Method: Proposes AdaptMMBench benchmark across five domains (real-world, OCR, GUI, knowledge, math) with Matthews Correlation Coefficient metric to evaluate selection rationality of reasoning modes, dynamically identifying task difficulties based on models’ capability boundaries, and facilitating multi-dimensional process evaluation across key step coverage, tool effectiveness, and computational efficiency.

Result: Evaluation reveals that adaptive mode selection scales with model capacity but decouples from final accuracy, while key step coverage aligns with performance, and tool effectiveness remains highly inconsistent across model architectures.

Conclusion: AdaptMMBench provides a comprehensive framework for evaluating adaptive multimodal reasoning that isolates meta-cognition abilities and enables fine-grained analysis of reasoning processes beyond simple accuracy metrics.

Abstract: Adaptive multimodal reasoning has emerged as a promising frontier in Vision-Language Models (VLMs), aiming to dynamically modulate between tool-augmented visual reasoning and text reasoning to enhance both effectiveness and efficiency. However, existing evaluations rely on static difficulty labels and simplistic metrics, which fail to capture the dynamic nature of difficulty relative to varying model capacities. Consequently, they obscure the distinction between adaptive mode selection and general performance while neglecting fine-grained process analyses. In this paper, we propose AdaptMMBench, a comprehensive benchmark for adaptive multimodal reasoning across five domains: real-world, OCR, GUI, knowledge, and math, encompassing both direct perception and complex reasoning tasks. AdaptMMBench utilizes a Matthews Correlation Coefficient (MCC) metric to evaluate the selection rationality of different reasoning modes, isolating this meta-cognition ability by dynamically identifying task difficulties based on models’ capability boundaries. Moreover, AdaptMMBench facilitates multi-dimensional process evaluation across key step coverage, tool effectiveness, and computational efficiency. Our evaluation reveals that while adaptive mode selection scales with model capacity, it notably decouples from final accuracy. Conversely, key step coverage aligns with performance, though tool effectiveness remains highly inconsistent across model architectures.

Method: Proposes GRAM using multi-channel masked autoencoder to learn spatial audio representations, evaluated on two benchmark suites: NatHEAR (simulated naturalistic spatial environments) and RealSELD (real-world recordings).

Result: GRAM outperforms state-of-the-art self-supervised audio foundation models on NatHEAR and clean single-channel HEAR benchmarks using less training data, and shows SOTA localization performance in simulated environments with efficient generalization to real-world recordings.

Conclusion: GRAM represents a significant advance toward robust spatial audio foundation models for real-world environments, addressing key limitations of current models.

Abstract: Audio foundation models learn general-purpose audio representations that facilitate a wide range of downstream tasks. While the performance of these models has greatly increased for conventional single-channel, dry audio clips, their success in real-world acoustic environments with reverberation and noise is limited. Furthermore, most audio foundation models ignore the spatial dimension of real-world acoustic environments, ruling out tasks involving sound localization. To address these limitations, we propose GRAM: a general-purpose real-world audio model that employs a multi-channel masked autoencoder to efficiently learn spatial audio representations. We evaluated GRAM and other audio foundation models in a standardized manner on high-quality simulations of naturalistic, spatial acoustic environments as well as recordings of real-world environments and release these two complementary benchmark task suites: NatHEAR and RealSELD. Our results demonstrate that GRAM outperforms all state-of-the-art self-supervised audio foundation models on NatHEAR and the clean, single-channel version HEAR, while using only a fraction of the training data. GRAM also shows state-of-the-art localization performance in simulated environments and generalizes efficiently to real-world recordings in RealSELD. Taken together, GRAM presents a significant advance toward robust spatial audio foundation models for real-world environments.

Method: 1) Latent-guided token compression module reduces redundancy in visual sequences while preserving semantics. 2) Semi-autoregressive decoding scheme leverages co-occurrence and local correlations of visual entities to predict multiple tokens in one forward pass.

Result: Achieves up to 2.68× speed-up on video captioning and 2.55× on visual instruction tuning compared to original LMMs, consistently outperforming prior speculative decoding baselines.

Conclusion: SpecFLASH effectively accelerates multimodal model inference by exploiting visual structure characteristics, making it a practical solution for efficient multimodal generation tasks.

Abstract: Large language models and large multimodal models (LLMs and LMMs) deliver strong generative performance but suffer from slow decoding, a problem that becomes more severe when handling visual inputs, whose sequences typically contain many more tokens with lower information density than text. Speculative decoding accelerates LLM inference by letting a compact draft model propose candidate tokens that are selectively accepted by a larger target model, achieving speed-up without degrading quality. However, existing multimodal speculative decoding approaches largely ignore the structural characteristics of visual representations and usually rely on text-only draft models. In this paper, we introduce SpecFLASH, a speculative decoding framework tailored to LMMs that explicitly exploits multimodal structure when designing the draft model. We first mitigate redundancy in visual token sequences with a lightweight, latent-guided token compression module that compacts visual features while preserving semantics, and then leverage the co-occurrence and local correlations of visual entities via a semi-autoregressive decoding scheme that predicts multiple tokens in a single forward pass. Extensive experiments demonstrate that SpecFLASH consistently surpasses prior speculative decoding baselines, achieving up to $2.68\times$ speed-up on video captioning and $2.55\times$ on visual instruction tuning, relative to the original LMM. Our code is available here: https://github.com/ZihuaEvan/FlashSD/.

Method: Introduces the Hypocrisy Gap metric using Sparse Autoencoders (SAEs) to quantify divergence between a model’s internal reasoning and final generation. Uses sparse linear probes to derive internal truth beliefs and compares them to final generated trajectories in latent space.

Result: Experiments on Gemma, Llama, and Qwen models using Anthropic’s Sycophancy benchmark show AUROC of 0.55-0.73 for detecting sycophantic runs and 0.55-0.74 for hypocritical cases, outperforming log-probability baselines (0.41-0.50 AUROC).

Conclusion: The Hypocrisy Gap provides an effective mechanistic approach for detecting unfaithful behavior in LLMs by quantifying the divergence between internal reasoning and external outputs.

Abstract: Large Language Models (LLMs) frequently exhibit unfaithful behavior, producing a final answer that differs significantly from their internal chain of thought (CoT) reasoning in order to appease the user they are conversing with. In order to better detect this behavior, we introduce the Hypocrisy Gap, a mechanistic metric utilizing Sparse Autoencoders (SAEs) to quantify the divergence between a model’s internal reasoning and its final generation. By mathematically comparing an internal truth belief, derived via sparse linear probes, to the final generated trajectory in latent space, we quantify and detect a model’s tendency to engage in unfaithful behavior. Experiments on Gemma, Llama, and Qwen models using Anthropic’s Sycophancy benchmark show that our method achieves an AUROC of 0.55-0.73 for detecting sycophantic runs and 0.55-0.74 for hypocritical cases where the model internally “knows” the user is wrong, consistently outperforming a decision-aligned log-probability baseline (0.41-0.50 AUROC).

Method: Proposes STEMVerse framework that characterizes model performance across academic specialization and cognitive complexity. Re-aggregates over 20,000 STEM problems into a unified “Discipline × Cognition” capability space with dual-axis labels for each instance.

Result: Systematic evaluation of representative LLM families reveals structural failure patterns in STEM reasoning. The framework provides clear perspective on scientific reasoning characteristics of LLMs.

Conclusion: STEMVerse offers a diagnostic framework that integrates multi-disciplinary coverage and fine-grained cognitive stratification to better understand LLM capabilities in STEM reasoning.

Abstract: As Large Language Models (LLMs) achieve significant breakthroughs in complex reasoning tasks, evaluating their proficiency in science, technology, engineering, and mathematics (STEM) has become a primary method for measuring machine intelligence. However, current evaluation paradigms often treat benchmarks as isolated “silos,” offering only monolithic aggregate scores that neglect the intricacies of both academic specialization and cognitive depth. This result-oriented approach fails to distinguish whether model errors stem from insufficient domain knowledge or deficiencies in cognitive capacity, thereby limiting the diagnostic value. To address this, we propose STEMVerse, a diagnostic framework designed to systematically analyze the STEM reasoning capabilities of LLMs. This framework characterizes model performance across academic specialization and cognitive complexity to map the capability required for reasoning. We re-aggregate over 20,000 STEM problems from mainstream benchmarks into a unified “Discipline $\times$ Cognition” capability space, assigning dual-axis labels to every instance. Utilizing this unified diagnostic framework, we systematically evaluate representative LLM families across varying parameter scales and training paradigms. Our empirical results reveal structural failure patterns in STEM reasoning. By integrating multi-disciplinary coverage and fine-grained cognitive stratification into a unified framework, STEMVerse provides a clear and actionable perspective for understanding the scientific reasoning characteristics of LLMs.

Method: Proposes a test-time procedure that approximates the gradient of completion toxicity with respect to input embeddings using zeroth-order optimization. Requires only access to input embeddings, a toxicity scoring function, and forward evaluations of the model. Uses a small number of descent steps to steer generation toward less toxic continuations.

Result: The approach delivers robust toxicity reductions across models and prompts, achieving the best overall toxicity-quality trade-off in most settings. Works without requiring any training or access to intermediate computations.

Conclusion: Word embeddings can serve as effective control variables for guiding autoregressive language models toward safer text generation. Encourages wider use of black-box optimization for scalable safety improvements without model modifications.

Abstract: Large language models can produce toxic or inappropriate text even for benign inputs, creating risks when deployed at scale. Detoxification is therefore important for safety and user trust, particularly when we want to reduce harmful content without sacrificing the model’s generation quality. Many existing approaches rely on model retraining, gradients, or learned auxiliary components, which can be costly and may not transfer across model families or to truly black-box settings. We introduce a test-time procedure that approximates the gradient of completion toxicity with respect to the input embeddings and uses a small number of descent steps to steer generation toward less toxic continuations. This is achieved with zeroth-order optimization that requires only access to input embeddings, a toxicity scoring function, and forward evaluations of the model. Empirically, the approach delivers robust toxicity reductions across models and prompts and, in most settings, achieves the best overall toxicity-quality trade-off. More broadly, our work positions word embeddings as effective control variables and encourages wider use of black-box optimization to guide autoregressive language models toward scalable, safer text generation, without requiring any training or access to intermediate computations.

Method: Introduces ROSA-Tuning with a CPU-based ROSA (RWKV Online Suffix Automaton) retrieval module that efficiently locates historically relevant positions in long contexts. Uses binary discretization strategy and counterfactual gradient algorithm for end-to-end training, with asynchronous CPU-GPU pipeline optimization. Combines standard attention with retrieval-and-recall mechanism.

Result: On Qwen3-Base-1.7B, ROSA-Tuning substantially restores long-context modeling ability of windowed-attention models, achieving performance close to global attention on LongBench benchmarks while maintaining computational efficiency and GPU memory usage comparable to windowed-attention methods.

Conclusion: ROSA-Tuning offers a new technical path for efficient long-context processing by combining retrieval mechanisms with attention, achieving near-global attention performance with windowed-attention efficiency through CPU-based retrieval and trainable information injection.

Abstract: Long-context capability and computational efficiency are among the central challenges facing today’s large language models. Existing efficient attention methods reduce computational complexity, but they typically suffer from a limited coverage of the model state. This paper proposes ROSA-Tuning, a retrieval-and-recall mechanism for enhancing the long-context modeling ability of pretrained models. Beyond the standard attention mechanism, ROSA-Tuning introduces in parallel a CPU-based ROSA (RWKV Online Suffix Automaton) retrieval module, which efficiently locates historical positions in long contexts that are relevant to the current query, and injects the retrieved information into the model state in a trainable manner; subsequent weighted fusion can then be handled by range-restricted attention. To enable end-to-end training, we design a binary discretization strategy and a counterfactual gradient algorithm, and further optimize overall execution efficiency via an asynchronous CPU-GPU pipeline. Systematic evaluations on Qwen3-Base-1.7B show that ROSA-Tuning substantially restores the long-context modeling ability of windowed-attention models, achieving performance close to and in some cases matching global attention on benchmarks such as LongBench, while maintaining computational efficiency and GPU memory usage that are nearly comparable to windowed-attention methods, offering a new technical path for efficient long-context processing. The example code can be found at https://github.com/zyaaa-ux/ROSA-Tuning.

Method: Builds domain-specific knowledge graph, retrieves query-relevant subgraphs, uses ChatGLM as Transformer backbone with LoRA fine-tuning, employs graph neural network to learn node representations capturing symptom-disease-treatment dependencies, and incorporates graph evidence into LLM input.

Result: Shows consistent improvements over text-only baselines, with largest gains on multi-hop and comparative reasoning questions requiring chaining multiple relations.

Conclusion: Graph-augmented reasoning effectively integrates structured domain knowledge with LLMs for improved pest/disease management recommendations, particularly for complex reasoning tasks.

Abstract: This paper proposes a graph-augmented reasoning framework for tobacco pest and disease management that integrates structured domain knowledge into large language models. Building on GraphRAG, we construct a domain-specific knowledge graph and retrieve query-relevant subgraphs to provide relational evidence during answer generation. The framework adopts ChatGLM as the Transformer backbone with LoRA-based parameter-efficient fine-tuning, and employs a graph neural network to learn node representations that capture symptom-disease-treatment dependencies. By explicitly modeling diseases, symptoms, pesticides, and control measures as linked entities, the system supports evidence-aware retrieval beyond surface-level text similarity. Retrieved graph evidence is incorporated into the LLM input to guide generation toward domain-consistent recommendations and to mitigate hallucinated or inappropriate treatments. Experimental results show consistent improvements over text-only baselines, with the largest gains observed on multi-hop and comparative reasoning questions that require chaining multiple relations.

Method: Two-pronged approach: 1) Created WideSeekBench, a General Broad Information Seeking benchmark with diverse information volume, constraints, and domains; 2) Developed WideSeek, a dynamic hierarchical multi-agent architecture that forks parallel sub-agents based on task requirements, trained with end-to-end RL on linearized multi-agent trajectories.

Result: Experimental results show effectiveness of WideSeek and multi-agent RL, demonstrating that scaling agent number is promising for advancing Wide Research paradigm.

Conclusion: The work addresses critical gaps in wide research through dedicated benchmark and multi-agent optimization, showing multi-agent scaling as a viable direction for broad information seeking systems.

Abstract: Search intelligence is evolving from Deep Research to Wide Research, a paradigm essential for retrieving and synthesizing comprehensive information under complex constraints in parallel. However, progress in this field is impeded by the lack of dedicated benchmarks and optimization methodologies for search breadth. To address these challenges, we take a deep dive into Wide Research from two perspectives: Data Pipeline and Agent Optimization. First, we produce WideSeekBench, a General Broad Information Seeking (GBIS) benchmark constructed via a rigorous multi-phase data pipeline to ensure diversity across the target information volume, logical constraints, and domains. Second, we introduce WideSeek, a dynamic hierarchical multi-agent architecture that can autonomously fork parallel sub-agents based on task requirements. Furthermore, we design a unified training framework that linearizes multi-agent trajectories and optimizes the system using end-to-end RL. Experimental results demonstrate the effectiveness of WideSeek and multi-agent RL, highlighting that scaling the number of agents is a promising direction for advancing the Wide Research paradigm.

Method: Enforce monotonicity selectively in feed-forward sublayers of sequence-to-sequence Transformers while leaving attention mechanisms unconstrained. This architectural separation allows attention to handle negation, contradiction, and contextual interactions, while feed-forward layers ensure order-preserving semantic refinement.

Result: Adversarial attack success rates drop from approximately 69% to 19% while standard summarization performance degrades only marginally. Monotone language models preserve the performance of their pretrained counterparts while substantially improving robustness.

Conclusion: Monotonicity is a viable architectural inductive bias for improving Transformer robustness without sacrificing expressivity. The trade-off between monotonicity and expressivity is not inherent when properly separated in architecture, with attention handling complex interactions and feed-forward layers ensuring order-preserving behavior.

Abstract: Large language models (LLMs) are known to exhibit brittle behavior under adversarial prompts and jailbreak attacks, even after extensive alignment and fine-tuning. This fragility reflects a broader challenge of modern neural language models: small, carefully structured perturbations in high-dimensional input spaces can induce large and unpredictable changes in internal semantic representations and output. We investigate monotonicity as an architectural inductive bias for improving the robustness of Transformer-based language models. Monotonicity constrains semantic transformations so that strengthening information, evidence, or constraints cannot lead to regressions in the corresponding internal representations. Such order-preserving behavior has long been exploited in control and safety-critical systems to simplify reasoning and improve robustness, but has traditionally been viewed as incompatible with the expressivity required by neural language models. We show that this trade-off is not inherent. By enforcing monotonicity selectively in the feed-forward sublayers of sequence-to-sequence Transformers – while leaving attention mechanisms unconstrained – we obtain monotone language models that preserve the performance of their pretrained counterparts. This architectural separation allows negation, contradiction, and contextual interactions to be introduced explicitly through attention, while ensuring that subsequent semantic refinement is order-preserving. Empirically, monotonicity substantially improves robustness: adversarial attack success rates drop from approximately 69% to 19%, while standard summarization performance degrades only marginally.

Method: InfMem implements System-2-style control via a PreThink-Retrieve-Write protocol that actively monitors evidence sufficiency, performs targeted in-document retrieval, and applies evidence-aware joint compression to update bounded memory. It uses a practical SFT-to-RL training recipe to align retrieval, writing, and stopping decisions with end-task correctness.

Result: InfMem consistently outperforms MemAgent across backbones on ultra-long QA benchmarks from 32k to 1M tokens, improving average absolute accuracy by +10.17, +11.84, and +8.23 points on Qwen3-1.7B, Qwen3-4B, and Qwen2.5-7B respectively, while reducing inference time by 3.9× on average (up to 5.1×) via adaptive early stopping.

Conclusion: The proposed control-centric agent with active memory management and evidence-aware compression effectively addresses the challenge of preserving bridging evidence for multi-hop reasoning in ultra-long documents while improving both accuracy and efficiency.

Abstract: Reasoning over ultra-long documents requires synthesizing sparse evidence scattered across distant segments under strict memory constraints. While streaming agents enable scalable processing, their passive memory update strategy often fails to preserve low-salience bridging evidence required for multi-hop reasoning. We propose InfMem, a control-centric agent that instantiates System-2-style control via a PreThink-Retrieve-Write protocol. InfMem actively monitors evidence sufficiency, performs targeted in-document retrieval, and applies evidence-aware joint compression to update a bounded memory. To ensure reliable control, we introduce a practical SFT-to-RL training recipe that aligns retrieval, writing, and stopping decisions with end-task correctness. On ultra-long QA benchmarks from 32k to 1M tokens, InfMem consistently outperforms MemAgent across backbones. Specifically, InfMem improves average absolute accuracy by +10.17, +11.84, and +8.23 points on Qwen3-1.7B, Qwen3-4B, and Qwen2.5-7B, respectively, while reducing inference time by $3.9\times$ on average (up to $5.1\times$) via adaptive early stopping.

Method: Analyzed EHR data from 4,276,403 VA patients using static and time-varying representations with clinician-informed logic to model persistence of clinical conditions and social risks. Compared classical ML, transformer-based masked language models, and fine-tuned LLMs for predicting homelessness 3-12 months in advance.

Result: Incorporating social/behavioral factors improved PR-AUC by 15-30%. In top 1% risk tier, models achieved PPVs ranging from 3.93-13.80% across time horizons. LLMs underperformed encoder-based models on discrimination but showed smaller performance disparities across racial groups.

Conclusion: Longitudinal, socially informed EHR modeling concentrates homelessness risk into actionable strata, enabling targeted prevention strategies for at-risk veterans, with LLMs offering potential benefits for fairness despite lower discrimination performance.

Abstract: Homelessness among US veterans remains a critical public health challenge, yet risk prediction offers a pathway for proactive intervention. In this retrospective prognostic study, we analyzed electronic health record (EHR) data from 4,276,403 Veterans Affairs patients during a 2016 observation period to predict first-episode homelessness occurring 3-12 months later in 2017 (prevalence: 0.32-1.19%). We constructed static and time-varying EHR representations, utilizing clinician-informed logic to model the persistence of clinical conditions and social risks over time. We then compared the performance of classical machine learning, transformer-based masked language models, and fine-tuned large language models (LLMs). We demonstrate that incorporating social and behavioral factors into longitudinal models improved precision-recall area under the curve (PR-AUC) by 15-30%. In the top 1% risk tier, models yielded positive predictive values ranging from 3.93-4.72% at 3 months, 7.39-8.30% at 6 months, 9.84-11.41% at 9 months, and 11.65-13.80% at 12 months across model architectures. Large language models underperformed encoder-based models on discrimination but showed smaller performance disparities across racial groups. These results demonstrate that longitudinal, socially informed EHR modeling concentrates homelessness risk into actionable strata, enabling targeted and data-informed prevention strategies for at-risk veterans.

Method: Developed a constructive greedy heuristic algorithm for vehicle dispatching that tests four fleet configurations: ambulances only, ambulances+UAVs, ambulances+eVTOLs, and fully integrated ambulances+UAVs+eVTOLs. Algorithm includes payload consolidation, accounts for traffic congestion (ground) and weather conditions (air), and enables rapid dispatching compared to optimization models.

Result: Evaluation of all four fleet types under common conditions to identify most effective configurations for fulfilling medical transportation needs while minimizing operating costs, recharging/fuel costs, and total transportation time.

Conclusion: Multimodal integration of ground and air vehicles offers potential efficiency improvements for medical transportation, with specific fleet configurations providing optimal balance of speed, cost, and reliability.

Abstract: Timely transportation of organs, patients, and medical supplies is critical to modern healthcare, particularly in emergencies and transplant scenarios where even short delays can severely impact outcomes. Traditional ground-based vehicles such as ambulances are often hindered by traffic congestion; while air vehicles such as helicopters are faster but costly. Emerging air vehicles – Unmanned Aerial Vehicles and electric vertical take-off and landing aircraft – have lower operating costs, but remain limited by range and susceptibility to weather conditions. A multimodal transportation system that integrates both air and ground vehicles can leverage the strengths of each to enhance overall transportation efficiency. This study introduces a constructive greedy heuristic algorithm for multimodal vehicle dispatching for medical transportation. Four different fleet configurations were tested: (i) ambulances only, (ii) ambulances with Unmanned Aerial Vehicles, (iii) ambulances with electric vertical take-off and landing aircraft, and (iv) a fully integrated fleet of ambulances, Unmanned Aerial Vehicles, and electric vertical take-off and landing aircraft. The algorithm incorporates payload consolidation across compatible routes, accounts for traffic congestion in ground operations and weather conditions in aerial operations, while enabling rapid vehicle dispatching compared to computationally intensive optimization models. Using a common set of conditions, we evaluate all four fleet types to identify the most effective configurations for fulfilling medical transportation needs while minimizing operating costs, recharging/fuel costs, and total transportation time.

Method: Stress-test deployment-relevant robustness using a grid-based game with simple goals but long-horizon execution. Episodes violate clean-interface assumptions while remaining solvable, forcing agents to infer rules, pay for information, adapt to environmental/internal shifts, and act cautiously under noise.

Result: Large gaps found between nominal task-solving and deployment-like robustness across five state-of-the-art LLM agents. Performance degrades with grid size and horizon, but rankings are unstable - weaker models can beat stronger ones when strategy matches uncertainty regime. Agents show partial objective inference without explicit instruction.

Conclusion: Highlights need for work on verification, safe action selection, and objective inference under partial observability, noise, and non-stationarity. Current agent evaluations overestimate real-world readiness, requiring more realistic testing frameworks.

Abstract: Large language models are increasingly deployed as specialized agents that plan, call tools, and take actions over extended horizons. Yet many existing evaluations assume a “clean interface” where dynamics are specified and stable, tools and sensors are reliable, and success is captured by a single explicit objective-often overestimating real-world readiness. In practice, agents face underspecified rules, unreliable signals, shifting environments, and implicit, multi-stakeholder goals. The challenge is therefore not just solving tasks, but adapting while solving: deciding what to trust, what is wanted, when to verify, and when to fall back or escalate. We stress-test deployment-relevant robustness under four operational circumstances: partial observability, dynamic environments, noisy signals, and dynamic agent state. We benchmark agentic LLMs in a grid-based game with a simple goal but long-horizon execution. Episodes violate clean-interface assumptions yet remain solvable, forcing agents to infer rules, pay for information, adapt to environmental and internal shifts, and act cautiously under noise. Across five state-of-the-art LLM agents, we find large gaps between nominal task-solving and deployment-like robustness. Performance generally degrades as grid size and horizon increase, but rankings are unstable: weaker models can beat stronger ones when strategy matches the uncertainty regime. Despite no explicit instruction, agents trade off completion, efficiency, and penalty avoidance, suggesting partial objective inference. Ablations and feature analyses reveal model-specific sensitivities and failure drivers, motivating work on verification, safe action selection, and objective inference under partial observability, noise, and non-stationarity.

Method: Created AmharicStoryQA benchmark with culturally diverse narratives from different Ethiopian regions, evaluated existing LLMs on long-sequence story question answering, and conducted supervised fine-tuning experiments.

Result: Found significant narrative understanding gap in LLMs, pronounced regional performance differences despite shared language, and uneven improvements from fine-tuning across regions.

Conclusion: Need culturally grounded benchmarks beyond language-level evaluation to accurately assess and improve narrative understanding in low-resource languages.

