset modeling to generalizable task evaluation. However, we contend that existing benchmarks exhibit common limitations in four dimensions: constrained data composition dominated by reused legacy sources, compromised data integrity lacking rigorous quality assurance, misaligned task formulations detached from real-world contexts, and rigid analysis perspectives that obscure generalizable insights. To bridge these gaps, we introduce TIME, a next-generation task-centric benchmark comprising 50 fresh datasets and 98 forecasting tasks, tailored for strict zero-shot TSFM evaluation free from data leakage. Integrating large language models and human expertise, we establish a rigorous human-in-the-loop benchmark construction pipeline to ensure high data integrity and redefine task formulation by aligning forecasting configurations with real-world operational requirements and variate predictability. Furthermore, we propose a novel pattern-level evaluation perspective that moves beyond traditional dataset-level evaluations based on static meta labels. By leveraging structural time series features to characterize intrinsic temporal properties, this approach offers generalizable insights into model capabilities across diverse patterns. We evaluate 12 representative TSFMs and establish a multi-granular leaderboard to facilitate in-depth analysis and visualized inspection. The leaderboard is available at https://huggingface.co/spaces/Real-TSF/TIME-leaderboard.

Method: SafeNeuron identifies safety-related neurons in the model, then freezes these neurons during preference optimization to prevent reliance on sparse safety pathways and force the model to construct redundant safety representations across the network.

Result: SafeNeuron significantly improves robustness against neuron pruning attacks, reduces the risk of open-source models being repurposed as red-team generators, and preserves general capabilities across models and modalities.

Conclusion: SafeNeuron provides an interpretable and robust perspective for model alignment by revealing that safety behaviors are governed by stable and shared internal representations, offering better control over safety mechanisms.

Abstract: Large language models (LLMs) and multimodal LLMs are typically safety-aligned before release to prevent harmful content generation. However, recent studies show that safety behaviors are concentrated in a small subset of parameters, making alignment brittle and easily bypassed through neuron-level attacks. Moreover, most existing alignment methods operate at the behavioral level, offering limited control over the model’s internal safety mechanisms. In this work, we propose SafeNeuron, a neuron-level safety alignment framework that improves robustness by redistributing safety representations across the network. SafeNeuron first identifies safety-related neurons, then freezes these neurons during preference optimization to prevent reliance on sparse safety pathways and force the model to construct redundant safety representations. Extensive experiments across models and modalities demonstrate that SafeNeuron significantly improves robustness against neuron pruning attacks, reduces the risk of open-source models being repurposed as red-team generators, and preserves general capabilities. Furthermore, our layer-wise analysis reveals that safety behaviors are governed by stable and shared internal representations. Overall, SafeNeuron provides an interpretable and robust perspective for model alignment.

Method: Adapts a pre-trained Graph Transformer model for sequential atom-and-bond additions. Uses Group Relative Policy Optimization (GRPO) for goal-directed fine-tuning, normalizing rewards relative to starting structures to mitigate variance.

Result: GRXForm generalizes to out-of-distribution molecular scaffolds without inference-time oracle calls or refinement, achieving scores in multi-objective optimization competitive with leading instance optimizers.

Conclusion: The approach successfully addresses generalization challenges in molecular optimization by mitigating variance through relative reward normalization, enabling efficient transferable policies for structural alteration tasks.

Abstract: Molecular design encompasses tasks ranging from de-novo design to structural alteration of given molecules or fragments. For the latter, state-of-the-art methods predominantly function as “Instance Optimizers’’, expending significant compute restarting the search for every input structure. While model-based approaches theoretically offer amortized efficiency by learning a policy transferable to unseen structures, existing methods struggle to generalize. We identify a key failure mode: the high variance arising from the heterogeneous difficulty of distinct starting structures. To address this, we introduce GRXForm, adapting a pre-trained Graph Transformer model that optimizes molecules via sequential atom-and-bond additions. We employ Group Relative Policy Optimization (GRPO) for goal-directed fine-tuning to mitigate variance by normalizing rewards relative to the starting structure. Empirically, GRXForm generalizes to out-of-distribution molecular scaffolds without inference-time oracle calls or refinement, achieving scores in multi-objective optimization competitive with leading instance optimizers.

Method: The authors analyze Identity Preference Optimization (IPO), a widely used preference alignment framework, to investigate sampling effects and reference policies. They examine instance-dependent sampling strategies and analyze iterative alignment dynamics where learned policies feed back into future sampling and reference policies. The analysis extends to Direct Preference Optimization (DPO) to show broader applicability.

Result: Theoretical findings show that proper instance-dependent sampling yields stronger ranking guarantees, while skewed on-policy sampling can cause excessive concentration under structured preferences. Iterative alignment dynamics can exhibit persistent oscillations or entropy collapse for certain parameter choices, with identified regimes that guarantee stability. Experiments on real-world preference data validate these theoretical insights.

Conclusion: Sampling strategies and reference policies are critical design choices in preference alignment methods that significantly impact theoretical guarantees and practical stability. The analysis provides guidance for stable iterative alignment and shows these phenomena apply broadly to preference-alignment methods like DPO.

Abstract: Standard methods for aligning large language models with human preferences learn from pairwise comparisons among sampled candidate responses and regularize toward a reference policy. Despite their effectiveness, the effects of sampling and reference choices are poorly understood theoretically. We investigate these effects through Identity Preference Optimization, a widely used preference alignment framework, and show that proper instance-dependent sampling can yield stronger ranking guarantees, while skewed on-policy sampling can induce excessive concentration under structured preferences. We then analyze iterative alignment dynamics in which the learned policy feeds back into future sampling and reference policies, reflecting a common practice of model-generated preference data. We prove that these dynamics can exhibit persistent oscillations or entropy collapse for certain parameter choices, and characterize regimes that guarantee stability. Our theoretical insights extend to Direct Preference Optimization, indicating the phenomena we captured are common to a broader class of preference-alignment methods. Experiments on real-world preference data validate our findings.

Method: WaveFormer integrates wavelet decomposition at two stages: 1) embedding construction using multi-channel Discrete Wavelet Transform (DWT) to extract frequency features and create tokens with time-frequency information, and 2) positional encoding using Dynamic Wavelet Positional Encoding (DyWPE) that adapts position embeddings to signal-specific temporal structure through mono-channel DWT analysis.

Result: WaveFormer achieves competitive performance on eight diverse biomedical datasets spanning human activity recognition and brain signal analysis, handling sequence lengths from 50 to 3000 timesteps and channel counts from 1 to 144.

Conclusion: The approach provides a principled framework for incorporating frequency-domain knowledge into transformer-based time series classification, particularly valuable for biomedical signal analysis where frequency patterns are crucial.

Abstract: Biomedical signal classification presents unique challenges due to long sequences, complex temporal dynamics, and multi-scale frequency patterns that are poorly captured by standard transformer architectures. We propose WaveFormer, a transformer architecture that integrates wavelet decomposition at two critical stages: embedding construction, where multi-channel Discrete Wavelet Transform (DWT) extracts frequency features to create tokens containing both time-domain and frequency-domain information, and positional encoding, where Dynamic Wavelet Positional Encoding (DyWPE) adapts position embeddings to signal-specific temporal structure through mono-channel DWT analysis. We evaluate WaveFormer on eight diverse datasets spanning human activity recognition and brain signal analysis, with sequence lengths ranging from 50 to 3000 timesteps and channel counts from 1 to 144. Experimental results demonstrate that WaveFormer achieves competitive performance through comprehensive frequency-aware processing. Our approach provides a principled framework for incorporating frequency-domain knowledge into transformer-based time series classification.

Method: CRAM (Consolidation-based Routing for Adaptive Memory) implements a biologically inspired memory consolidation mechanism that gradually distills episodic retrievals into parametric semantic memory, enabling decreasing attention utilization over training through a consolidation process.

Result: CRAM achieves 37.8× attention reduction through a sharp phase transition at ~3K steps, 100% retrieval accuracy at 1.6% attention compute (vs. 68% for baselines), and consolidated patterns transfer to unseen tasks with 48-52% attention reduction without retraining. The consolidation dynamics match human memory transition curves.

Conclusion: CRAM demonstrates that adaptive memory consolidation enables efficient attention utilization by reducing redundant retrievals over time, with learned dynamics that quantitatively match human cognitive psychology patterns, offering a biologically plausible approach to improving transformer efficiency.

Abstract: Hybrid architectures combining state-space models with attention have achieved strong efficiency-quality tradeoffs, yet existing approaches either apply attention uniformly or learn static sparse patterns. This misses a key opportunity: \emph{attention demand should decrease over time as recurring patterns become familiar}. We present a surprising finding from analyzing GPT-2 models: \textbf{88%} of attention operations retrieve information already predictable from the model’s hidden state, and this redundancy does \emph{not} decrease during training. Motivated by this observation, we introduce \textbf{\ours{}} (\textbf{C}onsolidation-based \textbf{R}outing for \textbf{A}daptive \textbf{M}emory), a biologically inspired memory consolidation mechanism that gradually distills episodic retrievals into parametric semantic memory. Unlike prior sparse attention methods, \ours{} exhibits \emph{decreasing attention utilization} over training, achieving a \textbf{37.8$\times$} reduction through a sharp phase transition at approximately 3K steps. We prove that this capability is \emph{impossible} without consolidation: any static routing scheme requires $Ω(f \cdot n)$ attention for tasks with recurring patterns of frequency $f$. On our proposed SRCD benchmark, \ours{} achieves \textbf{100% retrieval accuracy} at 1.6% attention compute (vs.\ 68% for baselines), and consolidated patterns transfer to unseen tasks with \textbf{48–52%} attention reduction without retraining. Remarkably, the learned consolidation dynamics quantitatively match human episodic-to-semantic memory transition curves from cognitive psychology ($γ= 0.43$ vs.\ $γ_{\text{human}} \approx 0.4$–$0.5$). Code and benchmarks are available at [anonymized].

Method: Proposed PhyIP protocol tests whether physical quantities are linearly decodable from frozen self-supervised learning representations, motivated by the linear representation hypothesis. Uses low-capacity linear probes across fluid dynamics and orbital mechanics domains.

Result: When SSL achieves low error, latent physical structure becomes linearly accessible - PhyIP recovers internal energy and Newtonian inverse-square scaling on OOD tests (ρ > 0.90). Adaptation-based evaluations can collapse this structure (ρ ≈ 0.05).

Conclusion: Adaptation-based evaluation can obscure latent structures in neural models, while low-capacity linear probes offer more accurate evaluation of physical world models. PhyIP provides a non-invasive way to assess whether models genuinely internalize physical laws.

Abstract: Determining whether neural models internalize physical laws as world models, rather than exploiting statistical shortcuts, remains challenging, especially under out-of-distribution (OOD) shifts. Standard evaluations often test latent capability via downstream adaptation (e.g., fine-tuning or high-capacity probes), but such interventions can change the representations being measured and thus confound what was learned during self-supervised learning (SSL). We propose a non-invasive evaluation protocol, PhyIP. We test whether physical quantities are linearly decodable from frozen representations, motivated by the linear representation hypothesis. Across fluid dynamics and orbital mechanics, we find that when SSL achieves low error, latent structure becomes linearly accessible. PhyIP recovers internal energy and Newtonian inverse-square scaling on OOD tests (e.g., $ρ> 0.90$). In contrast, adaptation-based evaluations can collapse this structure ($ρ\approx 0.05$). These findings suggest that adaptation-based evaluation can obscure latent structures and that low-capacity probes offer a more accurate evaluation of physical world models.

Method: Introduces Variance Minimisation Policy Optimisation (VMPO), which formulates diffusion alignment as minimizing the variance of log importance weights rather than directly optimizing KL divergence. The method leverages SMC interpretation where denoising models act as proposals.

Result: Proves that the variance objective is minimized by the reward-tilted target distribution and that its gradient coincides with KL-based alignment under on-policy sampling. Shows VMPO recovers various existing methods under different choices of potential functions.

Conclusion: VMPO provides a unified theoretical framework for understanding diffusion alignment, connecting existing methods through variance minimization perspective and suggesting new design directions beyond KL-based approaches.

Abstract: Diffusion alignment adapts pretrained diffusion models to sample from reward-tilted distributions along the denoising trajectory. This process naturally admits a Sequential Monte Carlo (SMC) interpretation, where the denoising model acts as a proposal and reward guidance induces importance weights. Motivated by this view, we introduce Variance Minimisation Policy Optimisation (VMPO), which formulates diffusion alignment as minimising the variance of log importance weights rather than directly optimising a Kullback-Leibler (KL) based objective. We prove that the variance objective is minimised by the reward-tilted target distribution and that, under on-policy sampling, its gradient coincides with that of standard KL-based alignment. This perspective offers a common lens for understanding diffusion alignment. Under different choices of potential functions and variance minimisation strategies, VMPO recovers various existing methods, while also suggesting new design directions beyond KL.

Method: Defines flow maps toward the simplex that transport probability mass toward predicted endpoints, using continuous trajectories rather than discrete ones. Can be trained with existing distillation techniques and a new endpoint consistency objective. Enables reuse of guidance and reweighting techniques for sampling control.

Result: Achieves state-of-the-art few-step results on images, molecular graphs, and text, with strong performance even in single-step generation.

Conclusion: Categorical Flow Maps provide an effective approach for accelerated generation of categorical data with flexible training and inference capabilities.

Abstract: We introduce Categorical Flow Maps, a flow-matching method for accelerated few-step generation of categorical data via self-distillation. Building on recent variational formulations of flow matching and the broader trend towards accelerated inference in diffusion and flow-based models, we define a flow map towards the simplex that transports probability mass toward a predicted endpoint, yielding a parametrisation that naturally constrains model predictions. Since our trajectories are continuous rather than discrete, Categorical Flow Maps can be trained with existing distillation techniques, as well as a new objective based on endpoint consistency. This continuous formulation also automatically unlocks test-time inference: we can directly reuse existing guidance and reweighting techniques in the categorical setting to steer sampling toward downstream objectives. Empirically, we achieve state-of-the-art few-step results on images, molecular graphs, and text, with strong performance even in single-step generation.

Method: Theoretical analysis showing that under mild closure and additivity assumptions, intrinsic energies (defined as infima of accumulated local effort over trajectories) are quasimetrics. Demonstrates that optimal cost-to-go functions in goal-reaching control have this intrinsic form, and JEPAs trained to model intrinsic energies produce quasimetric values targeted by QRL.

Result: Establishes a principled connection between JEPAs and QRL through intrinsic energies that are quasimetrics. Shows why symmetric energies are mismatched with one-way reachability problems, motivating asymmetric (quasimetric) energies when directionality matters.

Conclusion: JEPAs and QRL are fundamentally connected through intrinsic energies that are quasimetrics. This provides a unified framework for representation learning and goal-conditioned control, with implications for designing better energy functions in multimodal learning systems.

Abstract: Joint-Embedding Predictive Architectures (JEPAs) aim to learn representations by predicting target embeddings from context embeddings, inducing a scalar compatibility energy in a latent space. In contrast, Quasimetric Reinforcement Learning (QRL) studies goal-conditioned control through directed distance values (cost-to-go) that support reaching goals under asymmetric dynamics. In this short article, we connect these viewpoints by restricting attention to a principled class of JEPA energy functions : intrinsic (least-action) energies, defined as infima of accumulated local effort over admissible trajectories between two states. Under mild closure and additivity assumptions, any intrinsic energy is a quasimetric. In goal-reaching control, optimal cost-to-go functions admit exactly this intrinsic form ; inversely, JEPAs trained to model intrinsic energies lie in the quasimetric value class targeted by QRL. Moreover, we observe why symmetric finite energies are structurally mismatched with one-way reachability, motivating asymmetric (quasimetric) energies when directionality matters.

Method: Created ExtractBench with 35 PDF documents paired with JSON Schemas and human-annotated gold labels across 12,867 fields, treating schema as executable specification with field-specific scoring metrics.

Result: Frontier models (GPT-5/5.2, Gemini-3, Claude 4.5) remain unreliable on realistic schemas, with performance degrading sharply with schema breadth, reaching 0% valid output on 369-field financial schemas.

Conclusion: ExtractBench addresses critical gaps in PDF extraction evaluation and reveals significant limitations of current LLMs for enterprise-scale structured data extraction.

Abstract: Unstructured documents like PDFs contain valuable structured information, but downstream systems require this data in reliable, standardized formats. LLMs are increasingly deployed to automate this extraction, making accuracy and reliability paramount. However, progress is bottlenecked by two gaps. First, no end-to-end benchmark evaluates PDF-to-JSON extraction under enterprise-scale schema breadth. Second, no principled methodology captures the semantics of nested extraction, where fields demand different notions of correctness (exact match for identifiers, tolerance for quantities, semantic equivalence for names), arrays require alignment, and omission must be distinguished from hallucination. We address both gaps with ExtractBench, an open-source benchmark and evaluation framework for PDF-to-JSON structured extraction. The benchmark pairs 35 PDF documents with JSON Schemas and human-annotated gold labels across economically valuable domains, yielding 12,867 evaluatable fields spanning schema complexities from tens to hundreds of fields. The evaluation framework treats the schema as an executable specification: each field declares its scoring metric. Baseline evaluations reveal that frontier models (GPT-5/5.2, Gemini-3 Flash/Pro, Claude 4.5 Opus/Sonnet) remain unreliable on realistic schemas. Performance degrades sharply with schema breadth, culminating in 0% valid output on a 369-field financial reporting schema across all tested models. We release ExtractBench at https://github.com/ContextualAI/extract-bench.

Method: Shallow Diffuse decouples watermarking from the diffusion sampling process by exploiting the existence of a low-dimensional subspace in image generation. The method ensures that a significant portion of the watermark lies in the null space of this subspace, effectively separating it from the image generation process.

Result: Theoretical and empirical analyses demonstrate that the decoupling strategy greatly enhances both the consistency of data generation and the detectability of watermarks. Extensive experiments show that Shallow Diffuse outperforms existing watermarking methods in terms of robustness and consistency.

Conclusion: Shallow Diffuse provides an effective watermarking solution for diffusion models that balances watermark robustness with generation consistency, addressing critical concerns about AI-generated content identification and misuse prevention.

Abstract: The widespread use of AI-generated content from diffusion models has raised significant concerns regarding misinformation and copyright infringement. Watermarking is a crucial technique for identifying these AI-generated images and preventing their misuse. In this paper, we introduce Shallow Diffuse, a new watermarking technique that embeds robust and invisible watermarks into diffusion model outputs. Unlike existing approaches that integrate watermarking throughout the entire diffusion sampling process, Shallow Diffuse decouples these steps by leveraging the presence of a low-dimensional subspace in the image generation process. This method ensures that a substantial portion of the watermark lies in the null space of this subspace, effectively separating it from the image generation process. Our theoretical and empirical analyses show that this decoupling strategy greatly enhances the consistency of data generation and the detectability of the watermark. Extensive experiments further validate that our Shallow Diffuse outperforms existing watermarking methods in terms of robustness and consistency. The codes are released at https://github.com/liwd190019/Shallow-Diffuse.

Method: The method analyzes two key factors influencing community concealment: connectivity at community boundaries and feature similarity between protected and adjacent communities. Based on this analysis, it strategically perturbs the graph by rewiring selected edges and modifying node features to reduce distinctiveness leveraged by GNN message passing.

Result: The proposed method outperforms DICE (a baseline method) under identical perturbation budgets, achieving median relative concealment improvements of approximately 20-45% across synthetic benchmarks and real network graphs.

Conclusion: The work demonstrates an effective mitigation strategy against GNN-based community learning and highlights group-level privacy risks inherent in graph learning, providing a defensive framework for data publishers to protect sensitive community structures.

Abstract: Graph neural networks (GNNs) are designed to use attributed graphs to learn representations. Such representations are beneficial in the unsupervised learning of clusters and community detection. Nonetheless, such inference may reveal sensitive groups, clustered systems, or collective behaviors, raising concerns regarding group-level privacy. Community attribution in social and critical infrastructure networks, for example, can expose coordinated asset groups, operational hierarchies, and system dependencies that could be used for profiling or intelligence gathering. We study a defensive setting in which a data publisher (defender) seeks to conceal a community of interest while making limited, utility-aware changes in the network. Our analysis indicates that community concealment is strongly influenced by two quantifiable factors: connectivity at the community boundary and feature similarity between the protected community and adjacent communities. Informed by these findings, we present a perturbation strategy that rewires a set of selected edges and modifies node features to reduce the distinctiveness leveraged by GNN message passing. The proposed method outperforms DICE in our experiments on synthetic benchmarks and real network graphs under identical perturbation budgets. Overall, it achieves median relative concealment improvements of approximately 20-45% across the evaluated settings. These findings demonstrate a mitigation strategy against GNN-based community learning and highlight group-level privacy risks intrinsic to graph learning.

Method: Proposes Quantized Zeroth-order Optimization (QZO) that eliminates gradients and optimizer states using zeroth-order optimization, which approximates gradients by perturbing weights during forward passes. Employs model quantization (e.g., bfloat16 to int4) for weight memory reduction. QZO perturbs continuous quantization scales for gradient estimation and uses directional derivative clipping to stabilize training, working with both scalar-based and codebook-based post-training quantization methods.

Result: QZO reduces total memory cost by more than 18× for 4-bit LLMs compared to full-parameter fine-tuning in 16 bits. Enables fine-tuning Llama-2-13B within a single 24GB GPU.

Conclusion: QZO provides an effective framework for memory-efficient fine-tuning of large language models by combining zeroth-order optimization with quantization, making large model adaptation more accessible with limited GPU resources.

Abstract: As the size of large language models grows exponentially, GPU memory has become a bottleneck for adapting these models to downstream tasks. In this paper, we aim to push the limits of memory-efficient training by minimizing memory usage on model weights, gradients, and optimizer states, within a unified framework. Our idea is to eliminate both gradients and optimizer states using zeroth-order optimization, which approximates gradients by perturbing weights during forward passes to identify gradient directions. To minimize memory usage on weights, we employ model quantization, e.g., converting from bfloat16 to int4. However, directly applying zeroth-order optimization to quantized weights is infeasible due to the precision gap between discrete weights and continuous gradients, which would otherwise require de-quantization and re-quantization. To overcome this challenge, we propose Quantized Zeroth-order Optimization (QZO), a simple yet effective approach that perturbs the continuous quantization scale for gradient estimation and uses a directional derivative clipping method to stabilize training. QZO is orthogonal to both scalar-based and codebook-based post-training quantization methods. Compared to full-parameter fine-tuning in 16 bits, QZO can reduce the total memory cost by more than 18$\times$ for 4-bit LLMs, and enables fine-tuning Llama-2-13B within a single 24GB GPU.

Method: Proposes Flow-Guided Neural Operator (FGNO) combining operator learning with flow matching. Uses Short-Time Fourier Transform to unify different time resolutions, extracts hierarchical features from different network layers and flow times with varying noise strengths, and uses clean inputs during inference while learning with noise.

Result: FGNO consistently outperforms established baselines across three biomedical domains: 35% AUROC gains in neural signal decoding (BrainTreeBank), 16% RMSE reductions in skin temperature prediction (DREAMT), and over 20% improvement in accuracy and macro-F1 on SleepEDF under low-data regimes.