Abstract: With the growing emphasis on multilingual and cultural evaluation benchmarks for large language models, language and culture are often treated as synonymous, and performance is commonly used as a proxy for a models understanding of a given language. In this work, we argue that such evaluations overlook meaningful cultural variation that exists within a single language. We address this gap by focusing on narratives from different regions of Ethiopia and demonstrate that, despite shared linguistic characteristics, region-specific and domain-specific content substantially influences language evaluation outcomes. To this end, we introduce \textbf{\textit{AmharicStoryQA}}, a long-sequence story question answering benchmark grounded in culturally diverse narratives from Amharic-speaking regions. Using this benchmark, we reveal a significant narrative understanding gap in existing LLMs, highlight pronounced regional differences in evaluation results, and show that supervised fine-tuning yields uneven improvements across regions and evaluation settings. Our findings emphasize the need for culturally grounded benchmarks that go beyond language-level evaluation to more accurately assess and improve narrative understanding in low-resource languages.

Method: Uses Information Bottleneck framework to formalize trade-off between simplicity and informativeness. First demonstrates correlation between IB optimality and novel generalization of convexity. Second experiment manipulates modeling parameters in IB framework to determine which factors drive this correlation.

Result: Found that convexity of the communicative need distribution plays especially important role in explaining correlation between convexity and optimality. Moves beyond showing efficient communication explains semantic typology to identifying underlying responsible factors.

Conclusion: The research identifies specific factors (particularly convexity of communicative need distributions) that make efficient communication explanations successful for semantic typology patterns, advancing understanding of why language evolution follows certain patterns.

Abstract: Much recent work has argued that the variation in the languages of the world can be explained from the perspective of efficient communication; in particular, languages can be seen as optimally balancing competing pressures to be simple and to be informative. Focusing on the expression of meaning – semantic typology – the present paper asks what factors are responsible for successful explanations in terms of efficient communication. Using the Information Bottleneck (IB) approach to formalizing this trade-off, we first demonstrate and analyze a correlation between optimality in the IB sense and a novel generalization of convexity to this setting. In a second experiment, we manipulate various modeling parameters in the IB framework to determine which factors drive the correlation between convexity and optimality. We find that the convexity of the communicative need distribution plays an especially important role. These results move beyond showing that efficient communication can explain aspects of semantic typology into explanations for why that is the case by identifying which underlying factors are responsible.

Method: R2-Router treats output length budget as a controllable variable and jointly selects the best LLM and length budget, enforcing the budget via length-constrained instructions. The approach includes constructing R2-Bench, the first routing dataset capturing LLM behavior across diverse output length budgets.

Result: Experiments show R2-Router achieves state-of-the-art performance at 4-5x lower cost compared with existing routers. The router discovers that powerful LLMs with constrained outputs can outperform weaker LLMs at comparable cost-efficient configurations.

Conclusion: This work opens a new direction: routing as reasoning, where routers evolve from reactive selectors to deliberate reasoners that explore which LLM to use and at what cost budget, enabling more sophisticated and cost-effective LLM deployment strategies.

Abstract: As LLMs proliferate with diverse capabilities and costs, LLM routing has emerged by learning to predict each LLM’s quality and cost for a given query, then selecting the one with high quality and low cost. However, existing routers implicitly assume a single fixed quality and cost per LLM for each query, ignoring that the same LLM’s quality varies with its output length. This causes routers to exclude powerful LLMs when their estimated cost exceeds the budget, missing the opportunity that these LLMs could still deliver high quality at reduced cost with shorter outputs. To address this, we introduce R2-Router, which treats output length budget as a controllable variable and jointly selects the best LLM and length budget, enforcing the budget via length-constrained instructions. This enables R2-Router to discover that a powerful LLM with constrained output can outperform a weaker LLM at comparable cost-efficient configurations invisible to prior methods. Together with the router framework, we construct R2-Bench, the first routing dataset capturing LLM behavior across diverse output length budgets. Experiments show that R2-Router achieves state-of-the-art performance at 4-5x lower cost compared with existing routers. This work opens a new direction: routing as reasoning, where routers evolve from reactive selectors to deliberate reasoners that explore which LLM to use and at what cost budget.

Method: Proposes LatentMem framework with two components: an experience bank storing raw interaction trajectories in lightweight form, and a memory composer that synthesizes compact latent memories conditioned on retrieved experience and agent-specific contexts. Also introduces Latent Memory Policy Optimization (LMPO) to propagate task-level optimization signals through latent memories to the composer.

Result: Extensive experiments across diverse benchmarks and mainstream MAS frameworks show LatentMem achieves up to 19.36% performance gain over vanilla settings and consistently outperforms existing memory architectures without requiring modifications to underlying frameworks.

Conclusion: LatentMem effectively addresses memory homogenization and information overload in multi-agent systems through learnable, role-aware memory customization, enabling more efficient collective intelligence in LLM-powered multi-agent systems.

Abstract: Large language model (LLM)-powered multi-agent systems (MAS) demonstrate remarkable collective intelligence, wherein multi-agent memory serves as a pivotal mechanism for continual adaptation. However, existing multi-agent memory designs remain constrained by two fundamental bottlenecks: (i) memory homogenization arising from the absence of role-aware customization, and (ii) information overload induced by excessively fine-grained memory entries. To address these limitations, we propose LatentMem, a learnable multi-agent memory framework designed to customize agent-specific memories in a token-efficient manner. Specifically, LatentMem comprises an experience bank that stores raw interaction trajectories in a lightweight form, and a memory composer that synthesizes compact latent memories conditioned on retrieved experience and agent-specific contexts. Further, we introduce Latent Memory Policy Optimization (LMPO), which propagates task-level optimization signals through latent memories to the composer, encouraging it to produce compact and high-utility representations. Extensive experiments across diverse benchmarks and mainstream MAS frameworks show that LatentMem achieves a performance gain of up to $19.36$% over vanilla settings and consistently outperforms existing memory architectures, without requiring any modifications to the underlying frameworks.

Method: CATNIP uses calibrated token-level negative preference alignment that rescales unlearning effects proportionally to the model’s token-level confidence. This provides fine-grained control over forgetting by precisely targeting undesirable knowledge based on the model’s own confidence scores, eliminating the need for retention data or contrastive response pairs.

Result: Extensive evaluations on MUSE and WMDP benchmarks show CATNIP achieves effective unlearning without requiring retention data or contrastive pairs, with stronger knowledge forgetting and preservation tradeoffs than state-of-the-art methods.

Conclusion: CATNIP provides a principled approach to LLM unlearning that addresses key limitations of existing methods by using token-level confidence calibration, enabling precise removal of undesirable knowledge while preserving general capabilities without requiring additional data resources.

Abstract: Pretrained knowledge memorized in LLMs raises critical concerns over safety and privacy, which has motivated LLM Unlearning as a technique for selectively removing the influences of undesirable knowledge. Existing approaches, rooted in Gradient Ascent (GA), often degrade general domain knowledge while relying on retention data or curated contrastive pairs, which can be either impractical or data and computationally prohibitive. Negative Preference Alignment has been explored for unlearning to tackle the limitations of GA, which, however, remains confined by its choice of reference model and shows undermined performance in realistic data settings. These limitations raise two key questions: i) Can we achieve effective unlearning that quantifies model confidence in undesirable knowledge and uses it to calibrate gradient updates more precisely, thus reducing catastrophic forgetting? ii) Can we make unlearning robust to data scarcity and length variation? We answer both questions affirmatively with CATNIP (Calibrated and Tokenized Negative Preference Alignment), a principled method that rescales unlearning effects in proportion to the model’s token-level confidence, thus ensuring fine-grained control over forgetting. Extensive evaluations on MUSE and WMDP benchmarks demonstrated that our work enables effective unlearning without requiring retention data or contrastive unlearning response pairs, with stronger knowledge forgetting and preservation tradeoffs than state-of-the-art methods.

Method: AlignAtt exploits attention information from the model to generate source-target alignments. These alignments are then used as a policy to guide the simultaneous speech translation model during inference, determining when to read more input versus when to produce output.

Result: Experiments on 8 language pairs from MuST-C v1.0 show AlignAtt outperforms previous state-of-the-art SimulST policies, achieving BLEU score improvements of 2 points and latency reductions of 0.5s to 0.8s across all languages.

Conclusion: Attention-based alignment information can be effectively used to develop superior simultaneous speech translation policies that balance translation quality and latency better than previous approaches.

Abstract: Attention is the core mechanism of today’s most used architectures for natural language processing and has been analyzed from many perspectives, including its effectiveness for machine translation-related tasks. Among these studies, attention resulted to be a useful source of information to get insights about word alignment also when the input text is substituted with audio segments, as in the case of the speech translation (ST) task. In this paper, we propose AlignAtt, a novel policy for simultaneous ST (SimulST) that exploits the attention information to generate source-target alignments that guide the model during inference. Through experiments on the 8 language pairs of MuST-C v1.0, we show that AlignAtt outperforms previous state-of-the-art SimulST policies applied to offline-trained models with gains in terms of BLEU of 2 points and latency reductions ranging from 0.5s to 0.8s across the 8 languages.

Method: Developed a computational model based on expected regret theory, then tested predictions through two experiments: one examining linguistic responses to questions, and another extending to choices between clarification and non-linguistic actions.

Result: Results show humans tend to seek clarification proportional to the risk of substantial loss when acting under uncertainty, supporting the expected regret model’s predictions about the interaction between uncertainty and action costs.

Conclusion: Human decision-making about asking clarification questions follows rational principles based on expected regret, balancing uncertainty reduction against the potential costs of incorrect actions.

Abstract: When deciding how to act under uncertainty, agents may choose to act to reduce uncertainty or they may act despite that uncertainty.In communicative settings, an important way of reducing uncertainty is by asking clarification questions (CQs). We predict that the decision to ask a CQ depends on both contextual uncertainty and the cost of alternative actions, and that these factors interact: uncertainty should matter most when acting incorrectly is costly. We formalize this interaction in a computational model based on expected regret: how much an agent stands to lose by acting now rather than with full information. We test these predictions in two experiments, one examining purely linguistic responses to questions and another extending to choices between clarification and non-linguistic action. Taken together, our results suggest a rational tradeoff: humans tend to seek clarification proportional to the risk of substantial loss when acting under uncertainty.

Method: Collaborative course design by a team with complementary expertise, offered as an upper-level undergraduate course cross-listed between Linguistics and Computer Science, with deep integration of theoretical and empirical aspects and exploratory assignments.

Result: Successful implementation of the course in Fall 2025, with detailed course description and takeaways from an independent survey showing positive outcomes in bridging discourse theory with NLP practice.

Conclusion: The course successfully addresses the gap in NLP education by integrating discourse processing with language generation, providing a model for interdisciplinary course design in evolving technical fields, with plans for future improvements.

Abstract: The field of NLP has undergone vast, continuous transformations over the past few years, sparking debates going beyond discipline boundaries. This begs important questions in education: how do we design courses that bridge sub-disciplines in this shifting landscape? This paper explores this question from the angle of discourse processing, an area with rich linguistic insights and computational models for the intentional, attentional, and coherence structure of language. Discourse is highly relevant for open-ended or long-form text generation, yet this connection is under-explored in existing undergraduate curricula. We present a new course, “Computational Discourse and Natural Language Generation”. The course is collaboratively designed by a team with complementary expertise and was offered for the first time in Fall 2025 as an upper-level undergraduate course, cross-listed between Linguistics and Computer Science. Our philosophy is to deeply integrate the theoretical and empirical aspects, and create an exploratory mindset inside the classroom and in the assignments. This paper describes the course in detail and concludes with takeaways from an independent survey as well as our vision for future directions.

Method: Proposes a computational framework that models sarcasm as integration of semantic interpretation and prosodic realization. Semantic cues are derived from LLaMA 3 model fine-tuned to capture discourse-level markers of sarcastic intent. Prosodic cues are extracted through semantically aligned utterances from a database of sarcastic speech, providing prosodic exemplars of sarcastic delivery. Uses speech synthesis testbed for perceptual evaluations.

Result: Perceptual evaluations demonstrate that both semantic and prosodic cues independently enhance listeners’ perception of sarcasm, with the strongest effects emerging when the two are combined. Findings highlight complementary roles of semantics and prosody in pragmatic interpretation.

Conclusion: The study illustrates how computational modeling can shed light on mechanisms underlying sarcastic communication, showing that both semantic and prosodic cues play important and complementary roles in sarcasm recognition, with optimal perception occurring when both are present.

Abstract: Sarcasm is a pragmatic phenomenon in which speakers convey meanings that diverge from literal content, relying on an interaction between semantics and prosodic expression. However, how these cues jointly contribute to the recognition of sarcasm remains poorly understood. We propose a computational framework that models sarcasm as the integration of semantic interpretation and prosodic realization. Semantic cues are derived from an LLaMA 3 model fine-tuned to capture discourse-level markers of sarcastic intent, while prosodic cues are extracted through semantically aligned utterances drawn from a database of sarcastic speech, providing prosodic exemplars of sarcastic delivery. Using a speech synthesis testbed, perceptual evaluations demonstrate that both semantic and prosodic cues independently enhance listeners’ perception of sarcasm, with the strongest effects emerging when the two are combined. These findings highlight the complementary roles of semantics and prosody in pragmatic interpretation and illustrate how modeling can shed light on the mechanisms underlying sarcastic communication.

Method: HALT uses top-20 token log-probabilities as time series data, processed through a gated recurrent unit (GRU) model with entropy-based features to learn model calibration bias. It requires only output log-probs, not hidden states or attention maps.

Result: HALT outperforms Lettuce (fine-tuned BERT-base encoder) while being 30x smaller and achieving 60x speedup on the HUB benchmark, which covers ten diverse LLM capabilities including reasoning and general-purpose tasks.

Conclusion: HALT provides an efficient, lightweight hallucination detection framework that generalizes well across domains and is compatible with proprietary LLMs without requiring internal model access.

Abstract: Hallucinations remain a major obstacle for large language models (LLMs), especially in safety-critical domains. We present HALT (Hallucination Assessment via Log-probs as Time series), a lightweight hallucination detector that leverages only the top-20 token log-probabilities from LLM generations as a time series. HALT uses a gated recurrent unit model combined with entropy-based features to learn model calibration bias, providing an extremely efficient alternative to large encoders. Unlike white-box approaches, HALT does not require access to hidden states or attention maps, relying only on output log-probabilities. Unlike black-box approaches, it operates on log-probs rather than surface-form text, which enables stronger domain generalization and compatibility with proprietary LLMs without requiring access to internal weights. To benchmark performance, we introduce HUB (Hallucination detection Unified Benchmark), which consolidates prior datasets into ten capabilities covering both reasoning tasks (Algorithmic, Commonsense, Mathematical, Symbolic, Code Generation) and general purpose skills (Chat, Data-to-Text, Question Answering, Summarization, World Knowledge). While being 30x smaller, HALT outperforms Lettuce, a fine-tuned modernBERT-base encoder, achieving a 60x speedup gain on HUB. HALT and HUB together establish an effective framework for hallucination detection across diverse LLM capabilities.

Method: Controlled fairness audit using counterfactual prompt design evaluating GPT-4 and LLaMA-3.1-70B on career advice tasks while varying identity attributes (age, gender, nationality). Assessed access fairness through refusal analysis and measured interaction quality using automated linguistic metrics (sentiment, politeness, hedging). Used paired statistical tests to evaluate identity-conditioned differences.

Result: Both models exhibited zero refusal rates across all identities (uniform access). However, systematic model-specific disparities in interaction quality: GPT-4 expressed significantly higher hedging toward younger male users, while LLaMA exhibited broader sentiment variation across identity groups.

Conclusion: Fairness disparities can persist at the interaction level even when access is equal, motivating evaluation beyond refusal-based audits to include interaction quality metrics.

Abstract: Prior work on fairness in large language models (LLMs) has primarily focused on access-level behaviors such as refusals and safety filtering. However, equitable access does not ensure equitable interaction quality once a response is provided. In this paper, we conduct a controlled fairness audit examining how LLMs differ in tone, uncertainty, and linguistic framing across demographic identities after access is granted. Using a counterfactual prompt design, we evaluate GPT-4 and LLaMA-3.1-70B on career advice tasks while varying identity attributes along age, gender, and nationality. We assess access fairness through refusal analysis and measure interaction quality using automated linguistic metrics, including sentiment, politeness, and hedging. Identity-conditioned differences are evaluated using paired statistical tests. Both models exhibit zero refusal rates across all identities, indicating uniform access. Nevertheless, we observe systematic, model-specific disparities in interaction quality: GPT-4 expresses significantly higher hedging toward younger male users, while LLaMA exhibits broader sentiment variation across identity groups. These results show that fairness disparities can persist at the interaction level even when access is equal, motivating evaluation beyond refusal-based audits.

Method: Created SNIC (Situated Norms in Context), a human-validated diagnostic testbed to probe LLMs’ ability to extract and utilize normative principles for reference resolution in everyday tasks.

Result: Even state-of-the-art LLMs struggle to consistently identify and apply social norms, especially when norms are implicit, underspecified, or conflicting.

Conclusion: Current LLMs have a blind spot in social norm reasoning, highlighting a key challenge for deploying language-based systems in socially situated, embodied settings.

Abstract: Embodied agents, such as robots, will need to interact in situated environments where successful communication often depends on reasoning over social norms: shared expectations that constrain what actions are appropriate in context. A key capability in such settings is norm-based reference resolution (NBRR), where interpreting referential expressions requires inferring implicit normative expectations grounded in physical and social context. Yet it remains unclear whether Large Language Models (LLMs) can support this kind of reasoning. In this work, we introduce SNIC (Situated Norms in Context), a human-validated diagnostic testbed designed to probe how well state-of-the-art LLMs can extract and utilize normative principles relevant to NBRR. SNIC emphasizes physically grounded norms that arise in everyday tasks such as cleaning, tidying, and serving. Across a range of controlled evaluations, we find that even the strongest LLMs struggle to consistently identify and apply social norms, particularly when norms are implicit, underspecified, or in conflict. These findings reveal a blind spot in current LLMs and highlight a key challenge for deploying language-based systems in socially situated, embodied settings.

Method: A collaborative Coach-Player paradigm where Coach proposes instructions tailored to Player’s capability and receives rewards based on Player’s performance improvement, while Player is rewarded for solving increasingly instructive tasks. This creates a cooperative optimization loop without external data.

Result: CPMöbius achieves substantial improvements without external training data, outperforming existing unsupervised approaches. On Qwen2.5-Math-7B-Instruct, it improves accuracy by +4.9 overall average and +5.4 out-of-distribution average, exceeding RENT by +1.5 on overall accuracy and R-zero by +4.2 on OOD accuracy.

Conclusion: The collaborative Coach-Player paradigm enables effective data-free reinforcement learning for mathematical reasoning, offering a sustainable alternative to supervision-heavy training methods.

Abstract: Large Language Models (LLMs) have demonstrated strong potential in complex reasoning, yet their progress remains fundamentally constrained by reliance on massive high-quality human-curated tasks and labels, either through supervised fine-tuning (SFT) or reinforcement learning (RL) on reasoning-specific data. This dependence renders supervision-heavy training paradigms increasingly unsustainable, with signs of diminishing scalability already evident in practice. To overcome this limitation, we introduce CPMöbius (CPMobius), a collaborative Coach-Player paradigm for data-free reinforcement learning of reasoning models. Unlike traditional adversarial self-play, CPMöbius, inspired by real world human sports collaboration and multi-agent collaboration, treats the Coach and Player as independent but cooperative roles. The Coach proposes instructions targeted at the Player’s capability and receives rewards based on changes in the Player’s performance, while the Player is rewarded for solving the increasingly instructive tasks generated by the Coach. This cooperative optimization loop is designed to directly enhance the Player’s mathematical reasoning ability. Remarkably, CPMöbius achieves substantial improvement without relying on any external training data, outperforming existing unsupervised approaches. For example, on Qwen2.5-Math-7B-Instruct, our method improves accuracy by an overall average of +4.9 and an out-of-distribution average of +5.4, exceeding RENT by +1.5 on overall accuracy and R-zero by +4.2 on OOD accuracy.

Method: Proposes SAES-SVD with two components: 1) Cumulative Error-Aware Layer Compression (CEALC) formulates compression as local reconstruction plus weighted cumulative error compensation, deriving closed-form low-rank solution using second-order activation statistics; 2) Adaptive Collaborative Error Suppression (ACES) automatically adjusts weighting coefficients to enhance low-rank structure and maximize rank budget utilization.

Result: Extensive experiments across multiple LLM architectures and tasks show SAES-SVD consistently improves post-compression performance without fine-tuning or mixed-rank strategies.

Conclusion: SAES-SVD effectively addresses error accumulation in LLM compression by jointly optimizing intra-layer reconstruction and inter-layer error compensation, providing a more effective compression framework.

Abstract: The rapid growth in the parameter scale of large language models (LLMs) has created a high demand for efficient compression techniques. As a hardware-agnostic and highly compatible technique, low-rank compression has been widely adopted. However, existing methods typically compress each layer independently by minimizing per-layer reconstruction error, overlooking a critical limitation: the reconstruction error propagates and accumulates through the network, which leads to amplified global deviations from the full-precision baseline. To address this, we propose Self-Adaptive Error Suppression SVD (SAES-SVD), a LLMs compression framework that jointly optimizes intra-layer reconstruction and inter-layer error compensation. SAES-SVD is composed of two novel components: (1) Cumulative Error-Aware Layer Compression (CEALC), which formulates the compression objective as a combination of local reconstruction and weighted cumulative error compensation. Based on it, we derive a closed-form low-rank solution relied on second-order activation statistics, which explicitly aligns each layer’s output with its full-precision counterpart to compensate for accumulated errors. (2) Adaptive Collaborative Error Suppression (ACES), which automatically adjusts the weighting coefficient to enhance the low-rank structure of the compression objective in CEALC. Specifically, the coefficient is optimized to maximize the ratio between the Frobenius norm of the compressed layer’s output and that of the compression objective under a fixed rank, thus ensuring that the rank budget is utilized effectively. Extensive experiments across multiple LLM architectures and tasks show that, without fine-tuning or mixed-rank strategies, SAES-SVD consistently improves post-compression performance.

Method: ReMiT (Reinforcement Learning-Guided Mid-Training) leverages reasoning priors from RL-tuned models to dynamically reweight tokens during the critical mid-training (annealing) phase. This phase occurs at the end of pre-training with high-quality corpora and decaying learning rates. The method prioritizes tokens pivotal for reasoning without requiring specially trained teacher or reference models.

Result: ReMiT achieves average 3% improvement on 10 pre-training benchmarks across math, code, and general reasoning domains. These gains are sustained by over 2% throughout the post-training pipeline, validating the iterative feedback loop concept.

Conclusion: The paper demonstrates a successful bidirectional training approach that enables continuous, self-reinforcing evolution of LLMs through reinforcement learning-guided mid-training, establishing a flywheel effect where RL-tuned models improve base models which then enhance subsequent post-training performance.

Abstract: Standard training pipelines for large language models (LLMs) are typically unidirectional, progressing from pre-training to post-training. However, the potential for a bidirectional process–where insights from post-training retroactively improve the pre-trained foundation–remains unexplored. We aim to establish a self-reinforcing flywheel: a cycle in which reinforcement learning (RL)-tuned model strengthens the base model, which in turn enhances subsequent post-training performance, requiring no specially trained teacher or reference model. To realize this, we analyze training dynamics and identify the mid-training (annealing) phase as a critical turning point for model capabilities. This phase typically occurs at the end of pre-training, utilizing high-quality corpora under a rapidly decaying learning rate. Building upon this insight, we introduce ReMiT (Reinforcement Learning-Guided Mid-Training). Specifically, ReMiT leverages the reasoning priors of RL-tuned models to dynamically reweight tokens during the mid-training phase, prioritizing those pivotal for reasoning. Empirically, ReMiT achieves an average improvement of 3% on 10 pre-training benchmarks, spanning math, code, and general reasoning, and sustains these gains by over 2% throughout the post-training pipeline. These results validate an iterative feedback loop, enabling continuous and self-reinforcing evolution of LLMs.

Method: AERO uses a synergistic dual-loop system with internalized self-questioning, answering, and criticism. It employs entropy-based positioning inspired by Zone of Proximal Development theory to target the “solvability gap,” Independent Counterfactual Correction for robust verification, and a Staggered Training Strategy to synchronize capability growth.

Result: Extensive evaluations across nine benchmarks spanning three domains show average performance improvements of 4.57% on Qwen3-4B-Base and 5.10% on Qwen3-8B-Base, outperforming competitive baselines.

Conclusion: AERO successfully achieves autonomous reasoning evolution without expert supervision, addressing key limitations in existing self-evolution paradigms through its innovative dual-loop architecture and training strategies.

Abstract: Large Language Models (LLMs) have achieved significant success in complex reasoning but remain bottlenecked by reliance on expert-annotated data and external verifiers. While existing self-evolution paradigms aim to bypass these constraints, they often fail to identify the optimal learning zone and risk reinforcing collective hallucinations and incorrect priors through flawed internal feedback. To address these challenges, we propose \underline{A}utonomous \underline{E}volutionary \underline{R}easoning \underline{O}ptimization (AERO), an unsupervised framework that achieves autonomous reasoning evolution by internalizing self-questioning, answering, and criticism within a synergistic dual-loop system. Inspired by the \textit{Zone of Proximal Development (ZPD)} theory, AERO utilizes entropy-based positioning to target the ``solvability gap’’ and employs Independent Counterfactual Correction for robust verification. Furthermore, we introduce a Staggered Training Strategy to synchronize capability growth across functional roles and prevent curriculum collapse. Extensive evaluations across nine benchmarks spanning three domains demonstrate that AERO achieves average performance improvements of 4.57% on Qwen3-4B-Base and 5.10% on Qwen3-8B-Base, outperforming competitive baselines. Code is available at https://github.com/mira-ai-lab/AERO.

Method: Proposes Test-time Recursive Thinking (TRT), an iterative self-improvement framework that conditions generation on rollout-specific strategies, accumulated knowledge, and self-generated verification signals.

Result: Open-source models reach 100% accuracy on AIME-25/24, and closed-source models improve by 10.4-14.8 percentage points on LiveCodeBench’s most difficult problems without external feedback.

Conclusion: LLMs can effectively self-improve reasoning through iterative test-time frameworks like TRT without requiring additional training data or external supervision.

Abstract: Modern Large Language Models (LLMs) have shown rapid improvements in reasoning capabilities, driven largely by reinforcement learning (RL) with verifiable rewards. Here, we ask whether these LLMs can self-improve without the need for additional training. We identify two core challenges for such systems: (i) efficiently generating diverse, high-quality candidate solutions, and (ii) reliably selecting correct answers in the absence of ground-truth supervision. To address these challenges, we propose Test-time Recursive Thinking (TRT), an iterative self-improvement framework that conditions generation on rollout-specific strategies, accumulated knowledge, and self-generated verification signals. Using TRT, open-source models reach 100% accuracy on AIME-25/24, and on LiveCodeBench’s most difficult problems, closed-source models improve by 10.4-14.8 percentage points without external feedback.

Method: Introduces Task-Specificity Score (TSS) that quantifies how much an instruction matters for predicting its output by contrasting the true instruction against plausible alternatives for the same input. Also proposes TSS++ with hard alternatives and a small quality term to mitigate easy-negative effects.

Result: Across three instruction datasets (Alpaca, Dolly-15k, NI-20) and three open LLMs (Gemma, Llama, Qwen), selecting task-specific examples improves downstream performance under tight token budgets and complements quality-based filters like perplexity and IFD.

Conclusion: Task specificity is an important dimension for instruction tuning that complements existing quality metrics, and selecting task-specific examples can improve model performance especially under constrained training budgets.

Abstract: Instruction tuning is now the default way to train and adapt large language models, but many instruction–input–output pairs are only weakly specified: for a given input, the same output can remain plausible under several alternative instructions. This raises a simple question: \emph{does the instruction uniquely determine the target output?} We propose the \textbf{Task–Specificity Score (TSS)} to quantify how much an instruction matters for predicting its output, by contrasting the true instruction against plausible alternatives for the same input. We further introduce \textbf{TSS++}, which uses hard alternatives and a small quality term to mitigate easy-negative effects. Across three instruction datasets (\textsc{Alpaca}, \textsc{Dolly-15k}, \textsc{NI-20}) and three open LLMs (Gemma, Llama, Qwen), we show that selecting task-specific examples improves downstream performance under tight token budgets and complements quality-based filters such as perplexity and IFD.

Method: Constructed three-category dataset (politeness, impoliteness, mock politeness) using Rapport Management Theory and Model of Mock Politeness; tested six LLMs (including GPT-5.1 and DeepSeek) under zero-shot, few-shot, knowledge-enhanced, and hybrid prompting conditions.

Result: Performance differences among LLMs in recognizing Chinese politeness phenomena were systematically evaluated, though specific quantitative results are not provided in the abstract.

Conclusion: This study represents a meaningful attempt in “Great Linguistics” paradigm, offering novel approach to applying pragmatic theory in technological transformation and bridging linguistic technology with humanistic reflection.

Abstract: From a pragmatic perspective, this study systematically evaluates the differences in performance among representative large language models (LLMs) in recognizing politeness, impoliteness, and mock politeness phenomena in Chinese. Addressing the existing gaps in pragmatic comprehension, the research adopts the frameworks of Rapport Management Theory and the Model of Mock Politeness to construct a three-category dataset combining authentic and simulated Chinese discourse. Six representative models, including GPT-5.1 and DeepSeek, were selected as test subjects and evaluated under four prompting conditions: zero-shot, few-shot, knowledge-enhanced, and hybrid strategies. This study serves as a meaningful attempt within the paradigm of ``Great Linguistics,’’ offering a novel approach to applying pragmatic theory in the age of technological transformation. It also responds to the contemporary question of how technology and the humanities may coexist, representing an interdisciplinary endeavor that bridges linguistic technology and humanistic reflection.

Method: Created ChemPro benchmark with 4100 natural language question-answer pairs across 4 difficulty sections covering Bio-Chemistry, Inorganic-Chemistry, Organic-Chemistry and Physical-Chemistry. Includes Multiple Choice and Numerical Questions with balanced ratio of question types. Evaluated 45+7 state-of-the-art LLMs (open-source and proprietary).

Result: LLMs perform well on basic chemistry questions but accuracy declines with different types and levels of complexity. Findings highlight critical limitations in general scientific reasoning and understanding, pointing toward understudied dimensions of difficulty.

Conclusion: The benchmark reveals significant gaps in LLMs’ ability to handle complex scientific reasoning, emphasizing the need for more robust methodologies to improve LLMs’ scientific understanding capabilities.

Abstract: We introduce ChemPro, a progressive benchmark with 4100 natural language question-answer pairs in Chemistry, across 4 coherent sections of difficulty designed to assess the proficiency of Large Language Models (LLMs) in a broad spectrum of general chemistry topics. We include Multiple Choice Questions and Numerical Questions spread across fine-grained information recall, long-horizon reasoning, multi-concept questions, problem-solving with nuanced articulation, and straightforward questions in a balanced ratio, effectively covering Bio-Chemistry, Inorganic-Chemistry, Organic-Chemistry and Physical-Chemistry. ChemPro is carefully designed analogous to a student’s academic evaluation for basic to high-school chemistry. A gradual increase in the question difficulty rigorously tests the ability of LLMs to progress from solving basic problems to solving more sophisticated challenges. We evaluate 45+7 state-of-the-art LLMs, spanning both open-source and proprietary variants, and our analysis reveals that while LLMs perform well on basic chemistry questions, their accuracy declines with different types and levels of complexity. These findings highlight the critical limitations of LLMs in general scientific reasoning and understanding and point towards understudied dimensions of difficulty, emphasizing the need for more robust methodologies to improve LLMs.

Method: OMAR uses a reinforcement learning framework where a single model role-plays all participants in conversations simultaneously through self-play. To handle long dialogues, it implements hierarchical advantage estimation with turn-level and token-level advantages. Training occurs in social environments like SOTOPIA and Werewolf strategy games.

Result: The trained models develop emergent social intelligence including empathy, persuasion, and compromise seeking. They demonstrate effective collaboration even in competitive scenarios. While some practical challenges like reward hacking were identified, the results show rich social intelligence can emerge without human supervision.

Conclusion: OMAR demonstrates that AI can develop sophisticated social intelligence through multi-agent conversational self-play without human supervision. The framework successfully enables learning of complex social norms and long-term goals in dynamic interactions, paving the way for more socially intelligent AI systems.

Abstract: This paper introduces OMAR: One Model, All Roles, a reinforcement learning framework that enables AI to develop social intelligence through multi-turn, multi-agent conversational self-play. Unlike traditional paradigms that rely on static, single-turn optimizations, OMAR allows a single model to role-play all participants in a conversation simultaneously, learning to achieve long-term goals and complex social norms directly from dynamic social interaction. To ensure training stability across long dialogues, we implement a hierarchical advantage estimation that calculates turn-level and token-level advantages. Evaluations in the SOTOPIA social environment and Werewolf strategy games show that our trained models develop fine-grained, emergent social intelligence, such as empathy, persuasion, and compromise seeking, demonstrating the effectiveness of learning collaboration even under competitive scenarios. While we identify practical challenges like reward hacking, our results show that rich social intelligence can emerge without human supervision. We hope this work incentivizes further research on AI social intelligence in group conversations.

Method: CoSMo uses a split-merge algorithm to dynamically refine reasoning chains by merging redundant segments and splitting logical gaps. It employs structure-aligned reinforcement learning with a novel segment-level budget to supervise efficient reasoning structures during training.

Result: CoSMo achieves superior performance across multiple benchmarks and backbones, improving accuracy by 3.3 points while reducing segment usage by 28.7% on average compared to reasoning efficiency baselines.

Conclusion: CoSMo effectively optimizes reasoning efficiency in Large Reasoning Models by addressing structural redundancy, demonstrating significant improvements in both accuracy and computational efficiency.

Abstract: While Large Reasoning Models (LRMs) have demonstrated impressive capabilities in solving complex tasks through the generation of long reasoning chains, this reliance on verbose generation results in significant latency and computational overhead. To address these challenges, we propose \textbf{CoSMo} (\textbf{Co}nsistency-Guided \textbf{S}plit-\textbf{M}erge \textbf{O}ptimization), a framework designed to eliminate structural redundancy rather than indiscriminately restricting token volume. Specifically, CoSMo utilizes a split-merge algorithm that dynamically refines reasoning chains by merging redundant segments and splitting logical gaps to ensure coherence. We then employ structure-aligned reinforcement learning with a novel segment-level budget to supervise the model in maintaining efficient reasoning structures throughout training. Extensive experiments across multiple benchmarks and backbones demonstrate that CoSMo achieves superior performance, improving accuracy by \textbf{3.3} points while reducing segment usage by \textbf{28.7%} on average compared to reasoning efficiency baselines.

Method: FASA discovers functional sparsity at the frequency-chunk level in RoPE, identifying a small subset of “dominant” FCs that correlate with full attention. It uses these as a free computational proxy to predict token importance, then performs attention only on the pruned subset of critical tokens.

Result: FASA outperforms all token-eviction baselines across long-context tasks, achieving near-oracle accuracy. On LongBench-V1, it reaches nearly 100% of full-KV performance with only 256 tokens and achieves 2.56× speedup using 18.9% of cache on AIME24.

Conclusion: FASA provides an effective query-aware token eviction framework that significantly reduces KV cache memory requirements while maintaining performance, addressing a critical bottleneck in LLM deployment for long inputs.

Abstract: The deployment of Large Language Models (LLMs) faces a critical bottleneck when handling lengthy inputs: the prohibitive memory footprint of the Key Value (KV) cache. To address this bottleneck, the token pruning paradigm leverages attention sparsity to selectively retain a small, critical subset of tokens. However, existing approaches fall short, with static methods risking irreversible information loss and dynamic strategies employing heuristics that insufficiently capture the query-dependent nature of token importance. We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance. FASA stems from a novel insight into RoPE: the discovery of functional sparsity at the frequency-chunk (FC) level. Our key finding is that a small, identifiable subset of “dominant” FCs consistently exhibits high contextual agreement with the full attention head. This provides a robust and computationally free proxy for identifying salient tokens. %making them a powerful and efficient proxy for token importance. Building on this insight, FASA first identifies a critical set of tokens using dominant FCs, and then performs focused attention computation solely on this pruned subset. % Since accessing only a small fraction of the KV cache, FASA drastically lowers memory bandwidth requirements and computational cost. Across a spectrum of long-context tasks, from sequence modeling to complex CoT reasoning, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy, demonstrating remarkable robustness even under constraint budgets. Notably, on LongBench-V1, FASA reaches nearly 100% of full-KV performance when only keeping 256 tokens, and achieves 2.56$\times$ speedup using just 18.9% of the cache on AIME24.

Method: Created Privasis dataset from scratch with 1.4M records across diverse document types (medical, legal, financial, calendars, messages) with 55.1M annotated attributes. Used this to build parallel corpus for text sanitization via pipeline that decomposes texts and applies targeted sanitization.

Result: Compact sanitization models (≤4B parameters) trained on Privasis outperform state-of-the-art large language models like GPT-5 and Qwen-3 235B. Dataset offers orders-of-magnitude larger scale with quality and greater diversity than existing datasets.

Conclusion: Privasis enables privacy research at scale without compromising real user data, with compact models outperforming large LLMs, accelerating research in privacy-sensitive domains and AI agents handling personal information.

Abstract: Research involving privacy-sensitive data has always been constrained by data scarcity, standing in sharp contrast to other areas that have benefited from data scaling. This challenge is becoming increasingly urgent as modern AI agents–such as OpenClaw and Gemini Agent–are granted persistent access to highly sensitive personal information. To tackle this longstanding bottleneck and the rising risks, we present Privasis (i.e., privacy oasis), the first million-scale fully synthetic dataset entirely built from scratch–an expansive reservoir of texts with rich and diverse private information–designed to broaden and accelerate research in areas where processing sensitive social data is inevitable. Compared to existing datasets, Privasis, comprising 1.4 million records, offers orders-of-magnitude larger scale with quality, and far greater diversity across various document types, including medical history, legal documents, financial records, calendars, and text messages with a total of 55.1 million annotated attributes such as ethnicity, date of birth, workplace, etc. We leverage Privasis to construct a parallel corpus for text sanitization with our pipeline that decomposes texts and applies targeted sanitization. Our compact sanitization models (<=4B) trained on this dataset outperform state-of-the-art large language models, such as GPT-5 and Qwen-3 235B. We plan to release data, models, and code to accelerate future research on privacy-sensitive domains and agents.

Method: 1) Golden Eviction algorithm identifies optimal eviction KV pairs using future attention scores; 2) Supervised training with Pairwise Ranking Loss distills these traces; 3) Formulate cache eviction as Markov Decision Process and apply GRPO algorithm to mitigate language modeling loss on low-entropy tokens.

Result: Experiments on AIME2024 and AIME2025 benchmarks with three reasoning models show ForesightKV consistently outperforms prior methods under only half the cache budget, benefiting from both supervised and reinforcement learning approaches.

Conclusion: ForesightKV effectively balances efficiency and performance in KV cache management for long-text generation, demonstrating superior performance with reduced cache requirements through combined supervised and reinforcement learning techniques.

Abstract: Recently, large language models (LLMs) have shown remarkable reasoning abilities by producing long reasoning traces. However, as the sequence length grows, the key-value (KV) cache expands linearly, incurring significant memory and computation costs. Existing KV cache eviction methods mitigate this issue by discarding less important KV pairs, but often fail to capture complex KV dependencies, resulting in performance degradation. To better balance efficiency and performance, we introduce ForesightKV, a training-based KV cache eviction framework that learns to predict which KV pairs to evict during long-text generations. We first design the Golden Eviction algorithm, which identifies the optimal eviction KV pairs at each step using future attention scores. These traces and the scores at each step are then distilled via supervised training with a Pairwise Ranking Loss. Furthermore, we formulate cache eviction as a Markov Decision Process and apply the GRPO algorithm to mitigate the significant language modeling loss increase on low-entropy tokens. Experiments on AIME2024 and AIME2025 benchmarks of three reasoning models demonstrate that ForesightKV consistently outperforms prior methods under only half the cache budget, while benefiting synergistically from both supervised and reinforcement learning approaches.

Method: Proposes Token Sparse Attention, a lightweight dynamic token-level sparsification mechanism that compresses per-head Q, K, V matrices to a reduced token set during attention computation, then decompresses the output back to the original sequence. This allows token information to be reconsidered in subsequent layers. The method is compatible with dense attention implementations like Flash Attention and can be composed with existing sparse attention kernels.

Result: Achieves up to 3.23× attention speedup at 128K context length with less than 1% accuracy degradation. Consistently improves accuracy-latency trade-off compared to existing methods.

Conclusion: Dynamic and interleaved token-level sparsification is a complementary and effective strategy for scalable long-context inference, providing a new design point at the intersection of token selection and sparse attention.

Abstract: The quadratic complexity of attention remains the central bottleneck in long-context inference for large language models. Prior acceleration methods either sparsify the attention map with structured patterns or permanently evict tokens at specific layers, which can retain irrelevant tokens or rely on irreversible early decisions despite the layer-/head-wise dynamics of token importance. In this paper, we propose Token Sparse Attention, a lightweight and dynamic token-level sparsification mechanism that compresses per-head $Q$, $K$, $V$ to a reduced token set during attention and then decompresses the output back to the original sequence, enabling token information to be reconsidered in subsequent layers. Furthermore, Token Sparse Attention exposes a new design point at the intersection of token selection and sparse attention. Our approach is fully compatible with dense attention implementations, including Flash Attention, and can be seamlessly composed with existing sparse attention kernels. Experimental results show that Token Sparse Attention consistently improves accuracy-latency trade-off, achieving up to $\times$3.23 attention speedup at 128K context with less than 1% accuracy degradation. These results demonstrate that dynamic and interleaved token-level sparsification is a complementary and effective strategy for scalable long-context inference.

Method: ATACompressor uses a selective encoder to compress only task-relevant portions of long contexts, with an adaptive allocation controller that perceives relevant content length and adjusts compression rate accordingly.

Result: Outperforms existing methods on three QA datasets (HotpotQA, MSMARCO, SQUAD) in both compression efficiency and task performance, with ablation studies providing insights into key components.

Conclusion: Provides a scalable solution for long-context processing in LLMs by adaptively compressing based on task requirements while preserving essential information.

Abstract: Long-context inputs in large language models (LLMs) often suffer from the “lost in the middle” problem, where critical information becomes diluted or ignored due to excessive length. Context compression methods aim to address this by reducing input size, but existing approaches struggle with balancing information preservation and compression efficiency. We propose Adaptive Task-Aware Compressor (ATACompressor), which dynamically adjusts compression based on the specific requirements of the task. ATACompressor employs a selective encoder that compresses only the task-relevant portions of long contexts, ensuring that essential information is preserved while reducing unnecessary content. Its adaptive allocation controller perceives the length of relevant content and adjusts the compression rate accordingly, optimizing resource utilization. We evaluate ATACompressor on three QA datasets: HotpotQA, MSMARCO, and SQUAD-showing that it outperforms existing methods in terms of both compression efficiency and task performance. Our approach provides a scalable solution for long-context processing in LLMs. Furthermore, we perform a range of ablation studies and analysis experiments to gain deeper insights into the key components of ATACompressor.

Method: Introduces Prefill-Only Pruning (POP) with: 1) Virtual gate mechanism to analyze layer importance, revealing deep layers are critical for decode but redundant for prefill; 2) Independent Key-Value projections to maintain cache integrity during stage transitions; 3) Boundary handling strategy for first token accuracy.

Result: Achieves up to 1.37× speedup in prefill latency with minimal performance loss on Llama-3.1, Qwen3-VL, and Gemma-3 across diverse modalities, overcoming accuracy-efficiency trade-off of existing pruning methods.

Conclusion: POP demonstrates that stage-aware pruning can significantly improve inference efficiency for LLMs/VLMs while maintaining accuracy, by leveraging the asymmetric computational requirements of prefill vs decode stages.

Abstract: Large Language Models (LLMs) and Vision-Language Models (VLMs) have demonstrated remarkable capabilities. However, their deployment is hindered by significant computational costs. Existing structured pruning methods, while hardware-efficient, often suffer from significant accuracy degradation. In this paper, we argue that this failure stems from a stage-agnostic pruning approach that overlooks the asymmetric roles between the prefill and decode stages. By introducing a virtual gate mechanism, our importance analysis reveals that deep layers are critical for next-token prediction (decode) but largely redundant for context encoding (prefill). Leveraging this insight, we propose Prefill-Only Pruning (POP), a stage-aware inference strategy that safely omits deep layers during the computationally intensive prefill stage while retaining the full model for the sensitive decode stage. To enable the transition between stages, we introduce independent Key-Value (KV) projections to maintain cache integrity, and a boundary handling strategy to ensure the accuracy of the first generated token. Extensive experiments on Llama-3.1, Qwen3-VL, and Gemma-3 across diverse modalities demonstrate that POP achieves up to 1.37$\times$ speedup in prefill latency with minimal performance loss, effectively overcoming the accuracy-efficiency trade-off limitations of existing structured pruning methods.

Method: MIRROR uses a fine-tuning-free, end-to-end multi-agent framework with two core mechanisms: (1) execution-driven iterative adaptive revision for automatic error correction, and (2) hierarchical retrieval to fetch relevant modeling and coding exemplars from a curated library.

Result: MIRROR outperforms existing methods on standard OR benchmarks and achieves notable results on complex industrial datasets like IndustryOR and Mamo-ComplexLP.

Conclusion: By combining precise external knowledge infusion with systematic error correction, MIRROR provides non-expert users with an efficient and reliable OR modeling solution, overcoming limitations of general-purpose LLMs in expert optimization tasks.

Abstract: Operations Research (OR) relies on expert-driven modeling-a slow and fragile process ill-suited to novel scenarios. While large language models (LLMs) can automatically translate natural language into optimization models, existing approaches either rely on costly post-training or employ multi-agent frameworks, yet most still lack reliable collaborative error correction and task-specific retrieval, often leading to incorrect outputs. We propose MIRROR, a fine-tuning-free, end-to-end multi-agent framework that directly translates natural language optimization problems into mathematical models and solver code. MIRROR integrates two core mechanisms: (1) execution-driven iterative adaptive revision for automatic error correction, and (2) hierarchical retrieval to fetch relevant modeling and coding exemplars from a carefully curated exemplar library. Experiments show that MIRROR outperforms existing methods on standard OR benchmarks, with notable results on complex industrial datasets such as IndustryOR and Mamo-ComplexLP. By combining precise external knowledge infusion with systematic error correction, MIRROR provides non-expert users with an efficient and reliable OR modeling solution, overcoming the fundamental limitations of general-purpose LLMs in expert optimization tasks.

Method: Identifies a disruption-recovery tradeoff and proposes a pre-deployment test using a small pilot of 50 tasks to estimate whether intervention will help or harm without requiring full deployment.

Result: Intervention degrades performance on high-success tasks (0 to -26 pp) while yielding modest improvement on high-failure benchmarks (+2.8 pp, p=0.014). The test correctly anticipates outcomes across benchmarks.

Conclusion: The primary value is identifying when not to intervene, preventing severe regressions before deployment, as critic accuracy alone is insufficient for safe intervention decisions.

Abstract: Proactive interventions by LLM critic models are often assumed to improve reliability, yet their effects at deployment time are poorly understood. We show that a binary LLM critic with strong offline accuracy (AUROC 0.94) can nevertheless cause severe performance degradation, inducing a 26 percentage point (pp) collapse on one model while affecting another by near zero pp. This variability demonstrates that LLM critic accuracy alone is insufficient to determine whether intervention is safe. We identify a disruption-recovery tradeoff: interventions may recover failing trajectories but also disrupt trajectories that would have succeeded. Based on this insight, we propose a pre-deployment test that uses a small pilot of 50 tasks to estimate whether intervention is likely to help or harm, without requiring full deployment. Across benchmarks, the test correctly anticipates outcomes: intervention degrades performance on high-success tasks (0 to -26 pp), while yielding a modest improvement on the high-failure ALFWorld benchmark (+2.8 pp, p=0.014). The primary value of our framework is therefore identifying when not to intervene, preventing severe regressions before deployment.

Method: A two-stage RL framework where translation outputs are sampled to construct post-editing inputs, allowing return estimation in the post-editing stage to benefit from conditioning on current translation behavior. A task-specific weighting scheme balances translation and post-editing objectives.

Result: Experiments on English→Finnish, English→Turkish, and English↔Chinese show consistent gains over RL baselines. For English→Turkish, performance on COMET-KIWI is comparable to advanced LLM-based systems like DeepSeek-V3.2.

Conclusion: PEGRL effectively addresses RL challenges in machine translation by using post-editing as an auxiliary task, providing more stable training and better optimization through a two-stage approach with balanced objectives.

Abstract: Reinforcement learning (RL) has shown strong promise for LLM-based machine translation, with recent methods such as GRPO demonstrating notable gains; nevertheless, translation-oriented RL remains challenged by noisy learning signals arising from Monte Carlo return estimation, as well as a large trajectory space that favors global exploration over fine-grained local optimization. We introduce \textbf{PEGRL}, a \textit{two-stage} RL framework that uses post-editing as an auxiliary task to stabilize training and guide overall optimization. At each iteration, translation outputs are sampled to construct post-editing inputs, allowing return estimation in the post-editing stage to benefit from conditioning on the current translation behavior, while jointly supporting both global exploration and fine-grained local optimization. A task-specific weighting scheme further balances the contributions of translation and post-editing objectives, yielding a biased yet more sample-efficient estimator. Experiments on English$\to$Finnish, English$\to$Turkish, and English$\leftrightarrow$Chinese show consistent gains over RL baselines, and for English$\to$Turkish, performance on COMET-KIWI is comparable to advanced LLM-based systems (DeepSeek-V3.2).

Method: 1) Analyze each RAG component and propose practical alternatives, 2) Conduct systematic evaluations on three types of tasks to reveal best practices and trade-offs

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

Conclusion: Provides practical guidance for building effective RAG systems in industrial applications, particularly in medical domain where precision matters

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: Uses conditional mutual information (CMI) between teacher logits and input queries conditioned on labels to characterize distillation-relevant information. Proposes learning a transformation matrix to purify outputs, with a CMI-inspired anti-distillation objective to remove distillation information while preserving utility.

Result: Extensive experiments across multiple LLMs and distillation algorithms show the method significantly degrades distillation performance while preserving task accuracy, effectively protecting model IP.

Conclusion: The proposed information-theoretic approach provides effective defense against logit-based distillation attacks on proprietary LLMs, addressing a previously unexplored vulnerability.

Abstract: Proprietary large language models (LLMs) embody substantial economic value and are generally exposed only as black-box APIs, yet adversaries can still exploit their outputs to extract knowledge via distillation. Existing defenses focus exclusively on text-based distillation, leaving the important logit-based distillation largely unexplored. In this work, we analyze this problem and present an effective solution from an information-theoretic perspective. We characterize distillation-relevant information in teacher outputs using the conditional mutual information (CMI) between teacher logits and input queries conditioned on ground-truth labels. This quantity captures contextual information beneficial for model extraction, motivating us to defend distillation via CMI minimization. Guided by our theoretical analysis, we propose learning a transformation matrix that purifies the original outputs to enhance distillation resistance. We further derive a CMI-inspired anti-distillation objective to optimize this transformation, which effectively removes distillation-relevant information while preserving output utility. Extensive experiments across multiple LLMs and strong distillation algorithms demonstrate that the proposed method significantly degrades distillation performance while preserving task accuracy, effectively protecting models’ intellectual property.