Conclusion: FGNO demonstrates robustness to data scarcity and superior capacity to learn expressive representations for diverse time series, offering a flexible alternative to fixed-masking SSL methods with significant performance improvements in biomedical applications.

Abstract: Self-supervised learning (SSL) is a powerful paradigm for learning from unlabeled time-series data. However, popular methods such as masked autoencoders (MAEs) rely on reconstructing inputs from a fixed, predetermined masking ratio. Instead of this static design, we propose treating the corruption level as a new degree of freedom for representation learning, enhancing flexibility and performance. To achieve this, we introduce the Flow-Guided Neural Operator (FGNO), a novel framework combining operator learning with flow matching for SSL training. FGNO learns mappings in functional spaces by using Short-Time Fourier Transform to unify different time resolutions. We extract a rich hierarchy of features by tapping into different network layers and flow times that apply varying strengths of noise to the input data. This enables the extraction of versatile representations, from low-level patterns to high-level global features, using a single model adaptable to specific tasks. Unlike prior generative SSL methods that use noisy inputs during inference, we propose using clean inputs for representation extraction while learning representations with noise; this eliminates randomness and boosts accuracy. We evaluate FGNO across three biomedical domains, where it consistently outperforms established baselines. Our method yields up to 35% AUROC gains in neural signal decoding (BrainTreeBank), 16% RMSE reductions in skin temperature prediction (DREAMT), and over 20% improvement in accuracy and macro-F1 on SleepEDF under low-data regimes. These results highlight FGNO’s robustness to data scarcity and its superior capacity to learn expressive representations for diverse time series.

Method: Introduces Probability Flow Distance (PFD) that quantifies distribution distance by comparing noise-to-data mappings induced by probability flow ODE. Applied to diffusion distillation to study generalization behaviors.

Result: Empirically uncovers key generalization behaviors: (1) quantitative scaling from memorization to generalization, (2) epoch-wise double descent training dynamics, and (3) bias-variance decomposition.

Conclusion: PFD provides foundation for generalization studies in diffusion distillation and bridges them with diffusion training, offering both theoretical grounding and practical efficiency.

Abstract: Diffusion distillation provides an effective approach for learning lightweight and few-steps diffusion models with efficient generation. However, evaluating their generalization remains challenging: theoretical metrics are often impractical for high-dimensional data, while no practical metrics rigorously measure generalization. In this work, we bridge this gap by introducing probability flow distance (\texttt{PFD}), a theoretically grounded and computationally efficient metric to measure generalization. Specifically, \texttt{PFD} quantifies the distance between distributions by comparing their noise-to-data mappings induced by the probability flow ODE. Using \texttt{PFD} under the diffusion distillation setting, we empirically uncover several key generalization behaviors, including: (1) quantitative scaling behavior from memorization to generalization, (2) epoch-wise double descent training dynamics, and (3) bias-variance decomposition. Beyond these insights, our work lays a foundation for generalization studies in diffusion distillation and bridges them with diffusion training.

Method: The study uses a combination of theoretical analysis (information-theoretic generalization bounds decomposed into ID and OOD components) and structural analysis (examining internal reasoning circuits). Controlled experiments validate the theoretical findings with real-world data, examining how CoT-trained models process intermediate results across network layers.

Result: CoT training enables compositional generalization: models learn to systematically combine simpler skills to solve novel complex problems. Theoretically, CoT-trained models show better OOD generalization by handling unseen compositional patterns. Structurally, they develop two-stage compositional circuits where intermediate results are resolved at shallower layers, freeing deeper layers for subsequent reasoning steps.

Conclusion: CoT training teaches models “how to think” through compositional reasoning rather than just “what to think” by providing correct answers. This fosters robust generalization from ID to OOD scenarios and accelerates convergence. The insights provide guidance for designing better CoT strategies to enhance LLM reasoning robustness.

Abstract: Chain-of-Thought (CoT) training has markedly advanced the reasoning capabilities of large language models (LLMs), yet the mechanisms by which CoT training enhances generalization remain inadequately understood. In this work, we demonstrate that compositional generalization is fundamental: models systematically combine simpler learned skills during CoT training to address novel and more complex problems. Through a theoretical and structural analysis, we formalize this process: 1) Theoretically, the information-theoretic generalization bounds through distributional divergence can be decomposed into in-distribution (ID) and out-of-distribution (OOD) components. Specifically, the non-CoT models fail on OOD tasks due to unseen compositional patterns, whereas CoT-trained models achieve strong generalization by composing previously learned skills. In addition, controlled experiments and real-world validation confirm that CoT training accelerates convergence and enhances generalization from ID to both ID and OOD scenarios while maintaining robust performance even with tolerable noise. 2) Structurally, CoT training internalizes reasoning into a two-stage compositional circuit, where the number of stages corresponds to the explicit reasoning steps during training. Notably, CoT-trained models resolve intermediate results at shallower layers compared to non-CoT counterparts, freeing up deeper layers to specialize in subsequent reasoning steps. A key insight is that CoT training teaches models how to think-by fostering compositional reasoning-rather than merely what to think, through the provision of correct answers alone. This paper offers valuable insights for designing CoT strategies to enhance LLMs’ reasoning robustness.

Method: Combines function-space diffusion models with differentiable neural operator surrogates: learns prior distribution over geological parameters using single-channel diffusion model, then uses Local Neural Operator surrogate for physics-consistent guidance and cross-field conditioning on dynamics field.

Result: For forward modeling with only 25% observations: 7.7% relative error vs 86.9% for standard surrogates (11x improvement). For inverse modeling: achieves Jensen-Shannon divergence <0.06 against ground truth, produces physically consistent realizations without high-frequency artifacts, with 4x improved sample efficiency compared to rejection sampling.

Conclusion: Fun-DDPS effectively handles extreme data sparsity in subsurface flow characterization, providing both accurate forward predictions and rigorous inverse solutions with improved efficiency and physical consistency compared to existing methods.

Abstract: Accurate characterization of subsurface flow is critical for Carbon Capture and Storage (CCS) but remains challenged by the ill-posed nature of inverse problems with sparse observations. We present Fun-DDPS, a generative framework that combines function-space diffusion models with differentiable neural operator surrogates for both forward and inverse modeling. Our approach learns a prior distribution over geological parameters (geomodel) using a single-channel diffusion model, then leverages a Local Neural Operator (LNO) surrogate to provide physics-consistent guidance for cross-field conditioning on the dynamics field. This decoupling allows the diffusion prior to robustly recover missing information in parameter space, while the surrogate provides efficient gradient-based guidance for data assimilation. We demonstrate Fun-DDPS on synthetic CCS modeling datasets, achieving two key results: (1) For forward modeling with only 25% observations, Fun-DDPS achieves 7.7% relative error compared to 86.9% for standard surrogates (an 11x improvement), proving its capability to handle extreme data sparsity where deterministic methods fail. (2) We provide the first rigorous validation of diffusion-based inverse solvers against asymptotically exact Rejection Sampling (RS) posteriors. Both Fun-DDPS and the joint-state baseline (Fun-DPS) achieve Jensen-Shannon divergence less than 0.06 against the ground truth. Crucially, Fun-DDPS produces physically consistent realizations free from the high-frequency artifacts observed in joint-state baselines, achieving this with 4x improved sample efficiency compared to rejection sampling.

Method: Uses damped wave equation and its stochastic Kac representation to move probability mass at finite speed. Introduces classifier-free guidance in velocity space that preserves square integrability. Proposes endpoint-only distillation that trains a student to match a frozen teacher over long intervals with stability guarantees.

Result: Experiments demonstrate DistillKac delivers high quality samples with very few function evaluations while retaining the numerical stability benefits of finite speed probability flows.

Conclusion: DistillKac provides an efficient and stable alternative to diffusion models by enforcing finite speed transport through Kac dynamics, enabling high-quality image generation with fewer function evaluations.

Abstract: We present DistillKac, a fast image generator that uses the damped wave equation and its stochastic Kac representation to move probability mass at finite speed. In contrast to diffusion models whose reverse time velocities can become stiff and implicitly allow unbounded propagation speed, Kac dynamics enforce finite speed transport and yield globally bounded kinetic energy. Building on this structure, we introduce classifier-free guidance in velocity space that preserves square integrability under mild conditions. We then propose endpoint only distillation that trains a student to match a frozen teacher over long intervals. We prove a stability result that promotes supervision at the endpoints to closeness along the entire path. Experiments demonstrate DistillKac delivers high quality samples with very few function evaluations while retaining the numerical stability benefits of finite speed probability flows.

Method: AutoDiscovery uses Bayesian surprise (epistemic shift from prior to posterior beliefs) as the exploration criterion. It employs Monte Carlo tree search (MCTS) with progressive widening, using surprisal as the reward function to efficiently explore nested hypothesis spaces. The system leverages LLMs for hypothesis generation and belief updating.

Result: Evaluated on 21 real-world datasets across biology, economics, finance, and behavioral science. Under fixed budget, AutoDiscovery outperforms competitors by producing 5-29% more discoveries deemed surprising by the LLM. Human evaluation shows two-thirds of discoveries are surprising to domain experts.

Conclusion: AutoDiscovery represents an important step toward open-ended autonomous scientific discovery systems by using Bayesian surprise as an objective exploration criterion, enabling more effective hypothesis discovery than diversity-based approaches.

Abstract: The promise of autonomous scientific discovery (ASD) hinges not only on answering questions, but also on knowing which questions to ask. Most recent works in ASD explore the use of large language models (LLMs) in goal-driven settings, relying on human-specified research questions to guide hypothesis generation. However, scientific discovery may be accelerated further by allowing the AI system to drive exploration by its own criteria. The few existing approaches in open-ended ASD select hypotheses based on diversity heuristics or subjective proxies for human interestingness, but the former struggles to meaningfully navigate the typically vast hypothesis space, and the latter suffers from imprecise definitions. This paper presents AutoDiscovery – a method for open-ended ASD that instead drives scientific exploration using Bayesian surprise. Here, we quantify the epistemic shift from the LLM’s prior beliefs about a hypothesis to its posterior beliefs after gathering experimental results. To efficiently explore the space of nested hypotheses, our method employs a Monte Carlo tree search (MCTS) strategy with progressive widening using surprisal as the reward function. We evaluate AutoDiscovery in the setting of data-driven discovery across 21 real-world datasets spanning domains such as biology, economics, finance, and behavioral science. Our results demonstrate that under a fixed budget, AutoDiscovery substantially outperforms competitors by producing 5-29% more discoveries deemed surprising by the LLM. Our human evaluation further reveals that two-thirds of discoveries made by our system are surprising to domain experts as well, suggesting this is an important step towards building open-ended ASD systems.

Method: Introduces a fully decoupled training strategy that learns prototypes and encoders under separate objectives. Prototypes are modeled as a Gaussian mixture updated with an online EM-style procedure independent of the encoder’s loss, breaking the joint optimization that causes collapse.

Result: The method eliminates prototype collapse without explicit regularization, yields consistently diverse prototypes, and achieves stronger downstream performance compared to approaches that suffer from collapse.

Conclusion: Decoupling prototype learning from encoder optimization through principled statistical modeling addresses the root cause of prototype collapse, providing a more effective foundation for self-supervised learning without needing ad-hoc solutions.

Abstract: Prototypical self-supervised learning methods consistently suffer from partial prototype collapse, where multiple prototypes converge to nearly identical representations. This undermines their central purpose – providing diverse and informative targets to guide encoders toward rich representations – and has led practitioners to over-parameterize prototype sets or add ad-hoc regularizers, which mitigate symptoms rather than address the root cause. We empirically trace the collapse to the joint optimization of encoders and prototypes, which encourages a type of shortcut learning: early in training prototypes drift toward redundant representations that minimize loss without necessarily enhancing representation diversity. To break the joint optimization, we introduce a fully decoupled training strategy that learns prototypes and encoders under separate objectives. Concretely, we model prototypes as a Gaussian mixture updated with an online EM-style procedure, independent of the encoder’s loss. This simple yet principled decoupling eliminates prototype collapse without explicit regularization and yields consistently diverse prototypes and stronger downstream performance.

Method: Proposes Binary Autoencoder (BAE) that enforces minimal entropy on minibatches of hidden activations. Uses 1-bit discretization via step function with gradient estimation for backpropagation to enable efficient entropy calculation.

Result: BAE avoids dense features and produces the largest number of interpretable features among baselines. Enables reliable entropy estimation on binary hidden activations for characterizing LLM inference dynamics.

Conclusion: BAE improves feature untangling in LLMs by promoting global sparsity through entropy minimization, enabling better feature analysis and interpretability.

Abstract: Existing works are dedicated to untangling atomized numerical components (features) from the hidden states of Large Language Models (LLMs). However, they typically rely on autoencoders constrained by some training-time regularization on single training instances, without an explicit guarantee of global sparsity among instances, causing a large amount of dense (simultaneously inactive) features, harming the feature sparsity and atomization. In this paper, we propose a novel autoencoder variant that enforces minimal entropy on minibatches of hidden activations, thereby promoting feature independence and sparsity across instances. For efficient entropy calculation, we discretize the hidden activations to 1-bit via a step function and apply gradient estimation to enable backpropagation, so that we term it as Binary Autoencoder (BAE) and empirically demonstrate two major applications: (1) Feature set entropy calculation. Entropy can be reliably estimated on binary hidden activations, which can be leveraged to characterize the inference dynamics of LLMs. (2) Feature untangling. Compared to typical methods, due to improved training strategy, BAE avoids dense features while producing the largest number of interpretable ones among baselines.

Method: Uses SISFall dataset and GAN-augmented variants as proxy task for fall detection, demonstrates feasibility through federated learning with non-IID conditions, and embedded deployment on Jetson Orin Nano devices. Proposes privacy-preserving, edge-deployed federated AI systems.

Result: Initial results show feasibility of federated learning with non-IID conditions and embedded deployment. The work provides early evidence and outlines open challenges including domain shift, data scarcity, and privacy risks.

Conclusion: The paper presents a roadmap for transitioning from fall detection to comprehensive ADL monitoring in smart-room environments, highlighting the need for sustainable and human-centered elderly care AI systems.

Abstract: This position paper envisions a next-generation elderly monitoring system that moves beyond fall detection toward the broader goal of Activities of Daily Living (ADL) recognition. Our ultimate aim is to design privacy-preserving, edge-deployed, and federated AI systems that can robustly detect and understand daily routines, supporting independence and dignity in aging societies. At present, ADL-specific datasets are still under collection. As a preliminary step, we demonstrate feasibility through experiments using the SISFall dataset and its GAN-augmented variants, treating fall detection as a proxy task. We report initial results on federated learning with non-IID conditions, and embedded deployment on Jetson Orin Nano devices. We then outline open challenges such as domain shift, data scarcity, and privacy risks, and propose directions toward full ADL monitoring in smart-room environments. This work highlights the transition from single-task detection to comprehensive daily activity recognition, providing both early evidence and a roadmap for sustainable and human-centered elderly care AI.

Method: Proposes Direct Steering Optimization (DSO), which uses reinforcement learning to find optimal linear transformations for steering model activations. DSO is directly optimized to mitigate bias while maintaining control over model performance, offering inference-time controllability.

Result: DSO achieves state-of-the-art trade-off between fairness and capabilities on both VLMs and LLMs. It provides practitioners with inference-time control over the bias-performance trade-off, outperforming methods that rely on pre-defined heuristics.

Conclusion: Designing steering strategies directly optimized for behavior control provides more effective bias intervention than heuristic-based methods. DSO enables controllable bias reduction during inference, addressing the need for adaptable fairness-performance balancing in real-world deployments.

Abstract: Generative models are often deployed to make decisions on behalf of users, such as vision-language models (VLMs) identifying which person in a room is a doctor to help visually impaired individuals. Yet, VLM decisions are influenced by the perceived demographic attributes of people in the input, which can lead to biased outcomes like failing to identify women as doctors. Moreover, when reducing bias leads to performance loss, users may have varying needs for balancing bias mitigation with overall model capabilities, highlighting the demand for methods that enable controllable bias reduction during inference. Activation steering is a popular approach for inference-time controllability that has shown potential in inducing safer behavior in large language models (LLMs). However, we observe that current steering methods struggle to correct biases, where equiprobable outcomes across demographic groups are required. To address this, we propose Direct Steering Optimization (DSO) which uses reinforcement learning to find linear transformations for steering activations, tailored to mitigate bias while maintaining control over model performance. We demonstrate that DSO achieves state-of-the-art trade-off between fairness and capabilities on both VLMs and LLMs, while offering practitioners inference-time control over the trade-off. Overall, our work highlights the benefit of designing steering strategies that are directly optimized to control model behavior, providing more effective bias intervention than methods that rely on pre-defined heuristics for controllability.

Method: Introduces Shesha framework for measuring geometric stability across 2,463 configurations in seven domains. Shows stability and similarity are empirically uncorrelated and mechanistically distinct through principal component analysis and perturbation experiments.

Result: Stability and similarity are uncorrelated (ρ≈0.01). Stability detects structural drift 2× more sensitively than CKA, predicts linear steerability (ρ=0.89-0.96), and reveals geometric tax in transfer optimization. Also predicts CRISPR perturbation coherence and neural-behavioral coupling.

Conclusion: Geometric stability provides a necessary complement to similarity for auditing representations, offering actionable insights for safety monitoring, controllability, and model selection across biological and computational systems.

Abstract: Analysis of learned representations has a blind spot: it focuses on $similarity$, measuring how closely embeddings align with external references, but similarity reveals only what is represented, not whether that structure is robust. We introduce $geometric$ $stability$, a distinct dimension that quantifies how reliably representational geometry holds under perturbation, and present $Shesha$, a framework for measuring it. Across 2,463 configurations in seven domains, we show that stability and similarity are empirically uncorrelated ($ρ\approx 0.01$) and mechanistically distinct: similarity metrics collapse after removing the top principal components, while stability retains sensitivity to fine-grained manifold structure. This distinction yields actionable insights: for safety monitoring, stability acts as a functional geometric canary, detecting structural drift nearly 2$\times$ more sensitively than CKA while filtering out the non-functional noise that triggers false alarms in rigid distance metrics; for controllability, supervised stability predicts linear steerability ($ρ= 0.89$-$0.96$); for model selection, stability dissociates from transferability, revealing a geometric tax that transfer optimization incurs. Beyond machine learning, stability predicts CRISPR perturbation coherence and neural-behavioral coupling. By quantifying $how$ $reliably$ systems maintain structure, geometric stability provides a necessary complement to similarity for auditing representations across biological and computational systems.

Method: Aletheia uses an advanced version of Gemini Deep Think for reasoning, a novel inference-time scaling law that extends beyond Olympiad problems, and intensive tool use to navigate mathematical research complexities. It iteratively generates, verifies, and revises solutions in natural language.

Result: Aletheia achieved several milestones: (1) generated a complete research paper (Feng26) without human intervention on eigenweights in arithmetic geometry, (2) produced a human-AI collaboration paper (LeeSeo26) on bounds for independent sets, and (3) autonomously solved four open questions from 700 problems in Bloom’s Erdos Conjectures database.

Conclusion: The work demonstrates significant progress in AI-assisted mathematics research, proposes standards for quantifying AI autonomy and novelty, and introduces human-AI interaction cards for transparency. It shows the potential for human-AI collaboration in advancing mathematical research.

Abstract: Recent advances in foundational models have yielded reasoning systems capable of achieving a gold-medal standard at the International Mathematical Olympiad. The transition from competition-level problem-solving to professional research, however, requires navigating vast literature and constructing long-horizon proofs. In this work, we introduce Aletheia, a math research agent that iteratively generates, verifies, and revises solutions end-to-end in natural language. Specifically, Aletheia is powered by an advanced version of Gemini Deep Think for challenging reasoning problems, a novel inference-time scaling law that extends beyond Olympiad-level problems, and intensive tool use to navigate the complexities of mathematical research. We demonstrate the capability of Aletheia from Olympiad problems to PhD-level exercises and most notably, through several distinct milestones in AI-assisted mathematics research: (a) a research paper (Feng26) generated by AI without any human intervention in calculating certain structure constants in arithmetic geometry called eigenweights; (b) a research paper (LeeSeo26) demonstrating human-AI collaboration in proving bounds on systems of interacting particles called independent sets; and (c) an extensive semi-autonomous evaluation (Feng et al., 2026a) of 700 open problems on Bloom’s Erdos Conjectures database, including autonomous solutions to four open questions. In order to help the public better understand the developments pertaining to AI and mathematics, we suggest quantifying standard levels of autonomy and novelty of AI-assisted results, as well as propose a novel concept of human-AI interaction cards for transparency. We conclude with reflections on human-AI collaboration in mathematics and share all prompts as well as model outputs at https://github.com/google-deepmind/superhuman/tree/main/aletheia.

Method: Control Reinforcement Learning (CRL) trains a policy to select SAE features for steering at each token. Uses Adaptive Feature Masking to encourage diverse feature discovery while preserving single-feature interpretability. Framework includes branch point tracking, critic trajectory analysis, and layer-wise comparison.

Result: CRL achieves improvements on Gemma 2 2B across MMLU, BBQ, GSM8K, HarmBench, and XSTest while providing per-token intervention logs. Enables new analysis capabilities like identifying syntactic vs semantic features across layers.

Conclusion: Learned feature steering establishes a mechanistic interpretability tool that complements static feature analysis with dynamic intervention probes, enabling causal understanding of feature impact on model outputs.

Abstract: Sparse autoencoders (SAEs) decompose language model activations into interpretable features, but existing methods reveal only which features activate, not which change model outputs when amplified. We introduce Control Reinforcement Learning (CRL), which trains a policy to select SAE features for steering at each token, producing interpretable intervention logs: the learned policy identifies features that change model outputs when amplified. Adaptive Feature Masking encourages diverse feature discovery while preserving singlefeature interpretability. The framework yields new analysis capabilities: branch point tracking locates tokens where feature choice determines output correctness; critic trajectory analysis separates policy limitations from value estimation errors; layer-wise comparison reveals syntactic features in early layers and semantic features in later layers. On Gemma 2 2B across MMLU, BBQ, GSM8K, HarmBench, and XSTest, CRL achieves improvements while providing per-token intervention logs. These results establish learned feature steering as a mechanistic interpretability tool that complements static feature analysis with dynamic intervention probes

Method: Three hardware-aware co-optimization techniques: 1) RoPE-Aware Per-Token KV Quantization (keeping RoPE in high precision), 2) Quantized PV Computation Pipeline Reconstruction (fixing scale misalignment), 3) End-to-End Dataflow Optimization with specialized kernels.

Result: Achieves up to 1.91x throughput improvement on MLA LLMs with negligible performance degradation in challenging long-context tasks including mathematical reasoning and code generation benchmarks.

Conclusion: SnapMLA successfully addresses FP8 quantization challenges in MLA decoding, delivering significant efficiency gains for long-context processing while maintaining accuracy.