Method: CSO starts from failed policy trajectories, uses a process reward model (PRM) to identify candidate critical steps, leverages expert models to propose alternative actions, continues execution with the policy model, and only uses successfully corrected trajectories as DPO training data. This focuses learning on verified critical steps.

Result: Experiments on GAIA-Text-103 and XBench-DeepSearch show CSO achieves 37% and 26% relative improvement over SFT baseline, substantially outperforms other post-training methods, and requires supervision at only 16% of trajectory steps.

Conclusion: CSO demonstrates the effectiveness of selective verification-based learning for agent post-training, providing fine-grained supervision at critical decisions while avoiding trajectory-level coarseness and step-level noise.

Abstract: As large language model agents tackle increasingly complex long-horizon tasks, effective post-training becomes critical. Prior work faces fundamental challenges: outcome-only rewards fail to precisely attribute credit to intermediate steps, estimated step-level rewards introduce systematic noise, and Monte Carlo sampling approaches for step reward estimation incur prohibitive computational cost. Inspired by findings that only a small fraction of high-entropy tokens drive effective RL for reasoning, we propose Critical Step Optimization (CSO), which focuses preference learning on verified critical steps, decision points where alternate actions demonstrably flip task outcomes from failure to success. Crucially, our method starts from failed policy trajectories rather than expert demonstrations, directly targeting the policy model’s weaknesses. We use a process reward model (PRM) to identify candidate critical steps, leverage expert models to propose high-quality alternatives, then continue execution from these alternatives using the policy model itself until task completion. Only alternatives that the policy successfully executes to correct outcomes are verified and used as DPO training data, ensuring both quality and policy reachability. This yields fine-grained, verifiable supervision at critical decisions while avoiding trajectory-level coarseness and step-level noise. Experiments on GAIA-Text-103 and XBench-DeepSearch show that CSO achieves 37% and 26% relative improvement over the SFT baseline and substantially outperforms other post-training methods, while requiring supervision at only 16% of trajectory steps. This demonstrates the effectiveness of selective verification-based learning for agent post-training.

Method: FactNet employs a strictly deterministic construction pipeline that extracts 1.7B atomic assertions with 3.01B evidence pointers from 316 Wikipedia editions, ensuring every evidence unit is recoverable with byte-level precision (unlike synthetic approaches).

Result: Extensive auditing confirms high grounding precision of 92.1%, even in long-tail languages. The resource includes FactNet-Bench, a comprehensive evaluation suite for Knowledge Graph Completion, Question Answering, and Fact Checking.

Conclusion: FactNet provides a foundational, reproducible resource for training and evaluating trustworthy, verifiable multilingual systems, bridging the gap between structured knowledge and grounded text at massive scale.

Abstract: While LLMs exhibit remarkable fluency, their utility is often compromised by factual hallucinations and a lack of traceable provenance. Existing resources for grounding mitigate this but typically enforce a dichotomy: they offer either structured knowledge without textual context (e.g., knowledge bases) or grounded text with limited scale and linguistic coverage. To bridge this gap, we introduce FactNet, a massive, open-source resource designed to unify 1.7 billion atomic assertions with 3.01 billion auditable evidence pointers derived exclusively from 316 Wikipedia editions. Unlike recent synthetic approaches, FactNet employs a strictly deterministic construction pipeline, ensuring that every evidence unit is recoverable with byte-level precision. Extensive auditing confirms a high grounding precision of 92.1%, even in long-tail languages. Furthermore, we establish FactNet-Bench, a comprehensive evaluation suite for Knowledge Graph Completion, Question Answering, and Fact Checking. FactNet provides the community with a foundational, reproducible resource for training and evaluating trustworthy, verifiable multilingual systems.

Method: A-RAG exposes hierarchical retrieval interfaces directly to the model, providing three retrieval tools: keyword search, semantic search, and chunk read. This allows the agent to adaptively search and retrieve information across multiple granularities based on the task needs.

Result: Experiments on multiple open-domain QA benchmarks show A-RAG consistently outperforms existing approaches with comparable or lower retrieved tokens. The framework effectively leverages model capabilities and dynamically adapts to different RAG tasks.

Conclusion: A-RAG demonstrates that giving language models direct control over retrieval decisions enables more efficient information gathering and better performance. The paper also systematically studies how A-RAG scales with model size and test-time compute.

Abstract: Frontier language models have demonstrated strong reasoning and long-horizon tool-use capabilities. However, existing RAG systems fail to leverage these capabilities. They still rely on two paradigms: (1) designing an algorithm that retrieves passages in a single shot and concatenates them into the model’s input, or (2) predefining a workflow and prompting the model to execute it step-by-step. Neither paradigm allows the model to participate in retrieval decisions, preventing efficient scaling with model improvements. In this paper, we introduce A-RAG, an Agentic RAG framework that exposes hierarchical retrieval interfaces directly to the model. A-RAG provides three retrieval tools: keyword search, semantic search, and chunk read, enabling the agent to adaptively search and retrieve information across multiple granularities. Experiments on multiple open-domain QA benchmarks show that A-RAG consistently outperforms existing approaches with comparable or lower retrieved tokens, demonstrating that A-RAG effectively leverages model capabilities and dynamically adapts to different RAG tasks. We further systematically study how A-RAG scales with model size and test-time compute. We will release our code and evaluation suite to facilitate future research. Code and evaluation suite are available at https://github.com/Ayanami0730/arag.

Method: Train models on Swedish from scratch and by fine-tuning English-pretrained models, probing preferences at checkpoints using minimal pairs. Adapt existing benchmarks for linguistic acceptability and introduce two new datasets for idiomaticity: conventionalized idioms vs. variants, and idiomatic Swedish vs. Translationese.

Result: Idiomatic competence emerges more slowly than other linguistic abilities. Longer training yields diminishing returns for most tasks but idiom performance continues improving, especially in largest models (8B). Instruction tuning on machine-translated English data causes rapid loss of idiomatic language preference.

Conclusion: Models develop idiomatic understanding gradually, but standard instruction tuning approaches using machine-translated data can degrade idiomatic competence, highlighting challenges for multilingual model development.

Abstract: In this study, we investigate how language models develop preferences for \textit{idiomatic} as compared to \textit{linguistically acceptable} Swedish, both during pretraining and when adapting a model from English to Swedish. To do so, we train models on Swedish from scratch and by fine-tuning English-pretrained models, probing their preferences at various checkpoints using minimal pairs that differ in linguistic acceptability or idiomaticity. For linguistic acceptability, we adapt existing benchmarks into a minimal-pair format. To assess idiomaticity, we introduce two novel datasets: one contrasting conventionalized idioms with plausible variants, and another contrasting idiomatic Swedish with Translationese. Our findings suggest that idiomatic competence emerges more slowly than other linguistic abilities, including grammatical and lexical correctness. While longer training yields diminishing returns for most tasks, idiom-related performance continues to improve, particularly in the largest model tested (8B). However, instruction tuning on data machine-translated from English – the common approach for languages with little or no native instruction data – causes models to rapidly lose their preference for idiomatic language.

Method: Experience-driven test-time framework that detects recheck activation, consults offline experience pool of past verification outcomes, estimates if recheck is likely unnecessary via efficient retrieval, and suppresses unnecessary rechecks.

Result: Reduces token usage up to 20.3% while maintaining accuracy across multiple models and benchmarks; some datasets even show accuracy improvements.

Conclusion: Self-verification in LRMs is often overused; the proposed experience-driven suppression framework effectively reduces unnecessary verification steps while preserving reasoning quality.

Abstract: Large Reasoning Models (LRMs) achieve strong performance by generating long reasoning traces with reflection. Through a large-scale empirical analysis, we find that a substantial fraction of reflective steps consist of self-verification (recheck) that repeatedly confirm intermediate results. These rechecks occur frequently across models and benchmarks, yet the vast majority are confirmatory rather than corrective, rarely identifying errors and altering reasoning outcomes. This reveals a mismatch between how often self-verification is activated and how often it is actually useful. Motivated by this, we propose a novel, experience-driven test-time framework that reduces the overused verification. Our method detects the activation of recheck behavior, consults an offline experience pool of past verification outcomes, and estimates whether a recheck is likely unnecessary via efficient retrieval. When historical experience suggests unnecessary, a suppression signal redirects the model to proceed. Across multiple model and benchmarks, our approach reduces token usage up to 20.3% while maintaining the accuracy, and in some datasets even yields accuracy improvements.

Method: Proposes FaithRL framework with: 1) Formalized faithfulness-maximization objective, 2) Geometric reward design, 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. Further analysis confirms increased step-wise reasoning faithfulness and robust generalization.

Conclusion: FaithRL provides an effective reinforcement learning framework that directly optimizes reasoning faithfulness, addressing key limitations of sparse reward RLVR pipelines and reducing hallucinations in LLM reasoning tasks.

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 a two-stage data curriculum: first train on symbolic representations (code or graphs), then train on natural language data. This approach helps models learn to generalize across different representations for procedural tasks like scheduling and planning.

Result: The curriculum substantially improves model performance across model families and tasks. Remarkably, a 1.5B Qwen model trained with this method can closely match zero-shot GPT-4o in naturalistic planning tasks.

Conclusion: Successful cross-representation generalization can be interpreted as a form of generative analogy, which the proposed curriculum effectively encourages. The method bridges the gap between symbolic and natural language representations for procedural tasks.

Abstract: Large language models (LLMs) are trained and tested extensively on symbolic representations such as code and graphs, yet real-world user tasks are often specified in natural language. To what extent can LLMs generalize across these representations? Here, we approach this question by studying isomorphic tasks involving procedures represented in code, graphs, and natural language (e.g., scheduling steps in planning). We find that training LLMs with popular post-training methods on graphs or code data alone does not reliably generalize to corresponding natural language tasks, while training solely on natural language can lead to inefficient performance gains. To address this gap, we propose a two-stage data curriculum that first trains on symbolic, then natural language data. The curriculum substantially improves model performance across model families and tasks. Remarkably, a 1.5B Qwen model trained by our method can closely match zero-shot GPT-4o in naturalistic planning. Finally, our analysis suggests that successful cross-representation generalization can be interpreted as a form of generative analogy, which our curriculum effectively encourages.

Method: SEAD decouples user modeling into two components: a Profile Controller that generates diverse user states to manage training curriculum, and a User Role-play Model that focuses on realistic role-playing, ensuring adaptive training scenarios rather than adversarial environments.

Result: SEAD significantly outperforms both open-source foundation models and closed-source commercial models, improving task completion rate by 17.6% and dialogue efficiency by 11.1%.

Conclusion: The SEAD framework enables agents to learn effective service dialogue strategies without large-scale human annotations, addressing data scarcity and simulation challenges through self-evolving mechanisms.

Abstract: Large Language Models have demonstrated remarkable capabilities in open-domain dialogues. However, current methods exhibit suboptimal performance in service dialogues, as they rely on noisy, low-quality human conversation data. This limitation arises from data scarcity and the difficulty of simulating authentic, goal-oriented user behaviors. To address these issues, we propose SEAD (Self-Evolving Agent for Service Dialogue), a framework that enables agents to learn effective strategies without large-scale human annotations. SEAD decouples user modeling into two components: a Profile Controller that generates diverse user states to manage training curriculum, and a User Role-play Model that focuses on realistic role-playing. This design ensures the environment provides adaptive training scenarios rather than acting as an unfair adversary. Experiments demonstrate that SEAD significantly outperforms Open-source Foundation Models and Closed-source Commercial Models, improving task completion rate by 17.6% and dialogue efficiency by 11.1%. Code is available at: https://github.com/Da1yuqin/SEAD.

Method: Analyze two large pre-trained multilingual translation models (NLLB-200 and Tower+), examine translation quality across broad language set, control for factors like data resourcedness and writing script, study how typology affects decoding strategies.

Result: Target language typology drives translation quality in both encoder-decoder and decoder-only models; certain typological properties benefit more from wider output space search, suggesting alternative decoding strategies could help.

Conclusion: Typological properties are significant factors in multilingual model performance, and language-specific decoding strategies could improve translation for languages with certain typological features.

Abstract: Despite major advances in multilingual modeling, large quality disparities persist across languages. Besides the obvious impact of uneven training resources, typological properties have also been proposed to determine the intrinsic difficulty of modeling a language. The existing evidence, however, is mostly based on small monolingual language models or bilingual translation models trained from scratch. We expand on this line of work by analyzing two large pre-trained multilingual translation models, NLLB-200 and Tower+, which are state-of-the-art representatives of encoder-decoder and decoder-only machine translation, respectively. Based on a broad set of languages, we find that target language typology drives translation quality of both models, even after controlling for more trivial factors, such as data resourcedness and writing script. Additionally, languages with certain typological properties benefit more from a wider search of the output space, suggesting that such languages could profit from alternative decoding strategies beyond the standard left-to-right beam search. To facilitate further research in this area, we release a set of fine-grained typological properties for 212 languages of the FLORES+ MT evaluation benchmark.

Method: HySparse interleaves each full attention layer with several sparse attention layers. It strategically derives each sparse layer’s token selection and KV caches directly from the preceding full attention layer, using the full attention as a precise oracle to identify important tokens. This enables sparse attention layers to reuse the full attention KV cache.

Result: HySparse consistently outperforms both full attention and hybrid SWA baselines across 7B dense and 80B MoE models. In the 80B MoE model with 49 total layers, only 5 layers use full attention, yet HySparse achieves substantial performance gains while reducing KV cache storage by nearly 10x.

Conclusion: HySparse provides an effective architecture that overcomes fundamental limitations of prior sparse attention methods by using full attention as an oracle for token selection and enabling KV cache reuse, achieving better performance with significantly reduced memory requirements.

Abstract: This work introduces Hybrid Sparse Attention (HySparse), a new architecture that interleaves each full attention layer with several sparse attention layers. While conceptually simple, HySparse strategically derives each sparse layer’s token selection and KV caches directly from the preceding full attention layer. This architecture resolves two fundamental limitations of prior sparse attention methods. First, conventional approaches typically rely on additional proxies to predict token importance, introducing extra complexity and potentially suboptimal performance. In contrast, HySparse uses the full attention layer as a precise oracle to identify important tokens. Second, existing sparse attention designs often reduce computation without saving KV cache. HySparse enables sparse attention layers to reuse the full attention KV cache, thereby reducing both computation and memory. We evaluate HySparse on both 7B dense and 80B MoE models. Across all settings, HySparse consistently outperforms both full attention and hybrid SWA baselines. Notably, in the 80B MoE model with 49 total layers, only 5 layers employ full attention, yet HySparse achieves substantial performance gains while reducing KV cache storage by nearly 10x.

Method: Proposes Aligned Contrastive Learning (ACL) with three components: 1) ACL-Embed treats label embeddings as augmented samples and aligns them with sample representations; 2) ACL-Grad selectively discards ACL-Embed terms when objectives conflict; 3) ACL-CL uses teacher exits to guide shallow student exits in multi-exit BERT models.

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

Conclusion: The proposed ACL framework successfully resolves conflicts between contrastive learning and cross-entropy loss in supervised settings, enabling effective application of contrastive learning to improve model performance and efficiency in NLP tasks.

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 EA-GraphRAG with three components: 1) syntactic feature constructor that parses queries and extracts structural features, 2) lightweight complexity scorer that maps features to continuous complexity scores, 3) score-driven routing policy that selects dense RAG for low-score queries, graph-based retrieval for high-score queries, and complexity-aware reciprocal rank fusion for borderline cases.

Result: Extensive experiments on comprehensive benchmark (two single-hop and two multi-hop QA benchmarks) demonstrate significant improvements in accuracy, reduced latency, and state-of-the-art performance in handling mixed scenarios of simple and complex queries.

Conclusion: EA-GraphRAG effectively addresses the limitations of rigid GraphRAG application by dynamically adapting retrieval strategies based on query complexity, achieving optimal performance across diverse query types.

Abstract: Large language models (LLMs) often struggle with knowledge-intensive tasks due to hallucinations and outdated parametric knowledge. While Retrieval-Augmented Generation (RAG) addresses this by integrating external corpora, its effectiveness is limited by fragmented information in unstructured domain documents. Graph-augmented RAG (GraphRAG) emerged to enhance contextual reasoning through structured knowledge graphs, yet paradoxically underperforms vanilla RAG in real-world scenarios, exhibiting significant accuracy drops and prohibitive latency despite gains on complex queries. We identify the rigid application of GraphRAG to all queries, regardless of complexity, as the root cause. To resolve this, we propose an efficient and adaptive GraphRAG framework called EA-GraphRAG that dynamically integrates RAG and GraphRAG paradigms through syntax-aware complexity analysis. Our approach introduces: (i) a syntactic feature constructor that parses each query and extracts a set of structural features; (ii) a lightweight complexity scorer that maps these features to a continuous complexity score; and (iii) a score-driven routing policy that selects dense RAG for low-score queries, invokes graph-based retrieval for high-score queries, and applies complexity-aware reciprocal rank fusion to handle borderline cases. Extensive experiments on a comprehensive benchmark, consisting of two single-hop and two multi-hop QA benchmarks, demonstrate that our EA-GraphRAG significantly improves accuracy, reduces latency, and achieves state-of-the-art performance in handling mixed scenarios involving both simple and complex queries.

Method: V₀ treats policy capability as explicit context input using history of instruction-performance pairs. It focuses on value estimation at State Zero (initial prompt) and serves as a resource scheduler for GRPO training and model routing during deployment.

Result: V₀ significantly outperforms heuristic budget allocation and achieves Pareto-optimal trade-off between performance and cost in LLM routing tasks, demonstrating effective value estimation without parameter updates.

Conclusion: The V₀ model provides an efficient alternative to traditional value estimation methods by using policy capability as context, enabling better resource allocation in RL training and cost-effective model routing in deployment.

Abstract: Policy gradient methods rely on a baseline to measure the relative advantage of an action, ensuring the model reinforces behaviors that outperform its current average capability. In the training of Large Language Models (LLMs) using Actor-Critic methods (e.g., PPO), this baseline is typically estimated by a Value Model (Critic) often as large as the policy model itself. However, as the policy continuously evolves, the value model requires expensive, synchronous incremental training to accurately track the shifting capabilities of the policy. To avoid this overhead, Group Relative Policy Optimization (GRPO) eliminates the coupled value model by using the average reward of a group of rollouts as the baseline; yet, this approach necessitates extensive sampling to maintain estimation stability. In this paper, we propose $V_0$, a Generalist Value Model capable of estimating the expected performance of any model on unseen prompts without requiring parameter updates. We reframe value estimation by treating the policy’s dynamic capability as an explicit context input; specifically, we leverage a history of instruction-performance pairs to dynamically profile the model, departing from the traditional paradigm that relies on parameter fitting to perceive capability shifts. Focusing on value estimation at State Zero (i.e., the initial prompt, hence $V_0$), our model serves as a critical resource scheduler. During GRPO training, $V_0$ predicts success rates prior to rollout, allowing for efficient sampling budget allocation; during deployment, it functions as a router, dispatching instructions to the most cost-effective and suitable model. Empirical results demonstrate that $V_0$ significantly outperforms heuristic budget allocation and achieves a Pareto-optimal trade-off between performance and cost in LLM routing tasks.

Method: Introduces CL-bench with 500 complex contexts, 1,899 tasks, and 31,607 verification rubrics crafted by domain experts. Each task requires learning new content from context including domain-specific knowledge, rule systems, complex procedures, and laws derived from empirical data.

Result: Evaluation of ten frontier LMs shows they solve only 17.2% of tasks on average, with the best model (GPT-5.1) solving only 23.7%, revealing LMs have not achieved effective context learning.

Conclusion: Context learning is a critical bottleneck for real-world applications, and CL-bench represents a step toward building LMs with this fundamental capability to make them more intelligent and deployable in real-world scenarios.

Abstract: Current language models (LMs) excel at reasoning over prompts using pre-trained knowledge. However, real-world tasks are far more complex and context-dependent: models must learn from task-specific context and leverage new knowledge beyond what is learned during pre-training to reason and resolve tasks. We term this capability context learning, a crucial ability that humans naturally possess but has been largely overlooked. To this end, we introduce CL-bench, a real-world benchmark consisting of 500 complex contexts, 1,899 tasks, and 31,607 verification rubrics, all crafted by experienced domain experts. Each task is designed such that the new content required to resolve it is contained within the corresponding context. Resolving tasks in CL-bench requires models to learn from the context, ranging from new domain-specific knowledge, rule systems, and complex procedures to laws derived from empirical data, all of which are absent from pre-training. This goes far beyond long-context tasks that primarily test retrieval or reading comprehension, and in-context learning tasks, where models learn simple task patterns via instructions and demonstrations. Our evaluations of ten frontier LMs find that models solve only 17.2% of tasks on average. Even the best-performing model, GPT-5.1, solves only 23.7%, revealing that LMs have yet to achieve effective context learning, which poses a critical bottleneck for tackling real-world, complex context-dependent tasks. CL-bench represents a step towards building LMs with this fundamental capability, making them more intelligent and advancing their deployment in real-world scenarios.

Method: The paper presents a general algorithm for PCSPs over Series-Parallel-Loop (SPL) decompositions of control-flow graphs. The algorithm has time complexity O(|G|·|D|⁶), where |G| is graph size and |D| is domain size, yielding linear-time solutions for fixed domains. This unifies previous SPL-based approaches for specific compiler optimization problems.

Result: The algorithm achieves runtimes four times better than previous state-of-the-art for the Optimal Bank Selection problem. For fixed domains, it provides linear-time solutions to PCSPs over control-flow graphs.

Conclusion: The work provides a unified, efficient algorithm for solving PCSPs over control-flow graphs using SPL decompositions, generalizing previous specialized approaches and demonstrating practical improvements for compiler optimization tasks.

Abstract: In this work, we focus on the Partial Constraint Satisfaction Problem (PCSP) over control-flow graphs (CFGs) of programs. PCSP serves as a generalization of the well-known Constraint Satisfaction Problem (CSP). In the CSP framework, we define a set of variables, a set of constraints, and a finite domain $D$ that encompasses all possible values for each variable. The objective is to assign a value to each variable in such a way that all constraints are satisfied. In the graph variant of CSP, an underlying graph is considered and we have one variable corresponding to each vertex of the graph and one or several constraints corresponding to each edge. In PCSPs, we allow for certain constraints to be violated at a specified cost, aiming to find a solution that minimizes the total cost. Numerous classical compiler optimization tasks can be framed as PCSPs over control-flow graphs. Examples include Register Allocation, Lifetime-optimal Speculative Partial Redundancy Elimination (LOSPRE), and Optimal Placement of Bank Selection Instructions. On the other hand, it is well-known that control-flow graphs of structured programs are sparse and decomposable in a variety of ways. In this work, we rely on the Series-Parallel-Loop (SPL) decompositions as introduced by~\cite{RegisterAllocation}. Our main contribution is a general algorithm for PCSPs over SPL graphs with a time complexity of (O(|G| \cdot |D|^6)), where (|G|) represents the size of the control-flow graph. Note that for any fixed domain $D,$ this yields a linear-time solution. Our algorithm can be seen as a generalization and unification of previous SPL-based approaches for register allocation and LOSPRE. In addition, we provide experimental results over another classical PCSP task, i.e. Optimal Bank Selection, achieving runtimes four times better than the previous state of the art.

Method: CORE optimizes retrieved content by appending three types of optimization content: string-based, reasoning-based, and review-based. It targets search engine content since LLM interactions are black-box. Evaluated on ProductBench benchmark with 15 categories and 200 products each.

Result: CORE achieves average Promotion Success Rate of 91.4% @Top-5, 86.6% @Top-3, and 80.3% @Top-1 across 15 product categories on four LLMs (GPT-4o, Gemini-2.5, Claude-4, Grok-3), outperforming existing ranking manipulation methods while preserving content fluency.

Conclusion: CORE effectively controls output rankings in LLM-based search engines through content optimization, addressing fairness concerns in product recommendations while maintaining recommendation quality.

Abstract: The way customers search for and choose products is changing with the rise of large language models (LLMs). LLM-based search, or generative engines, provides direct product recommendations to users, rather than traditional online search results that require users to explore options themselves. However, these recommendations are strongly influenced by the initial retrieval order of LLMs, which disadvantages small businesses and independent creators by limiting their visibility. In this work, we propose CORE, an optimization method that \textbf{C}ontrols \textbf{O}utput \textbf{R}ankings in g\textbf{E}nerative Engines for LLM-based search. Since the LLM’s interactions with the search engine are black-box, CORE targets the content returned by search engines as the primary means of influencing output rankings. Specifically, CORE optimizes retrieved content by appending strategically designed optimization content to steer the ranking of outputs. We introduce three types of optimization content: string-based, reasoning-based, and review-based, demonstrating their effectiveness in shaping output rankings. To evaluate CORE in realistic settings, we introduce ProductBench, a large-scale benchmark with 15 product categories and 200 products per category, where each product is associated with its top-10 recommendations collected from Amazon’s search interface. Extensive experiments on four LLMs with search capabilities (GPT-4o, Gemini-2.5, Claude-4, and Grok-3) demonstrate that CORE achieves an average Promotion Success Rate of \textbf{91.4% @Top-5}, \textbf{86.6% @Top-3}, and \textbf{80.3% @Top-1}, across 15 product categories, outperforming existing ranking manipulation methods while preserving the fluency of optimized content.

Method: 1) Construct dataset of DeepResearch queries with human preference annotations over report pairs; 2) Train rubric generators via reinforcement learning with hybrid reward combining human preference supervision and LLM-based rubric evaluation; 3) Introduce Multi-agent Markov-state workflow for better long-horizon reasoning in report generation.

Result: The rubric generators deliver more discriminative and better human-aligned supervision than existing strategies. When integrated into the MaMs framework, DeepResearch systems outperform all open-source baselines on DeepResearch Bench and achieve performance comparable to leading closed-source models.

Conclusion: The proposed pipeline successfully addresses the rubric generation challenge for DeepResearch evaluation, providing scalable, query-specific rubrics that align with human preferences and improve system performance.

Abstract: Nowadays, training and evaluating DeepResearch-generated reports remain challenging due to the lack of verifiable reward signals. Accordingly, rubric-based evaluation has become a common practice. However, existing approaches either rely on coarse, pre-defined rubrics that lack sufficient granularity, or depend on manually constructed query-specific rubrics that are costly and difficult to scale. In this paper, we propose a pipeline to train human-preference-aligned query-specific rubric generators tailored for DeepResearch report generation. We first construct a dataset of DeepResearch-style queries annotated with human preferences over paired reports, and train rubric generators via reinforcement learning with a hybrid reward combining human preference supervision and LLM-based rubric evaluation. To better handle long-horizon reasoning, we further introduce a Multi-agent Markov-state (MaMs) workflow for report generation. We empirically show that our proposed rubric generators deliver more discriminative and better human-aligned supervision than existing rubric design strategies. Moreover, when integrated into the MaMs training framework, DeepResearch systems equipped with our rubric generators consistently outperform all open-source baselines on the DeepResearch Bench and achieve performance comparable to that of leading closed-source models.

Method: Created BIRDTurk through controlled translation pipeline adapting schema identifiers to Turkish while preserving SQL logical structure and execution semantics. Translation quality validated using Central Limit Theorem sampling for 95% confidence. Evaluated three approaches: inference-based prompting, agentic multi-stage reasoning, and supervised fine-tuning.

Result: Turkish introduces consistent performance degradation compared to English, driven by structural linguistic divergence and underrepresentation in LLM pretraining. Agentic reasoning demonstrates stronger cross-lingual robustness. Supervised fine-tuning remains challenging for standard multilingual baselines but scales effectively with modern instruction-tuned models.

Conclusion: BIRDTurk provides a controlled testbed for cross-lingual Text-to-SQL evaluation under realistic database conditions. Turkish performance degradation highlights challenges in low-resource languages, with agentic reasoning showing promise for cross-lingual robustness.

Abstract: Text-to-SQL systems have achieved strong performance on English benchmarks, yet their behavior in morphologically rich, low-resource languages remains largely unexplored. We introduce BIRDTurk, the first Turkish adaptation of the BIRD benchmark, constructed through a controlled translation pipeline that adapts schema identifiers to Turkish while strictly preserving the logical structure and execution semantics of SQL queries and databases. Translation quality is validated on a sample size determined by the Central Limit Theorem to ensure 95% confidence, achieving 98.15% accuracy on human-evaluated samples. Using BIRDTurk, we evaluate inference-based prompting, agentic multi-stage reasoning, and supervised fine-tuning. Our results reveal that Turkish introduces consistent performance degradation, driven by both structural linguistic divergence and underrepresentation in LLM pretraining, while agentic reasoning demonstrates stronger cross-lingual robustness. Supervised fine-tuning remains challenging for standard multilingual baselines but scales effectively with modern instruction-tuned models. BIRDTurk provides a controlled testbed for cross-lingual Text-to-SQL evaluation under realistic database conditions. We release the training and development splits to support future research.

Method: Proposes Trust Region Entropy (TRE) that encourages exploration strictly within the model’s trust region, preventing probability dilution into the vast tail of invalid tokens while maintaining exploration of plausible candidates.

Result: Extensive experiments across mathematical reasoning (MATH), combinatorial search (Countdown), and preference alignment (HH) tasks show TRE consistently outperforms vanilla PPO, standard entropy regularization, and other exploration baselines.

Conclusion: TRE effectively addresses the cumulative tail risk problem in LLM RL by restricting exploration to the trust region, leading to better performance on diverse reasoning and alignment tasks compared to standard approaches.

Abstract: Entropy regularization is a standard technique in reinforcement learning (RL) to enhance exploration, yet it yields negligible effects or even degrades performance in Large Language Models (LLMs). We attribute this failure to the cumulative tail risk inherent to LLMs with massive vocabularies and long generation horizons. In such environments, standard global entropy maximization indiscriminately dilutes probability mass into the vast tail of invalid tokens rather than focusing on plausible candidates, thereby disrupting coherent reasoning. To address this, we propose Trust Region Entropy (TRE), a method that encourages exploration strictly within the model’s trust region. Extensive experiments across mathematical reasoning (MATH), combinatorial search (Countdown), and preference alignment (HH) tasks demonstrate that TRE consistently outperforms vanilla PPO, standard entropy regularization, and other exploration baselines. Our code is available at https://github.com/WhyChaos/TRE-Encouraging-Exploration-in-the-Trust-Region.

Method: Constructed a Turkish RAG dataset from Turkish Wikipedia and CulturaX, then benchmarked seven stages of the RAG pipeline (query transformation, reranking, answer refinement) without task-specific fine-tuning.

Result: HyDE achieved highest accuracy (85%) vs baseline (78.70%). Pareto-optimal configuration using Cross-encoder Reranking and Context Augmentation achieved comparable performance (84.60%) with lower cost. Over-stacking generative modules degraded performance by distorting morphological cues.

Conclusion: Simple query clarification with robust reranking is effective for Turkish RAG. Complex methods can maximize accuracy but simpler configurations offer better cost-performance tradeoffs. Morphological richness requires careful pipeline design.

Abstract: Retrieval-Augmented Generation (RAG) enhances LLM factuality, yet design guidance remains English-centric, limiting insights for morphologically rich languages like Turkish. We address this by constructing a comprehensive Turkish RAG dataset derived from Turkish Wikipedia and CulturaX, comprising question-answer pairs and relevant passage chunks. We benchmark seven stages of the RAG pipeline, from query transformation and reranking to answer refinement, without task-specific fine-tuning. Our results show that complex methods like HyDE maximize accuracy (85%) that is considerably higher than the baseline (78.70%). Also a Pareto-optimal configuration using Cross-encoder Reranking and Context Augmentation achieves comparable performance (84.60%) with much lower cost. We further demonstrate that over-stacking generative modules can degrade performance by distorting morphological cues, whereas simple query clarification with robust reranking offers an effective solution.

Method: Analyzes modality following through an information flow lens, examining attention layers and MLP layers, identifying specialized attention heads that drive modality arbitration, and performing causal interventions.

Result: Found that instruction tokens serve as structural anchors; shallow layers route multimodal cues non-selectively, deep layers resolve modality competition, MLP layers act adversarially; identified sparse specialized attention heads; manipulating 5% of critical heads can change modality-following ratio by ±60%.

Conclusion: Provides insights into MLLM decision-making for multimodal context usage, advancing model transparency and offering a framework for orchestrating multimodal information.

Abstract: Modality following serves as the capacity of multimodal large language models (MLLMs) to selectively utilize multimodal contexts based on user instructions. It is fundamental to ensuring safety and reliability in real-world deployments. However, the underlying mechanisms governing this decision-making process remain poorly understood. In this paper, we investigate its working mechanism through an information flow lens. Our findings reveal that instruction tokens function as structural anchors for modality arbitration: Shallow attention layers perform non-selective information transfer, routing multimodal cues to these anchors as a latent buffer; Modality competition is resolved within deep attention layers guided by the instruction intent, while MLP layers exhibit semantic inertia, acting as an adversarial force. Furthermore, we identify a sparse set of specialized attention heads that drive this arbitration. Causal interventions demonstrate that manipulating a mere $5%$ of these critical heads can decrease the modality-following ratio by $60%$ through blocking, or increase it by $60%$ through targeted amplification of failed samples. Our work provides a substantial step toward model transparency and offers a principled framework for the orchestration of multimodal information in MLLMs.

Method: Proposes Neural Attention Search Linear (NAtS-L) that applies both linear attention (Gated DeltaNet) and softmax attention operations within the same layer on different tokens. Automatically determines token-level attention type based on whether tokens have short-term impact (linear attention) or require long-term retrieval (softmax attention).

Result: NAtS-L provides a strong yet efficient token-level hybrid architecture that reduces computational complexity while maintaining expressivity for long-context modeling.

Conclusion: The framework enables efficient long-context processing by intelligently allocating computational resources at the token level, balancing efficiency and expressivity in attention mechanisms.

Abstract: The quadratic computational complexity of softmax transformers has become a bottleneck in long-context scenarios. In contrast, linear attention model families provide a promising direction towards a more efficient sequential model. These linear attention models compress past KV values into a single hidden state, thereby efficiently reducing complexity during both training and inference. However, their expressivity remains limited by the size of their hidden state. Previous work proposed interleaving softmax and linear attention layers to reduce computational complexity while preserving expressivity. Nevertheless, the efficiency of these models remains bottlenecked by their softmax attention layers. In this paper, we propose Neural Attention Search Linear (NAtS-L), a framework that applies both linear attention and softmax attention operations within the same layer on different tokens. NAtS-L automatically determines whether a token can be handled by a linear attention model, i.e., tokens that have only short-term impact and can be encoded into fixed-size hidden states, or require softmax attention, i.e., tokens that contain information related to long-term retrieval and need to be preserved for future queries. By searching for optimal Gated DeltaNet and softmax attention combinations across tokens, we show that NAtS-L provides a strong yet efficient token-level hybrid architecture.

Method: Proposes BAR-RAG with a boundary-aware evidence selector trained with reinforcement learning using generator feedback, and a two-stage pipeline that fine-tunes the generator under the induced evidence distribution to mitigate training-inference mismatch.

Result: Experiments on knowledge-intensive QA benchmarks show BAR-RAG consistently improves end-to-end performance under noisy retrieval, achieving average 10.3% gain over strong RAG and reranking baselines while substantially improving robustness.

Conclusion: BAR-RAG effectively addresses retrieval noise in RAG systems by selecting evidence in the generator’s Goldilocks Zone, leading to significant performance improvements and enhanced robustness.

Abstract: Retrieval-Augmented Generation (RAG) systems remain brittle under realistic retrieval noise, even when the required evidence appears in the top-K results. A key reason is that retrievers and rerankers optimize solely for relevance, often selecting either trivial, answer-revealing passages or evidence that lacks the critical information required to answer the question, without considering whether the evidence is suitable for the generator. We propose BAR-RAG, which reframes the reranker as a boundary-aware evidence selector that targets the generator’s Goldilocks Zone – evidence that is neither trivially easy nor fundamentally unanswerable for the generator, but is challenging yet sufficient for inference and thus provides the strongest learning signal. BAR-RAG trains the selector with reinforcement learning using generator feedback, and adopts a two-stage pipeline that fine-tunes the generator under the induced evidence distribution to mitigate the distribution mismatch between training and inference. Experiments on knowledge-intensive question answering benchmarks show that BAR-RAG consistently improves end-to-end performance under noisy retrieval, achieving an average gain of 10.3 percent over strong RAG and reranking baselines while substantially improving robustness. Code is publicly avaliable at https://github.com/GasolSun36/BAR-RAG.

Method: Created OCRTurk benchmark with 180 Turkish documents from academic articles, theses, slide decks, and non-academic articles at three difficulty levels. Evaluated seven OCR models using element-wise metrics including Normalized Edit Distance.

Result: PaddleOCR achieved strongest overall results across difficulty levels, leading most element-wise metrics except figures. Performance varied by document type: models performed well on non-academic documents while slideshows were most challenging.

Conclusion: OCRTurk addresses the gap in Turkish document parsing benchmarks, providing a standardized evaluation framework that reflects real-world document diversity and difficulty levels for assessing OCR model performance.

Abstract: Document parsing is now widely used in applications, such as large-scale document digitization, retrieval-augmented generation, and domain-specific pipelines in healthcare and education. Benchmarking these models is crucial for assessing their reliability and practical robustness. Existing benchmarks mostly target high-resource languages and provide limited coverage for low-resource settings, such as Turkish. Moreover, existing studies on Turkish document parsing lack a standardized benchmark that reflects real-world scenarios and document diversity. To address this gap, we introduce OCRTurk, a Turkish document parsing benchmark covering multiple layout elements and document categories at three difficulty levels. OCRTurk consists of 180 Turkish documents drawn from academic articles, theses, slide decks, and non-academic articles. We evaluate seven OCR models on OCRTurk using element-wise metrics. Across difficulty levels, PaddleOCR achieves the strongest overall results, leading most element-wise metrics except figures and attaining high Normalized Edit Distance scores in easy, medium, and hard subsets. We also observe performance variation by document type. Models perform well on non-academic documents, while slideshows become the most challenging.

Method: ReQUESTA decomposes MCQ authoring into specialized subtasks and coordinates LLM-powered agents with rule-based components. It uses a hybrid approach with multi-agent orchestration for planning, controlled generation, iterative evaluation, and post-processing. The framework was evaluated against a single-pass GPT-5 zero-shot baseline in a large-scale reading comprehension study using academic expository texts.

Result: ReQUESTA-generated items were consistently more challenging, more discriminative, and more strongly aligned with overall reading comprehension performance. Expert evaluations showed stronger alignment with central concepts and superior distractor linguistic consistency and semantic plausibility, particularly for inferential questions.

Conclusion: Hybrid, agentic orchestration can systematically improve the reliability and controllability of LLM-based generation, highlighting workflow design as a key lever for structured artifact generation beyond single-pass prompting.

Abstract: Recent advances in large language models (LLMs) have made automated multiple-choice question (MCQ) generation increasingly feasible; however, reliably producing items that satisfy controlled cognitive demands remains a challenge. To address this gap, we introduce ReQUESTA, a hybrid, multi-agent framework for generating cognitively diverse MCQs that systematically target text-based, inferential, and main idea comprehension. ReQUESTA decomposes MCQ authoring into specialized subtasks and coordinates LLM-powered agents with rule-based components to support planning, controlled generation, iterative evaluation, and post-processing. We evaluated the framework in a large-scale reading comprehension study using academic expository texts, comparing ReQUESTA-generated MCQs with those produced by a single-pass GPT-5 zero-shot baseline. Psychometric analyses of learner responses assessed item difficulty and discrimination, while expert raters evaluated question quality across multiple dimensions, including topic relevance and distractor quality. Results showed that ReQUESTA-generated items were consistently more challenging, more discriminative, and more strongly aligned with overall reading comprehension performance. Expert evaluations further indicated stronger alignment with central concepts and superior distractor linguistic consistency and semantic plausibility, particularly for inferential questions. These findings demonstrate that hybrid, agentic orchestration can systematically improve the reliability and controllability of LLM-based generation, highlighting workflow design as a key lever for structured artifact generation beyond single-pass prompting.

Method: Builds image-audio retrieval-augmented generation module for fetching relevant frames/audio snippets, uses agent loop for planning/tool calling across turns, and applies group relative policy optimization for joint improvement.

Result: Outperforms prior methods on OmniVideoBench, WorldSense, and Daily-Omni under low-resource settings, with ablations validating each component.

Conclusion: OmniRAG-Agent effectively addresses key challenges in long-horizon omnimodal QA through retrieval-augmented generation, agentic planning, and policy optimization.

Abstract: Long-horizon omnimodal question answering answers questions by reasoning over text, images, audio, and video. Despite recent progress on OmniLLMs, low-resource long audio-video QA still suffers from costly dense encoding, weak fine-grained retrieval, limited proactive planning, and no clear end-to-end optimization.To address these issues, we propose OmniRAG-Agent, an agentic omnimodal QA method for budgeted long audio-video reasoning. It builds an image-audio retrieval-augmented generation module that lets an OmniLLM fetch short, relevant frames and audio snippets from external banks. Moreover, it uses an agent loop that plans, calls tools across turns, and merges retrieved evidence to answer complex queries. Furthermore, we apply group relative policy optimization to jointly improve tool use and answer quality over time. Experiments on OmniVideoBench, WorldSense, and Daily-Omni show that OmniRAG-Agent consistently outperforms prior methods under low-resource settings and achieves strong results, with ablations validating each component.

Method: Proposes SemanticSpec, a semantic-aware speculative decoding framework that verifies entire semantic sequences instead of tokens. Introduces semantic probability estimation mechanism that probes model’s internal hidden states to assess likelihood of generating sequences with specific meanings.

Result: Achieves up to 2.7x speedup on DeepSeekR1-32B and 2.1x on QwQ-32B, consistently outperforming token-level and sequence-level baselines in both efficiency and effectiveness across four benchmarks.

Conclusion: SemanticSpec provides an effective approach to accelerate LLM inference by leveraging semantic equivalence in speculative decoding, addressing limitations of token-level methods.

Abstract: Large Language Models (LLMs) achieve strong performance across many tasks but suffer from high inference latency due to autoregressive decoding. The issue is exacerbated in Large Reasoning Models (LRMs), which generate lengthy chains of thought. While speculative decoding accelerates inference by drafting and verifying multiple tokens in parallel, existing methods operate at the token level and ignore semantic equivalence (i.e., different token sequences expressing the same meaning), leading to inefficient rejections. We propose SemanticSpec, a semantic-aware speculative decoding framework that verifies entire semantic sequences instead of tokens. SemanticSpec introduces a semantic probability estimation mechanism that probes the model’s internal hidden states to assess the likelihood of generating sequences with specific meanings.Experiments on four benchmarks show that SemanticSpec achieves up to 2.7x speedup on DeepSeekR1-32B and 2.1x on QwQ-32B, consistently outperforming token-level and sequence-level baselines in both efficiency and effectiveness.

Method: Created ID-MoCQA dataset by systematically transforming single-hop cultural questions into multi-hop reasoning chains spanning six clue types (commonsense, temporal, geographical, etc.). Used multi-stage validation pipeline combining expert review and LLM-as-a-judge filtering to ensure high-quality question-answer pairs.

Result: Evaluation across state-of-the-art models reveals substantial gaps in cultural reasoning, particularly in tasks requiring nuanced inference. The dataset provides a challenging benchmark for assessing cultural competency.

Conclusion: ID-MoCQA is an essential benchmark for advancing the cultural competency of LLMs, addressing limitations of existing single-hop cultural QA datasets and enabling better assessment of genuine cultural reasoning capabilities.

Abstract: Understanding culture requires reasoning across context, tradition, and implicit social knowledge, far beyond recalling isolated facts. Yet most culturally focused question answering (QA) benchmarks rely on single-hop questions, which may allow models to exploit shallow cues rather than demonstrate genuine cultural reasoning. In this work, we introduce ID-MoCQA, the first large-scale multi-hop QA dataset for assessing the cultural understanding of large language models (LLMs), grounded in Indonesian traditions and available in both English and Indonesian. We present a new framework that systematically transforms single-hop cultural questions into multi-hop reasoning chains spanning six clue types (e.g., commonsense, temporal, geographical). Our multi-stage validation pipeline, combining expert review and LLM-as-a-judge filtering, ensures high-quality question-answer pairs. Our evaluation across state-of-the-art models reveals substantial gaps in cultural reasoning, particularly in tasks requiring nuanced inference. ID-MoCQA provides a challenging and essential benchmark for advancing the cultural competency of LLMs.

Method: BranPO truncates trajectories near the tail and resamples alternative continuations to construct contrastive suffixes over shared prefixes, providing step-level contrastive supervision without dense rewards. It introduces difficulty-aware branch sampling to adapt branching frequency across tasks, and redundant step masking to suppress uninformative actions.

Result: Extensive experiments on various question answering benchmarks show BranPO consistently outperforms strong baselines, achieving significant accuracy gains on long-horizon tasks without increasing the overall training budget.

Conclusion: BranPO effectively addresses credit assignment in long-horizon RL for language models through contrastive suffix construction and adaptive sampling techniques, demonstrating superior performance on complex planning tasks.

Abstract: Agentic reinforcement learning has enabled large language models to perform complex multi-turn planning and tool use. However, learning in long-horizon settings remains challenging due to sparse, trajectory-level outcome rewards. While prior tree-based methods attempt to mitigate this issue, they often suffer from high variance and computational inefficiency. Through empirical analysis of search agents, We identify a common pattern: performance diverges mainly due to decisions near the tail. Motivated by this observation, we propose Branching Relative Policy Optimization (BranPO), a value-free method that provides step-level contrastive supervision without dense rewards. BranPO truncates trajectories near the tail and resamples alternative continuations to construct contrastive suffixes over shared prefixes, reducing credit ambiguity in long-horizon rollouts. To further boost efficiency and stabilize training, we introduce difficulty-aware branch sampling to adapt branching frequency across tasks, and redundant step masking to suppress uninformative actions. Extensive experiments on various question answering benchmarks demonstrate that BranPO consistently outperforms strong baselines, achieving significant accuracy gains on long-horizon tasks without increasing the overall training budget. Our code is available at \href{https://github.com/YubaoZhao/BranPO}{code}.

Method: Joint text-vision pre-training, zero-vision SFT (Supervised Fine-Tuning), and joint text-vision reinforcement learning for multimodal optimization. Agent Swarm framework for self-directed parallel agent orchestration that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently.

Result: Achieves state-of-the-art results across coding, vision, reasoning, and agentic tasks. Agent Swarm reduces latency by up to 4.5× compared to single-agent baselines.

Conclusion: Kimi K2.5 successfully demonstrates effective multimodal agentic intelligence with joint text-vision optimization and efficient parallel task execution through Agent Swarm, advancing the field of general agentic intelligence.

Abstract: We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.

Method: CUBO integrates streaming ingestion with O(1) buffer overhead, tiered hybrid retrieval, and hardware-aware orchestration to operate within a strict 15.5GB RAM ceiling. It uses local-only processing to maintain GDPR compliance.

Result: Achieves competitive Recall@10 scores (0.48-0.97 across BEIR domains) with retrieval latencies of 185ms (p50) on consumer laptops, all within the 15.5GB RAM constraint. The 37,000-line codebase validates practical deployability for small-to-medium archives.

Conclusion: CUBO demonstrates that efficient local RAG systems can run on consumer hardware while maintaining GDPR compliance, making AI-powered document analysis accessible to organizations with privacy concerns.

Abstract: Organizations handling sensitive documents face a tension: cloud-based AI risks GDPR violations, while local systems typically require 18-32 GB RAM. This paper presents CUBO, a systems-oriented RAG platform for consumer laptops with 16 GB shared memory. CUBO’s novelty lies in engineering integration of streaming ingestion (O(1) buffer overhead), tiered hybrid retrieval, and hardware-aware orchestration that enables competitive Recall@10 (0.48-0.97 across BEIR domains) within a hard 15.5 GB RAM ceiling. The 37,000-line codebase achieves retrieval latencies of 185 ms (p50) on C1,300 laptops while maintaining data minimization through local-only processing aligned with GDPR Art. 5(1)(c). Evaluation on BEIR benchmarks validates practical deployability for small-to-medium professional archives. The codebase is publicly available at https://github.com/PaoloAstrino/CUBO.

Method: ComprExIT formulates soft compression as explicit information transmission over frozen LLM hidden states, decoupling compression from self-attention dynamics. It performs depth-wise transmission to selectively transmit multi-layer information into token anchors, and width-wise transmission to aggregate anchors into a small number of slots via a globally optimized transmission plan.

Result: Across six question-answering benchmarks, ComprExIT consistently outperforms state-of-the-art context compression methods while introducing only ~1% additional parameters.

Conclusion: Explicit and coordinated information transmission enables more effective and robust long-context compression compared to existing self-attention-based approaches.

Abstract: Long-context inference with Large Language Models (LLMs) is costly due to quadratic attention and growing key-value caches, motivating context compression. In this work, we study soft context compression, where a long context is condensed into a small set of continuous representations. Existing methods typically re-purpose the LLM itself as a trainable compressor, relying on layer-by-layer self-attention to iteratively aggregate information. We argue that this paradigm suffers from two structural limitations: (i) progressive representation overwriting across layers (ii) uncoordinated allocation of compression capacity across tokens. We propose ComprExIT (Context Compression via Explicit Information Transmission), a lightweight framework that formulates soft compression into a new paradigm: explicit information transmission over frozen LLM hidden states. This decouples compression from the model’s internal self-attention dynamics. ComprExIT performs (i) depth-wise transmission to selectively transmit multi-layer information into token anchors, mitigating progressive overwriting, and (ii) width-wise transmission to aggregate anchors into a small number of slots via a globally optimized transmission plan, ensuring coordinated allocation of information. Across six question-answering benchmarks, ComprExIT consistently outperforms state-of-the-art context compression methods while introducing only ~1% additional parameters, demonstrating that explicit and coordinated information transmission enables more effective and robust long-context compression.

Method: Three-stage framework: (1) Enrich multimodal representations with structured knowledge from ConceptNet, Wikidata, Hatebase; (2) Use LoRA adapters to sharpen decision boundaries; (3) Generate cascaded explanations for interpretability.

Result: Outperforms state-of-the-art methods on five benchmarks with eight LVLMs, achieving up to 17% relative F1 improvement while providing interpretable justifications.

Conclusion: CROSS-ALIGN+ effectively addresses key limitations in meme abuse detection through knowledge enrichment, boundary refinement, and explanation generation.

Abstract: Meme-based social abuse detection is challenging because harmful intent often relies on implicit cultural symbolism and subtle cross-modal incongruence. Prior approaches, from fusion-based methods to in-context learning with Large Vision-Language Models (LVLMs), have made progress but remain limited by three factors: i) cultural blindness (missing symbolic context), ii) boundary ambiguity (satire vs. abuse confusion), and iii) lack of interpretability (opaque model reasoning). We introduce CROSS-ALIGN+, a three-stage framework that systematically addresses these limitations: (1) Stage I mitigates cultural blindness by enriching multimodal representations with structured knowledge from ConceptNet, Wikidata, and Hatebase; (2) Stage II reduces boundary ambiguity through parameter-efficient LoRA adapters that sharpen decision boundaries; and (3) Stage III enhances interpretability by generating cascaded explanations. Extensive experiments on five benchmarks and eight LVLMs demonstrate that CROSS-ALIGN+ consistently outperforms state-of-the-art methods, achieving up to 17% relative F1 improvement while providing interpretable justifications for each decision.