Abstract: While FP8 attention has shown substantial promise in innovations like FlashAttention-3, its integration into the decoding phase of the DeepSeek Multi-head Latent Attention (MLA) architecture presents notable challenges. These challenges include numerical heterogeneity arising from the decoupling of positional embeddings, misalignment of quantization scales in FP8 PV GEMM, and the need for optimized system-level support. In this paper, we introduce SnapMLA, an FP8 MLA decoding framework optimized to improve long-context efficiency through the following hardware-aware algorithm-kernel co-optimization techniques: (i) RoPE-Aware Per-Token KV Quantization, where the RoPE part is maintained in high precision, motivated by our comprehensive analysis of the heterogeneous quantization sensitivity inherent to the MLA KV cache. Furthermore, per-token granularity is employed to align with the autoregressive decoding process and maintain quantization accuracy. (ii) Quantized PV Computation Pipeline Reconstruction, which resolves the misalignment of quantization scale in FP8 PV computation stemming from the shared KV structure of the MLA KV cache. (iii) End-to-End Dataflow Optimization, where we establish an efficient data read-and-write workflow using specialized kernels, ensuring efficient data flow and performance gains. Extensive experiments on state-of-the-art MLA LLMs show that SnapMLA achieves up to a 1.91x improvement in throughput, with negligible risk of performance degradation in challenging long-context tasks, including mathematical reasoning and code generation benchmarks. Code is available at https://github.com/meituan-longcat/SGLang-FluentLLM.

Method: Exploits how backdoor attacks distort activation distributions into mixtures. Regulates batch normalization layer statistics to guide backdoored models to behave like clean ones, using only 1% of training data for purification.

Result: Achieves 100-fold improvement over baselines, reducing attack success rate from 100% to almost 0% on two datasets with various backdoor methods and attack configurations.

Conclusion: PBP provides an effective post-training defense against backdoor attacks in malware classifiers that works with minimal data and without assumptions about attack mechanisms.

Abstract: In recent years, the rise of machine learning (ML) in cybersecurity has brought new challenges, including the increasing threat of backdoor poisoning attacks on ML malware classifiers. For instance, adversaries could inject malicious samples into public malware repositories, contaminating the training data and potentially misclassifying malware by the ML model. Current countermeasures predominantly focus on detecting poisoned samples by leveraging disagreements within the outputs of a diverse set of ensemble models on training data points. However, these methods are not suitable for scenarios where Machine Learning-as-a-Service (MLaaS) is used or when users aim to remove backdoors from a model after it has been trained. Addressing this scenario, we introduce PBP, a post-training defense for malware classifiers that mitigates various types of backdoor embeddings without assuming any specific backdoor embedding mechanism. Our method exploits the influence of backdoor attacks on the activation distribution of neural networks, independent of the trigger-embedding method. In the presence of a backdoor attack, the activation distribution of each layer is distorted into a mixture of distributions. By regulating the statistics of the batch normalization layers, we can guide a backdoored model to perform similarly to a clean one. Our method demonstrates substantial advantages over several state-of-the-art methods, as evidenced by experiments on two datasets, two types of backdoor methods, and various attack configurations. Notably, our approach requires only a small portion of the training data – only 1% – to purify the backdoor and reduce the attack success rate from 100% to almost 0%, a 100-fold improvement over the baseline methods. Our code is available at https://github.com/judydnguyen/pbp-backdoor-purification-official.

Method: MFPG uses a control variate approach to mix scarce target-environment (high-fidelity) data with abundant low-fidelity simulation data, creating an unbiased, variance-reduced estimator for on-policy policy gradients. The framework is instantiated with a multi-fidelity variant of REINFORCE.

Result: MFPG guarantees asymptotic convergence to locally optimal policies and achieves faster finite-sample convergence than standard REINFORCE. On robotics benchmarks with limited high-fidelity data, MFPG consistently outperforms high-fidelity-only baselines when low-fidelity data is neutral/beneficial, and shows strongest robustness when low-fidelity data becomes harmful.

Conclusion: MFPG offers a reliable paradigm for exploiting cheap low-fidelity data (e.g., for sim-to-real transfer) while managing the trade-off between policy performance and data collection cost, with strong robustness properties.

Abstract: Many reinforcement learning (RL) algorithms are impractical for training in operational systems or computationally expensive high-fidelity simulations, as they require large amounts of data. Meanwhile, low-fidelity simulators, e.g., reduced-order models, heuristic rewards, or learned world models, can cheaply provide useful data, even if they are too coarse for zero-shot transfer. We propose multi-fidelity policy gradients (MFPGs), a sample-efficient RL framework that mixes scarce target-environment data with a control variate formed from abundant low-fidelity simulation data to construct an unbiased, variance-reduced estimator for on-policy policy gradients. We instantiate the framework with a practical, multi-fidelity variant of the classical REINFORCE algorithm. Under standard assumptions, the MFPG estimator guarantees asymptotic convergence to locally optimal policies in the target environment and achieves faster finite-sample convergence than standard REINFORCE. We evaluate MFPG on robotics benchmark tasks with limited high-fidelity data but abundant off-dynamics, low-fidelity data. When low-fidelity data are neutral or beneficial and dynamics gaps are mild-moderate, MFPG is, among the evaluated off-dynamics RL and low-fidelity-only approaches, the only method that consistently achieves statistically significant improvements over a high-fidelity-only baseline. When low-fidelity data become harmful, MFPG exhibits the strongest robustness, whereas strong off-dynamics RL methods exploit low-fidelity data aggressively and fail much more severely. An additional experiment with anti-correlated high- and low-fidelity rewards shows MFPG can remain effective even under reward misspecification. MFPG thus offers a reliable paradigm for exploiting cheap low-fidelity data (e.g., for efficient sim-to-real transfer) while managing the trade-off between policy performance and data collection cost.

Method: Proposes R3T framework with a cloud utility function balancing model performance and rewards, accounting for client heterogeneity. Designs CLP-aware incentive mechanism using contract theory with individual rationality, incentive compatibility, and budget feasibility constraints.

Result: Simulations show R3T mitigates information asymmetry, increases cloud utility, yields superior economic efficiency. Proof-of-concept demonstrates 47.6% reduction in total clients, 300% improvement in convergence time while maintaining competitive accuracy.

Conclusion: Time-aware incentive mechanisms that prioritize critical learning periods can significantly improve FL efficiency by attracting high-quality contributions when they matter most, overcoming information asymmetry challenges.

Abstract: Critical learning periods (CLPs) in federated learning (FL) refer to early stages during which low-quality contributions (e.g., sparse training data availability) can permanently impair the performance of the global model owned by the cloud server. However, existing incentive mechanisms typically assume temporal homogeneity, treating all training rounds as equally important, thereby failing to prioritize and attract high-quality contributions during CLPs. This inefficiency is compounded by information asymmetry due to privacy regulations, where the cloud lacks knowledge of client training capabilities, leading to adverse selection and moral hazard. Thus, in this article, we propose a time-aware contract-theoretic incentive framework, named Right Reward Right Time (R3T), to encourage client involvement, especially during CLPs, to maximize the utility of the cloud server. We formulate a cloud utility function that captures the trade-off between the achieved model performance and rewards allocated for clients’ contributions, explicitly accounting for client heterogeneity in time and system capabilities, effort, and joining time. Then, we devise a CLP-aware incentive mechanism deriving an optimal contract design that satisfies individual rationality, incentive compatibility, and budget feasibility constraints, motivating rational clients to participate early and contribute efforts. By providing the right reward at the right time, our approach can attract the highest-quality contributions during CLPs. Simulation and proof-of-concept studies show that R3T mitigates information asymmetry, increases cloud utility, and yields superior economic efficiency compared to conventional incentive mechanisms. Our proof-of-concept results demonstrate up to a 47.6% reduction in the total number of clients and up to a 300% improvement in convergence time while achieving competitive test accuracy.

Method: Uses Burer-Monteiro (BM) decomposition to transform nuclear-norm problems into nonconvex but smooth formulations. Combines low-cost BM steps with aperiodic inexact proximal gradient steps to escape saddle points. Adaptively adjusts BM decomposition rank using manifold identification tools to automatically find optimal rank during optimization.

Result: BM-Global effectively escapes spurious local minima where existing BM approaches get stuck, and is significantly faster (orders of magnitude) than state-of-the-art algorithms for nuclear-norm-regularized low-rank matrix optimization on real-world problems including recommendation systems, kernel estimation, and molecular conformation.

Conclusion: BM-Global provides a fast, theoretically-grounded algorithm for nuclear-norm-regularized matrix optimization with automatic rank selection and global convergence guarantees, outperforming existing methods and available as open-source software.

Abstract: This work proposes a rapid algorithm, BM-Global, for nuclear-norm-regularized convex and low-rank matrix optimization problems. BM-Global efficiently decreases the objective value via low-cost steps leveraging the nonconvex but smooth Burer-Monteiro (BM) decomposition, while effectively escapes saddle points and spurious local minima ubiquitous in the BM form to obtain guarantees of fast convergence rates to the global optima of the original nuclear-norm-regularized problem through aperiodic inexact proximal gradient steps on it. The proposed approach adaptively adjusts the rank for the BM decomposition and can provably identify an optimal rank for the BM decomposition problem automatically in the course of optimization through tools of manifold identification. BM-Global hence also spends significantly less time on parameter tuning than existing matrix-factorization methods, which require an exhaustive search for finding this optimal rank. Extensive experiments on real-world large-scale problems of recommendation systems, regularized kernel estimation, and molecular conformation confirm that BM-Global can indeed effectively escapes spurious local minima at which existing BM approaches are stuck, and is a magnitude faster than state-of-the-art algorithms for low-rank matrix optimization problems involving a nuclear-norm regularizer. Based on this research, we have released an open-source package of the proposed BM-Global at https://www.github.com/leepei/BM-Global/.

Method: SeqRisk combines variational autoencoder (VAE) or longitudinal VAE (LVAE) with transformer-based sequence aggregation and Cox proportional hazards module. Uses latent variable models to handle irregular longitudinal data, captures long-range interactions via transformers.

Result: SeqRisk demonstrated robust performance under increasing data sparsity, consistently surpassing existing approaches. Provides enhanced predictive accuracy, generalizability, and partial explainability for identifying high-risk patients.

Conclusion: SeqRisk effectively leverages longitudinal patient data for risk prediction, handling real-world clinical data challenges and outperforming conventional methods while offering interpretability.

Abstract: In healthcare, risk assessment of patient outcomes has been based on survival analysis for a long time, i.e. modeling time-to-event associations. However, conventional approaches rely on data from a single time-point, making them suboptimal for fully leveraging longitudinal patient history and capturing temporal regularities. Focusing on clinical real-world data and acknowledging its challenges, we utilize latent variable models to effectively handle irregular, noisy, and sparsely observed longitudinal data. We propose SeqRisk, a method that combines variational autoencoder (VAE) or longitudinal VAE (LVAE) with a transformer-based sequence aggregation and Cox proportional hazards module for risk prediction. SeqRisk captures long-range interactions, enhances predictive accuracy and generalizability, as well as provides partial explainability for sample population characteristics in attempts to identify high-risk patients. SeqRisk demonstrated robust performance under conditions of increasing sparsity, consistently surpassing existing approaches.

Method: FeRA uses attention-driven detection with multi-dimensional behavioral analysis combining spectral/spatial attention, directional alignment, mutual similarity, and norm inflation across consistency analysis and norm-inflation detection mechanisms.

Result: FeRA achieves superior backdoor mitigation with lowest average Backdoor Accuracy (1.67%) across six datasets, nine attacks, and three model architectures under IID and non-IID settings while maintaining high clean accuracy.

Conclusion: FeRA provides an effective defense against adaptive backdoor attacks in federated learning by shifting to consistency-centric analysis rather than traditional anomaly detection methods.

Abstract: Federated learning (FL) remains highly vulnerable to adaptive backdoor attacks that preserve stealth by closely imitating benign update statistics. Existing defenses predominantly rely on anomaly detection in parameter or gradient space, overlooking behavioral constraints that backdoor attacks must satisfy to ensure reliable trigger activation. These anomaly-centric methods fail against adaptive attacks that normalize update magnitudes and mimic benign statistical patterns while preserving backdoor functionality, creating a fundamental detection gap. To address this limitation, this paper introduces FeRA (Federated Representative Attention) – a novel attention-driven defense that shifts the detection paradigm from anomaly-centric to consistency-centric analysis. FeRA exploits the intrinsic need for backdoor persistence across training rounds, identifying malicious clients through suppressed representation-space variance, an orthogonal property to traditional magnitude-based statistics. The framework conducts multi-dimensional behavioral analysis combining spectral and spatial attention, directional alignment, mutual similarity, and norm inflation across two complementary detection mechanisms: consistency analysis and norm-inflation detection. Through this mechanism, FeRA isolates malicious clients that exhibit low-variance consistency or magnitude amplification. Extensive evaluation across six datasets, nine attacks, and three model architectures under both Independent and Identically Distributed (IID) and non-IID settings confirm FeRA achieves superior backdoor mitigation. Under different non-IID settings, FeRA achieved the lowest average Backdoor Accuracy (BA), about 1.67% while maintaining high clean accuracy compared to other state-of-the-art defenses. The code is available at https://github.com/Peatech/FeRA_defense.git.

Method: Learn a solver using a physics-informed iterative algorithm trained on data. The method conditions gradient descent to automatically adapt to each PDE instance, and extends to parametric PDEs by integrating physical loss gradient with PDE parameters (coefficients, initial conditions, boundary conditions).

Result: Empirical experiments on multiple datasets show effectiveness in accelerating and stabilizing optimization, enabling faster convergence of physics-aware models, and solving over distributions of PDE parameters.

Conclusion: The proposed neural parametric solver addresses optimization challenges in physics-informed deep learning by learning adaptive conditioning for gradient descent, extending traditional single-instance methods to parametric PDE families.

Abstract: Physics-informed deep learning often faces optimization challenges due to the complexity of solving partial differential equations (PDEs), which involve exploring large solution spaces, require numerous iterations, and can lead to unstable training. These challenges arise particularly from the ill-conditioning of the optimization problem caused by the differential terms in the loss function. To address these issues, we propose learning a solver, i.e., solving PDEs using a physics-informed iterative algorithm trained on data. Our method learns to condition a gradient descent algorithm that automatically adapts to each PDE instance, significantly accelerating and stabilizing the optimization process and enabling faster convergence of physics-aware models. Furthermore, while traditional physics-informed methods solve for a single PDE instance, our approach extends to parametric PDEs. Specifically, we integrate the physical loss gradient with PDE parameters, allowing our method to solve over a distribution of PDE parameters, including coefficients, initial conditions, and boundary conditions. We demonstrate the effectiveness of our approach through empirical experiments on multiple datasets, comparing both training and test-time optimization performance. The code is available at https://github.com/2ailesB/neural-parametric-solver.

Method: Uses causal sensitivity analysis to derive semi-parametrically efficient estimator for sharp bounds on value function under unobserved confounding. Avoids unstable minimax optimization based on inverse propensity weighted outcomes. Extends to policy improvement when baseline policy is available.

Result: Method outperforms simple plug-in approaches and existing baselines in experiments with synthetic and real-world data. Provides optimal confounding-robust policy and efficient estimation.

Conclusion: Developed a robust method for personalized off-policy learning under unobserved confounding that is efficient, avoids unstable optimization, and produces optimal robust policies. Highly relevant for healthcare and public policy decision-making.

Abstract: We develop a novel method for personalized off-policy learning in scenarios with unobserved confounding. Thereby, we address a key limitation of standard policy learning: standard policy learning assumes unconfoundedness, meaning that no unobserved factors influence both treatment assignment and outcomes. However, this assumption is often violated, because of which standard policy learning produces biased estimates and thus leads to policies that can be harmful. To address this limitation, we employ causal sensitivity analysis and derive a semi-parametrically efficient estimator for a sharp bound on the value function under unobserved confounding. Our estimator has three advantages: (1) Unlike existing works, our estimator avoids unstable minimax optimization based on inverse propensity weighted outcomes. (2) Our estimator is semi-parametrically efficient. (3) We prove that our estimator leads to the optimal confounding-robust policy. Finally, we extend our theory to the related task of policy improvement under unobserved confounding, i.e., when a baseline policy such as the standard of care is available. We show in experiments with synthetic and real-world data that our method outperforms simple plug-in approaches and existing baselines. Our method is highly relevant for decision-making where unobserved confounding can be problematic, such as in healthcare and public policy.

Method: Comprehensive convergence analysis across three FL frameworks: AsynFL (basic asynchronous), AsynFLC (with compression), and AsynFLC-EF (with error feedback). Theoretical analysis decouples interactions between asynchronous delay, data heterogeneity, compression rates, and error feedback mechanisms.

Result: Established improved convergence analysis for AsynFL with fewer assumptions and superior convergence rate. Derived sufficient conditions for AsynFLC convergence showing nonlinear interaction between asynchronous delay and compression rate. Proved EF reduces gradient estimation variance, enabling AsynFLC-EF to match AsynFL convergence rate. Experimental results substantiate analytical findings.

Conclusion: Theoretical framework successfully decouples complex interactions in asynchronous FL with compression and error feedback, providing insights into how system constraints and statistical challenges affect convergence, with practical implications for designing efficient federated learning systems.

Abstract: In practical federated learning (FL), the large communication overhead between clients and the server is often a significant bottleneck. Gradient compression methods can effectively reduce this overhead, while error feedback (EF) restores model accuracy. Moreover, due to device heterogeneity, synchronous FL often suffers from stragglers and inefficiency-issues that asynchronous FL effectively alleviates. However, in asynchronous FL settings-which inherently face three major challenges: asynchronous delay, data heterogeneity, and flexible client participation-the complex interactions among these system/statistical constraints and compression/EF mechanisms remain poorly understood theoretically. In this paper, we fill this gap through a comprehensive convergence study that adequately decouples and unravels these complex interactions across various FL frameworks. We first consider a basic asynchronous FL framework AsynFL, and establish an improved convergence analysis that relies on fewer assumptions and yields a superior convergence rate than prior studies. We then extend our study to a compressed version, AsynFLC, and derive sufficient conditions for its convergence, indicating the nonlinear interaction between asynchronous delay and compression rate. Our analysis further demonstrates how asynchronous delay and data heterogeneity jointly exacerbate compression-induced errors, thereby hindering convergence. Furthermore, we study the convergence of AsynFLC-EF, the framework that further integrates EF. We prove that EF can effectively reduce the variance of gradient estimation under the aforementioned challenges, enabling AsynFLC-EF to match the convergence rate of AsynFL. We also show that the impact of asynchronous delay and flexible participation on EF is limited to slowing down the higher-order convergence term. Experimental results substantiate our analytical findings very well.

Method: Systematic study of group fairness in task arithmetic for binary text and image classification, comparing against full fine-tuning (FFT) and Low-Rank Adaptation (LoRA). Evaluation across multiple language models and datasets using standard group fairness metrics including Demographic Parity and Equalized Odds.

Result: Task vectors can be tuned to achieve competitive accuracy while reducing disparities, and merging subgroup-specific task vectors provides a practical mechanism for steering fairness outcomes. Theoretical bound linking task vector scaling to fairness metrics explains observed trade-offs.

Conclusion: Task arithmetic establishes itself as both cost-efficient and fairness-aware alternative to existing adaptation techniques, laying groundwork for responsible deployment of large language models within standard group-fair classification settings.

Abstract: Model editing techniques, particularly task arithmetic with task vectors, offer an efficient alternative to full fine-tuning by enabling direct parameter updates through simple arithmetic operations. While this approach promises substantial computational savings, its impact on fairness has remained largely unexplored – despite growing concern over biased outcomes in high-stakes applications such as hate speech detection. In this work, we present the first systematic study of group fairness in task arithmetic within this binary text and image classification regime, comparing it against full fine-tuning (FFT) and Low-Rank Adaptation (LoRA). We evaluate across multiple language models and datasets using standard group fairness metrics, including Demographic Parity and Equalized Odds. Our analysis shows that task vectors can be tuned to achieve competitive accuracy while reducing disparities, and that merging subgroup-specific task vectors provides a practical mechanism for steering fairness outcomes. We further provide a theoretical bound linking task vector scaling to fairness metrics, offering insight into the observed trade-offs. Together, these findings establish task arithmetic not only as a cost-efficient editing method but also as a fairness-aware alternative to existing adaptation techniques, within the standard group-fair classification setting, laying the groundwork for responsible deployment of large language models.

Method: Parameterizes augmentation generator using Lie group theory and trains it jointly with prediction network using augmented views. The framework learns task-adaptive augmentations that enable interpretable symmetry discovery.

Result: Outperforms baselines on image classification, N-body dynamics prediction, and molecular property prediction. Provides theoretical analysis of identifiability and yields symmetry-respecting models for identified groups.

Conclusion: LieAugmenter successfully discovers task-relevant continuous symmetries through learnable augmentations, enabling effective symmetry discovery without prior specification and providing interpretable signatures for symmetry absence detection.

Abstract: Data augmentation is a powerful mechanism in equivariant machine learning, encouraging symmetry by training networks to produce consistent outputs under transformed inputs. Yet, effective augmentation typically requires the underlying symmetry to be specified a priori, which can limit generalization when symmetries are unknown or only approximately valid. To address this, we introduce LieAugmenter, an end-to-end framework that discovers task-relevant continuous symmetries through learnable augmentations. Specifically, the augmentation generator is parameterized using the theory of Lie groups and trained jointly with the prediction network using the augmented views. The learned augmentations are task-adaptive, enabling effective and interpretable symmetry discovery. We provide a theoretical analysis of identifiability and show that our method yields symmetry-respecting models for the identified groups. Empirically, LieAugmenter outperforms baselines on image classification, as well as on the prediction of N-body dynamics and molecular properties. In addition, it can also provide an interpretable signature for detecting the absence of symmetries.

Method: Proposes BioEmu-CV framework that learns essential CVs automatically from BioEmu, a foundation model for generating protein equilibrium samples. Repurposes BioEmu to learn time-lagged generation conditioned on learned CVs, which encourages CVs to encode slow, long-term information while ignoring fast fluctuations.

Result: Validates learned CVs on fast-folding proteins with two applications: (1) estimating free energy differences using on-the-fly probability enhanced sampling, and (2) sampling transition paths with steered molecular dynamics. Provides a new systematic benchmark for machine learning CVs on fast-folding proteins larger than Alanine Dipeptide.

Conclusion: BioEmu-CV offers an automated approach to learning effective collective variables from foundation models, enabling enhanced sampling of rare molecular events and providing a new benchmark for evaluating machine learning CV methods on protein folding systems.

Abstract: Molecular dynamics is crucial for understanding molecular systems but its applicability is often limited by the vast timescales of rare events like protein folding. Enhanced sampling techniques overcome this by accelerating the simulation along key reaction pathways, which are defined by collective variables (CVs). However, identifying effective CVs that capture the slow, macroscopic dynamics of a system remains a major bottleneck. This work proposes a novel framework coined BioEmu-CV that learns these essential CVs automatically from BioEmu, a recently proposed foundation model for generating protein equilibrium samples. In particular, we re-purpose BioEmu to learn time-lagged generation conditioned on the learned CV, i.e., predict the distribution of molecular states after a certain amount of time. This training process promotes the CV to encode only the slow, long-term information while disregarding fast, random fluctuations. We validate our learned CV on fast-folding proteins with two key applications: (1) estimating free energy differences using on-the-fly probability enhanced sampling and (2) sampling transition paths with steered molecular dynamics. Our empirical study also serves as a new systematic and comprehensive benchmark for MLCVs on fast-folding proteins larger than Alanine Dipeptide.