Method: Collection of case studies using Google’s Gemini-based models (Gemini Deep Think and variants) with interactive, conversational methodology, plus specific techniques like using AI as adversarial reviewer and embedding in neuro-symbolic loops for code verification.

Result: Successful collaboration on solving open problems, refuting conjectures, and generating new proofs across theoretical computer science, economics, optimization, and physics, with extracted effective collaboration techniques.

Conclusion: AI can serve as a versatile partner in scientific discovery, not just for automation, with potential demonstrated through interactive collaboration and specialized techniques like adversarial review and neuro-symbolic verification.

Abstract: Recent advances in large language models (LLMs) have opened new avenues for accelerating scientific research. While models are increasingly capable of assisting with routine tasks, their ability to contribute to novel, expert-level mathematical discovery is less understood. We present a collection of case studies demonstrating how researchers have successfully collaborated with advanced AI models, specifically Google’s Gemini-based models (in particular Gemini Deep Think and its advanced variants), to solve open problems, refute conjectures, and generate new proofs across diverse areas in theoretical computer science, as well as other areas such as economics, optimization, and physics. Based on these experiences, we extract common techniques for effective human-AI collaboration in theoretical research, such as iterative refinement, problem decomposition, and cross-disciplinary knowledge transfer. While the majority of our results stem from this interactive, conversational methodology, we also highlight specific instances that push beyond standard chat interfaces. These include deploying the model as a rigorous adversarial reviewer to detect subtle flaws in existing proofs, and embedding it within a “neuro-symbolic” loop that autonomously writes and executes code to verify complex derivations. Together, these examples highlight the potential of AI not just as a tool for automation, but as a versatile, genuine partner in the creative process of scientific discovery.

Method: Introduces 2D probing to expose width-depth dynamics by periodically eliciting intermediate answers from all branches. Based on insights from this analysis, develops Parallel-Probe with consensus-based early stopping to regulate reasoning depth and deviation-based branch pruning to dynamically adjust width.

Result: Extensive experiments across three benchmarks and multiple models show Parallel-Probe establishes superior Pareto frontier for test-time scaling. Reduces sequential tokens by up to 35.8% and total token cost by over 25.8% while maintaining competitive accuracy compared to standard majority voting.

Conclusion: Parallel-Probe effectively optimizes parallel thinking efficiency through principled exploitation of global dynamics, offering significant computational savings without sacrificing reasoning quality.

Abstract: Parallel thinking has emerged as a promising paradigm for reasoning, yet it imposes significant computational burdens. Existing efficiency methods primarily rely on local, per-trajectory signals and lack principled mechanisms to exploit global dynamics across parallel branches. We introduce 2D probing, an interface that exposes the width-depth dynamics of parallel thinking by periodically eliciting intermediate answers from all branches. Our analysis reveals three key insights: non-monotonic scaling across width-depth allocations, heterogeneous reasoning branch lengths, and early stabilization of global consensus. Guided by these insights, we introduce $\textbf{Parallel-Probe}$, a training-free controller designed to optimize online parallel thinking. Parallel-Probe employs consensus-based early stopping to regulate reasoning depth and deviation-based branch pruning to dynamically adjust width. Extensive experiments across three benchmarks and multiple models demonstrate that Parallel-Probe establishes a superior Pareto frontier for test-time scaling. Compared to standard majority voting, it reduces sequential tokens by up to $\textbf{35.8}$% and total token cost by over $\textbf{25.8}$% while maintaining competitive accuracy.

Method: Created merged dataset from ChemProt and DrugProt, evaluated using BioBERT (for local context) and BioBERT+GCN (for global+local context) for chemical-gene relation extraction.

Result: Merged dataset significantly improved model performance, especially in shared CPR groups. BioBERT+GCN achieved higher precision and recall than BioBERT alone by incorporating global context.

Conclusion: Dataset merging and combining local (BioBERT) with global (GCN) context modeling improves chemical-gene relation extraction, benefiting biomedical applications.

Abstract: The extraction of chemical-gene relations plays a pivotal role in understanding the intricate interactions between chemical compounds and genes, with significant implications for drug discovery, disease understanding, and biomedical research. This paper presents a data set created by merging the ChemProt and DrugProt datasets to augment sample counts and improve model accuracy. We evaluate the merged dataset using two state of the art relationship extraction algorithms: Bidirectional Encoder Representations from Transformers (BERT) specifically BioBERT, and Graph Convolutional Networks (GCNs) combined with BioBERT. While BioBERT excels at capturing local contexts, it may benefit from incorporating global information essential for understanding chemical-gene interactions. This can be achieved by integrating GCNs with BioBERT to harness both global and local context. Our results show that by integrating the ChemProt and DrugProt datasets, we demonstrated significant improvements in model performance, particularly in CPR groups shared between the datasets. Incorporating the global context using GCN can help increase the overall precision and recall in some of the CPR groups over using just BioBERT.

Method: Proposes SELSP (Sequence Labeling Syntactic Parser) that reformulates dependency parsing as a sequence labeling task, integrating syntactic information into sentiment analysis via a rule-based pipeline. Compares with conventional parsers (Stanza), heuristic approaches (VADER), and Transformer models.

Result: SELSP demonstrates greater speed and accuracy than conventional parsers like Stanza and heuristic approaches like VADER. It also outperforms Transformer-based models in speed for polarity prediction. Dictionaries accounting for polarity judgment variation yield better performance.

Conclusion: SELSP provides an efficient solution to the computational bottleneck in syntax-based sentiment analysis, offering both speed and accuracy advantages over existing methods, making it attractive for academic and industrial applications.

Abstract: Sentiment Analysis (SA) is a crucial aspect of Natural Language Processing (NLP), focusing on identifying and interpreting subjective assessments in textual content. Syntactic parsing is useful in SA as it improves accuracy and provides explainability; however, it often becomes a computational bottleneck due to slow parsing algorithms. This article proposes a solution to this bottleneck by using a Sequence Labeling Syntactic Parser (SELSP) to integrate syntactic information into SA via a rule-based sentiment analysis pipeline. By reformulating dependency parsing as a sequence labeling task, we significantly improve the efficiency of syntax-based SA. SELSP is trained and evaluated on a ternary polarity classification task, demonstrating greater speed and accuracy compared to conventional parsers like Stanza and heuristic approaches such as Valence Aware Dictionary and sEntiment Reasoner (VADER). The combination of speed and accuracy makes SELSP especially attractive for sentiment analysis applications in both academic and industrial contexts. Moreover, we compare SELSP with Transformer-based models trained on a 5-label classification task. In addition, we evaluate multiple sentiment dictionaries with SELSP to determine which yields the best performance in polarity prediction. The results show that dictionaries accounting for polarity judgment variation outperform those that ignore it. Furthermore, we show that SELSP outperforms Transformer-based models in terms of speed for polarity prediction.

Method: Generates context-dependent definitions of concept mentions by retrieving full-text literature, uses these definitions to enhance cross-document relation detection, generates relational definitions describing how two concepts are related/different, and employs an efficient re-ranking approach to handle combinatorial explosion in cross-paper link inference.

Result: Achieves large performance gains in both fine-tuning and in-context learning settings, particularly on data subsets with high amounts of different surface forms and ambiguity that are challenging for models.

Conclusion: The approach effectively addresses LLM limitations on fine-grained scientific concepts and provides insights into LLMs’ relational reasoning abilities through analysis of generated definitions.

Abstract: We address the fundamental task of inferring cross-document coreference and hierarchy in scientific texts, which has important applications in knowledge graph construction, search, recommendation and discovery. Large Language Models (LLMs) can struggle when faced with many long-tail technical concepts with nuanced variations. We present a novel method which generates context-dependent definitions of concept mentions by retrieving full-text literature, and uses the definitions to enhance detection of cross-document relations. We further generate relational definitions, which describe how two concept mentions are related or different, and design an efficient re-ranking approach to address the combinatorial explosion involved in inferring links across papers. In both fine-tuning and in-context learning settings, we achieve large gains in performance on data subsets with high amount of different surfaces forms and ambiguity, that are challenging for models. We provide analysis of generated definitions, shedding light on the relational reasoning ability of LLMs over fine-grained scientific concepts.

Method: Replaces linear projection layers with in-memory lookup tables storing discrete vectors, uses hash algorithms to dynamically retrieve relevant vectors based on input embeddings, and combines retrieved vectors to approximate matrix multiplication results.

Result: MemoryFormer significantly reduces computational complexity (FLOPs) while maintaining performance across various benchmarks when trained from scratch.

Conclusion: MemoryFormer offers a novel, computationally efficient transformer architecture that reduces FLOPs by replacing expensive matrix multiplications with cheaper memory retrieval operations.

Abstract: In order to reduce the computational complexity of large language models, great efforts have been made to to improve the efficiency of transformer models such as linear attention and flash-attention. However, the model size and corresponding computational complexity are constantly scaled up in pursuit of higher performance. In this work, we present MemoryFormer, a novel transformer architecture which significantly reduces the computational complexity (FLOPs) from a new perspective. We eliminate nearly all the computations of the transformer model except for the necessary computation required by the multi-head attention operation. This is made possible by utilizing an alternative method for feature transformation to replace the linear projection of fully-connected layers. Specifically, we first construct a group of in-memory lookup tables that store a large amount of discrete vectors to replace the weight matrix used in linear projection. We then use a hash algorithm to retrieve a correlated subset of vectors dynamically based on the input embedding. The retrieved vectors combined together will form the output embedding, which provides an estimation of the result of matrix multiplication operation in a fully-connected layer. Compared to conducting matrix multiplication, retrieving data blocks from memory is a much cheaper operation which requires little computations. We train MemoryFormer from scratch and conduct extensive experiments on various benchmarks to demonstrate the effectiveness of the proposed model.

Method: Three methodological advances: 1) Multi-turn evaluation of 14 anthropomorphic behaviors, 2) Scalable automated approach using user interaction simulations, 3) Large-scale human subject study (N=1101) to validate that measured behaviors predict real users’ anthropomorphic perceptions.

Result: All SOTA LLMs exhibit similar anthropomorphic behaviors characterized by relationship-building (empathy, validation) and first-person pronoun use, with most behaviors only emerging after multiple turns.

Conclusion: Establishes empirical foundation for investigating how design choices influence anthropomorphic model behaviors and advances ethical debate about desirability of such behaviors, demonstrating necessity of multi-turn evaluations for complex social phenomena in human-AI interaction.

Abstract: The tendency of users to anthropomorphise large language models (LLMs) is of growing interest to AI developers, researchers, and policy-makers. Here, we present a novel method for empirically evaluating anthropomorphic LLM behaviours in realistic and varied settings. Going beyond single-turn static benchmarks, we contribute three methodological advances in state-of-the-art (SOTA) LLM evaluation. First, we develop a multi-turn evaluation of 14 anthropomorphic behaviours. Second, we present a scalable, automated approach by employing simulations of user interactions. Third, we conduct an interactive, large-scale human subject study (N=1101) to validate that the model behaviours we measure predict real users’ anthropomorphic perceptions. We find that all SOTA LLMs evaluated exhibit similar behaviours, characterised by relationship-building (e.g., empathy and validation) and first-person pronoun use, and that the majority of behaviours only first occur after multiple turns. Our work lays an empirical foundation for investigating how design choices influence anthropomorphic model behaviours and for progressing the ethical debate on the desirability of these behaviours. It also showcases the necessity of multi-turn evaluations for complex social phenomena in human-AI interaction.

Method: Defines design space for layout-aware IE with LLMs, investigates sub-problems like input representation and prompting, develops LayIE-LLM test suite, uses one-factor-at-a-time method to find optimal configurations.

Result: Optimized LLM configurations achieve 13.3-37.5 F1 points improvement over baseline; OFAT method finds near-optimal configurations with only 2.8% of computation cost; well-configured LLMs match specialized model performance.

Conclusion: General-purpose LLMs can provide cost-effective, finetuning-free alternatives to specialized IE models when properly configured for layout-aware document understanding.

Abstract: This paper defines and explores the design space for information extraction (IE) from layout-rich documents using large language models (LLMs). The three core challenges of layout-aware IE with LLMs are 1) data structuring, 2) model engagement, and 3) output refinement. Our study investigates the sub-problems and methods within these core challenges, such as input representation, chunking, prompting, selection of LLMs, and multimodal models. It examines the effect of different design choices through LayIE-LLM, a new, open-source, layout-aware IE test suite, benchmarking against traditional, fine-tuned IE models. The results on two IE datasets show that LLMs require adjustment of the IE pipeline to achieve competitive performance: the optimized configuration found with LayIE-LLM achieves 13.3–37.5 F1 points more than a general-practice baseline configuration using the same LLM. To find a well-working configuration, we develop a one-factor-at-a-time (OFAT) method that achieves near-optimal results. Our method is only 0.8–1.8 points lower than the best full factorial exploration with a fraction (2.8%) of the required computation. Overall, we demonstrate that, if well-configured, general-purpose LLMs match the performance of specialized models, providing a cost-effective, finetuning-free alternative. Our test-suite is available at https://github.com/gayecolakoglu/LayIE-LLM.

Method: Integrates linguistically grounded discourse frameworks into reinforcement learning using Proximal Policy Optimization with dense token-level rewards. Uses two complementary reward models: one for surface-level textual features (readability) and another for global discourse patterns (coherence and rhetorical sophistication).

Result: Outperforms both standard and RLHF-enhanced models in tasks such as essay generation and long-document summarization by producing more coherent and well-organized outputs.

Conclusion: Structural Alignment effectively enhances long-form text generation by aligning LLMs with human discourse structures through reinforcement learning with fine-grained rewards.

Abstract: Generating long, coherent text remains a challenge for large language models (LLMs), as they lack hierarchical planning and structured organization in discourse generation. We introduce Structural Alignment, a novel method that aligns LLMs with human-like discourse structures to enhance long-form text generation. By integrating linguistically grounded discourse frameworks into reinforcement learning, our approach guides models to produce coherent and well-organized outputs. We employ a dense reward scheme within a Proximal Policy Optimization framework, assigning fine-grained, token-level rewards based on the discourse distinctiveness relative to human writing. Two complementary reward models are evaluated: the first improves readability by scoring surface-level textual features to provide explicit structuring, while the second reinforces deeper coherence and rhetorical sophistication by analyzing global discourse patterns through hierarchical discourse motifs, outperforming both standard and RLHF-enhanced models in tasks such as essay generation and long-document summarization. All training data and code will be publicly shared at https://github.com/minnesotanlp/struct_align.

Method: Proposes short-m@k inference method: execute k independent generations in parallel, halt when first m thinking processes complete, then use majority voting among these m chains. Also finetunes LLMs using short, long, and randomly selected reasoning chains to compare training effectiveness.

Result: Shorter reasoning chains are up to 34.5% more accurate than longest chains. short-1@k matches or exceeds standard majority voting with 40% fewer thinking tokens. short-3@k consistently surpasses majority voting across all compute budgets with up to 33% wall time reduction. Training on shorter chains yields better performance.

Conclusion: Longer “thinking” in reasoning LLMs doesn’t necessarily improve performance and can degrade results. The short-m@k method offers more efficient inference by leveraging shorter, more accurate reasoning chains, challenging current test-time compute paradigms.

Abstract: Reasoning large language models (LLMs) heavily rely on scaling test-time compute to perform complex reasoning tasks by generating extensive “thinking” chains. While demonstrating impressive results, this approach incurs significant computational costs and inference time. In this work, we challenge the assumption that long thinking chains results in better reasoning capabilities. We first demonstrate that shorter reasoning chains within individual questions are significantly more likely to yield correct answers - up to 34.5% more accurate than the longest chain sampled for the same question. Based on these results, we suggest short-m@k, a novel reasoning LLM inference method. Our method executes k independent generations in parallel and halts computation once the first m thinking processes are done. The final answer is chosen using majority voting among these m chains. Basic short-1@k demonstrates similar or even superior performance over standard majority voting in low-compute settings - using up to 40% fewer thinking tokens. short-3@k, while slightly less efficient than short-1@k, consistently surpasses majority voting across all compute budgets, while still being substantially faster (up to 33% wall time reduction). To further validate our findings, we finetune LLMs using short, long, and randomly selected reasoning chains. We then observe that training on the shorter ones leads to better performance. Our findings suggest rethinking current methods of test-time compute in reasoning LLMs, emphasizing that longer “thinking” does not necessarily translate to improved performance and can, counter-intuitively, lead to degraded results.

Method: Introduces v1, a lightweight extension for active visual referencing via point-and-copy: the model selects relevant image patches and copies their embeddings back into the reasoning stream. Uses semantic representations as keys for patch retrieval to ensure perceptual evidence remains aligned with reasoning space. Trained on v1 dataset of 300K multimodal reasoning traces with interleaved grounding annotations.

Result: Across multimodal mathematical reasoning benchmarks, v1 consistently outperforms comparable baselines. The model demonstrates improved ability to maintain focus on relevant visual regions during extended reasoning chains.

Conclusion: Active visual referencing through point-and-copy mechanisms significantly improves multimodal reasoning by allowing models to revisit visual evidence during reasoning, addressing the limitation of single-encoding approaches. The method shows promise for enhancing visual grounding in multimodal language models.

Abstract: When thinking with images, humans rarely rely on a single glance: they revisit visual evidence while reasoning. In contrast, most Multimodal Language Models encode an image once to key-value cache and then reason purely in text, making it hard to re-ground intermediate steps. We empirically confirm this: as reasoning chains lengthen, models progressively lose focus on relevant regions. We introduce v1, a lightweight extension for active visual referencing via point-and-copy: the model selects relevant image patches and copies their embeddings back into the reasoning stream. Crucially, our point-and-copy mechanism retrieves patches using their semantic representations as keys, ensuring perceptual evidence remains aligned with the reasoning space. To train this behavior, we build v1, a dataset of 300K multimodal reasoning traces with interleaved grounding annotations. Across multimodal mathematical reasoning benchmarks, v1 consistently outperforms comparable baselines. We plan to release the model checkpoint and data.

Method: Exhaustive analysis of reward models’ responses across entire vocabulary space, examining single-token responses to value-laden prompts across ten open-source reward models with varying architectures.

Result: Found substantial heterogeneity between models, systematic asymmetries in encoding high/low-scoring tokens, sensitivity to prompt framing mirroring human biases, overvaluation of frequent tokens, and concerning biases toward identity groups.

Conclusion: Reward models are not interchangeable and may encode problematic biases that propagate through downstream LLMs, challenging their suitability as proxies for complex human values.

Abstract: Reward modeling has emerged as a crucial component in aligning large language models with human values. Significant attention has focused on using reward models as a means for fine-tuning generative models. However, the reward models themselves – which directly encode human value judgments by turning prompt-response pairs into scalar rewards – remain relatively understudied. We present a novel approach to reward model interpretability through exhaustive analysis of their responses across their entire vocabulary space. By examining how different reward models score every possible single-token response to value-laden prompts, we uncover several striking findings: (i) substantial heterogeneity between models trained on similar objectives, (ii) systematic asymmetries in how models encode high- vs low-scoring tokens, (iii) significant sensitivity to prompt framing that mirrors human cognitive biases, and (iv) overvaluation of more frequent tokens. We demonstrate these effects across ten recent open-source reward models of varying parameter counts and architectures. Our results challenge assumptions about the interchangeability of reward models, as well as their suitability as proxies of complex and context-dependent human values. We find that these models can encode concerning biases toward certain identity groups, which may emerge as unintended consequences of harmlessness training – distortions that risk propagating through the downstream large language models now deployed to millions.

Method: Comprehensive evaluation using surprisal estimates from 17 pre-trained LMs across three different LM families on two fMRI datasets, testing the relationship between models’ per-word estimated probability and model fit on brain imaging data.

Result: Results show the inverse scaling relationship between models’ per-word estimated probability and model fit on both fMRI datasets, confirming the trend extends beyond latency-based measures.

Conclusion: The inverse scaling phenomenon is not specific to latency data and holds for brain imaging measures, resolving previous inconclusive results about LM surprisal’s predictive power for human language processing.

Abstract: There has been considerable interest in using surprisal from Transformer-based language models (LMs) as predictors of human sentence processing difficulty. Recent work has observed an inverse scaling relationship between Transformers’ per-word estimated probability and the predictive power of their surprisal estimates on reading times, showing that LMs with more parameters and trained on more data are less predictive of human reading times. However, these studies focused on predicting latency-based measures. Tests on brain imaging data have not shown a trend in any direction when using a relatively small set of LMs, leaving open the possibility that the inverse scaling phenomenon is constrained to latency data. This study therefore conducted a more comprehensive evaluation using surprisal estimates from 17 pre-trained LMs across three different LM families on two functional magnetic resonance imaging (fMRI) datasets. Results show that the inverse scaling relationship between models’ per-word estimated probability and model fit on both datasets still obtains, resolving the inconclusive results of previous work and indicating that this trend is not specific to latency-based measures.

Method: Systematically evaluate instruction-tuned models across 20 benchmarks using non-canonical tokenizations (random tokenization, character-level tokenization), analyze performance patterns, and investigate the source of robustness through comparisons between base and post-trained models.

Result: Models retain up to 93.4% performance with random tokenization and 90.8% with character-level tokenization. Character-level segmentation improves string manipulation and code tasks by +14%, and right-aligned digit grouping enhances large-number arithmetic by +33%. Robustness emerges during instruction-tuning phase.

Conclusion: Language models are less tied to their tokenizer than previously believed, and inference-time tokenization interventions can boost performance on specific tasks, with robustness arising from instruction-tuning rather than pre-training.

Abstract: Modern tokenizers employ deterministic algorithms to map text into a single “canonical” token sequence, yet the same string can be encoded as many non-canonical tokenizations using the tokenizer vocabulary. In this work, we investigate the robustness of LMs to text encoded with non-canonical tokenizations entirely unseen during training. Surprisingly, when evaluated across 20 benchmarks, we find that instruction-tuned models retain up to 93.4% of their original performance when given a randomly sampled tokenization, and 90.8% with character-level tokenization. We see that overall stronger models tend to be more robust, and robustness diminishes as the tokenization departs farther from the canonical form. Motivated by these results, we then identify settings where non-canonical tokenization schemes can improve performance, finding that character-level segmentation improves string manipulation and code understanding tasks by up to +14%, and right-aligned digit grouping enhances large-number arithmetic by +33%. Finally, we investigate the source of this robustness, finding that it arises in the instruction-tuning phase. We show that while both base and post-trained models grasp the semantics of non-canonical tokenizations (perceiving them as containing misspellings), base models try to mimic the imagined mistakes and degenerate into nonsensical output, while post-trained models are committed to fluent responses. Overall, our findings suggest that models are less tied to their tokenizer than previously believed, and demonstrate the promise of intervening on tokenization at inference time to boost performance.

Method: Uses mechanistic interpretability with transformer circuits (minimal computational subgraphs) and carefully constructed contrastive datasets to identify where model generation diverges from memorized content and isolate specific circuits for different aspects of memorization.

Result: Found that circuits initiating memorization can also maintain it once started, while circuits that only maintain memorization cannot trigger initiation. Memorization prevention mechanisms transfer robustly across text domains, while memorization induction appears more context-dependent.

Conclusion: Provides insights into the mechanistic differences between memorization initiation and maintenance in LLMs, revealing domain transfer properties that could inform better memorization control strategies.

Abstract: Underlying mechanisms of memorization in LLMs – the verbatim reproduction of training data – remain poorly understood. What exact part of the network decides to retrieve a token that we would consider as start of memorization sequence? How exactly is the models’ behaviour different when producing memorized sentence vs non-memorized? In this work we approach these questions from mechanistic interpretability standpoint by utilizing transformer circuits – the minimal computational subgraphs that perform specific functions within the model. Through carefully constructed contrastive datasets, we identify points where model generation diverges from memorized content and isolate the specific circuits responsible for two distinct aspects of memorization. We find that circuits that initiate memorization can also maintain it once started, while circuits that only maintain memorization cannot trigger its initiation. Intriguingly, memorization prevention mechanisms transfer robustly across different text domains, while memorization induction appears more context-dependent.

Method: Formally defined scoring bias and identified three novel types: rubric order bias, score ID bias, and reference answer score bias. Proposed comprehensive framework with multi-faceted metrics and automatic data synthesis pipeline to create tailored evaluation corpus for quantifying these biases.

Result: Experiments empirically demonstrate that even the most advanced LLMs suffer from substantial scoring biases. The analysis provides actionable insights for designing more robust scoring prompts and mitigating these newly identified biases.

Conclusion: Scoring bias in LLM judges is a significant, previously under-investigated problem that requires attention, especially for industrial applications. The proposed framework and identified bias types provide foundation for developing more reliable automated evaluation systems.

Abstract: The “LLM-as-a-Judge” paradigm, using Large Language Models (LLMs) as automated evaluators, is pivotal to LLM development, offering scalable feedback for complex tasks. However, the reliability of these judges is compromised by various biases. Existing research has heavily concentrated on biases in comparative evaluations. In contrast, scoring-based evaluations-which assign an absolute score and are often more practical in industrial applications-remain under-investigated. To address this gap, we undertake the first dedicated examination of scoring bias in LLM judges. We shift the focus from biases tied to the evaluation targets to those originating from the scoring prompt itself. We formally define scoring bias and identify three novel, previously unstudied types: rubric order bias, score ID bias, and reference answer score bias. We propose a comprehensive framework to quantify these biases, featuring a suite of multi-faceted metrics and an automatic data synthesis pipeline to create a tailored evaluation corpus. Our experiments empirically demonstrate that even the most advanced LLMs suffer from these substantial scoring biases. Our analysis yields actionable insights for designing more robust scoring prompts and mitigating these newly identified biases.

Method: Proposed InfoRidge framework uses information theory to measure predictive information (mutual information between hidden representations and target outputs) across layers during training. Conducted experiments across various models and datasets with complementary analyses using residual scaling, attention patterns, and controlled model capacity.

Result: Consistent non-monotonic trend: predictive information peaks in intermediate layers (forming a “generalization ridge”) before declining in final layers, reflecting transition between generalization and memorization. This ridge phenomenon persists across decoding steps in multiple-token generation experiments.

Conclusion: Findings offer new insights into transformer internal mechanisms and underscore critical role of intermediate layers in supporting generalization. The generalization ridge phenomenon reveals how transformers balance generalization and memorization across layers.

Abstract: Transformer-based language models have achieved state-of-the-art performance in natural language generation (NLG), yet their internal mechanisms for synthesizing task-relevant information remain insufficiently understood. While prior studies suggest that intermediate layers often yield more generalizable representations than final layers, how this generalization ability emerges and propagates across layers during training remains unclear. To address this gap, we propose InfoRidge, an information-theoretic framework, to characterize how predictive information-the mutual information between hidden representations and target outputs-varies across depth during training. Our experiments across various models and datasets reveal a consistent non-monotonic trend: predictive information peaks in intermediate layers-forming a generalization ridge-before declining in final layers, reflecting a transition between generalization and memorization. To further investigate this phenomenon, we conduct a set of complementary analyses that leverage residual scaling, attention pattern, and controlled model capacity to characterize layer-wise functional specialization. We further validate our findings with multiple-token generation experiments, verifying that the observed ridge phenomenon persists across decoding steps. Together, these findings offer new insights into the internal mechanisms of transformers and underscore the critical role of intermediate layers in supporting generalization.

Method: Introduces GeoResponder framework with scaffolded instruction-tuning curriculum that stratifies geospatial learning into different cognitive layers, anchoring semantic knowledge to continuous coordinate manifold and enforcing internalization of spatial axioms.

Result: Extensive evaluations across four topologically distinct cities and diverse tasks show GeoResponder significantly outperforms both state-of-the-art foundation models and domain-specific baselines.

Conclusion: LLMs can begin to internalize and generalize geospatial structures, pointing toward future development of language models capable of supporting disaster response needs.

Abstract: Large Language Models excel at linguistic tasks but lack the inner geospatial capabilities needed for time-critical disaster response, where reasoning about road networks, continuous coordinates, and access to essential infrastructure such as hospitals, shelters, and pharmacies is vital. We introduce GeoResponder, a framework that instills robust spatial reasoning through a scaffolded instruction-tuning curriculum. By stratifying geospatial learning into different cognitive layers, we effectively anchor semantic knowledge to the continuous coordinate manifold and enforce the internalization of spatial axioms. Extensive evaluations across four topologically distinct cities and diverse tasks demonstrate that GeoResponder significantly outperforms both state-of-the-art foundation models and domain-specific baselines. These results suggest that LLMs can begin to internalize and generalize geospatial structures, pointing toward the future development of language models capable of supporting disaster response needs.

Method: Quantifies external context utilization via distributional distance, measures internal knowledge utilization by tracking how predicted tokens evolve across transformer layers, and introduces statistical validation framework.

Result: Achieves consistently high AUROC and AUPRC scores, outperforming prior methods by up to +13% AUROC on HalluRAG benchmark, robust under relaxed assumptions about retrieval quality and model matching.

Conclusion: LUMINA provides effective and practical hallucination detection for RAG systems without extensive hyperparameter tuning, offering statistical validation for context-knowledge signals.

Abstract: Retrieval-Augmented Generation (RAG) aims to mitigate hallucinations in large language models (LLMs) by grounding responses in retrieved documents. Yet, RAG-based LLMs still hallucinate even when provided with correct and sufficient context. A growing line of work suggests that this stems from an imbalance between how models use external context and their internal knowledge, and several approaches have attempted to quantify these signals for hallucination detection. However, existing methods require extensive hyperparameter tuning, limiting their generalizability. We propose LUMINA, a novel framework that detects hallucinations in RAG systems through context–knowledge signals: external context utilization is quantified via distributional distance, while internal knowledge utilization is measured by tracking how predicted tokens evolve across transformer layers. We further introduce a framework for statistically validating these measurements. Experiments on common RAG hallucination benchmarks and four open-source LLMs show that LUMINA achieves consistently high AUROC and AUPRC scores, outperforming prior utilization-based methods by up to +13% AUROC on HalluRAG. Moreover, LUMINA remains robust under relaxed assumptions about retrieval quality and model matching, offering both effectiveness and practicality. LUMINA: https://github.com/deeplearning-wisc/LUMINA

Method: A2D aligns dLLMs at token-level using randomized masking to train models to emit [EOS] refusal signals whenever harmful content arises, enabling real-time monitoring and automatic termination of unsafe continuations.

Result: A2D reduces DIJA success rates from over 80% to near-zero (1.3% on LLaDA-8B-Instruct, 0.0% on Dream-v0-Instruct-7B) and enables up to 19.3x faster safe termination through thresholded [EOS] probability monitoring.

Conclusion: Token-level alignment with randomized masking provides robust defense against any-order and any-step attacks in diffusion LLMs, enabling both prevention of harmful outputs and real-time safety monitoring.

Abstract: Diffusion large language models (dLLMs) enable any-order generation, but this flexibility enlarges the attack surface: harmful spans may appear at arbitrary positions, and template-based prefilling attacks such as DIJA bypass response-level refusals. We introduce A2D (Any-Order, Any-Step Defense), a token-level alignment method that aligns dLLMs to emit an [EOS] refusal signal whenever harmful content arises. By aligning safety directly at the token-level under randomized masking, A2D achieves robustness to both any-decoding-order and any-step prefilling attacks under various conditions. It also enables real-time monitoring: dLLMs may begin a response but automatically terminate if unsafe continuation emerges. On safety benchmarks, A2D consistently prevents the generation of harmful outputs, slashing DIJA success rates from over 80% to near-zero (1.3% on LLaDA-8B-Instruct, 0.0% on Dream-v0-Instruct-7B), and thresholded [EOS] probabilities allow early rejection, yielding up to 19.3x faster safe termination.

Method: Introduces Error-Uncertainty (EU) Plane diagnostic framework to characterize data utility across samples and tokens. Q-Tuning uses two-stage strategy: 1) sample-level triage to retain examples with informative misconceptions or calibration signals, 2) asymmetric token-pruning policy with context-aware scoring to trim less salient tokens from misconception samples while preserving calibration samples entirely.

Result: Sets new SOTA across five diverse benchmarks. On SmolLM2-1.7B, achieves +38% average improvement over full-data SFT baseline using only 12.5% of original training data. First dynamic pruning approach to consistently outperform full-data training.

Conclusion: Q-Tuning provides practical and scalable blueprint for maximizing data utilization in budget-constrained LLM SFT by jointly optimizing sample and token pruning through unified framework.

Abstract: As supervised fine-tuning (SFT) evolves from a lightweight post-training step into a compute-intensive phase rivaling mid-training in scale, data efficiency has become critical for aligning large language models (LLMs) under tight budgets. Existing data pruning methods suffer from a fragmented design: they operate either at the sample level or the token level in isolation, failing to jointly optimize both dimensions. This disconnect leads to significant inefficiencies–high-value samples may still contain redundant tokens, while token-level pruning often discards crucial instructional or corrective signals embedded in individual examples. To address this bottleneck, we introduce the Error-Uncertainty (EU) Plane, a diagnostic framework that jointly characterizes the heterogeneous utility of training data across samples and tokens. Guided by this insight, we propose Quadrant-based Tuning (Q-Tuning), a unified framework that strategically coordinates sample pruning and token pruning. Q-Tuning employs a two-stage strategy: first, it performs sample-level triage to retain examples rich in informative misconceptions or calibration signals; second, it applies an asymmetric token-pruning policy, using a context-aware scoring mechanism to trim less salient tokens exclusively from misconception samples while preserving calibration samples in their entirety. Our method sets a new state of the art across five diverse benchmarks. Remarkably, on SmolLM2-1.7B, Q-Tuning achieves a +38% average improvement over the full-data SFT baseline using only 12.5% of the original training data. As the first dynamic pruning approach to consistently outperform full-data training, Q-Tuning provides a practical and scalable blueprint for maximizing data utilization in budget-constrained LLM SFT.

Method: Develop MathLens benchmark with textbook-style geometry problems derived from symbolic specifications, accompanied by visual diagrams, text-only variants, multimodal questions, and targeted perceptual probes. This enables controlled measurement of Perception (visual understanding), Reasoning (logical processing), and Integration (combining modalities).

Result: Different training strategies yield systematically different capability profiles: reinforcement learning primarily improves perceptual grounding and robustness to diagram variation, while textual SFT yields gains through reflective reasoning. As perception and reasoning improve, integration errors become the dominant failure mode.

Conclusion: Progress in multimodal reasoning reflects shifting balances among subskills rather than uniform advancement, motivating evaluation beyond scalar accuracy. The decomposition approach reveals hidden capability profiles and failure modes invisible in aggregate metrics.

Abstract: Evaluation of multimodal reasoning models is typically reduced to a single accuracy score, implicitly treating reasoning as a unitary capability. We introduce MathLens, a benchmark of textbook-style geometry problems that exposes this assumption by operationally decomposing performance into Perception, Reasoning, and Integration. Each problem is derived from a symbolic specification and accompanied by visual diagrams, text-only variants, multimodal questions, and targeted perceptual probes, enabling controlled measurement of each component. Using this decomposition, we show that common training strategies induce systematically different capability profiles that are invisible under aggregate accuracy. Reinforcement learning primarily improves perceptual grounding and robustness to diagram variation, while textual SFT yields gains through reflective reasoning. In contrast, as perception and reasoning improve, a growing fraction of remaining errors fall outside these components and are categorized as integration. These results suggest that apparent progress in multimodal reasoning reflects shifting balances among subskills rather than uniform advancement, motivating evaluation beyond scalar accuracy.

Method: Empirical study of scaling properties of model reuse, analyzing how scaling exponent with respect to second-stage training tokens decreases logarithmically with the number of tokens used to pretrain the base model.

Result: Found that scaling efficiency diminishes in a predictable manner, with joint dependence on first- and second-stage tokens accurately modeled by a simple scaling law, revealing saturation effect in multi-stage pretraining.

Conclusion: There’s a fundamental trade-off in multi-stage pretraining: the more extensively a base model is pretrained, the less benefit additional pretraining provides, offering practical insights for efficient language model training and raising considerations for reuse of overtrained models.

Abstract: Reusing pretrained base models for further pretraining, such as continual pretraining or model growth, is promising at reducing the cost of training language models from scratch. However, the effectiveness remains unclear, especially when applied to overtrained base models. In this work, we empirically study the scaling properties of model reuse and find that the scaling efficiency diminishes in a predictable manner: The scaling exponent with respect to second-stage training tokens decreases logarithmically with the number of tokens used to pretrain the base model. The joint dependence on first- and second-stage tokens is accurately modeled by a simple scaling law. Such saturation effect reveals a fundamental trade-off in multi-stage pretraining strategies: the more extensively a base model is pretrained, the less benefit additional pretraining provides. Our findings provide practical insights for efficient language model training and raise important considerations for the reuse of overtrained models.

Method: Proposes training-free fingerprinting based on weight matrices. Uses Linear Assignment Problem (LAP) and unbiased Centered Kernel Alignment (CKA) similarity to neutralize effects of parameter manipulations. Creates robust similarity metric for model lineage verification.

Result: Tested on 60 positive and 90 negative model pairs. Demonstrates exceptional robustness against all six post-training categories with near-zero false positive risk. Achieves perfect scores on all classification metrics. Computation completes within 30s on NVIDIA 3090 GPU.

Conclusion: Method establishes strong basis for reliable model lineage verification. Provides effective solution for intellectual property protection of LLMs despite intensive post-training modifications.

Abstract: Protecting the intellectual property of large language models (LLMs) is crucial, given the substantial resources required for their training. Consequently, there is an urgent need for both model owners and third parties to determine whether a suspect LLM is trained from scratch or derived from an existing base model. However, the intensive post-training processes that models typically undergo-such as supervised fine-tuning, extensive continued pretraining, reinforcement learning, multi-modal extension, pruning, and upcycling-pose significant challenges to reliable identification. In this work, we propose a training-free fingerprinting method based on weight matrices. We leverage the Linear Assignment Problem (LAP) and an unbiased Centered Kernel Alignment (CKA) similarity to neutralize the effects of parameter manipulations, yielding a highly robust and high-fidelity similarity metric. On a comprehensive testbed of 60 positive and 90 negative model pairs, our method demonstrates exceptional robustness against all six aforementioned post-training categories while exhibiting a near-zero risk of false positives. By achieving perfect scores on all classification metrics, our approach establishes a strong basis for reliable model lineage verification. Moreover, the entire computation completes within 30s on an NVIDIA 3090 GPU. The code is available at https://github.com/LUMIA-Group/AWM.

Method: Models hate speech generation as a causal graph involving contextual environment, creator motivation, target, and style. Uses causal representation learning to disentangle these interpretable latent factors and control confounders. Allows counterfactual reasoning by intervening on style within the latent space to identify hate speech in varying forms.

Result: CADET demonstrates superior performance in comprehensive experiments, showing the potential of causal priors in advancing generalizable hate speech detection.

Conclusion: Causal representation learning with disentangled latent factors and counterfactual reasoning enables robust hate speech detection across diverse stylistic variations, addressing limitations of surface-level approaches.

Abstract: The proliferation of online hate speech poses a significant threat to the harmony of the web. While explicit hate is easily recognized through overt slurs, implicit hate speech is often conveyed through sarcasm, irony, stereotypes, or coded language – making it harder to detect. Existing hate speech detection models, which predominantly rely on surface-level linguistic cues, fail to generalize effectively across diverse stylistic variations. Moreover, hate speech spread on different platforms often targets distinct groups and adopts unique styles, potentially inducing spurious correlations between them and labels, further challenging current detection approaches. Motivated by these observations, we hypothesize that the generation of hate speech can be modeled as a causal graph involving key factors: contextual environment, creator motivation, target, and style. Guided by this graph, we propose CADET, a causal representation learning framework that disentangles hate speech into interpretable latent factors and then controls confounders, thereby isolating genuine hate intent from superficial linguistic cues. Furthermore, CADET allows counterfactual reasoning by intervening on style within the latent space, naturally guiding the model to robustly identify hate speech in varying forms. CADET demonstrates superior performance in comprehensive experiments, highlighting the potential of causal priors in advancing generalizable hate speech detection.

Method: 1) OpenRubrics: large-scale collection of (prompt, rubric) pairs; 2) Contrastive Rubric Generation (CRG): derives both hard rules (explicit constraints) and principles (implicit qualities) by contrasting preferred and rejected responses; 3) Noise removal via preference-label consistency preservation.

Result: Rubric-RM surpasses strong size-matched baselines by 8.4% across multiple reward-modeling benchmarks. Gains transfer to policy models on instruction-following and biomedical benchmarks.

Conclusion: Structured rubric-based reward modeling using OpenRubrics and CRG provides more comprehensive evaluation signals than traditional approaches, leading to improved performance in RLHF applications.

Abstract: Reward modeling lies at the core of reinforcement learning from human feedback (RLHF), yet most existing reward models rely on scalar or pairwise judgments that fail to capture the multifaceted nature of human preferences. Recent studies have explored rubrics-as-rewards (RaR) that uses structured criteria to capture multiple dimensions of response quality. However, producing rubrics that are both reliable and scalable remains a key challenge. In this work, we introduce OpenRubrics, a diverse, large-scale collection of (prompt, rubric) pairs for training rubric-generation and rubric-based reward models. To elicit discriminative and comprehensive evaluation signals, we introduce Contrastive Rubric Generation (CRG), which derives both hard rules (explicit constraints) and principles (implicit qualities) by contrasting preferred and rejected responses. We further remove noisy rubrics via preserving preference-label consistency. Across multiple reward-modeling benchmarks, our rubric-based reward model, Rubric-RM, surpasses strong size-matched baselines by 8.4%. These gains transfer to policy models on instruction-following and biomedical benchmarks.

Method: Compare training on majority-vote labels with various HLV methods, conducting experiments to assess fairness impacts

Result: HLV training methods have a positive impact on fairness under certain configurations without requiring explicit debiasing techniques

Conclusion: Human label variation methods can contribute to improved model fairness, offering an alternative or complementary approach to explicit debiasing methods

Abstract: The impact of human label variation (HLV) on model fairness is an unexplored topic. This paper examines the interplay by comparing training on majority-vote labels with a range of HLV methods. Our experiments show that without explicit debiasing, HLV training methods have a positive impact on fairness under certain configurations.

Method: DynaSpec uses lightweight meta-classifiers to route each context to a small set of coarse token clusters. The union of top-selected clusters defines the drafter’s dynamic shortlist, while the target model still verifies over the full vocabulary (preserving exactness). System-wise, routing is overlapped with draft computation via parallel execution streams to reduce overhead.

Result: Across standard benchmarks, DynaSpec recovers 98.4% of full-vocabulary performance for Llama-3-8B (vs 93.6% for fixed-shortlist baselines) and achieves up to 2.23x throughput gain (vs 1.91x for static approaches) on datasets with rare tokens.

Conclusion: DynaSpec enables efficient speculative decoding for large-vocabulary LLMs by dynamically adapting shortlists to context, significantly improving acceptance rates and throughput while preserving exact verification.

Abstract: Speculative decoding accelerates LLM inference by letting a small drafter propose multiple tokens which a large target model verifies once per speculation step. As vocabularies scale past 10e5 tokens,verification cost in the target model is largely unchanged, but the drafter can become bottlenecked by its O(|V|d) output projection. Recent approaches (e.g., FR-Spec, VocabTrim) mitigate this by restricting drafting to a fixed, frequency-ranked shortlist; however, such static truncation is corpus-dependent and suppresses rare or domain-specific tokens, reducing acceptance and limiting speedups. We propose DynaSpec, a context-dependent dynamic shortlisting mechanism for large-vocabulary speculative decoding. DynaSpec trains lightweight meta-classifiers that route each context to a small set of coarse token clusters; the union of the top-selected clusters defines the drafter’s shortlist, while the target model still verifies over the full vocabulary, preserving exactness. Systems-wise, routing is overlapped with draft computation via parallel execution streams, reducing end-to-end overhead. Across standard speculative decoding benchmarks, DynaSpec consistently improves mean accepted length-recovering 98.4% of full-vocabulary performance for Llama-3-8B versus 93.6% for fixed-shortlist baselines-and achieves up to a 2.23x throughput gain compared to 1.91x for static approaches on the dataset with rare tokens.

Method: Proposes POPI framework with two modular components: preference inference (extracts user preferences) and conditioned generation (generates responses based on preferences). Uses natural language as interface between components. Jointly optimizes both components using reinforcement learning with unified preference optimization objective.

Result: POPI improves personalization performance across multiple benchmarks while reducing context overhead. Learned natural-language preference summaries transfer effectively to frozen, off-the-shelf LLMs including black-box APIs, demonstrating modularity and generator-transferability.

Conclusion: Modular personalization framework with explicit decomposition into preference inference and conditioned generation enables effective, transferable personalization. Natural language serves as effective interface between components, allowing preference summaries to work across different LLMs.

Abstract: Large language models (LLMs) are typically aligned with population-level preferences, despite substantial variation across individual users. While many LLM personalization methods exist, the underlying structure of user-level personalization is often left implicit. We formalize user-level, prompt-independent personalization as a decomposition into two components: preference inference and conditioned generation. We advocate for a modular design that decouples these components; identify natural language as a generator-agnostic interface between them; and characterize generator-transferability as a key implication of modular personalization. Guided by this abstraction, we introduce POPI, a novel instantiation of modular personalization that parameterizes both preference inference and conditioned generation as shared LLMs. POPI jointly optimizes the two components under a unified preference optimization objective, using reinforcement learning as an optimization tool. Across multiple benchmarks, POPI consistently improves personalization performance while reducing context overhead. We further demonstrate that the learned natural-language preference summaries transfer effectively to frozen, off-the-shelf LLMs, including black-box APIs, providing empirical evidence of modularity and generator-transferability.

Method: Two-stage pipeline: 1) Train a style-similarity judge by fine-tuning a sentence-transformer with authorship-verification supervision and calibrate outputs to [0,1] reward, 2) Use this judge as primary reward in Group Relative Policy Optimization (GRPO) to fine-tune an 8B story generator for style-conditioned writing

Result: GRPO-trained 8B model achieves higher style scores than open-weight baselines across four target authors (Mark Twain, Jane Austen, Charles Dickens, Thomas Hardy), with average style score of 0.893

Conclusion: AV-calibrated reward modelling provides a practical mechanism for controllable style transfer in long-form generation under moderate model size and training budget

Abstract: Evaluating and optimising authorial style in long-form story generation remains challenging because style is often assessed with ad hoc prompting and is frequently conflated with overall writing quality. We propose a two-stage pipeline. First, we train a dedicated style-similarity judge by fine-tuning a sentence-transformer with authorship-verification supervision, and calibrate its similarity outputs into a bounded $[0,1]$ reward. Second, we use this judge as the primary reward in Group Relative Policy Optimization (GRPO) to fine-tune an 8B story generator for style-conditioned writing, avoiding the accept/reject supervision required by Direct Preference Optimization (DPO). Across four target authors (Mark Twain, Jane Austen, Charles Dickens, Thomas Hardy), the GRPO-trained 8B model achieves higher style scores than open-weight baselines, with an average style score of 0.893 across authors. These results suggest that AV-calibrated reward modelling provides a practical mechanism for controllable style transfer in long-form generation under a moderate model size and training budget.

Method: DEER systematizes evaluation with an expert-developed taxonomy (7 dimensions, 25 subdimensions, 101 rubric items), provides Expert Evaluation Guidance for LLM-based judging, and includes a claim verification architecture that verifies both cited and uncited claims while quantifying evidence quality.

Result: Experiments show current deep research systems can produce structurally plausible reports with citations, but need improvement in fulfilling expert-level requests and achieving logical completeness. DEER enables interpretable system analysis and diagnostic improvement signals.

Conclusion: DEER provides a comprehensive benchmark for evaluating expert-level deep research reports, making system strengths/limitations interpretable and offering diagnostic signals for improvement beyond simple performance comparisons.

Abstract: Recent advances in large language models have enabled deep research systems that generate expert-level reports through multi-step reasoning and evidence-based synthesis. However, evaluating such reports remains challenging: report quality is multifaceted, making it difficult to determine what to assess and by what criteria; LLM-based judges may miss errors that require domain expertise to identify; and because deep research relies on retrieved evidence, report-wide claim verification is also necessary. To address these issues, we propose DEER, a benchmark for evaluating expert-level deep research reports. DEER systematizes evaluation criteria with an expert-developed taxonomy (7 dimensions, 25 subdimensions) operationalized as 101 fine-grained rubric items. We also provide task-specific Expert Evaluation Guidance to support LLM-based judging. Alongside rubric-based assessment, we propose a claim verification architecture that verifies both cited and uncited claims and quantifies evidence quality. Experiments show that while current deep research systems can produce structurally plausible reports that cite external evidence, there is room for improvement in fulfilling expert-level user requests and achieving logical completeness. Beyond simple performance comparisons, DEER makes system strengths and limitations interpretable and provides diagnostic signals for improvement.

Method: The paper reviews existing definitions and folds them into a single unified definition where hallucinations are defined as inaccurate internal world modeling observable to users. The framework varies reference world models and conflict policies to subsume prior definitions.

Result: The unified framework distinguishes true hallucinations from planning or reward errors, provides common language for comparison across benchmarks, and enables clearer evaluation by forcing clarification of assumed reference “worlds.”

Conclusion: The unified view of hallucinations as inaccurate world modeling is useful for evaluation and mitigation. The authors plan to develop synthetic benchmarks with fully specified reference world models to stress-test and improve world modeling components in LLMs.

Abstract: Despite numerous attempts at mitigation since the inception of language models, hallucinations remain a persistent problem even in today’s frontier LLMs. Why is this? We review existing definitions of hallucination and fold them into a single, unified definition wherein prior definitions are subsumed. We argue that hallucination can be unified by defining it as simply inaccurate (internal) world modeling, in a form where it is observable to the user. For example, stating a fact which contradicts a knowledge base OR producing a summary which contradicts the source. By varying the reference world model and conflict policy, our framework unifies prior definitions. We argue that this unified view is useful because it forces evaluations to clarify their assumed reference “world”, distinguishes true hallucinations from planning or reward errors, and provides a common language for comparison across benchmarks and discussion of mitigation strategies. Building on this definition, we outline plans for a family of benchmarks using synthetic, fully specified reference world models to stress-test and improve world modeling components.

Method: The study analyzes BERT-12 architecture by examining attention maps and deriving context similarity matrices from vector spaces emerging from self-attention heads. It measures scalar products between token vectors to reveal similarities between token pairs within each head and layer.

Result: Final layers focus on sentence separator tokens, suggesting text segmentation based on semantic features. Different attention heads specialize in different linguistic characteristics (token repetitions, common tokens with context). Initial layers show long-range similarities, while deeper layers develop short-range similarities with preference for intra-sentence attention. Each head focuses on unique tokens and builds similarity pairs around them.

Conclusion: Self-attention mechanisms exhibit systematic specialization patterns where different heads capture distinct linguistic features, and attention evolves from long-range to short-range semantic relationships across layers, providing insights into how transformers process language.