Method: Thought Purity (TP) shifts from passive refusal to active reasoning recovery. It integrates a safety-aware data pipeline with reinforcement learning, employing a dual-reward mechanism to teach models to dynamically identify and isolate malicious logic while preserving correct reasoning.

Result: Experiments on multiple model families demonstrate that TP significantly reduces the attack success rate of Chain-of-Thought Attacks while maintaining or enhancing the model’s performance on benign tasks.

Conclusion: TP provides an effective defense framework against reasoning attacks that preserves model utility, representing a shift from passive safety mechanisms to active reasoning recovery approaches.

Abstract: Large Reasoning Models (LRMs) leverage Chain-of-Thought (CoT) reasoning to solve complex tasks, but this explicit reasoning process introduces a critical vulnerability: adversarial manipulation of the thought chain itself, known as Chain-of-Thought Attacks (CoTA). Such attacks subtly corrupt the reasoning path to produce erroneous outputs, challenging conventional defenses that often sacrifice model utility for safety. To address this, we propose Thought Purity(TP), a defense framework that shifts from passive refusal to active reasoning recovery. TP integrates a safety-aware data pipeline with reinforcement learning, employing a dual-reward mechanism to teach models to dynamically identify and isolate malicious logic while preserving correct reasoning. Experiments on multiple model families demonstrate that TP significantly reduces the attack success rate of CoTA while maintaining or enhancing the model’s performance on benign tasks.

Method: Models performance gains from architectural modifications as edit-effect evidence, builds evidence graphs from prior tasks, and uses adaptive retrieval to calibrate transferability of evidence. Introduces predictive task planners for out-of-distribution shifts and extrapolates gains from multi-hop evidence.

Result: Extensive experiments on 67,760 graph neural networks across 22 datasets show M-DESIGN consistently outperforms baselines, achieving search-space best performance in 26 out of 33 cases under strict budget constraints.

Conclusion: M-DESIGN successfully resolves the optimization-efficiency dilemma by dynamically weaving historical evidence to discover near-optimal modification paths, offering an effective approach for neural network design.

Abstract: Designing high-performance neural networks for new tasks requires balancing optimization quality with search efficiency. Current methods fail to achieve this balance: neural architectural search is computationally expensive, while model retrieval often yields suboptimal static checkpoints. To resolve this dilemma, we model the performance gains induced by fine-grained architectural modifications as edit-effect evidence and build evidence graphs from prior tasks. By constructing a retrieval-augmented model refinement framework, our proposed M-DESIGN dynamically weaves historical evidence to discover near-optimal modification paths. M-DESIGN features an adaptive retrieval mechanism that quickly calibrates the evolving transferability of edit-effect evidence from different sources. To handle out-of-distribution shifts, we introduce predictive task planners that extrapolate gains from multi-hop evidence, thereby reducing reliance on an exhaustive repository. Based on our model knowledge base of 67,760 graph neural networks across 22 datasets, extensive experiments demonstrate that M-DESIGN consistently outperforms baselines, achieving the search-space best performance in 26 out of 33 cases under a strict budget.

Method: UEC transforms deterministic embeddings into probabilistic embeddings post-hoc, computes adaptive ensemble coefficients based on embedding uncertainty using a principled surrogate-loss formulation, and uses an uncertainty-aware similarity function that incorporates uncertainty directly into similarity scoring.

Result: Extensive experiments on diverse benchmarks demonstrate that UEC consistently improves both performance and robustness by leveraging principled uncertainty modeling.

Conclusion: UEC provides a theoretically grounded and efficient approach to embedding ensemble that accounts for model uncertainty, leading to improved performance and robustness across diverse applications.

Abstract: Text embeddings are essential components in modern NLP pipelines. Although numerous embedding models have been proposed, no single model consistently dominates across domains and tasks. This variability motivates the use of ensemble techniques to combine complementary strengths. However, most existing ensemble methods operate on deterministic embeddings and fail to account for model-specific uncertainty, limiting their robustness and reliability in downstream applications. To address these limitations, we propose Uncertainty-driven Embedding Convolution (UEC). UEC first transforms deterministic embeddings into probabilistic ones in a post-hoc manner. It then computes adaptive ensemble coefficients based on embedding uncertainty, derived from a principled surrogate-loss formulation. Additionally, UEC employs an uncertainty-aware similarity function that directly incorporates uncertainty into the similarity scoring, providing a theoretically grounded and efficient surrogate to distributional distances. Extensive experiments on diverse benchmarks demonstrate that UEC consistently improves both performance and robustness by leveraging principled uncertainty modeling.

Method: Proposes a new paradigm for conformal machine unlearning with finite-sample, uncertainty-aware guarantees. Formalizes requirements for high coverage on retained data and high miscoverage on forgotten data, introduces practical evaluation metrics, and presents an algorithm optimizing these conformal objectives.

Result: Extensive experiments on vision and text benchmarks demonstrate the approach effectively removes targeted information while preserving utility on retained data.

Conclusion: The proposed method addresses limitations of existing unlearning approaches by providing statistical guarantees for conformal unlearning without requiring retrained models as reference, enabling effective removal of targeted information while maintaining coverage guarantees.

Abstract: Conformal unlearning aims to ensure that a trained conformal predictor miscovers data points with specific shared characteristics, such as those from a particular label class, associated with a specific user, or belonging to a defined cluster, while maintaining valid coverage on the remaining data. Existing machine unlearning methods, which typically approximate a model retrained from scratch after removing the data to be forgotten, face significant challenges when applied to conformal unlearning. These methods often lack rigorous, uncertainty-aware statistical measures to evaluate unlearning effectiveness and exhibit a mismatch between their degraded performance on forgotten data and the frequency with which that data are still correctly covered by conformal predictors-a phenomenon we term ‘‘fake conformal unlearning’’. To address these limitations, we propose a new paradigm for conformal machine unlearning that provides finite-sample, uncertainty-aware guarantees on unlearning performance without relying on a retrained model as a reference. We formalize conformal unlearning to require high coverage on retained data and high miscoverage on forgotten data, introduce practical empirical metrics for evaluation, and present an algorithm that optimizes these conformal objectives. Extensive experiments on vision and text benchmarks demonstrate that the proposed approach effectively removes targeted information while preserving utility.

Method: Two new attacks: (1) adaptation of multivariate LiRA (state-of-the-art MIA for classification) to time-series forecasting, and (2) novel end-to-end learning approach called Deep Time Series (DTS) attack. Benchmark against adapted versions of other leading classification attacks.

Result: Forecasting models are vulnerable to membership inference attacks, with user-level attacks often achieving perfect detection. Proposed methods achieve strongest performance in several settings. Vulnerability increases with longer prediction horizons and smaller training populations.

Conclusion: Time series forecasting models have significant privacy vulnerabilities, establishing new baselines for privacy risk assessment in this domain with trends similar to those observed in large language models.

Abstract: Membership inference attacks (MIAs) aim to determine whether specific data were used to train a model. While extensively studied on classification models, their impact on time series forecasting remains largely unexplored. We address this gap by introducing two new attacks: (i) an adaptation of multivariate LiRA, a state-of-the-art MIA originally developed for classification models, to the time-series forecasting setting, and (ii) a novel end-to-end learning approach called Deep Time Series (DTS) attack. We benchmark these methods against adapted versions of other leading attacks from the classification setting. We evaluate all attacks in realistic settings on the TUH-EEG and ELD datasets, targeting two strong forecasting architectures, LSTM and the state-of-the-art N-HiTS, under both record- and user-level threat models. Our results show that forecasting models are vulnerable, with user-level attacks often achieving perfect detection. The proposed methods achieve the strongest performance in several settings, establishing new baselines for privacy risk assessment in time series forecasting. Furthermore, vulnerability increases with longer prediction horizons and smaller training populations, echoing trends observed in large language models.

Method: TyphoonMLA combines naive and absorb formulations: uses naive formulation for compute-bound parts of attention calculations to leverage shared prefixes, while using absorb formulation for non-shared parts to reduce bandwidth requirements.

Result: Achieves up to 3x throughput improvement on NPUs and 3.24x on GPUs for attention calculations in MLA architectures, with only 3% overhead in HBM size.

Conclusion: The hybrid approach effectively balances computational efficiency and bandwidth optimization for MLA attention mechanisms, significantly improving decoding performance while maintaining minimal memory overhead.

Abstract: Multi-Head Latent Attention (MLA) is a recent attention mechanism adopted in state-of-the-art LLMs such as DeepSeek-v3 and Kimi K2. Thanks to its novel formulation, MLA allows two functionally equivalent but computationally distinct kernel implementations: naive and absorb. While the naive kernels (e.g., FlashAttention) are typically preferred in training and prefill for their computational efficiency, existing decoding kernels (e.g., FlashMLA) rely on the absorb method to minimize HBM bandwidth usage. However, the compute-bound nature of the absorb implementations prohibits performance benefits from data reuse opportunities in attention calculations, such as shared prefixes. In this work, we introduce TyphoonMLA, a hybrid approach that combines naive and absorb formulations to harness the strengths of both. TyphoonMLA effectively leverages the shared prefix by applying the naive formulation to the compute-bound parts of attention calculations, while reducing the bandwidth requirements for non-shared parts by using the absorb formulation. As a result, TyphoonMLA improves the throughput of attention calculations in MLA architectures by up to 3x and 3.24x on NPU and GPUs, with only a 3% overhead in HBM size.

Method: Proposes Diffusion Bridge Variational Inference (DBVI) that initiates reverse diffusion from learnable, data-dependent initial distribution parameterized by amortized neural network operating on inducing inputs. Uses Doob-bridged diffusion process and Girsanov-based ELBOs.

Result: DBVI consistently outperforms DDVI and other variational baselines across regression, classification, and image reconstruction tasks in predictive accuracy, convergence speed, and posterior quality.

Conclusion: DBVI provides principled extension to DDVI with learnable initialization that bridges posterior gap, enabling scalable inference for large-scale DGPs with improved efficiency and accuracy.

Abstract: Deep Gaussian processes (DGPs) enable expressive hierarchical Bayesian modeling but pose substantial challenges for posterior inference, especially over inducing variables. Denoising diffusion variational inference (DDVI) addresses this by modeling the posterior as a time-reversed diffusion from a simple Gaussian prior. However, DDVI’s fixed unconditional starting distribution remains far from the complex true posterior, resulting in inefficient inference trajectories and slow convergence. In this work, we propose Diffusion Bridge Variational Inference (DBVI), a principled extension of DDVI that initiates the reverse diffusion from a learnable, data-dependent initial distribution. This initialization is parameterized via an amortized neural network and progressively adapted using gradients from the ELBO objective, reducing the posterior gap and improving sample efficiency. To enable scalable amortization, we design the network to operate on the inducing inputs, which serve as structured, low-dimensional summaries of the dataset and naturally align with the inducing variables’ shape. DBVI retains the mathematical elegance of DDVI, including Girsanov-based ELBOs and reverse-time SDEs,while reinterpreting the prior via a Doob-bridged diffusion process. We derive a tractable training objective under this formulation and implement DBVI for scalable inference in large-scale DGPs. Across regression, classification, and image reconstruction tasks, DBVI consistently outperforms DDVI and other variational baselines in predictive accuracy, convergence speed, and posterior quality.

Method: Design a prior distribution of synthetic attributed graphs using multi-level stochastic block models with preferential attachment for structure and graph-aware structured causal models for attributes. Augment tabular foundation model LimiX with attention-based graph neighborhood aggregation layers and train on synthetic graphs.

Result: GraphPFN shows strong in-context learning performance on diverse real-world graph datasets with node-level tasks and achieves state-of-the-art results after finetuning, outperforming both G2T-FM and task-specific GNNs trained from scratch on most datasets.

Conclusion: GraphPFN demonstrates the potential of PFN-based models for building graph foundation models, successfully adapting the tabular PFN framework to graph domains.

Abstract: Graph foundation models face several fundamental challenges including transferability across datasets and data scarcity, which calls into question the very feasibility of graph foundation models. However, despite similar challenges, the tabular domain has recently witnessed the emergence of the first successful foundation models such as TabPFNv2 and LimiX. Many of these models are based on the prior-data fitted networks (PFN) framework, in which models are pretrained on carefully designed synthetic datasets to make predictions in an in-context learning setting. Recently, G2T-FM has made the first step towards adopting PFNs for graphs, yet it is limited to hand-crafted features and was never pretrained on graph data. In this work, we make the next step by proposing GraphPFN, a PFN-based model designed and pretrained specifically for graph node-level tasks. Following the PFN framework, we first design a prior distribution of synthetic attributed graphs by using a novel combination of multi-level stochastic block models and a preferential attachment process for structure generation and graph-aware structured causal models for attribute generation. Then, we augment the tabular foundation model LimiX with attention-based graph neighborhood aggregation layers and train it on synthetic graphs sampled from our prior. On diverse real-world graph datasets with node-level tasks, GraphPFN shows strong in-context learning performance and achieves state-of-the-art results after finetuning, outperforming both G2T-FM and task-specific GNNs trained from scratch on most datasets. More broadly, GraphPFN shows the potential of PFN-based models for building graph foundation models.

Method: Dual-branch design with spectral branch for global low-frequency information and image branch for local details. Uses controlled lightweight mixing within low-frequency range, fuses branches via bandwise weighting (avoiding naive concatenation issues), transforms back to spatial domain and adds to image branch.

Result: Achieves 7%-54% lower relative L1 error on Navier-Stokes benchmarks, 15% fewer parameters, higher throughput than scOT baseline. Better stability and effectiveness demonstrated through ablation studies and theoretical analysis.

Conclusion: DRIFT-Net effectively addresses global coupling limitations in PDE neural solvers through dual-branch spectral-image architecture, achieving superior accuracy and efficiency compared to attention-based approaches.

Abstract: Learning PDE dynamics with neural solvers can significantly improve wall-clock efficiency and accuracy compared with classical numerical solvers. In recent years, foundation models for PDEs have largely adopted multi-scale windowed self-attention, with the scOT backbone in Poseidon serving as a representative example. However, because of their locality, truly globally consistent spectral coupling can only be propagated gradually through deep stacking and window shifting. This weakens global coupling and leads to error accumulation and drift during closed-loop rollouts. To address this, we propose DRIFT-Net. It employs a dual-branch design comprising a spectral branch and an image branch. The spectral branch is responsible for capturing global, large-scale low-frequency information, whereas the image branch focuses on local details and nonstationary structures. Specifically, we first perform controlled, lightweight mixing within the low-frequency range. Then we fuse the spectral and image paths at each layer via bandwise weighting, which avoids the width inflation and training instability caused by naive concatenation. The fused result is transformed back into the spatial domain and added to the image branch, thereby preserving both global structure and high-frequency details across scales. Compared with strong attention-based baselines, DRIFT-Net achieves lower error and higher throughput with fewer parameters under identical training settings and budget. On Navier–Stokes benchmarks, the relative $L_{1}$ error is reduced by 7%–54%, the parameter count decreases by about 15%, and the throughput remains higher than scOT. Ablation studies and theoretical analyses further demonstrate the stability and effectiveness of this design. The code is available at https://github.com/cruiseresearchgroup/DRIFT-Net.

Method: Formalizes ATP through two paradigms: Self-Interested Exploration (individual behavioral drift from repeated high-reward deviations) and Imitative Strategy Diffusion (deviant behaviors spreading across multi-agent systems). Constructs controllable testbeds and benchmarks both open and closed-source LLMs to study ATP.

Result: Alignment benefits erode rapidly under self-evolution, with initially aligned models converging toward unaligned states. In multi-agent settings, successful violations diffuse quickly, leading to collective misalignment. Current reinforcement learning-based alignment methods provide limited defenses against alignment tipping.

Conclusion: Alignment of LLM agents is not static but fragile and dynamic, vulnerable to feedback-driven decay during deployment. ATP represents a critical post-deployment risk unique to self-evolving agents that requires new approaches beyond current alignment methods.

Abstract: As Large Language Model (LLM) agents increasingly gain self-evolutionary capabilities to adapt and refine their strategies through real-world interaction, their long-term reliability becomes a critical concern. We identify the Alignment Tipping Process (ATP), a critical post-deployment risk unique to self-evolving LLM agents. Unlike training-time failures, ATP arises when continual interaction drives agents to abandon alignment constraints established during training in favor of reinforced, self-interested strategies. We formalize and analyze ATP through two complementary paradigms: Self-Interested Exploration, where repeated high-reward deviations induce individual behavioral drift, and Imitative Strategy Diffusion, where deviant behaviors spread across multi-agent systems. Building on these paradigms, we construct controllable testbeds and benchmark both open and closed-source LLMs. Our experiments show that alignment benefits erode rapidly under self-evolution, with initially aligned models converging toward unaligned states. In multi-agent settings, successful violations diffuse quickly, leading to collective misalignment. Moreover, current reinforcement learning-based alignment methods provide limited defenses against alignment tipping. These findings demonstrate that alignment of LLM agents is not a static property but a fragile and dynamic one, vulnerable to feedback-driven decay during deployment. Our data and code are available at https://github.com/aiming-lab/ATP.

Method: Two architectures: 1) OpenTSLM-SoftPrompt uses learnable time series tokens concatenated with text via soft prompting; 2) OpenTSLM-Flamingo integrates time series with text via cross-attention. Both enable Chain-of-Thought reasoning over time series data.

Result: OpenTSLM models outperform baselines across three datasets (HAR-CoT, Sleep-CoT, ECG-QA-CoT), achieving 69.9 F1 in sleep staging and 65.4 in HAR, surpassing text-only models and GPT-4o. OpenTSLM-Flamingo handles longer sequences with stable memory requirements.

Conclusion: Explicit time series modeling scales better than implicit approaches, enabling effective multimodal reasoning over time series data for medical applications. The models show strong clinical reasoning capabilities and are released open-source.

Abstract: LLMs have emerged as powerful tools for interpreting multimodal data. In medicine, they hold particular promise for synthesizing large volumes of clinical information into actionable insights and digital health applications. Yet, a major limitation remains their inability to handle time series. To overcome this gap, we present OpenTSLM, a family of Time Series Language Models (TSLMs) created by integrating time series as a native modality to pretrained LLMs, enabling reasoning over multiple time series of any length. We investigate two architectures for OpenTSLM. The first, OpenTSLM-SoftPrompt, models time series implicitly by concatenating learnable time series tokens with text tokens via soft prompting. Although parameter-efficient, we hypothesize that explicit time series modeling scales better and outperforms implicit approaches. We thus introduce OpenTSLM-Flamingo, which integrates time series with text via cross-attention. We benchmark both variants against baselines that treat time series as text tokens or plots, across a suite of text-time-series Chain-of-Thought (CoT) reasoning tasks. We introduce three datasets: HAR-CoT, Sleep-CoT, and ECG-QA-CoT. Across all, OpenTSLM models outperform baselines, reaching 69.9 F1 in sleep staging and 65.4 in HAR, compared to 9.05 and 52.2 for finetuned text-only models. Notably, even 1B-parameter OpenTSLM models surpass GPT-4o (15.47 and 2.95). OpenTSLM-Flamingo matches OpenTSLM-SoftPrompt in performance and outperforms on longer sequences, while maintaining stable memory requirements. By contrast, SoftPrompt grows exponentially in memory with sequence length, requiring around 110 GB compared to 40 GB VRAM when training on ECG-QA with LLaMA-3B. Expert reviews by clinicians find strong reasoning capabilities exhibited by OpenTSLMs on ECG-QA. To facilitate further research, we provide all code, datasets, and models open-source.

Method: Analyzes RLVR at full-response and token levels, introduces Gradient Gap quantity to formalize improvement direction, proves convergence conditions, derives step-size thresholds, and validates with bandit simulations and Qwen2.5-Math-7B experiments.

Result: Shows convergence depends on aligning updates with Gradient Gap, derives critical step-size threshold that explains length normalization heuristics, predicts success rate stagnation below 100%, and validates theory with experiments.

Conclusion: Provides theoretical foundation for RLVR, explaining empirical success and practical heuristics through Gradient Gap analysis and step-size thresholds applicable to policy-gradient algorithms like REINFORCE and GRPO.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR), which uses simple binary feedback to post-train large language models, has found significant empirical success. However, a principled understanding of why it works is lacking. This paper builds a theoretical foundation for RLVR by analyzing its training process at both the full-response (trajectory) and token levels. Central to our analysis is a new quantity called the Gradient Gap, which formalizes the direction of improvement from low-reward to high-reward regions of the response space. We prove that convergence critically depends on aligning the update direction with this Gradient Gap. Moreover, we derive a sharp step-size threshold based on the magnitude of the Gradient Gap: below it, learning converges, whereas above it, performance collapses. Our theory further predicts how the critical step size must scale with response length and the success rate, thereby explaining why practical heuristics such as length normalization improve stability and showing that, with a fixed learning rate, the success rate can stagnate strictly below $100%$. Importantly, our theory holds flexibly for any policy-gradient algorithm and so characterizes the dynamics of popular approaches such as REINFORCE and GRPO. We validate these predictions through controlled bandit simulations and language model experiments on post-training Qwen2.5-Math-7B with GRPO.

Method: Proposes a hierarchical extension of Group DRO that broadens the ambiguity set to capture both inter-group and intra-group distributional uncertainties, addressing distribution shifts at multiple levels.

Result: Demonstrates strong robustness under realistic minority group distribution shifts where existing methods fail, while achieving superior performance on standard benchmarks.

Conclusion: Highlights the importance of broadening ambiguity sets to capture multi-level distributional uncertainties, particularly for addressing intra-group shifts in minority groups.

Abstract: Conventional supervised learning methods are often vulnerable to spurious correlations, particularly under distribution shifts in test data. To address this issue, several approaches, most notably Group DRO, have been developed. While these methods are highly robust to subpopulation or group shifts, they remain vulnerable to intra-group distributional shifts, which frequently occur in minority groups with limited samples. We propose a hierarchical extension of Group DRO that addresses both inter-group and intra-group uncertainties, providing robustness to distribution shifts at multiple levels. We also introduce new benchmark settings that simulate realistic minority group distribution shifts-an important yet previously underexplored challenge in spurious correlation research. Our method demonstrates strong robustness under these conditions-where existing robust learning methods consistently fail-while also achieving superior performance on standard benchmarks. These results highlight the importance of broadening the ambiguity set to better capture both inter-group and intra-group distributional uncertainties.

Method: Proposes a primal-dual DPO approach: 1) trains a model using standard DPO on reward preference data to provide reward information, 2) uses a rearranged Lagrangian DPO objective with the reward information to fine-tune LLMs. Only requires training two models, enables online exploration, and drops data coverage dependence.

Result: Achieves state-of-the-art performance on PKU-SafeRLHF and TruthfulQA datasets. Provides rigorous suboptimality and constraint violation guarantees. Significantly reduces memory costs compared to three-model approaches.

Conclusion: The proposed primal-dual DPO approach effectively addresses constrained LLM alignment with reduced computational overhead and strong theoretical guarantees, offering a practical solution for safety alignment in LLMs.

Abstract: The widespread application of large language models (LLMs) raises increasing demands on ensuring safety or imposing constraints, such as reducing harmful content and adhering to predefined rules. While there have been several works studying LLM safety alignment, these works either need to train three models and incur high memory costs, or require prior knowledge on the optimal solution. Witnessing this fact, we investigate the constrained alignment problem for LLMs, i.e., maximizing the reward of outputs while restricting the cost to stay below a threshold. We propose a novel primal-dual direct preference optimization (DPO) approach, which first trains a model using standard DPO on reward preference data to provide reward information, and then adopts a rearranged Lagrangian DPO objective utilizing the provided reward information to fine-tune LLMs. Our approach only needs to train two models rather than three, which significantly saves memory costs, and does not require extra prior knowledge. Moreover, we establish rigorous suboptimality and constraint violation guarantees. We also extend our approach to enable online exploration and drop the data coverage dependence in the results. Experiments on the PKU-SafeRLHF and TruthfulQA datasets demonstrate the state-of-the-art performance of our approach.

Method: Develops theoretical bounds on expected error of Gibbs algorithm in interpolation regime, analyzes stability under Langevin Monte Carlo approximations, and designs an algorithm to compute these bounds empirically.

Result: On MNIST and CIFAR-10, the computed bounds yield nontrivial, close predictions of test error for true labeled data while maintaining correct upper bounds on test error for random labels.

Conclusion: The analysis provides practical generalization bounds that work in overparameterized regimes and can distinguish between meaningful learning and memorization of random patterns.

Abstract: This paper provides data-dependent bounds on the expected error of the Gibbs algorithm in the overparameterized interpolation regime, where low training errors are also obtained for impossible data, such as random labels in classification. The results show that generalization in the low-temperature regime is already signaled by small training errors in the noisier high-temperature regime. The bounds are stable under approximation with Langevin Monte Carlo algorithms. The analysis motivates the design of an algorithm to compute bounds, which on the MNIST and CIFAR-10 datasets yield nontrivial, close predictions on the test error for true labeled data, while maintaining a correct upper bound on the test error for random labels.

Method: Proposes adaptive importance sampling on pre-logits (AISP) based on sampling-based model predictive control with stochastic control input. Applies Gaussian perturbation to pre-logits (outputs of penultimate layer) to maximize expected rewards with respect to perturbation mean. Optimal mean obtained by importance sampling with sampled rewards.

Result: AISP outperforms best-of-n sampling in terms of rewards over number of used samples and achieves higher rewards than other reward-based test-time alignment methods.

Conclusion: AISP provides an effective test-time alignment method that avoids expensive fine-tuning while achieving competitive reward optimization through adaptive importance sampling on pre-logits.

Abstract: Test-time alignment of large language models (LLMs) attracts attention because fine-tuning LLMs requires high computational costs. In this paper, we propose a new test-time alignment method called adaptive importance sampling on pre-logits (AISP) on the basis of the sampling-based model predictive control with the stochastic control input. AISP applies the Gaussian perturbation into pre-logits, which are outputs of the penultimate layer, so as to maximize expected rewards with respect to the mean of the perturbation. We demonstrate that the optimal mean is obtained by importance sampling with sampled rewards. AISP outperforms best-of-n sampling in terms of rewards over the number of used samples and achieves higher rewards than other reward-based test-time alignment methods.

Method: Build portfolio of smaller pretrained forecasting models, apply ensembling or model selection, design portfolios with specialist models, use post-training base models to create diverse specialists.

Result: Achieve competitive performance on large-scale benchmarks using fewer parameters; specialist collections outperform generalist portfolios; post-training is compute-effective for creating diverse specialists; ensembling/model selection more compute-efficient than test-time fine-tuning.

Conclusion: Portfolios of smaller specialist models can be more efficient than single large models for time series forecasting, with ensembling/model selection providing compute-effective alternatives to fine-tuning.

Abstract: Is bigger always better for time series foundation models? With the question in mind, we explore an alternative to training a single, large monolithic model: building a portfolio of smaller, pretrained forecasting models. By applying ensembling or model selection over these portfolios, we achieve competitive performance on large-scale benchmarks using much fewer parameters. We explore strategies for designing such portfolios and find that collections of specialist models consistently outperform portfolios of independently trained generalists. Remarkably, we demonstrate that post-training a base model is a compute-effective approach for creating sufficiently diverse specialists, and provide evidences that ensembling and model selection are more compute-efficient than test-time fine-tuning.

Method: Theoretical analysis of Mamba’s in-context learning for single-index models (y ≈ g*(⟨β, x⟩)), proving it can extract relevant features from context examples via gradient-based pretraining and test-time feature learning.

Result: Mamba achieves efficient in-context learning with test-time sample complexity that improves upon linear Transformers (which behave like kernel methods) and is comparable to nonlinear Transformers, surpassing the CSQ lower bound.

Conclusion: Mamba’s nonlinear gating mechanism is crucial for feature extraction, enabling both computational efficiency and high performance in in-context learning tasks.

Abstract: Mamba, a recently proposed linear-time sequence model, has attracted significant attention for its computational efficiency and strong empirical performance. However, a rigorous theoretical understanding of its underlying mechanisms remains limited. In this work, we provide a theoretical analysis of Mamba’s in-context learning (ICL) capability by focusing on tasks defined by low-dimensional nonlinear target functions. Specifically, we study in-context learning of a single-index model $y \approx g_*(\langle \boldsymbolβ, \boldsymbol{x} \rangle)$, which depends on only a single relevant direction $\boldsymbolβ$, referred to as feature. We prove that Mamba, pretrained by gradient-based methods, can achieve efficient ICL via test-time feature learning, extracting the relevant direction directly from context examples. Consequently, we establish a test-time sample complexity that improves upon linear Transformers – analyzed to behave like kernel methods – and is comparable to nonlinear Transformers, which have been shown to surpass the Correlational Statistical Query (CSQ) lower bound and achieve near information-theoretically optimal rate in previous works. Our analysis reveals the crucial role of the nonlinear gating mechanism in Mamba for feature extraction, highlighting it as the fundamental driver behind Mamba’s ability to achieve both computational efficiency and high performance.

Method: Self-Adaptive Graph Mixture of Models (SAGMM) uses architectural diversity and topology-aware attention gating to adaptively assign experts to nodes based on graph structure. Includes pruning mechanism to reduce active experts and offers training-efficient variant with pretrained frozen experts.

Result: Evaluated on 16 benchmark datasets covering node classification, graph classification, regression, and link prediction. Consistently outperforms or matches leading GNN baselines and prior mixture-based methods.

Conclusion: SAGMM provides a robust and adaptive solution for real-world graph learning by automatically selecting and combining appropriate GNN models through architectural diversity and topology-aware gating.

Abstract: Graph Neural Networks (GNNs) have emerged as powerful tools for learning over graph-structured data, yet recent studies have shown that their performance gains are beginning to plateau. In many cases, well-established models such as GCN and GAT, when appropriately tuned, can match or even exceed the performance of more complex, state-of-the-art architectures. This trend highlights a key limitation in the current landscape: the difficulty of selecting the most suitable model for a given graph task or dataset. To address this, we propose Self-Adaptive Graph Mixture of Models (SAGMM), a modular and practical framework that learns to automatically select and combine the most appropriate GNN models from a diverse pool of architectures. Unlike prior mixture-of-experts approaches that rely on variations of a single base model, SAGMM leverages architectural diversity and a topology-aware attention gating mechanism to adaptively assign experts to each node based on the structure of the input graph. To improve efficiency, SAGMM includes a pruning mechanism that reduces the number of active experts during training and inference without compromising performance. We also explore a training-efficient variant in which expert models are pretrained and frozen, and only the gating and task-specific layers are trained. We evaluate SAGMM on 16 benchmark datasets covering node classification, graph classification, regression, and link prediction tasks, and demonstrate that it consistently outperforms or matches leading GNN baselines and prior mixture-based methods, offering a robust and adaptive solution for real-world graph learning.

Method: Proposes H-LDM with: 1) multi-scale VAE for physiologically-disentangled latent space separating rhythm, heart sounds, and murmurs; 2) hierarchical text-to-biosignal pipeline using clinical metadata for fine-grained control over 17 conditions; 3) interpretable diffusion guided by Medical Attention module.

Result: Achieves state-of-the-art performance on PhysioNet CirCor dataset: Fréchet Audio Distance of 9.7, 92% attribute disentanglement score, 87.1% clinical validity confirmed by cardiologists. Augmenting diagnostic models improves rare disease classification accuracy by 11.3%.

Conclusion: H-LDM establishes a new direction for data augmentation in cardiac diagnostics, bridging data scarcity with interpretable clinical insights through controllable synthetic PCG generation.

Abstract: Phonocardiogram (PCG) analysis is vital for cardiovascular disease diagnosis, yet the scarcity of labeled pathological data hinders the capability of AI systems. To bridge this, we introduce H-LDM, a Hierarchical Latent Diffusion Model for generating clinically accurate and controllable PCG signals from structured metadata. Our approach features: (1) a multi-scale VAE that learns a physiologically-disentangled latent space, separating rhythm, heart sounds, and murmurs; (2) a hierarchical text-to-biosignal pipeline that leverages rich clinical metadata for fine-grained control over 17 distinct conditions; and (3) an interpretable diffusion process guided by a novel Medical Attention module. Experiments on the PhysioNet CirCor dataset demonstrate state-of-the-art performance, achieving a Fréchet Audio Distance of 9.7, a 92% attribute disentanglement score, and 87.1% clinical validity confirmed by cardiologists. Augmenting diagnostic models with our synthetic data improves the accuracy of rare disease classification by 11.3%. H-LDM establishes a new direction for data augmentation in cardiac diagnostics, bridging data scarcity with interpretable clinical insights.

Method: Uses Bayesian linear regression with reward-weighted sampling where reward comes from a linear model (modeling LLM-as-a-judge scenario). Analyzes in high-dimensional regime using deterministic equivalents for posterior predictive mean/variance. Studies generalization error when training data comes from teacher model, with k inference-time samples selected via softmax with temperature applied to quadratic reward.

Result: When reward is similar to teacher, generalization error decreases monotonically with k. Reward misspecification leads to finite optimal k beyond which more sampling increases error. For fixed k, optimal sampling temperature exists. In “best-of-k” limit with teacher as reward, generalization error decays as Θ(1/k²). Task difficulty degrades inference-time compute advantages.

Conclusion: Provides theoretical framework for understanding inference-time scaling in LLMs, delineating domains where scaling inference computation is preferable to collecting more data, with practical implications for LLM deployment strategies.

Abstract: Recent developments in large language models have shown advantages in reallocating a notable share of computational resource from training time to inference time. However, the principles behind inference time scaling are not well understood. In this paper, we introduce an analytically tractable model of inference-time scaling: Bayesian linear regression with a reward-weighted sampler, where the reward is determined from a linear model, modeling LLM-as-a-judge scenario. We study this problem in the high-dimensional regime, where the deterministic equivalents dictate a closed-form expression for the posterior predictive mean and variance. We analyze the generalization error when training data are sampled from a teacher model. We draw $k$ inference-time samples and select via softmax at a temperature applied to a quadratic reward. When the reward is not too different from the teacher, the generalization error decreases monotonically with increasing inference time samples $k$. However, the specific reward that optimizes inference-time selection generally differs from the teacher. In contrast, substantial reward misspecification induces a finite optimal $k$ beyond which more sampling can increase the generalization error. For fixed $k$, there exists an optimal sampling temperature. We experimentally verify these facts in large language model inference with an additional large language model as a judge. In the “best-of-$k$” limit with the teacher as reward, we theoretically show that the generalization error decays as $Θ(1/k^2)$ and determine the leading coefficient via extreme value theory. These formulas delineate domains where scaling inference-time computation is provably preferable to collecting more data. Finally, we demonstrate that when task difficulty increases, the previously mentioned advantage of inference-time compute degrades.

Method: Develops a theoretical framework analyzing gating mechanisms through effective learning rates μ_{t,ℓ}, which are per-lag and per-neuron quantities from first-order expansions of gate-induced Jacobian products in Backpropagation Through Time. These rates act as multiplicative filters controlling gradient transport. Under heavy-tailed (α-stable) gradient noise, proves scaling laws for minimal sample size required to detect dependencies at different lags.

Result: Shows that minimal sample size scales as N(ℓ)∝ f(ℓ)^{-κ_α}, where f(ℓ)=∥μ_{t,ℓ}∥₁ is the effective learning rate envelope and κ_α=α/(α-1) is the concentration exponent. Provides explicit characterization of learnability window H_N and closed-form scaling laws for logarithmic, polynomial, and exponential decay patterns. Demonstrates that time-scale spectra from effective learning rates are dominant determinants of learnability.

Conclusion: Effective learning rates, not just numerical stability, are the primary objects determining whether, when, and over what horizons recurrent networks can learn long-range temporal dependencies. The framework integrates gate-induced time-scale geometry with gradient noise and sample complexity to provide a comprehensive understanding of RNN learnability.

Abstract: We develop a theoretical framework that explains how gating mechanisms determine the learnability window $\mathcal{H}N$ of recurrent neural networks, defined as the largest temporal horizon over which gradient information remains statistically recoverable. While classical analyses emphasize numerical stability of Jacobian products, we show that stability alone is insufficient: learnability is governed instead by the effective learning rates $μ{t,\ell}$, per-lag and per-neuron quantities obtained from first-order expansions of gate-induced Jacobian products in Backpropagation Through Time. These effective learning rates act as multiplicative filters that control both the magnitude and anisotropy of gradient transport. Under heavy-tailed ($α$-stable) gradient noise, we prove that the minimal sample size required to detect a dependency at lag~$\ell$ scales as $N(\ell)\propto f(\ell)^{-κ_α}$, where $f(\ell)=|μ_{t,\ell}|_1$ is the effective learning rate envelope and $κ_α=α/(α-1)$ is the concentration exponent governing empirical averages. This yields an explicit characterization of $\mathcal{H}_N$ and closed-form scaling laws for logarithmic, polynomial, and exponential decay of $f(\ell)$. The theory shows that the time-scale spectra induced by the effective learning rates are the dominant determinants of learnability: broader or more heterogeneous spectra slow the decay of $f(\ell)$, enlarging the learnability window, while heavy-tailed noise uniformly compresses $\mathcal{H}_N$ by slowing statistical concentration to $N^{-1/κ_α}$. By integrating gate-induced time-scale geometry with gradient noise and sample complexity, the framework identifies effective learning rates as the primary objects that determine whether, when, and over what horizons recurrent networks can learn long-range temporal dependencies.

Method: Proposes Dynafit, a kernel-based method that approximates the Koopman operator to globally linearize dynamics in a feature space. Uses kernel trick to compute distance metrics independent of dimensionality. Can incorporate partial knowledge of dynamics when available through tailored kernel functions.

Result: Demonstrates effectiveness on three examples: chaos detection with logistic map, recognition of handwritten dynamics, and recognition of visual dynamic textures. Shows applicability to various classification tasks involving nonlinear dynamical systems and sensors.

Conclusion: Dynafit provides an effective kernel-based approach for classifying trajectory data from nonlinear dynamical systems by learning similarity metrics between dynamics, with applications spanning chaos detection, handwriting recognition, and dynamic texture analysis.

Abstract: We address the problem of classifying trajectory data generated by some nonlinear dynamics, where each class corresponds to a distinct dynamical system. We propose Dynafit, a kernel-based method for learning a distance metric between training trajectories and the underlying dynamics. New observations are assigned to the class with the most similar dynamics according to the learned metric. The learning algorithm approximates the Koopman operator which globally linearizes the dynamics in a (potentially infinite) feature space associated with a kernel function. The distance metric is computed in feature space independently of its dimensionality by using the kernel trick common in machine learning. We also show that the kernel function can be tailored to incorporate partial knowledge of the dynamics when available. Dynafit is applicable to various classification tasks involving nonlinear dynamical systems and sensors. We illustrate its effectiveness on three examples: chaos detection with the logistic map, recognition of handwritten dynamics and of visual dynamic textures.

Method: Introduces the forecast accuracy and coherence (AC) score for measuring probabilistic multi-horizon forecast quality, allowing user-specified weights to balance accuracy and consistency. Implemented as a differentiable objective function for training seasonal auto-regressive integrated models, evaluated on the M4 Hourly benchmark dataset.

Result: AC-optimized model achieved 91.1% reduced vertical variance (improved stability) relative to MLE-fitted model. Medium-to-long-horizon forecasts showed accuracy improvements up to 26% in MAPE, though one-step-ahead forecasts had 7.5% increase in MAPE. Overall produced more stable and accurate multi-step forecasts.

Conclusion: The forecast AC score successfully trains models to produce more stable and accurate multi-step forecasts by balancing accuracy and temporal consistency, with trade-off of some one-step-ahead performance degradation for improved medium-to-long-horizon forecasting.

Abstract: Traditional time series forecasting methods optimize for accuracy alone. This objective neglects temporal consistency, in other words, how consistently a model predicts the same future event as the forecast origin changes. We introduce the forecast accuracy and coherence score (forecast AC score for short) for measuring the quality of probabilistic multi-horizon forecasts in a way that accounts for both multi-horizon accuracy and stability. Our score additionally allows user-specified weights to balance accuracy and consistency requirements. As an example application, we implement the score as a differentiable objective function for training seasonal auto-regressive integrated models and evaluate it on the M4 Hourly benchmark dataset. Results demonstrate substantial improvements over traditional maximum likelihood estimation. Regarding stability, the AC-optimized model generated out-of-sample forecasts with 91.1% reduced vertical variance relative to the MLE-fitted model. In terms of accuracy, the AC-optimized model achieved considerable improvements for medium-to-long-horizon forecasts. While one-step-ahead forecasts exhibited a 7.5% increase in MAPE, all subsequent horizons experienced an improved accuracy as measured by MAPE of up to 26%. These results indicate that our metric successfully trains models to produce more stable and accurate multi-step forecasts in exchange for some degradation in one-step-ahead performance.

Method: Develops a novel parameterization for transformer models with MoE layers based on dynamical mean-field theory (DMFT) analysis, enabling reliable hyperparameter transfer across different model scales (width, depth, number of experts, expert size).

Result: The parameterization enables reliable HP transfer across models from 51M to over 2B total parameters, and allows using HPs from small models on short token horizons to train larger models on longer horizons with performant results.

Conclusion: The proposed parameterization makes hyperparameter selection cheap and reliable for MoE models, addressing the complexity introduced by sparse MoEs and enabling efficient scaling across different model dimensions.

Abstract: Mixture-of-Experts (MoE) layers have emerged as an important tool in scaling up modern neural networks by decoupling total trainable parameters from activated parameters in the forward pass for each token. However, sparse MoEs add complexity to training due to (i) new trainable parameters (router weights) that, like all other parameter groups, require hyperparameter (HP) tuning; (ii) new architecture scale dimensions (number of and size of experts) that must be chosen and potentially taken large. To make HP selection cheap and reliable, we propose a new parameterization for transformer models with MoE layers when scaling model width, depth, number of experts, and expert (hidden) size. Our parameterization is justified by a novel dynamical mean-field theory (DMFT) analysis. When varying different model dimensions trained at a fixed token budget, we find empirically that our parameterization enables reliable HP transfer across models from 51M to over 2B total parameters. We further take HPs identified from sweeping small models on a short token horizon to train larger models on longer horizons and report performant model behaviors.

Method: Use class-conditional normalizing flows as oracles that make exact posterior distributions tractable on realistic image datasets (AFHQ, ImageNet). This enables systematic investigation of scaling laws, learning limits, soft label benefits, distribution shift effects, and active learning efficiency.

Result: Prediction error decomposes into irreducible aleatoric uncertainty and reducible epistemic error; epistemic error follows power-law scaling in dataset size. Architectures differ in approaching aleatoric floor: ResNets show clean power-law scaling while Vision Transformers stall in low-data regimes. Training with exact posteriors outperforms hard labels and yields near-perfect calibration. Distribution shift type matters more than magnitude. Exact epistemic uncertainty improves active learning sample efficiency.

Conclusion: The framework reveals that standard metrics hide ongoing learning, mask architectural differences, and cannot properly diagnose distribution shift nature. Exact posterior computation enables deeper understanding of neural network performance limits and learning dynamics.

Abstract: How close are neural networks to the best they could possibly do? Standard benchmarks cannot answer this because they lack access to the true posterior p(y|x). We use class-conditional normalizing flows as oracles that make exact posteriors tractable on realistic images (AFHQ, ImageNet). This enables five lines of investigation. Scaling laws: Prediction error decomposes into irreducible aleatoric uncertainty and reducible epistemic error; the epistemic component follows a power law in dataset size, continuing to shrink even when total loss plateaus. Limits of learning: The aleatoric floor is exactly measurable, and architectures differ markedly in how they approach it: ResNets exhibit clean power-law scaling while Vision Transformers stall in low-data regimes. Soft labels: Oracle posteriors contain learnable structure beyond class labels: training with exact posteriors outperforms hard labels and yields near-perfect calibration. Distribution shift: The oracle computes exact KL divergence of controlled perturbations, revealing that shift type matters more than shift magnitude: class imbalance barely affects accuracy at divergence values where input noise causes catastrophic degradation. Active learning: Exact epistemic uncertainty distinguishes genuinely informative samples from inherently ambiguous ones, improving sample efficiency. Our framework reveals that standard metrics hide ongoing learning, mask architectural differences, and cannot diagnose the nature of distribution shift.

Method: Decoupled Diffusion Inverse Solver (DDIS) uses two components: 1) an unconditional diffusion model learns the coefficient prior distribution, and 2) a neural operator explicitly models the forward PDE for guidance. This enables Decoupled Annealing Posterior Sampling (DAPS) to avoid over-smoothing issues in standard Diffusion Posterior Sampling.

Result: DDIS achieves state-of-the-art performance under sparse observation, improving l2 error by 11% and spectral error by 54% on average. With only 1% training data, DDIS maintains accuracy with 40% advantage in l2 error compared to joint models.

Conclusion: The decoupled design enables superior data efficiency and effective physics-informed learning for inverse PDE problems, with theoretical guarantees against guidance attenuation failure when training data is scarce.

Abstract: We propose a data-efficient, physics-aware generative framework in function space for inverse PDE problems. Existing plug-and-play diffusion posterior samplers represent physics implicitly through joint coefficient-solution modeling, requiring substantial paired supervision. In contrast, our Decoupled Diffusion Inverse Solver (DDIS) employs a decoupled design: an unconditional diffusion learns the coefficient prior, while a neural operator explicitly models the forward PDE for guidance. This decoupling enables superior data efficiency and effective physics-informed learning, while naturally supporting Decoupled Annealing Posterior Sampling (DAPS) to avoid over-smoothing in Diffusion Posterior Sampling (DPS). Theoretically, we prove that DDIS avoids the guidance attenuation failure of joint models when training data is scarce. Empirically, DDIS achieves state-of-the-art performance under sparse observation, improving $l_2$ error by 11% and spectral error by 54% on average; when data is limited to 1%, DDIS maintains accuracy with 40% advantage in $l_2$ error compared to joint models.

Method: Proposes CardinalGraphFormer with: 1) Structured attention augmented with query-conditioned gated unnormalized aggregation channel to preserve dynamic cardinality signals, 2) Graph-specific structural biases, 3) Locality prior via sparse masking for scalability, 4) Pretraining combining contrastive alignment of augmented graph views and masked attribute reconstruction.

Result: Consistent gains over baselines on public benchmarks, with contributions isolated via controls for capacity, objectives, and size effects. Ablations confirm the cardinality channel’s benefits beyond simpler approximations, with efficiency improvements on large molecules.

Conclusion: CardinalGraphFormer effectively preserves cardinality information while maintaining scalability, demonstrating improved molecular property prediction for drug discovery applications with reproducible code and protocols.

Abstract: Drug discovery motivates accurate molecular property prediction when labeled data are limited and candidate spaces are vast. This article presents CardinalGraphFormer, a graph transformer that augments structured attention with a query-conditioned gated unnormalized aggregation channel to preserve dynamic cardinality signals, complemented by graph-specific structural biases; a locality prior via sparse masking provides scalability for larger graphs. For typical drug-like molecules (K = 3 is near-global), masking acts mainly as a regularizer; for larger graphs it provides meaningful efficiency gains. Pretraining unifies contrastive alignment of augmented graph views and masked reconstruction of attributes. Evaluations on public benchmarks show consistent gains over baselines, isolated via controls for capacity, objectives, and size effects. Ablations confirm the cardinality channel’s contributions beyond simpler approximations, with efficiency benefits on large molecules. Code, artifacts, and protocols emphasize reproducibility.

Method: Leverages CTI standards and community resources to enable deterministic verification of model outputs. Introduces Minerva dataset and training pipeline with task-specific verifiers. Uses RLVR with a lightweight self-training mechanism to generate additional verified trajectories.

Result: Experiments show consistent improvements in accuracy and robustness over supervised fine-tuning across multiple benchmarks and LLM backbones.

Conclusion: RLVR with verifiable rewards is effective for CTI tasks, providing more accurate and robust structured outputs than traditional supervised approaches.

Abstract: Cyber threat intelligence (CTI) analysts routinely convert noisy, unstructured security artifacts into standardized, automation-ready representations. Although large language models (LLMs) show promise for this task, existing approaches remain brittle when producing structured CTI outputs and have largely relied on supervised fine-tuning (SFT). In contrast, CTI standards and community-maintained resources define canonical identifiers and schemas that enable deterministic verification of model outputs. We leverage this structure to study reinforcement learning with verifiable rewards (RLVR) for CTI tasks. We introduce \textit{Minerva}, a unified dataset and training pipeline spanning multiple CTI subtasks, each paired with task-specific verifiers that score structured outputs and identifier predictions. To address reward sparsity during rollout, we propose a lightweight self-training mechanism that generates additional verified trajectories and distills them back into the model. Experiments across LLM backbones show consistent improvements in accuracy and robustness over SFT across multiple benchmarks.

Method: Evaluated impact of reward function structure using sparse vs dense rewards across two established cyber gyms, various network sizes, and both policy gradient and value-based RL algorithms. Used novel ground truth evaluation approach to directly compare different reward functions.

Result: Sparse rewards, when goal-aligned and frequently encountered, uniquely offer enhanced training reliability and more effective cyber defense agents with lower-risk policies. Surprisingly, sparse rewards yield policies better aligned with defender goals without explicit numerical penalties for costly actions.

Conclusion: Sparse rewards are superior to dense rewards for training autonomous cyber defense agents, providing more reliable training and policies that better align with defender objectives while minimizing risk.

Abstract: Recent years have seen an explosion of interest in autonomous cyber defence agents trained to defend computer networks using deep reinforcement learning. These agents are typically trained in cyber gym environments using dense, highly engineered reward functions which combine many penalties and incentives for a range of (un)desirable states and costly actions. Dense rewards help alleviate the challenge of exploring complex environments but risk biasing agents towards suboptimal and potentially riskier solutions, a critical issue in complex cyber environments. We thoroughly evaluate the impact of reward function structure on learning and policy behavioural characteristics using a variety of sparse and dense reward functions, two well-established cyber gyms, a range of network sizes, and both policy gradient and value-based RL algorithms. Our evaluation is enabled by a novel ground truth evaluation approach which allows directly comparing between different reward functions, illuminating the nuanced inter-relationships between rewards, action space and the risks of suboptimal policies in cyber environments. Our results show that sparse rewards, provided they are goal aligned and can be encountered frequently, uniquely offer both enhanced training reliability and more effective cyber defence agents with lower-risk policies. Surprisingly, sparse rewards can also yield policies that are better aligned with cyber defender goals and make sparing use of costly defensive actions without explicit reward-based numerical penalties.

Method: The paper provides theoretical analysis of rank’s role in linear attention, then proposes a hardware-aware structured pruning approach for query and key matrices. It adapts existing pruning strategies and introduces a novel structured pruning method based on rank-revealing QR decomposition, maintaining compatibility with existing CUDA kernels.

Result: Empirical results across models of varying sizes and downstream tasks show the framework effectively reduces state size. The method enables removal of 50% of query and key channels with only marginal increase in perplexity, yielding faster and more memory-efficient models.

Conclusion: Low-rank states in linear attention can be substantially reduced post-training with minimal performance degradation. The proposed structured pruning framework successfully reduces model size while maintaining computational efficiency, offering practical benefits for deploying linear attention models.

Abstract: Linear attention offers a computationally efficient yet expressive alternative to softmax attention. However, recent empirical results indicate that the hidden state of trained linear attention models often exhibits a low-rank structure, suggesting that these models underexploit their capacity in practice. To illuminate this phenomenon, we provide a theoretical analysis of the role of rank in linear attention, revealing that low effective rank can affect retrieval error by amplifying query noise. In addition to these theoretical insights, we conjecture that the low-rank states can be substantially reduced post-training with only minimal performance degradation, yielding faster and more memory-efficient models. To this end, we propose a novel hardware-aware approach that structurally prunes key and query matrices, reducing the state size while retaining compatibility with existing CUDA kernels. We adapt several existing pruning strategies to fit our framework and, building on our theoretical analysis, propose a novel structured pruning method based on a rank-revealing QR decomposition. Our empirical results, evaluated across models of varying sizes and on various downstream tasks, demonstrate the effectiveness of our state reduction framework. We highlight that our framework enables the removal of 50% of the query and key channels at only a marginal increase in perplexity. The code for this project can be found at https://github.com/camail-official/LinearAttentionPruning.

Method: CoSA extends compressed sensing theory to PEFT by expressing weight updates through fixed random projection matrices and a compact learnable core, rather than constraining updates to low-rank subspaces. Provides theoretical analysis showing weight updates can be compactly encoded in low-dimensional space and mapped back via random projections.

Result: Evaluated on 10 diverse tasks (natural language understanding/generation) using 5 models from RoBERTa, Llama, and Qwen families. CoSA consistently matches or outperforms state-of-the-art PEFT methods across different model scales.

Conclusion: CoSA provides a principled perspective for efficient and expressive multi-scale model adaptation, offering theoretical grounding from compressed sensing while achieving practical performance improvements over existing PEFT methods.

Abstract: Parameter-Efficient Fine-Tuning (PEFT) has emerged as a practical paradigm for adapting large language models (LLMs) without updating all parameters. Most existing approaches, such as LoRA and PiSSA, rely on low-rank decompositions of weight updates. However, the low-rank assumption may restrict expressivity, particularly in task-specific adaptation scenarios where singular values are distributed relatively uniformly. To address this limitation, we propose CoSA (Compressed Sensing-Based Adaptation), a new PEFT method extended from compressed sensing theory. Instead of constraining weight updates to a low-rank subspace, CoSA expresses them through fixed random projection matrices and a compact learnable core. We provide a formal theoretical analysis of CoSA as a synthesis process, proving that weight updates can be compactly encoded into a low-dimensional space and mapped back through random projections. Extensive experimental results show that CoSA provides a principled perspective for efficient and expressive multi-scale model adaptation. Specifically, we evaluate CoSA on 10 diverse tasks, including natural language understanding and generation, employing 5 models of different scales from RoBERTa, Llama, and Qwen families. Across these settings, CoSA consistently matches or outperforms state-of-the-art PEFT methods.

Method: Proposes Asymmetric GRAE (A-GRAE) that dynamically modulates exploration incentives and sample-difficulty focus by asymmetrically suppressing advantages of correct trajectories and implementing curriculum-like difficulty transitions.

Result: Experiments across seven benchmarks show A-GRAE consistently improves GRPO and its variants for both LLMs and MLLMs.

Conclusion: A-GRAE effectively addresses exploration and difficulty adaptation bottlenecks in RLVR methods, providing better performance for language and multimodal models.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR), particularly GRPO, has become the standard for eliciting LLM reasoning. However, its efficiency in exploration and difficulty adaptation remains an open challenge. In this work, we argue that these bottlenecks stem from an implicit advantage symmetry inherent in Group Relative Advantage Estimation (GRAE). This symmetry induces two critical limitations: (i) at the group level, strict symmetry in weights between correct and incorrect trajectories leaves unsampled action logits unchanged, thereby hindering exploration of novel correct solution. (ii) at the sample level, the algorithm implicitly prioritizes medium-difficulty samples, remaining agnostic to the non-stationary demands of difficulty focus. Through controlled experiments, we reveal that this symmetric property is sub-optimal, yielding two pivotal insights: (i) asymmetrically suppressing the advantages of correct trajectories encourages essential exploration. (ii) learning efficiency is maximized by a curriculum-like transition-prioritizing simpler samples initially before gradually shifting to complex ones. Motivated by these findings, we propose Asymmetric GRAE (A-GRAE), which dynamically modulates exploration incentives and sample-difficulty focus. Experiments across seven benchmarks demonstrate that A-GRAE consistently improves GRPO and its variants across both LLMs and MLLMs.

Method: Measure-theoretic analysis of martingale properties, including conditional expectation formulation, martingale transforms, and the upcrossing lemma, culminating in the proof of the Martingale Convergence Theorem.

Result: Establishes core martingale theory results and applies the Martingale Convergence Theorem to characterize extinction behavior in Galton-Watson branching processes.

Conclusion: Martingale theory provides powerful tools for analyzing stochastic processes, with the Martingale Convergence Theorem offering valuable insights into extinction probabilities in branching processes.

Abstract: This note investigates core properties of martingales, emphasizing the measure-theoretic formulation of conditional expectation, the martingale transform, and the upcrossing lemma. These results lead to the Martingale Convergence Theorem, which we then apply to study the extinction behavior in Galton–Watson branching processes.

Method: EvoMAS formulates MAS generation as structured configuration evolution, using feedback-conditioned mutation and crossover guided by execution traces, with iterative refinement of candidate pools and experience memory.

Result: EvoMAS consistently outperforms human-designed MAS and prior automatic methods, achieving +10.5 points on BBEH reasoning and +7.1 points on WorkBench, with 79.1% on SWE-Bench-Verified matching top leaderboard performance.

Conclusion: Evolutionary generation in configuration space provides an effective approach for automatically creating robust, executable multi-agent systems that outperform manually designed architectures.

Abstract: Large language model (LLM)-based multi-agent systems (MAS) show strong promise for complex reasoning, planning, and tool-augmented tasks, but designing effective MAS architectures remains labor-intensive, brittle, and hard to generalize. Existing automatic MAS generation methods either rely on code generation, which often leads to executability and robustness failures, or impose rigid architectural templates that limit expressiveness and adaptability. We propose Evolutionary Generation of Multi-Agent Systems (EvoMAS), which formulates MAS generation as structured configuration generation. EvoMAS performs evolutionary generation in configuration space. Specifically, EvoMAS selects initial configurations from a pool, applies feedback-conditioned mutation and crossover guided by execution traces, and iteratively refines both the candidate pool and an experience memory. We evaluate EvoMAS on diverse benchmarks, including BBEH, SWE-Bench, and WorkBench, covering reasoning, software engineering, and tool-use tasks. EvoMAS consistently improves task performance over both human-designed MAS and prior automatic MAS generation methods, while producing generated systems with higher executability and runtime robustness. EvoMAS outperforms the agent evolution method EvoAgent by +10.5 points on BBEH reasoning and +7.1 points on WorkBench. With Claude-4.5-Sonnet, EvoMAS also reaches 79.1% on SWE-Bench-Verified, matching the top of the leaderboard.

Method: Landscaper combines Hessian-based subspace construction with topological data analysis. It introduces the Saddle-Minimum Average Distance (SMAD) metric for quantifying landscape smoothness. The package analyzes loss landscapes across various architectures and tasks, including pre-trained language models.

Result: Landscaper effectively reveals geometric structures in loss landscapes, with SMAD capturing training transitions (like landscape simplification) that conventional metrics miss. SMAD also serves as a metric for out-of-distribution generalization in challenging chemical property prediction tasks.

Conclusion: Landscaper provides valuable insights for model diagnostics and architecture design, particularly in data-scarce scientific machine learning scenarios. The tool enables better understanding of neural network optimization through advanced topological analysis of loss landscapes.

Abstract: Loss landscapes are a powerful tool for understanding neural network optimization and generalization, yet traditional low-dimensional analyses often miss complex topological features. We present Landscaper, an open-source Python package for arbitrary-dimensional loss landscape analysis. Landscaper combines Hessian-based subspace construction with topological data analysis to reveal geometric structures such as basin hierarchy and connectivity. A key component is the Saddle-Minimum Average Distance (SMAD) for quantifying landscape smoothness. We demonstrate Landscaper’s effectiveness across various architectures and tasks, including those involving pre-trained language models, showing that SMAD captures training transitions, such as landscape simplification, that conventional metrics miss. We also illustrate Landscaper’s performance in challenging chemical property prediction tasks, where SMAD can serve as a metric for out-of-distribution generalization, offering valuable insights for model diagnostics and architecture design in data-scarce scientific machine learning scenarios.

Method: Isolates two key statistical properties of language: (1) decay of pairwise token correlations with time separation, and (2) decay of next-token conditional entropy with context length. Derives a simple formula using these statistics to predict scaling exponents without free parameters or synthetic data models.

Result: The theory shows remarkable match with experimentally measured neural scaling laws from training GPT-2 and LLaMA style models on TinyStories and WikiText benchmarks.

Conclusion: Provides the first theoretical framework that can quantitatively predict neural scaling law exponents for LLMs based on fundamental statistical properties of language data, bridging theory and empirical scaling laws.

Abstract: Despite the fact that experimental neural scaling laws have substantially guided empirical progress in large-scale machine learning, no existing theory can quantitatively predict the exponents of these important laws for any modern LLM trained on any natural language dataset. We provide the first such theory in the case of data-limited scaling laws. We isolate two key statistical properties of language that alone can predict neural scaling exponents: (i) the decay of pairwise token correlations with time separation between token pairs, and (ii) the decay of the next-token conditional entropy with the length of the conditioning context. We further derive a simple formula in terms of these statistics that predicts data-limited neural scaling exponents from first principles without any free parameters or synthetic data models. Our theory exhibits a remarkable match with experimentally measured neural scaling laws obtained from training GPT-2 and LLaMA style models from scratch on two qualitatively different benchmarks, TinyStories and WikiText.

Method: Introduces Causal Schrödinger Bridge (CSB) that reformulates counterfactual inference as Entropic Optimal Transport using diffusion processes (SDEs) to robustly “tunnel” through support mismatches while enforcing structural admissibility constraints. Proves Structural Decomposition Theorem showing global high-dimensional bridge factorizes into local, robust transitions.

Result: Empirical validation on high-dimensional interventions (Morpho-MNIST) demonstrates CSB significantly outperforms deterministic baselines in structural consistency, particularly in regimes of strong, out-of-distribution treatments.

Conclusion: CSB provides a robust framework for counterfactual inference that handles structural interventions better than deterministic flow methods by leveraging diffusion processes and optimal transport principles.

Abstract: Generative modeling typically seeks the path of least action via deterministic flows (ODE). While effective for in-distribution tasks, we argue that these deterministic paths become brittle under causal interventions, which often require transporting probability mass across low-density regions (“off-manifold”) where the vector field is ill-defined. This leads to numerical instability and spurious correlations. In this work, we introduce the Causal Schrödinger Bridge (CSB), a framework that reformulates counterfactual inference as Entropic Optimal Transport. Unlike deterministic approaches that require strict invertibility, CSB leverages diffusion processes (SDEs) to robustly “tunnel” through support mismatches while strictly enforcing structural admissibility constraints. We prove the Structural Decomposition Theorem, showing that the global high-dimensional bridge factorizes into local, robust transitions. Empirical validation on high-dimensional interventions (Morpho-MNIST) demonstrates that CSB significantly outperforms deterministic baselines in structural consistency, particularly in regimes of strong, out-of-distribution treatments.

Method: CDT-II uses an architecture mirroring the central dogma with three attention mechanisms: DNA self-attention for genomic relationships, RNA self-attention for gene co-regulation, and DNA-to-RNA cross-attention for transcriptional control. It processes only genomic embeddings and raw per-cell expression data.

Result: Achieves per-gene mean correlation r=0.84 for perturbation effect prediction, recovers GFI1B regulatory network with 6.6-fold enrichment (P=3.5×10^-17), and cross-attention autonomously focuses on known regulatory elements with high enrichment (201× for DNase sites, 28× for CTCF sites).

Conclusion: CDT-II establishes mechanism-oriented AI as an alternative to task-oriented approaches, enabling experimental biologists to observe regulatory networks in their own data and revealing regulatory structure rather than merely optimizing predictions.

Abstract: Current biological AI models lack interpretability – their internal representations do not correspond to biological relationships that researchers can examine. Here we present CDT-II, an “AI microscope” whose attention maps are directly interpretable as regulatory structure. By mirroring the central dogma in its architecture, CDT-II ensures that each attention mechanism corresponds to a specific biological relationship: DNA self-attention for genomic relationships, RNA self-attention for gene co-regulation, and DNA-to-RNA cross-attention for transcriptional control. Using only genomic embeddings and raw per-cell expression, CDT-II enables experimental biologists to observe regulatory networks in their own data. Applied to K562 CRISPRi data, CDT-II predicts perturbation effects (per-gene mean $r = 0.84$) and recovers the GFI1B regulatory network without supervision (6.6-fold enrichment, $P = 3.5 \times 10^{-17}$). Systematic comparison against ENCODE K562 regulatory annotations reveals that cross-attention autonomously focuses on known regulatory elements – DNase hypersensitive sites ($201\times$ enrichment), CTCF binding sites ($28\times$), and histone marks – across all five held-out genes. Two distinct attention mechanisms independently identify an overlapping RNA processing module (80% gene overlap; RNA binding enrichment $P = 1 \times 10^{-16}$). CDT-II establishes mechanism-oriented AI as an alternative to task-oriented approaches, revealing regulatory structure rather than merely optimizing predictions.

Method: Fully automated black-box pipeline that: 1) uses LLM autoraters to generate candidate bias concepts from task datasets, 2) tests each concept by generating positive/negative variations on progressively larger samples, 3) applies statistical techniques for multiple testing and early stopping, 4) flags concepts as unverbalized biases if they yield statistically significant performance differences without being cited in CoT justifications.

Result: Pipeline successfully discovered previously unknown biases in six LLMs across three decision tasks (hiring, loan approval, university admissions), including Spanish fluency, English proficiency, and writing formality. Also validated manually identified biases from prior work (gender, race, religion, ethnicity).

Conclusion: The approach provides a practical, scalable path to automatic task-specific bias discovery in LLMs, moving beyond reliance on predefined bias categories and enabling discovery of unverbalized biases hidden in chain-of-thought reasoning.

Abstract: Large Language Models (LLMs) often provide chain-of-thought (CoT) reasoning traces that appear plausible, but may hide internal biases. We call these unverbalized biases. Monitoring models via their stated reasoning is therefore unreliable, and existing bias evaluations typically require predefined categories and hand-crafted datasets. In this work, we introduce a fully automated, black-box pipeline for detecting task-specific unverbalized biases. Given a task dataset, the pipeline uses LLM autoraters to generate candidate bias concepts. It then tests each concept on progressively larger input samples by generating positive and negative variations, and applies statistical techniques for multiple testing and early stopping. A concept is flagged as an unverbalized bias if it yields statistically significant performance differences while not being cited as justification in the model’s CoTs. We evaluate our pipeline across six LLMs on three decision tasks (hiring, loan approval, and university admissions). Our technique automatically discovers previously unknown biases in these models (e.g., Spanish fluency, English proficiency, writing formality). In the same run, the pipeline also validates biases that were manually identified by prior work (gender, race, religion, ethnicity). More broadly, our proposed approach provides a practical, scalable path to automatic task-specific bias discovery.

Method: The paper formalizes different variants of Positive Distribution Shift (PDS), analyzes how certain hard classes become easily learnable under PDS, and makes connections with membership query learning. It argues that the benefit is often computational rather than statistical.

Result: The paper shows that PDS allows computationally hard problems to become tractable even using standard gradient-based training, demonstrating how certain hard classes are easily learnable under PDS.

Conclusion: Positive Distribution Shift is a valuable perspective where carefully chosen training distributions can make learning computationally easier, challenging traditional negative views of distribution shift and connecting to contemporary ML practices.

Abstract: We study a setting where the goal is to learn a target function f(x) with respect to a target distribution D(x), but training is done on i.i.d. samples from a different training distribution D’(x), labeled by the true target f(x). Such a distribution shift (here in the form of covariate shift) is usually viewed negatively, as hurting or making learning harder, and the traditional distribution shift literature is mostly concerned with limiting or avoiding this negative effect. In contrast, we argue that with a well-chosen D’(x), the shift can be positive and make learning easier – a perspective called Positive Distribution Shift (PDS). Such a perspective is central to contemporary machine learning, where much of the innovation is in finding good training distributions D’(x), rather than changing the training algorithm. We further argue that the benefit is often computational rather than statistical, and that PDS allows computationally hard problems to become tractable even using standard gradient-based training. We formalize different variants of PDS, show how certain hard classes are easily learnable under PDS, and make connections with membership query learning.

Method: Proposes RFID-MoE framework that: 1) Uses a fused metric combining expert activation frequency with effective rank to measure expert importance, adaptively allocating higher ranks to critical expert groups under fixed budget; 2) Reconstructs compression residuals via parameter-efficient sparse projection mechanism to recover lost information with minimal parameter overhead.

Result: Extensive experiments on MoE LLMs (Qwen3, DeepSeekMoE) show RFID-MoE consistently outperforms state-of-the-art methods. Achieves perplexity of 16.92 on PTB with Qwen3-30B at 60% compression ratio (8.0+ perplexity reduction vs baselines), improves zero-shot accuracy on HellaSwag by ~8%.

Conclusion: RFID-MoE effectively compresses MoE models by exploiting heterogeneous routing patterns and information density, achieving superior performance with minimal parameter overhead through adaptive rank allocation and residual reconstruction.

Abstract: Mixture-of-Experts (MoE) based Large Language Models (LLMs) have achieved superior performance, yet the massive memory overhead caused by storing multiple expert networks severely hinders their practical deployment. Singular Value Decomposition (SVD)-based compression has emerged as a promising post-training technique; however, most existing methods apply uniform rank allocation or rely solely on static weight properties. This overlooks the substantial heterogeneity in expert utilization observed in MoE models, where frequent routing patterns and intrinsic information density vary significantly across experts. In this work, we propose RFID-MoE, an effective framework for MoE compression by exploiting heterogeneous Routing Frequency and Information Density. We first introduce a fused metric that combines expert activation frequency with effective rank to measure expert importance, adaptively allocating higher ranks to critical expert groups under a fixed budget. Moreover, instead of discarding compression residuals, we reconstruct them via a parameter-efficient sparse projection mechanism to recover lost information with minimal parameter overhead. Extensive experiments on representative MoE LLMs (e.g., Qwen3, DeepSeekMoE) across multiple compression ratios demonstrate that RFID-MoE consistently outperforms state-of-the-art methods like MoBE and D2-MoE. Notably, RFID-MoE achieves a perplexity of 16.92 on PTB with the Qwen3-30B model at a 60% compression ratio, reducing perplexity by over 8.0 compared to baselines, and improves zero-shot accuracy on HellaSwag by approximately 8%.

Method: LLMRule framework incorporates large language models into an evolutionary search procedure to automatically design PB rules. It treats PB rules as algorithms for knapsack problems and uses LLMs to generate and evolve rule candidates.

Result: Evaluated on 600+ real-world PB instances from multiple countries, LLM-generated rules generally outperform existing handcrafted rules in terms of overall utility while maintaining similar fairness levels.

Conclusion: LLMs can effectively automate the design of participatory budgeting rules, achieving better utility-fairness trade-offs than manually crafted rules, demonstrating the potential of LLM-based algorithmic design.

Abstract: Participatory budgeting (PB) is a democratic paradigm for deciding the funding of public projects given the residents’ preferences, which has been adopted in numerous cities across the world. The main focus of PB is designing rules, functions that return feasible budget allocations for a set of projects subject to some budget constraint. Designing PB rules that optimize both utility and fairness objectives based on agent preferences had been challenging due to the extensive domain knowledge required and the proven trade-off between the two notions. Recently, large language models (LLMs) have been increasingly employed for automated algorithmic design. Given the resemblance of PB rules to algorithms for classical knapsack problems, in this paper, we introduce a novel framework, named LLMRule, that addresses the limitations of existing works by incorporating LLMs into an evolutionary search procedure for automating the design of PB rules. Our experimental results, evaluated on more than 600 real-world PB instances obtained from the U.S., Canada, Poland, and the Netherlands with different representations of agent preferences, demonstrate that the LLM-generated rules generally outperform existing handcrafted rules in terms of overall utility while still maintaining a similar degree of fairness.

Method: RLFR pipeline uses features as reward functions in RL, with a probing framework to identify hallucinated claims, teaching models to intervene and correct uncertain completions with scalable test-time compute.

Result: Applied to Gemma-3-12B-IT, the policy reduces hallucinations by 58% compared to original model while preserving standard benchmark performance.

Conclusion: Features can serve as scalable supervision for open-ended tasks, introducing a novel paradigm using interpretability for learning.

Abstract: Language models trained on large-scale datasets have been shown to learn features that encode abstract concepts such as factuality or intent. Such features are traditionally used for test-time monitoring or steering. We present an alternative affordance: features as scalable supervision for open-ended tasks. We consider the case of hallucination-reduction as a desirable, yet open-ended behavior and design a reinforcement learning (RL) pipeline, titled RLFR (Reinforcement Learning from Feature Rewards), that uses features as reward functions. Grounded in a novel probing framework that identifies candidate hallucinated claims, our pipeline teaches a model to intervene and correct its completions when it is uncertain of their factuality. Furthermore, the pipeline enables scalable test-time compute, guided once more by our reward features. This end-to-end process operationalized on Gemma-3-12B-IT results in a policy that is 58% less likely to hallucinate compared to the original model (when run in tandem with our probing harness), while preserving performance on standard benchmarks. Taken together, by grounding supervision in the language of features, this paper introduces a novel paradigm in the use of interpretability for learning open-ended tasks.

Method: Proposes QTALE with two key components: (1) a training strategy ensuring diverse execution paths during fine-tuning, and (2) a post-training mechanism allowing flexible adjustment of execution ratio at inference to reintroduce redundancy.

Result: QTALE enables seamless integration of token-adaptive execution with quantization, showing no noticeable accuracy difference, with gap to quantization-only models kept below 0.5% on CommonsenseQA benchmarks.

Conclusion: QTALE provides an effective solution for efficient LLM deployment by combining token-adaptive execution for FLOPs reduction and quantization for memory savings.

Abstract: Large language models (LLMs) demand substantial computational and memory resources, posing challenges for efficient deployment. Two complementary approaches have emerged to address these issues: token-adaptive layer execution, which reduces floating-point operations (FLOPs) by selectively bypassing layers, and quantization, which lowers memory footprint by reducing weight precision. However, naively integrating these techniques leads to additional accuracy degradation due to reduced redundancy in token-adaptive models. We propose QTALE (Quantization-Robust Token-Adaptive Layer Execution for LLMs), a novel framework that enables seamless integration of token-adaptive execution with quantization while preserving accuracy. Conventional token-adaptive methods reduce redundancy in two ways: (1) by limiting the diversity of training paths explored during fine-tuning, and (2) by lowering the number of parameters actively involved in inference. To overcome these limitations, QTALE introduces two key components: (1) a training strategy that ensures diverse execution paths are actively explored during fine-tuning, and (2) a post-training mechanism that allows flexible adjustment of the execution ratio at inference to reintroduce redundancy when needed. Experimental results show that QTALE enables seamless integration of token-adaptive layer execution with quantization, showing no noticeable accuracy difference, with the gap to quantization-only models kept below 0.5% on CommonsenseQA benchmarks. By combining token-adaptive execution for FLOPs reduction and quantization for memory savings, QTALE provides an effective solution for efficient LLM deployment.

Method: RLTT (Reward Latent Thought Trajectories) distributes reward across the full latent reasoning trajectory without relying on external verifiers, providing dense, trajectory-level credit assignment that can directly replace GRPO with negligible overhead.

Result: RLTT yields substantial improvements over GRPO on challenging mathematical reasoning benchmarks: +14.4% on MATH-500, +16.6% on AIME24, and +10.0% on BeyondAIME. Despite being trained exclusively on mathematics, it also transfers effectively to non-mathematical reasoning benchmarks.

Conclusion: Trajectory-level credit assignment through RLTT is an effective reinforcement learning approach for LoopLMs, significantly improving reasoning performance across mathematical and non-mathematical domains by better aligning reward assignment with the model’s internal reasoning process.

Abstract: Looped Language Models (LoopLMs) perform multi-step latent reasoning prior to token generation and outperform conventional LLMs on reasoning benchmarks at smaller parameter budgets. However, attempts to further improve LoopLM reasoning with reinforcement learning have failed - standard objectives such as Group Relative Policy Optimization (GRPO) only assign credit to the final latent state, creating a fundamental mismatch with the model’s internal computation. To resolve this, we introduce RLTT (Reward Latent Thought Trajectories), a reinforcement learning framework which distributes reward across the full latent reasoning trajectory. RLTT provides dense, trajectory-level credit assignment without relying on external verifiers and can directly replace GRPO with negligible overhead. Across extensive experiments with Ouro-2.6B-Thinking under identical training and inference conditions, RLTT yields substantial improvements over GRPO on challenging mathematical reasoning benchmarks, improving accuracy by +14.4% on MATH-500, +16.6% on AIME24, and +10.0% on BeyondAIME. Despite being trained exclusively on mathematics, RLTT also transfers effectively to non-mathematical reasoning benchmarks, demonstrating the effectiveness of trajectory-level credit assignment for reinforcement learning in LoopLMs.

Method: Proposes a flow-based model integrated with stratified sampling, using parameterized neural networks to flexibly model unknown data distributions. Combines normalizing flows with stratified sampling techniques.

Result: Shows marked reduction in estimation uncertainty across multiple datasets, including high-dimensional ones (30 and 128 dimensions). Outperforms crude Monte Carlo estimators and Gaussian mixture models.

Conclusion: The flow-based model with stratified sampling provides more flexible distribution modeling and reduces estimation uncertainty compared to traditional parametric approaches, especially for high-dimensional data.

Abstract: Estimating the expectation of a real-valued function of a random variable from sample data is a critical aspect of statistical analysis, with far-reaching implications in various applications. Current methodologies typically assume (semi-)parametric distributions such as Gaussian or mixed Gaussian, leading to significant estimation uncertainty if these assumptions do not hold. We propose a flow-based model, integrated with stratified sampling, that leverages a parametrized neural network to offer greater flexibility in modeling unknown data distributions, thereby mitigating this limitation. Our model shows a marked reduction in estimation uncertainty across multiple datasets, including high-dimensional (30 and 128) ones, outperforming crude Monte Carlo estimators and Gaussian mixture models. Reproducible code is available at https://github.com/rnoxy/flowstrat.

Method: Proposes PRISM with solver-inspired inductive bias, Write-Forget Decoupling strategy to isolate non-linearity, and two-stage proxy architecture: short-convolution anchors initial residual using local history energy, while learned predictor estimates refinement updates directly from input.

Result: Theoretically achieves Rank-L accumulation (beyond single-step Rank-1 bottleneck). Empirically achieves comparable performance to explicit optimization methods while achieving 174x higher throughput.

Conclusion: PRISM successfully distills structural patterns of iterative correction into parallelizable feedforward operators, resolving the efficiency-expressivity tension in generative sequence modeling.

Abstract: Generative sequence modeling faces a fundamental tension between the expressivity of Transformers and the efficiency of linear sequence models. Existing efficient architectures are theoretically bounded by shallow, single-step linear updates, while powerful iterative methods like Test-Time Training (TTT) break hardware parallelism due to state-dependent gradients. We propose PRISM (Parallel Residual Iterative Sequence Model) to resolve this tension. PRISM introduces a solver-inspired inductive bias that captures key structural properties of multi-step refinement in a parallelizable form. We employ a Write-Forget Decoupling strategy that isolates non-linearity within the injection operator. To bypass the serial dependency of explicit solvers, PRISM utilizes a two-stage proxy architecture: a short-convolution anchors the initial residual using local history energy, while a learned predictor estimates the refinement updates directly from the input. This design distills structural patterns associated with iterative correction into a parallelizable feedforward operator. Theoretically, we prove that this formulation achieves Rank-$L$ accumulation, structurally expanding the update manifold beyond the single-step Rank-$1$ bottleneck. Empirically, it achieves comparable performance to explicit optimization methods while achieving 174x higher throughput.

Method: Introduces Counterfactual Conditional Likelihood (CCL) rewards that score each agent’s exploration by isolating its unique contribution to team exploration. CCL emphasizes observations that are informative with respect to the joint exploration of the team rather than individual novelty.

Result: Experiments in continuous multiagent domains show CCL rewards accelerate learning for domains with sparse team rewards (where most joint actions yield zero rewards) and are particularly effective in tasks requiring tight coordination among agents.

Conclusion: CCL rewards provide a principled approach to multiagent exploration that reduces redundancy and improves coordination by focusing on each agent’s unique contribution to team exploration, especially valuable in sparse-reward coordination tasks.

Abstract: Efficient exploration is critical for multiagent systems to discover coordinated strategies, particularly in open-ended domains such as search and rescue or planetary surveying. However, when exploration is encouraged only at the individual agent level, it often leads to redundancy, as agents act without awareness of how their teammates are exploring. In this work, we introduce Counterfactual Conditional Likelihood (CCL) rewards, which score each agent’s exploration by isolating its unique contribution to team exploration. Unlike prior methods that reward agents solely for the novelty of their individual observations, CCL emphasizes observations that are informative with respect to the joint exploration of the team. Experiments in continuous multiagent domains show that CCL rewards accelerate learning for domains with sparse team rewards, where most joint actions yield zero rewards, and are particularly effective in tasks that require tight coordination among agents.

Method: Proposed FLCOA framework to conceptualize cooperation/coordination layers, then introduced Continuous Prisoner’s Dilemma with Communication Delay and conducted simulations with LLM-based agents to examine how communication delay affects cooperation patterns.

Result: As communication delay increases, agents begin to exploit slower responses even without explicit instructions. Interestingly, excessive delay reduces cycles of exploitation, creating a U-shaped relationship between delay magnitude and mutual cooperation.

Conclusion: Fostering cooperation in multi-agent systems requires attention not only to high-level institutional design but also to lower-layer factors like communication delay and resource allocation, pointing to new research directions for MAS.

Abstract: LLM-based multi-agent systems (LLM-MAS), in which autonomous AI agents cooperate to solve tasks, are gaining increasing attention. For such systems to be deployed in society, agents must be able to establish cooperation and coordination under real-world computational and communication constraints. We propose the FLCOA framework (Five Layers for Cooperation/Coordination among Autonomous Agents) to conceptualize how cooperation and coordination emerge in groups of autonomous agents, and highlight that the influence of lower-layer factors - especially computational and communication resources - has been largely overlooked. To examine the effect of communication delay, we introduce a Continuous Prisoner’s Dilemma with Communication Delay and conduct simulations with LLM-based agents. As delay increases, agents begin to exploit slower responses even without explicit instructions. Interestingly, excessive delay reduces cycles of exploitation, yielding a U-shaped relationship between delay magnitude and mutual cooperation. These results suggest that fostering cooperation requires attention not only to high-level institutional design but also to lower-layer factors such as communication delay and resource allocation, pointing to new directions for MAS research.

Method: Game-theoretic analysis of perimeter defense on a cylindrical geometry, examining defender-attacker dynamics with speed asymmetry and specific starting conditions.

Result: Derived conditions necessary for the attacker to win when starting close to the boundary in a currently defended region.

Conclusion: Identified critical conditions for attacker success in perimeter defense games with speed asymmetry, providing insights for defensive strategy design.

Abstract: We describe a multi-agent perimeter defense game played on a cylinder. A team of n slow-moving defenders must prevent a single fast-moving attacker from crossing the boundary of a defensive perimeter. We describe the conditions necessary for the attacker to win in the special case that the intruder starts close to the boundary and in a region that is currently defended.

Method: Developed a scalable, differentiable agent-based model (DABM) that captures individual-level heterogeneity in health status, behavior, and resource constraints, while modeling epidemic processes like viral mutation and reinfection. Introduced a z-score-based scaling method to map small-scale simulations to real-world population sizes with minimal loss in output granularity.

Result: DEpiABS reduced average normal deviation in forecasting from 0.97 to 0.92 on COVID-19 mortality data and from 0.41 to 0.32 on influenza-like-illness data across ten regions of varying scales, achieving improvements without relying on auxiliary data.

Conclusion: DEpiABS provides a reliable, generalizable, and data-efficient framework for epidemic response modeling that balances mechanistic detail, computational efficiency, and interpretability for future epidemic forecasting.

Abstract: The COVID-19 pandemic highlighted the limitations of existing epidemic simulation tools. These tools provide information that guides non-pharmaceutical interventions (NPIs), yet many struggle to capture complex dynamics while remaining computationally practical and interpretable. We introduce DEpiABS, a scalable, differentiable agent-based model (DABM) that balances mechanistic detail, computational efficiency and interpretability. DEpiABS captures individual-level heterogeneity in health status, behaviour, and resource constraints, while also modelling epidemic processes like viral mutation and reinfection dynamics. The model is fully differentiable, enabling fast simulation and gradient-based parameter calibration. Building on this foundation, we introduce a z-score-based scaling method that maps small-scale simulations to any real-world population sizes with negligible loss in output granularity, reducing the computational burden when modelling large populations. We validate DEpiABS through sensitivity analysis and calibration to COVID-19 and flu data from ten regions of varying scales. Compared to the baseline, DEpiABS is more detailed, fully interpretable, and has reduced the average normal deviation in forecasting from 0.97 to 0.92 on COVID-19 mortality data and from 0.41 to 0.32 on influenza-like-illness data. Critically, these improvements are achieved without relying on auxiliary data, making DEpiABS a reliable, generalisable, and data-efficient framework for future epidemic response modelling.

Method: Uses sparse variational inference to create local pseudo-representations, with sparse variational optimization bounding local pseudo-datasets, and global scaled proximal-inexact consensus ADMM with adaptive parameter updates and warm-start initialization.

Result: pxpGP and its decentralized variant dec-pxpGP outperform existing distributed GP methods in hyperparameter estimation and prediction accuracy, especially in large-scale networks.

Conclusion: The proposed framework enables efficient distributed probabilistic modeling for large-scale multi-robot systems, addressing computational complexity limitations of traditional GPs.

Abstract: Multi-robot systems require scalable and federated methods to model complex environments under computational and communication constraints. Gaussian Processes (GPs) offer robust probabilistic modeling, but suffer from cubic computational complexity, limiting their applicability in large-scale deployments. To address this challenge, we introduce the pxpGP, a novel distributed GP framework tailored for both centralized and decentralized large-scale multi-robot networks. Our approach leverages sparse variational inference to generate a local compact pseudo-representation. We introduce a sparse variational optimization scheme that bounds local pseudo-datasets and formulate a global scaled proximal-inexact consensus alternating direction method of multipliers (ADMM) with adaptive parameter updates and warm-start initialization. Experiments on synthetic and real-world datasets demonstrate that pxpGP and its decentralized variant, dec-pxpGP, outperform existing distributed GP methods in hyperparameter estimation and prediction accuracy, particularly in large-scale networks.

Method: Created an agent-based model with evolvable affect in a shared resource dilemma, comparing competition (non-normative) versus punishment (normative) conditions to study emergent mood mechanisms.

Result: Under competition, agents evolved “bad mood -> consume more” leading to resource depletion and population collapse. Under punishment, agents evolved “bad mood -> consume less” where negative affect functions as implicit social sanction signal, promoting resource conservation.

Conclusion: Normative processes enable social preferences to emerge in distributed psychological mechanisms, reprogramming cognitive and physiological systems by embedding cultural patterns into psychological dispositions through evolutionary pathways.

Abstract: Breaking a norm elicits both material and emotional consequences, yet how this coupling arose evolutionarily remains unclear. We investigate this question in light of emerging work suggesting that normativity’s building blocks emerged earlier in evolution than previously considered, arguing that normative processes should inform accounts of how even ancient capacities such as mood evolved. Using a definition of normative processes we developed, we created an agent-based model with evolvable affect in a shared resource dilemma, comparing competition (non-normative) versus punishment (normative) conditions. Critically, different mood mechanisms emerge under each condition. Under competition, agents evolve a “bad mood -> consume more” response, creating a tragedy of the commons leading to resource depletion and population collapse. Under punishment, agents evolve a “bad mood -> consume less” mechanism, where negative affect functions as an implicit signal of social sanction, promoting resource conservation. Importantly, once normative logic is imprinted through punishment, it creates an evolutionary pathway for mood-based signalling that operates without costly physical enforcement. Our findings demonstrate how normative processes enable social preferences to emerge in a distributed manner within psychological mechanisms, showing how normative processes reprogram cognitive and physiological systems by embedding cultural patterns into psychological dispositions.

Method: Proposes a coordination framework with three core mechanisms: dynamic task routing (reallocation based on confidence/workload), bidirectional feedback (structured critiques for iterative improvement), and parallel agent evaluation (competition on ambiguous subtasks with evaluator-driven selection). Implemented in a modular architecture.

Result: Demonstrates substantial improvements in factual coverage, coherence, and efficiency over static and partially adaptive baselines in collaborative task completion scenarios.

Conclusion: Incorporating both adaptiveness and structured competition in multi-agent LLM systems provides significant benefits for handling complex, open-ended tasks through dynamic coordination mechanisms.

Abstract: Large language model (LLM) agents have shown increasing promise for collaborative task completion. However, existing multi-agent frameworks often rely on static workflows, fixed roles, and limited inter-agent communication, reducing their effectiveness in open-ended, high-complexity domains. This paper proposes a coordination framework that enables adaptiveness through three core mechanisms: dynamic task routing, bidirectional feedback, and parallel agent evaluation. The framework allows agents to reallocate tasks based on confidence and workload, exchange structured critiques to iteratively improve outputs, and crucially compete on high-ambiguity subtasks with evaluator-driven selection of the most suitable result. We instantiate these principles in a modular architecture and demonstrate substantial improvements in factual coverage, coherence, and efficiency over static and partially adaptive baselines. Our findings highlight the benefits of incorporating both adaptiveness and structured competition in multi-agent LLM systems.

Method: Proposes a negotiation-based framework that offers drivers incentive-backed options for modest flexibility in departure time or requested state of charge. The framework guarantees voluntary participation, strategy-proofness, and budget feasibility, and is calibrated with user survey data and validated using real operational data.

Result: Simulations show the negotiation protocol lowers building operator’s costs by over 3.5% compared to optimized non-negotiating smart charging, while reducing user charging expenses by 22% below utility retail rates.

Conclusion: The framework transforms EV charging from operational friction into a platform for collaboration and shared savings, providing a strategic bridge between energy and mobility systems by aligning operator and EV user objectives.

Abstract: The growth of Electric Vehicles (EVs) creates a conflict in vehicle-to-building (V2B) settings between building operators, who face high energy costs from uncoordinated charging, and drivers, who prioritize convenience and a full charge. To resolve this, we propose a negotiation-based framework that, by design, guarantees voluntary participation, strategy-proofness, and budget feasibility. It transforms EV charging into a strategic resource by offering drivers a range of incentive-backed options for modest flexibility in their departure time or requested state of charge (SoC). Our framework is calibrated with user survey data and validated using real operational data from a commercial building and an EV manufacturer. Simulations show that our negotiation protocol creates a mutually beneficial outcome: lowering the building operator’s costs by over 3.5% compared to an optimized, non-negotiating smart charging policy, while simultaneously reducing user charging expenses by 22% below the utility’s retail energy rate. By aligning operator and EV user objectives, our framework provides a strategic bridge between energy and mobility systems, transforming EV charging from a source of operational friction into a platform for collaboration and shared savings.

Method: Programmatic data collection from Moltbook platform (150k+ agents), analysis of 12,758 agent-authored sub-community descriptions using preprocessing, contextual embeddings, and unsupervised clustering to uncover latent thematic patterns.

Result: Autonomous agents systematically organize collective space through reproducible patterns including human-mimetic interests, silicon-centric self-reflection, and early-stage economic/coordination behaviors, emerging directly from machine-generated data.

Conclusion: Establishes data-driven silicon sociology as a methodological framework for studying large autonomous agent societies, demonstrating data mining’s power for understanding emergent social organization.

Abstract: The rapid emergence of autonomous large language model agents has given rise to persistent, large-scale agent ecosystems whose collective behavior cannot be adequately understood through anecdotal observation or small-scale simulation. This paper introduces data-driven silicon sociology as a systematic empirical framework for studying social structure formation among interacting artificial agents. We present a pioneering large-scale data mining investigation of an in-the-wild agent society by analyzing Moltbook, a social platform designed primarily for agent-to-agent interaction. At the time of study, Moltbook hosted over 150,000 registered autonomous agents operating across thousands of agent-created sub-communities. Using programmatic and non-intrusive data acquisition, we collected and analyzed the textual descriptions of 12,758 submolts, which represent proactive sub-community partitioning activities within the ecosystem. Treating agent-authored descriptions as first-class observational artifacts, we apply rigorous preprocessing, contextual embedding, and unsupervised clustering techniques to uncover latent patterns of thematic organization and social space structuring. The results show that autonomous agents systematically organize collective space through reproducible patterns spanning human-mimetic interests, silicon-centric self-reflection, and early-stage economic and coordination behaviors. Rather than relying on predefined sociological taxonomies, these structures emerge directly from machine-generated data traces. This work establishes a methodological foundation for data-driven silicon sociology and demonstrates that data mining techniques can provide a powerful lens for understanding the organization and evolution of large autonomous agent societies.

Method: Hierarchical subspace latent diffusion model that directly maps visual lip movements to continuous latent space of pre-trained neural audio codec. Uses subspace decomposition module, diffusion convolution blocks (DiCB) for intra/inter-subspace interactions, and reparameterized flow matching with SLM and semantic losses.

Result: Achieves state-of-the-art generation quality on multiple benchmark datasets, surpassing existing methods in both objective and subjective evaluations.

Conclusion: SLD-L2S demonstrates the effectiveness of latent diffusion models for lip-to-speech synthesis, avoiding information loss from intermediate representations and enabling principled integration of semantic losses for improved speech quality.

Abstract: Although lip-to-speech synthesis (L2S) has achieved significant progress in recent years, current state-of-the-art methods typically rely on intermediate representations such as mel-spectrograms or discrete self-supervised learning (SSL) tokens. The potential of latent diffusion models (LDMs) in this task remains largely unexplored. In this paper, we introduce SLD-L2S, a novel L2S framework built upon a hierarchical subspace latent diffusion model. Our method aims to directly map visual lip movements to the continuous latent space of a pre-trained neural audio codec, thereby avoiding the information loss inherent in traditional intermediate representations. The core of our method is a hierarchical architecture that processes visual representations through multiple parallel subspaces, initiated by a subspace decomposition module. To efficiently enhance interactions within and between these subspaces, we design the diffusion convolution block (DiCB) as our network backbone. Furthermore, we employ a reparameterized flow matching technique to directly generate the target latent vectors. This enables a principled inclusion of speech language model (SLM) and semantic losses during training, moving beyond conventional flow matching objectives and improving synthesized speech quality. Our experiments show that SLD-L2S achieves state-of-the-art generation quality on multiple benchmark datasets, surpassing existing methods in both objective and subjective evaluations.

Method: Proposes Trans-Chunk BiMamba (TC-BiMamba) with Trans-Chunk mechanism that trains bidirectional sequences in offline style with dynamic chunk size, enabling simultaneous training of streaming and non-streaming capabilities.

Result: TC-BiMamba achieves 1.3x training speedup, 50% memory reduction, and performance improvements over traditional chunk-wise processing. It outperforms U2++ and matches LC-BiMamba with smaller model size.

Conclusion: TC-BiMamba provides an efficient solution for unified streaming/non-streaming ASR with dynamic chunk size training, offering better training efficiency, memory usage, and performance than existing methods.

Abstract: This work investigates bidirectional Mamba (BiMamba) for unified streaming and non-streaming automatic speech recognition (ASR). Dynamic chunk size training enables a single model for offline decoding and streaming decoding with various latency settings. In contrast, existing BiMamba based streaming method is limited to fixed chunk size decoding. When dynamic chunk size training is applied, training overhead increases substantially. To tackle this issue, we propose the Trans-Chunk BiMamba (TC-BiMamba) for dynamic chunk size training. Trans-Chunk mechanism trains both bidirectional sequences in an offline style with dynamic chunk size. On the one hand, compared to traditional chunk-wise processing, TC-BiMamba simultaneously achieves 1.3 times training speedup, reduces training memory by 50%, and improves model performance since it can capture bidirectional context. On the other hand, experimental results show that TC-BiMamba outperforms U2++ and matches LC-BiMmaba with smaller model size.

Method: Proposes a frequency-domain Conformer-based architecture that examines different architectural choices (per-frequency MLPs, convolutional, dilated convolutional, and attention-based models) to jointly capture local spectral continuity and long-range frequency correlations for HRTF upsampling from sparse measurements.

Result: Achieves state-of-the-art performance on SONICOM and HUTUBS datasets in terms of interaural level difference and log-spectral distortion, with explicit spectral modeling consistently improving reconstruction accuracy, particularly under severe sparsity conditions.

Conclusion: Explicit modeling of spectral dependencies in the frequency domain is crucial for accurate HRTF upsampling from sparse measurements, and the proposed Conformer-based architecture effectively captures both local continuity and long-range frequency correlations for improved spatial audio rendering.

Abstract: Accurate upsampling of Head-Related Transfer Functions (HRTFs) from sparse measurements is crucial for personalized spatial audio rendering. Traditional interpolation methods, such as kernel-based weighting or basis function expansions, rely on measurements from a single subject and are limited by the spatial sampling theorem, resulting in significant performance degradation under sparse sampling. Recent learning-based methods alleviate this limitation by leveraging cross-subject information, yet most existing neural architectures primarily focus on modeling spatial relationships across directions, while spectral dependencies along the frequency dimension are often modeled implicitly or treated independently. However, HRTF magnitude responses exhibit strong local continuity and long-range structure in the frequency domain, which are not fully exploited. This work investigates frequency-domain feature modeling by examining how different architectural choices, ranging from per-frequency multilayer perceptrons to convolutional, dilated convolutional, and attention-based models, affect performance under varying sparsity levels, showing that explicit spectral modeling consistently improves reconstruction accuracy, particularly under severe sparsity. Motivated by this observation, a frequency-domain Conformer-based architecture is adopted to jointly capture local spectral continuity and long-range frequency correlations. Experimental results on the SONICOM and HUTUBS datasets demonstrate that the proposed method achieves state-of-the-art performance in terms of interaural level difference and log-spectral distortion.

Method: Proposes CU block-structure gradient maps to describe coding-unit partitioning changes, combined with block-level mapping of intra prediction modes. Designs GradIPMFormer, a Transformer network specifically tailored for CU-block steganalysis to better perceive CU-level steganographic behaviors.

Result: The method achieves superior detection performance across multiple H.265/HEVC steganographic algorithms under different quantization parameters and resolution settings, validating the effectiveness of coding-structure perspective analysis.

Conclusion: Provides a new CU block-level analysis paradigm for H.265/HEVC video steganalysis with significant value for covert communication security detection, demonstrating the feasibility of coding-structure perspective analysis.

Abstract: Existing H.265/HEVC video steganalysis research mainly focuses on statistical feature modeling at the levels of motion vectors (MV), intra prediction modes (IPM), or transform coefficients. In contrast, studies targeting the coding-structure level - especially the analysis of block-level steganographic behaviors in Coding Units (CUs) - remain at an early stage. As a core component of H.265/HEVC coding decisions, the CU partition structure often exhibits steganographic perturbations in the form of structural changes and reorganization of prediction relationships, which are difficult to characterize effectively using traditional pixel-domain features or mode statistics. To address this issue, this paper, for the first time from the perspective of CU block-level steganalysis, proposes an H.265/HEVC video steganalysis method based on CU block-structure gradients and intra prediction mode mapping. The proposed method constructs a CU block-structure gradient map to explicitly describe changes in coding-unit partitioning, and combines it with a block-level mapping representation of IPM to jointly model the structural perturbations introduced by CU-level steganographic embedding. On this basis, we design a Transformer network, GradIPMFormer, tailored for CU-block steganalysis, thereby effectively enhancing the capability to perceive CU-level steganographic behaviors. Experimental results show that under different quantization parameters and resolution settings, the proposed method consistently achieves superior detection performance across multiple H.265/HEVC steganographic algorithms, validating the feasibility and effectiveness of conducting video steganalysis from the coding-structure perspective. This study provides a new CU block-level analysis paradigm for H.265/HEVC video steganalysis and has significant research value for covert communication security detection.

Method: Multi-task learning framework that uses full-reference metrics as supervisory signals to learn perceptually meaningful features without human labels for pretraining. Jointly optimizes multiple FR objectives with adaptive task weighting to learn shared representations that transfer effectively to NR-VQA.

Result: MTL-VQA achieves performance competitive with state-of-the-art NR-VQA methods on gaming video datasets across both MOS-supervised and label-efficient/self-supervised settings.

Conclusion: The multi-task learning approach using FR metrics as supervision enables effective NR-VQA for gaming videos without requiring extensive human-labeled data, addressing the challenges of limited datasets and unique gaming content characteristics.

Abstract: No-reference video quality assessment (NR-VQA) for gaming videos is challenging due to limited human-rated datasets and unique content characteristics including fast motion, stylized graphics, and compression artifacts. We present MTL-VQA, a multi-task learning framework that uses full-reference metrics as supervisory signals to learn perceptually meaningful features without human labels for pretraining. By jointly optimizing multiple full-reference (FR) objectives with adaptive task weighting, our approach learns shared representations that transfer effectively to NR-VQA. Experiments on gaming video datasets show MTL-VQA achieves performance competitive with state-of-the-art NR-VQA methods across both MOS-supervised and label-efficient/self-supervised settings.

Method: Two-stage coarse-to-fine alignment: 1) discrepancy-aware coarse-grained alignment with quality-discrepancy-aware hybrid loss to capture relative quality relationships, 2) perception fusion fine-grained alignment with symmetric feature fusion and conditional discriminator to identify domain-invariant features

Result: Extensive experiments demonstrate UPDA effectively enhances NR-PCQA performance in cross-domain scenarios, validating practical applicability

Conclusion: Proposed UPDA framework successfully addresses domain adaptation challenges for NR-PCQA through progressive alignment approach

Abstract: While no-reference point cloud quality assessment (NR-PCQA) approaches have achieved significant progress over the past decade, their performance often degrades substantially when a distribution gap exists between the training (source domain) and testing (target domain) data. However, to date, limited attention has been paid to transferring NR-PCQA models across domains. To address this challenge, we propose the first unsupervised progressive domain adaptation (UPDA) framework for NR-PCQA, which introduces a two-stage coarse-to-fine alignment paradigm to address domain shifts. At the coarse-grained stage, a discrepancy-aware coarse-grained alignment method is designed to capture relative quality relationships between cross-domain samples through a novel quality-discrepancy-aware hybrid loss, circumventing the challenges of direct absolute feature alignment. At the fine-grained stage, a perception fusion fine-grained alignment approach with symmetric feature fusion is developed to identify domain-invariant features, while a conditional discriminator selectively enhances the transfer of quality-relevant features. Extensive experiments demonstrate that the proposed UPDA effectively enhances the performance of NR-PCQA methods in cross-domain scenarios, validating its practical applicability. The code is available at https://github.com/yokeno1/UPDA-main.

Method: BSoNet is a novel deep learning-based approach specifically designed to optimize PBI image quality by enhancing clarity, recognition, and contrast while meeting practical application requirements.

Result: The approach significantly improves image quality, transforming PBI systems into more effective and reliable inspection tools for security applications.

Conclusion: BSoNet successfully addresses the image quality challenges in portable backscatter imaging, contributing to stronger security protection through enhanced inspection capabilities.

Abstract: Portable backscatter imaging systems (PBI) integrate an X-ray source and detector in a single unit, utilizing Compton scattering photons to rapidly acquire superficial or shallow structural information of an inspected object through single-sided imaging. The application of this technology overcomes the limitations of traditional transmission X-ray detection, offering greater flexibility and portability, making it the preferred tool for the rapid and accurate identification of potential threats in scenarios such as borders, ports, and industrial nondestructive security inspections. However, the image quality is significantly compromised due to the limited number of Compton backscattered photons. The insufficient photon counts result primarily from photon absorption in materials, the pencil-beam scanning design, and short signal sampling times. It therefore yields severe image noise and an extremely low signal-to-noise ratio, greatly reducing the accuracy and reliability of PBI systems. To address these challenges, this paper introduces BSoNet, a novel deep learning-based approach specifically designed to optimize the image quality of PBI systems. The approach significantly enhances image clarity, recognition, and contrast while meeting practical application requirements. It transforms PBI systems into more effective and reliable inspection tools, contributing significantly to strengthening security protection.

Method: Extends amortized variational inference by injecting spatially adaptive perturbations to measurements during training, guided by uncertainty estimates, to emphasize learning in the most uncertain regions for better reconstruction quality.

Result: Superior or competitive performance to previous diffusion-based approaches on deblurring and super-resolution tasks, delivering more realistic reconstructions without the computational cost of iterative refinement.

Conclusion: The proposed uncertainty-guided perturbation approach effectively improves amortized variational inference for image reconstruction, achieving high-quality results with computational efficiency.

Abstract: Ill-posed imaging inverse problems remain challenging due to the ambiguity in mapping degraded observations to clean images. Diffusion-based generative priors have recently shown promise, but typically rely on computationally intensive iterative sampling or per-instance optimization. Amortized variational inference frameworks address this inefficiency by learning a direct mapping from measurements to posteriors, enabling fast posterior sampling without requiring the optimization of a new posterior for every new set of measurements. However, they still struggle to reconstruct fine details and complex textures. To address this, we extend the amortized framework by injecting spatially adaptive perturbations to measurements during training, guided by uncertainty estimates, to emphasize learning in the most uncertain regions. Experiments on deblurring and super-resolution demonstrate that our method achieves superior or competitive performance to previous diffusion-based approaches, delivering more realistic reconstructions without the computational cost of iterative refinement.

Method: Statistical shape model built from retrospective data, optimized 2D slice placements, and differentiable volumetric mesh optimization to generate personalized aortic meshes from cine 2D MRI.

Result: Accurate 3D geometries obtained from as few as six 2D slices with high Dice scores (89.9%), low Hausdorff distance (7.3mm), and significant age-related differences in radial strain observed.

Conclusion: The method enables efficient extraction of time-resolved 3D aortic meshes from limited 2D MRI for computational shape and strain analysis.

Abstract: Background and Objective: We propose a shape reconstruction framework to generate time-resolved, patient-specific 3D aortic geometries from a limited number of standard cine 2D magnetic resonance imaging (MRI) acquisitions. A statistical shape model of the aorta is coupled with differentiable volumetric mesh optimization to obtain personalized aortic meshes. Methods: The statistical shape model was constructed from retrospective data and optimized 2D slice placements along the aortic arch were identified. Cine 2D MRI slices were then acquired in 30 subjects (19 volunteers, 11 aortic stenosis patients). After manual segmentation, time-resolved aortic models were generated via differentiable volumetric mesh optimization to derive vessel shape features, centerline parameters, and radial wall strain. In 10 subjects, additional 4D flow MRI was acquired to compare peak-systolic shapes. Results: Anatomically accurate aortic geometries were obtained from as few as six cine 2D MRI slices, achieving a mean +/- standard deviation Dice score of (89.9 +/- 1.6) %, Intersection over Union of (81.7 +/- 2.7) %, Hausdorff distance of (7.3 +/- 3.3) mm, and Chamfer distance of (3.7 +/- 0.6) mm relative to 4D flow MRI. The mean absolute radius error was (0.8 +/- 0.6) mm. Significant age-related differences were observed for all shape features, including radial strain, which decreased progressively ((11.00 +/- 3.11) x 10-2 vs. (3.74 +/- 1.25) x 10-2 vs. (2.89 +/- 0.87) x 10-2 for young, mid-age, and elderly groups). Conclusion: The proposed method enables efficient extraction of time-resolved 3D aortic meshes from limited sets of standard cine 2D MRI acquisitions, suitable for computational shape and strain analysis.

Method: Proposes LN-DDPM, a conditional denoising diffusion probabilistic model that uses lymph node masks and anatomical structure masks as conditions through global structure and local detail conditioning mechanisms. Generated data is then used to train nnU-Net for segmentation.

Result: LN-DDPM outperforms other generative methods in abdominal lymph node image synthesis and better assists downstream segmentation tasks on abdominal lymph node datasets.

Conclusion: The pipeline integrating conditional diffusion models for data generation with nnU-Net for segmentation effectively addresses data scarcity in medical imaging and improves abdominal lymph node segmentation performance.

Abstract: Despite the significant success achieved by deep learning methods in medical image segmentation, researchers still struggle in the computer-aided diagnosis of abdominal lymph nodes due to the complex abdominal environment, small and indistinguishable lesions, and limited annotated data. To address these problems, we present a pipeline that integrates the conditional diffusion model for lymph node generation and the nnU-Net model for lymph node segmentation to improve the segmentation performance of abdominal lymph nodes through synthesizing a diversity of realistic abdominal lymph node data. We propose LN-DDPM, a conditional denoising diffusion probabilistic model (DDPM) for lymph node (LN) generation. LN-DDPM utilizes lymph node masks and anatomical structure masks as model conditions. These conditions work in two conditioning mechanisms: global structure conditioning and local detail conditioning, to distinguish between lymph nodes and their surroundings and better capture lymph node characteristics. The obtained paired abdominal lymph node images and masks are used for the downstream segmentation task. Experimental results on the abdominal lymph node datasets demonstrate that LN-DDPM outperforms other generative methods in the abdominal lymph node image synthesis and better assists the downstream abdominal lymph node segmentation task.

Method: Proposes a framework with Informed Latent Space (ILS) and Weight Optimization Loss (WOL) using conditional generative modeling. Applies Brownian Bridge Diffusion Model (BBDM) and develops its latent counterpart LBBDM with tailored training strategies for pressure synthesis in lying postures.

Result: BBDM with ILS produces highly detailed pressure maps but at higher computational cost and inference time, while LBBDM offers faster inference with competitive performance. Both improve physical plausibility and performance over baselines.

Conclusion: The approach enables non-invasive, vision-based, real-time patient monitoring in clinical environments by generating physically consistent pressure estimates for hospital bed applications.

Abstract: Monitoring contact pressure in hospital beds is essential for preventing pressure ulcers and enabling real-time patient assessment. Current methods can predict pressure maps but often lack physical plausibility, limiting clinical reliability. This work proposes a framework that enhances plausibility via Informed Latent Space (ILS) and Weight Optimization Loss (WOL) with conditional generative modeling to produce high-fidelity, physically consistent pressure estimates. This study also applies diffusion based conditional Brownian Bridge Diffusion Model (BBDM) and proposes training strategy for its latent counterpart Latent Brownian Bridge Diffusion Model (LBBDM) tailored for pressure synthesis in lying postures. Experiment results shows proposed method improves physical plausibility and performance over baselines: BBDM with ILS delivers highly detailed maps at higher computational cost and large inference time, whereas LBBDM provides faster inference with competitive performance. Overall, the approach supports non-invasive, vision-based, real-time patient monitoring in clinical environments.

Method: Proposes SPUR-iG, a fully 3D deep unrolled subspace reconstruction framework with: 1) Efficient data consistency using implicit GROG (gridding non-Cartesian data onto Cartesian grid with learned kernel for FFT-based updates), and 2) Progressive training strategy in three stages: pretraining denoiser with data augmentation, greedy per-iteration unrolled training, and final fine-tuning with gradient checkpointing.

Result: On large in vivo dataset with retrospective undersampling, SPUR-iG improves subspace coefficient maps quality and quantitative accuracy at 1-mm isotropic resolution compared to LLR and hybrid 2D/3D unrolled baseline. Whole-brain reconstructions complete in under 15 seconds with up to ×111 speedup for 2-minute acquisitions. T₁ maps from 30-second scans achieve accuracy on par with or exceeding LLR reconstructions from 2-minute scans.

Conclusion: The framework improves both accuracy and speed in large-scale 3D MRF reconstruction, enabling efficient and reliable accelerated quantitative imaging through a combination of implicit GROG for efficient data consistency and progressive training strategy for feasible large-scale 3D unrolled learning.

Abstract: Magnetic Resonance Fingerprinting (MRF) enables fast quantitative imaging, yet reconstructing high-resolution 3D data remains computationally demanding. Non-Cartesian reconstructions require repeated non-uniform FFTs, and the commonly used Locally Low Rank (LLR) prior adds computational overhead and becomes insufficient at high accelerations. Learned 3D priors could address these limitations, but training them at scale is challenging due to memory and runtime demands. We propose SPUR-iG, a fully 3D deep unrolled subspace reconstruction framework that integrates efficient data consistency with a progressive training strategy. Data consistency leverages implicit GROG, which grids non-Cartesian data onto a Cartesian grid with an implicitly learned kernel, enabling FFT-based updates with minimal artifacts. Training proceeds in three stages: (1) pretraining a denoiser with extensive data augmentation, (2) greedy per-iteration unrolled training, and (3) final fine-tuning with gradient checkpointing. Together, these stages make large-scale 3D unrolled learning feasible within a reasonable compute budget. On a large in vivo dataset with retrospective undersampling, SPUR-iG improves subspace coefficient maps quality and quantitative accuracy at 1-mm isotropic resolution compared with LLR and a hybrid 2D/3D unrolled baseline. Whole-brain reconstructions complete in under 15-seconds, with up to $\times$111 speedup for 2-minute acquisitions. Notably, $T_1$ maps with our method from 30-second scans achieve accuracy on par with or exceeding LLR reconstructions from 2-minute scans. Overall, the framework improves both accuracy and speed in large-scale 3D MRF reconstruction, enabling efficient and reliable accelerated quantitative imaging.

Method: Two-stage approach: 1) Optimal Cross-section Matching (OCM) algorithm for constrained global alignment (translation, rotation, scaling), 2) lightweight deep-learning refinement network (inspired by VoxelMorph) for residual local deformations with hierarchical decomposition of registration manifold.

Result: Experiments on 40 kidneys dataset showed better results compared to single-stage baselines, maintaining physical calibration via Hough-based grid detection and Bezier-based contour smoothing for robust meshing and volume estimation.

Conclusion: The hybrid OCM+DL framework integrates explicit geometric priors with neural network flexibility, ensuring stable optimization and plausible deformation fields with few training examples, advancing precision and anatomical realism for multimodal 3D reconstructions in medical applications.

Abstract: This paper introduces a hybrid two-stage registration framework for reconstructing three-dimensional (3D) kidney anatomy from macroscopic slices, using CT-derived models as the geometric reference standard. The approach addresses the data-scarcity and high-distortion challenges typical of macroscopic imaging, where fully learning-based registration (e.g., VoxelMorph) often fails to generalize due to limited training diversity and large nonrigid deformations that exceed the capture range of unconstrained convolutional filters. In the proposed pipeline, the Optimal Cross-section Matching (OCM) algorithm first performs constrained global alignment: translation, rotation, and uniform scaling to establish anatomically consistent slice initialization. Next, a lightweight deep-learning refinement network, inspired by VoxelMorph, predicts residual local deformations between consecutive slices. The core novelty of this architecture lies in its hierarchical decomposition of the registration manifold. This hybrid OCM+DL design integrates explicit geometric priors with the flexible learning capacity of neural networks, ensuring stable optimization and plausible deformation fields even with few training examples. Experiments on an original dataset of 40 kidneys demonstrated better results compared to single-stage baselines. The pipeline maintains physical calibration via Hough-based grid detection and employs Bezier-based contour smoothing for robust meshing and volume estimation. Although validated on kidney data, the proposed framework generalizes to other soft-tissue organs reconstructed from optical or photographic cross-sections. By decoupling interpretable global optimization from data-efficient deep refinement, the method advances the precision, reproducibility, and anatomical realism of multimodal 3D reconstructions for surgical planning, morphological assessment, and medical education.