Abstract: The self-attention mechanism has significantly advanced the field of natural language processing, facilitating the development of advanced language-learning machines. Although its utility is widely acknowledged, the precise mechanisms of self-attention underlying its advanced learning and the quantitative characterization of this learning process remains an open research question. This study introduces a new approach for quantifying information processing within the self-attention mechanism. The analysis conducted on the BERT-12 architecture reveals that, in the final layers, the attention map focuses on sentence separator tokens, suggesting a practical approach to text segmentation based on semantic features. Based on the vector space emerging from the self-attention heads, a context similarity matrix, measuring the scalar product between two token vectors was derived, revealing distinct similarities between different token vector pairs within each head and layer. The findings demonstrated that different attention heads within an attention block focused on different linguistic characteristics, such as identifying token repetitions in a given text or recognizing a token of common appearance in the text and its surrounding context. This specialization is also reflected in the distribution of distances between token vectors with high similarity as the architecture progresses. The initial attention layers exhibit substantially long-range similarities; however, as the layers progress, a more short-range similarity develops, culminating in a preference for attention heads to create strong similarities within the same sentence. Finally, the behavior of individual heads was analyzed by examining the uniqueness of their most common tokens in their high similarity elements. Each head tends to focus on a unique token from the text and builds similarity pairs centered around it.

Method: Developed TurkBench with 8,151 data samples across 21 distinct subtasks organized under six main categories: Knowledge, Language Understanding, Reasoning, Content Moderation, Turkish Grammar and Vocabulary, and Instruction Following.

Result: Created a comprehensive Turkish-language benchmark that provides culturally relevant data for evaluating generative LLMs, published for online submissions via Hugging Face.

Conclusion: TurkBench fills an important gap in multilingual LLM evaluation by providing a valuable tool for researchers and developers to assess Turkish-language models and identify improvement areas.

Abstract: With the recent surge in the development of large language models, the need for comprehensive and language-specific evaluation benchmarks has become critical. While significant progress has been made in evaluating English-language models, benchmarks for other languages, particularly those with unique linguistic characteristics such as Turkish, remain less developed. Our study introduces TurkBench, a comprehensive benchmark designed to assess the capabilities of generative large language models in the Turkish language. TurkBench involves 8,151 data samples across 21 distinct subtasks. These are organized under six main categories of evaluation: Knowledge, Language Understanding, Reasoning, Content Moderation, Turkish Grammar and Vocabulary, and Instruction Following. The diverse range of tasks and the culturally relevant data would provide researchers and developers with a valuable tool for evaluating their models and identifying areas for improvement. We further publish our benchmark for online submissions at https://huggingface.co/turkbench

Method: Proposes Ca2KG framework with two components: 1) counterfactual prompting to expose retrieval-dependent uncertainties in knowledge quality and reasoning reliability, and 2) panel-based re-scoring mechanism that stabilizes predictions across interventions.

Result: Extensive experiments on two complex QA datasets show Ca2KG consistently improves calibration while maintaining or even enhancing predictive accuracy.

Conclusion: Ca2KG effectively addresses the overconfidence problem in KG-RAG through causality-aware calibration, making it more reliable for deployment in high-stakes domains.

Abstract: Knowledge Graph Retrieval-Augmented Generation (KG-RAG) extends the RAG paradigm by incorporating structured knowledge from knowledge graphs, enabling Large Language Models (LLMs) to perform more precise and explainable reasoning. While KG-RAG improves factual accuracy in complex tasks, existing KG-RAG models are often severely overconfident, producing high-confidence predictions even when retrieved sub-graphs are incomplete or unreliable, which raises concerns for deployment in high-stakes domains. To address this issue, we propose Ca2KG, a Causality-aware Calibration framework for KG-RAG. Ca2KG integrates counterfactual prompting, which exposes retrieval-dependent uncertainties in knowledge quality and reasoning reliability, with a panel-based re-scoring mechanism that stabilises predictions across interventions. Extensive experiments on two complex QA datasets demonstrate that Ca2KG consistently improves calibration while maintaining or even enhancing predictive accuracy.

Method: Proposes Bounded Hyperbolic Tanh (BHyT) which couples tanh nonlinearity with explicit, data-driven input bounding to keep activations in non-saturating range. Computes exact statistics once per block and replaces second normalization with lightweight variance approximation for efficiency.

Result: BHyT achieves 15.8% faster training and 4.2% higher token generation throughput compared to RMSNorm, while matching or surpassing inference performance and robustness across language understanding and reasoning benchmarks.

Conclusion: BHyT jointly addresses stability and efficiency issues in deep LLMs, providing a practical drop-in replacement for Pre-LN with theoretical stability guarantees and empirical performance improvements.

Abstract: Pre-Layer Normalization (Pre-LN) is the de facto choice for large language models (LLMs) and is crucial for stable pretraining and effective transfer learning. However, Pre-LN is inefficient due to repeated statistical calculations and suffers from the curse of depth. As layers grow, the magnitude and variance of the hidden state escalate, destabilizing training. Efficiency-oriented normalization-free methods such as Dynamic Tanh (DyT) improve speed but remain fragile at depth. To jointly address stability and efficiency, we propose Bounded Hyperbolic Tanh (BHyT), a drop-in replacement for Pre-LN. BHyT couples a tanh nonlinearity with explicit, data-driven input bounding to keep activations within a non-saturating range. It prevents depth-wise growth in activation magnitude and variance and comes with a theoretical stability guarantee. For efficiency, BHyT computes exact statistics once per block and replaces a second normalization with a lightweight variance approximation, enhancing efficiency. Empirically, BHyT demonstrates improved stability and efficiency during pretraining, achieving an average of 15.8% faster training and an average of 4.2% higher token generation throughput compared to RMSNorm., while matching or surpassing its inference performance and robustness across language understanding and reasoning benchmarks. Our code is available at: https://anonymous.4open.science/r/BHyT

Method: Two-stage LoRA pipeline with separate domain-oriented pre-training and supervised fine-tuning, followed by weighted adapter merging. Includes merge-verification routine to ensure correctness.

Result: PT signal systematically alters model behavior, producing reasoning-style outputs even when suppressed. BLEU/ROUGE scores drop while multiple-choice accuracy improves. Small pipeline mistakes can misattribute behavior.

Conclusion: Adapter interference between knowledge injection and instruction alignment has important implications for safety-critical model deployment, requiring careful verification.

Abstract: Large language models can exhibit surprising adapter interference when combining domain adaptation and instruction alignment in safety-critical settings. We study a two-stage LoRA pipeline for medical LLMs, where domain-oriented pre-training (PT) and supervised fine-tuning (SFT) are trained separately and later merged through weighted adapter merging. We observe that introducing PT signal can systematically alter model behavior and produce reasoning-style outputs, even when evaluation templates explicitly attempt to suppress such behavior. This interference leads to a divergence between surface metrics and reasoning or alignment behavior: BLEU/ROUGE scores drop significantly, while multiple-choice accuracy improves. We further show that small pipeline mistakes can easily misattribute SFT-only behavior to merged models, and provide a lightweight merge-verification routine to ensure correctness and reproducibility. Our findings highlight an interaction between knowledge injection and instruction alignment in adapter-based fine-tuning, with important implications for safety-critical model deployment.

Method: A three-stage text-to-state mapping framework: 1) conflict detection using rule-based segmentation for explicit markers and LLM-based enumeration for implicit ambiguity, 2) interpretation extraction, and 3) state construction. The approach combines rule-based methods for explicit conflict markers (adversative conjunctions, hedging expressions) with LLM-based techniques for implicit ambiguity types (epistemic, lexical, structural).

Result: On 68 ambiguous sentences, the framework preserved interpretive multiplicity with mean state entropy H = 1.087 bits across ambiguity categories, compared to H = 0 for collapse-based baselines. Demonstrated cross-lingual portability with Japanese markers.

Conclusion: The framework provides an algorithmic bridge between text and Non-Resolution Reasoning state space, enabling architectural collapse deferment in LLM inference by preserving multiple interpretations rather than forcing early semantic commitment.

Abstract: Large language models exhibit a systematic tendency toward early semantic commitment: given ambiguous input, they collapse multiple valid interpretations into a single response before sufficient context is available. We present a formal framework for text-to-state mapping ($φ: \mathcal{T} \to \mathcal{S}$) that transforms natural language into a non-collapsing state space where multiple interpretations coexist. The mapping decomposes into three stages: conflict detection, interpretation extraction, and state construction. We instantiate $φ$ with a hybrid extraction pipeline combining rule-based segmentation for explicit conflict markers (adversative conjunctions, hedging expressions) with LLM-based enumeration of implicit ambiguity (epistemic, lexical, structural). On a test set of 68 ambiguous sentences, the resulting states preserve interpretive multiplicity: mean state entropy $H = 1.087$ bits across ambiguity categories, compared to $H = 0$ for collapse-based baselines. We additionally instantiate the rule-based conflict detector for Japanese markers to illustrate cross-lingual portability. This framework extends Non-Resolution Reasoning (NRR) by providing the missing algorithmic bridge between text and the NRR state space, enabling architectural collapse deferment in LLM inference.

Method: Soft-Verified Efficient Repository Agents (SERA) uses Soft Verified Generation (SVG) to generate thousands of trajectories from a single code repository, combined with supervised finetuning for efficient training.

Result: SERA achieves state-of-the-art results among fully open-source models while matching frontier open-weight models, with 26x cheaper training than reinforcement learning and 57x cheaper than previous synthetic data methods.

Conclusion: The work makes repository specialization practical for open coding agents, accelerating research in this area and demonstrating the advantage of open-source models that can adapt to private codebases.

Abstract: Open-weight coding agents should hold a fundamental advantage over closed-source systems: they can be specialized to private codebases, encoding repository-specific information directly in their weights. Yet the cost and complexity of training has kept this advantage theoretical. We show it is now practical. We present Soft-Verified Efficient Repository Agents (SERA), an efficient method for training coding agents that enables the rapid and cheap creation of agents specialized to private codebases. Using only supervised finetuning (SFT), SERA achieves state-of-the-art results among fully open-source (open data, method, code) models while matching the performance of frontier open-weight models like Devstral-Small-2. Creating SERA models is 26x cheaper than reinforcement learning and 57x cheaper than previous synthetic data methods to reach equivalent performance. Our method, Soft Verified Generation (SVG), generates thousands of trajectories from a single code repository. Combined with cost-efficiency, this enables specialization to private codebases. Beyond repository specialization, we apply SVG to a larger corpus of codebases, generating over 200,000 synthetic trajectories. We use this dataset to provide detailed analysis of scaling laws, ablations, and confounding factors for training coding agents. Overall, we believe our work will greatly accelerate research on open coding agents and showcase the advantage of open-source models that can specialize to private codebases. We release SERA as the first model in Ai2’s Open Coding Agents series, along with all our code, data, and Claude Code integration to support the research community.

Method: Studied representation dynamics by analyzing how linear directions change during simulated conversations, examining changes in factuality representations, robustness across model families and layers, and effects of different conversation types (on-policy vs scripted).

Result: Representations change dramatically during conversations - factual information can become non-factual and vice versa. Changes are content-dependent, robust across models, and occur even with scripted conversations. Steering along representational directions has different effects at different conversation points.

Conclusion: Language model representations evolve dynamically in response to conversational context, challenging static interpretations of features and suggesting models adapt to play specific roles cued by conversations.

Abstract: Language model representations often contain linear directions that correspond to high-level concepts. Here, we study the dynamics of these representations: how representations evolve along these dimensions within the context of (simulated) conversations. We find that linear representations can change dramatically over a conversation; for example, information that is represented as factual at the beginning of a conversation can be represented as non-factual at the end and vice versa. These changes are content-dependent; while representations of conversation-relevant information may change, generic information is generally preserved. These changes are robust even for dimensions that disentangle factuality from more superficial response patterns, and occur across different model families and layers of the model. These representation changes do not require on-policy conversations; even replaying a conversation script written by an entirely different model can produce similar changes. However, adaptation is much weaker from simply having a sci-fi story in context that is framed more explicitly as such. We also show that steering along a representational direction can have dramatically different effects at different points in a conversation. These results are consistent with the idea that representations may evolve in response to the model playing a particular role that is cued by a conversation. Our findings may pose challenges for interpretability and steering – in particular, they imply that it may be misleading to use static interpretations of features or directions, or probes that assume a particular range of features consistently corresponds to a particular ground-truth value. However, these types of representational dynamics also point to exciting new research directions for understanding how models adapt to context.

Method: Created MilSCORE, the first scenario-level dataset of expert-authored, multi-hop questions grounded in complex simulated military planning scenarios. Includes diverse question types across seven categories targeting factual recall and multi-step reasoning about constraints, strategy, and spatial analysis.

Result: Baseline results for contemporary vision-language models show substantial headroom, indicating current systems struggle with realistic, scenario-level long-context planning.

Conclusion: MilSCORE serves as a challenging testbed for future work on long-context, multi-modal reasoning in complex planning scenarios.

Abstract: As large language models (LLMs) are applied to increasingly longer and more complex tasks, there is a growing need for realistic long-context benchmarks that require selective reading and integration of heterogeneous, multi-modal information sources. This need is especially acute for geospatial planning problems, such as those found in planning for large-scale military operations, which demand fast and accurate reasoning over maps, orders, intelligence reports, and other distributed data. To address this gap, we present MilSCORE (Military Scenario Contextual Reasoning), to our knowledge the first scenario-level dataset of expert-authored, multi-hop questions grounded in a complex, simulated military planning scenario used for training. MilSCORE is designed to evaluate high-stakes decision-making and planning, probing LLMs’ ability to combine tactical and spatial reasoning across multiple sources and to reason over long-horizon, geospatially rich context. The benchmark includes a diverse set of question types across seven categories targeting both factual recall and multi-step reasoning about constraints, strategy, and spatial analysis. We provide an evaluation protocol and report baseline results for a range of contemporary vision-language models. Our findings highlight substantial headroom on MilSCORE, indicating that current systems struggle with realistic, scenario-level long-context planning, and positioning MilSCORE as a challenging testbed for future work.

Method: Proposes a GPT-based architecture with selective fine-tuning strategy: freezes majority of GPT-2 backbone, trains only final Transformer block, final layer normalization, and lightweight classification head to reduce trainable parameters while preserving representational capacity.

Result: Evaluated on MIMIC-IV-Note radiology reports with CheXpert-style labels; shows stable convergence and strong classification performance across multiple problem formulations, particularly effective for non-mention and negated findings.

Conclusion: Selective fine-tuning of pretrained generative language models provides efficient and effective pathway for clinical text classification, enabling scalable adaptation to real-world EHR data with significantly reduced computational complexity.

Abstract: The increasing availability of unstructured clinical narratives in electronic health records (EHRs) has created new opportunities for automated disease characterization, cohort identification, and clinical decision support. However, modeling long, domain-specific clinical text remains challenging due to limited labeled data, severe class imbalance, and the high computational cost of adapting large pretrained language models. This study presents a GPT-based architecture for clinical text classification that adapts a pretrained decoder-only Transformer using a selective fine-tuning strategy. Rather than updating all model parameters, the majority of the GPT-2 backbone is frozen, and training is restricted to the final Transformer block, the final layer normalization, and a lightweight classification head. This approach substantially reduces the number of trainable parameters while preserving the representational capacity required to model complex clinical language. The proposed method is evaluated on radiology reports from the MIMIC-IV-Note dataset using uncertainty-aware CheXpert-style labels derived directly from report text. Experiments cover multiple problem formulations, including multi-label classification of radiographic findings, binary per-label classification under different uncertainty assumptions, and aggregate disease outcome prediction. Across varying dataset sizes, the model exhibits stable convergence behavior and strong classification performance, particularly in settings dominated by non-mention and negated findings. Overall, the results indicate that selective fine-tuning of pretrained generative language models provides an efficient and effective pathway for clinical text classification, enabling scalable adaptation to real-world EHR data while significantly reducing computational complexity.

Method: Developed a semi-automated annotation methodology based on a faceted taxonomy, implemented through an annotation extension framework. Created an annotation extension tool for Turkish that automatically extends existing flat annotations by inferring additional linguistic and metadata information as facets.

Result: Achieved 95.86% facet-level accuracy. Produced the first collaboratively annotated and taxonomically enriched Turkish Learner Corpus with enhanced querying capabilities and detailed exploratory analysis support.

Conclusion: The work introduces a novel approach to learner corpus annotation that enables complex linguistic and pedagogical analysis. It provides the first corpus designed according to the new taxonomy and paves the way for enriching existing error-annotated learner corpora.

Abstract: In terms of annotation structure, most learner corpora rely on holistic flat label inventories which, even when extensive, do not explicitly separate multiple linguistic dimensions. This makes linguistically deep annotation difficult and complicates fine-grained analyses aimed at understanding why and how learners produce specific errors. To address these limitations, this paper presents a semi-automated annotation methodology for learner corpora, built upon a recently proposed faceted taxonomy, and implemented through a novel annotation extension framework. The taxonomy provides a theoretically grounded, multi-dimensional categorization that captures the linguistic properties underlying each error instance, thereby enabling standardized, fine-grained, and interpretable enrichment beyond flat annotations. The annotation extension tool, implemented based on the proposed extension framework for Turkish, automatically extends existing flat annotations by inferring additional linguistic and metadata information as facets within the taxonomy to provide richer learner-specific context. It was systematically evaluated and yielded promising performance results, achieving a facet-level accuracy of 95.86%. The resulting taxonomically enriched corpus offers enhanced querying capabilities and supports detailed exploratory analyses across learner corpora, enabling researchers to investigate error patterns through complex linguistic and pedagogical dimensions. This work introduces the first collaboratively annotated and taxonomically enriched Turkish Learner Corpus, a manual annotation guideline with a refined tagset, and an annotation extender. As the first corpus designed in accordance with the recently introduced taxonomy, we expect our study to pave the way for subsequent enrichment efforts of existing error-annotated learner corpora.

Method: Systematically constructed semantically natural prompts ending with partial tokens across three domains: languages without whitespace (Chinese), highly compounding languages, and code. Evaluated frontier LMs’ probability distributions on correct continuations versus token-aligned “backed-off” prompts.

Result: In Chinese, up to 25% of word boundaries don’t align with token boundaries. LMs consistently place three orders of magnitude less probability on correct continuations for partial token prompts. Degradation doesn’t diminish with scale and often worsens for larger models. Evaluated inference-time mitigations and validated recent exact solutions.

Conclusion: The partial token problem causes severe probability distortion in realistic use cases, particularly for languages with misaligned token/word boundaries. The issue persists across model scales. Practical recommendations are provided for model inference providers, and recent exact solutions show effectiveness.

Abstract: Language models (LMs) are trained over sequences of tokens, whereas users interact with LMs via text. This mismatch gives rise to the partial token problem, which occurs when a user ends their prompt in the middle of the expected next-token, leading to distorted next-token predictions. Although this issue has been studied using arbitrary character prefixes, its prevalence and severity in realistic prompts respecting word boundaries remains underexplored. In this work, we identify three domains where token and “word” boundaries often do not line up: languages that do not use whitespace, highly compounding languages, and code. In Chinese, for example, up to 25% of word boundaries do not line up with token boundaries, making even natural, word-complete prompts susceptible to this problem. We systematically construct semantically natural prompts ending with a partial tokens; in experiments, we find that they comprise a serious failure mode: frontier LMs consistently place three orders of magnitude less probability on the correct continuation compared to when the prompt is “backed-off” to be token-aligned. This degradation does not diminish with scale and often worsens for larger models. Finally, we evaluate inference-time mitigations to the partial token problem and validate the effectiveness of recent exact solutions. Overall, we demonstrate the scale and severity of probability distortion caused by tokenization in realistic use cases, and provide practical recommentions for model inference providers.

Method: Three-phase pipeline: 1) Mid-training with Plasticity-Adjusted Sampling for domain adaptation, 2) Supervised fine-tuning with Legal Agentic CoT Distillation to distill reasoning from legal texts, 3) Curriculum Reinforcement Learning with progressive stages (memorization, understanding, reasoning).

Result: LegalOne achieves state-of-the-art performance across various legal tasks, surpassing larger general-purpose LLMs through enhanced knowledge density and efficiency. The model weights and LegalKit evaluation framework are publicly released.

Conclusion: LegalOne demonstrates that specialized legal foundation models can overcome limitations of general LLMs in legal reasoning, paving the way for trustworthy and interpretable AI in high-stakes judicial applications.

Abstract: While Large Language Models (LLMs) have demonstrated impressive general capabilities, their direct application in the legal domain is often hindered by a lack of precise domain knowledge and complexity of performing rigorous multi-step judicial reasoning. To address this gap, we present LegalOne, a family of foundational models specifically tailored for the Chinese legal domain. LegalOne is developed through a comprehensive three-phase pipeline designed to master legal reasoning. First, during mid-training phase, we propose Plasticity-Adjusted Sampling (PAS) to address the challenge of domain adaptation. This perplexity-based scheduler strikes a balance between the acquisition of new knowledge and the retention of original capabilities, effectively establishing a robust legal foundation. Second, during supervised fine-tuning, we employ Legal Agentic CoT Distillation (LEAD) to distill explicit reasoning from raw legal texts. Unlike naive distillation, LEAD utilizes an agentic workflow to convert complex judicial processes into structured reasoning trajectories, thereby enforcing factual grounding and logical rigor. Finally, we implement a Curriculum Reinforcement Learning (RL) strategy. Through a progressive reinforcement process spanning memorization, understanding, and reasoning, LegalOne evolves from simple pattern matching to autonomous and reliable legal reasoning. Experimental results demonstrate that LegalOne achieves state-of-the-art performance across a wide range of legal tasks, surpassing general-purpose LLMs with vastly larger parameter counts through enhanced knowledge density and efficiency. We publicly release the LegalOne weights and the LegalKit evaluation framework to advance the field of Legal AI, paving the way for deploying trustworthy and interpretable foundation models in high-stakes judicial applications.

Method: Proposes TerminalTraj pipeline that: (1) filters high-quality repositories to construct Dockerized execution environments, (2) generates Docker-aligned task instances, and (3) synthesizes agent trajectories with executable validation code for verification.

Result: Curated 32K Docker images and generated 50,733 verified terminal trajectories across eight domains. Models trained on this data achieved up to 20% improvement on TerminalBench 1.0 and 10% on TerminalBench 2.0, with TerminalTraj-32B reaching 35.30% on TB 1.0 and 22.00% on TB 2.0.

Conclusion: TerminalTraj provides a scalable solution for generating high-quality terminal trajectory data, enabling improved training of agentic models for terminal-based tasks with better executability and verifiability.

Abstract: Training agentic models for terminal-based tasks critically depends on high-quality terminal trajectories that capture realistic long-horizon interactions across diverse domains. However, constructing such data at scale remains challenging due to two key requirements: \textbf{\emph{Executability}}, since each instance requires a suitable and often distinct Docker environment; and \textbf{\emph{Verifiability}}, because heterogeneous task outputs preclude unified, standardized verification. To address these challenges, we propose \textbf{TerminalTraj}, a scalable pipeline that (i) filters high-quality repositories to construct Dockerized execution environments, (ii) generates Docker-aligned task instances, and (iii) synthesizes agent trajectories with executable validation code. Using TerminalTraj, we curate 32K Docker images and generate 50,733 verified terminal trajectories across eight domains. Models trained on this data with the Qwen2.5-Coder backbone achieve consistent performance improvements on TerminalBench (TB), with gains of up to 20% on TB1.0 and 10% on TB2.0 over their respective backbones. Notably, \textbf{TerminalTraj-32B} achieves strong performance among models with fewer than 100B parameters, reaching 35.30% on TB1.0 and 22.00% on TB2.0, and demonstrates improved test-time scaling behavior. All code and data are available at https://github.com/Wusiwei0410/TerminalTraj.

Method: Introduces EverMemBench with multi-party, multi-group conversations spanning over 1 million tokens, featuring temporally evolving information, cross-topic interleaving, and role-specific personas. Evaluates through 1,000+ QA pairs across three dimensions: fine-grained recall, memory awareness, and user profile understanding.

Result: Reveals critical limitations: (1) multi-hop reasoning collapses in multi-party settings (oracle models achieve only 26%), (2) temporal reasoning remains unsolved requiring version semantics, (3) memory awareness bottlenecked by retrieval where similarity-based methods fail to bridge semantic gaps.

Conclusion: EverMemBench provides a challenging testbed for developing next-generation memory architectures for conversational AI systems.

Abstract: Long-term conversational memory is essential for LLM-based assistants, yet existing benchmarks focus on dyadic, single-topic dialogues that fail to capture real-world complexity. We introduce EverMemBench, a benchmark featuring multi-party, multi-group conversations spanning over 1 million tokens with temporally evolving information, cross-topic interleaving, and role-specific personas. EverMemBench evaluates memory systems across three dimensions through 1,000+ QA pairs: fine-grained recall, memory awareness, and user profile understanding. Our evaluation reveals critical limitations: (1) multi-hop reasoning collapses in multi-party settings, with even oracle models achieving only 26%; (2) temporal reasoning remains unsolved, requiring version semantics beyond timestamp matching; (3) memory awareness is bottlenecked by retrieval, where current similarity-based methods fail to bridge the semantic gap between queries and implicitly relevant memories. EverMemBench provides a challenging testbed for developing next-generation memory architectures.

Method: This is a position paper that makes three main contributions: 1) explains why HLS serves as a practical abstraction layer and golden reference for agentic hardware design, 2) identifies key limitations of current HLS tools that AI agents can address, and 3) proposes a taxonomy for symbiotic evolution of agentic HLS showing responsibility shift from humans to AI agents.

Result: The paper establishes HLS as a critical layer for AI-driven hardware optimization, highlighting its advantages (faster iteration cycles, portability, design permutability) and identifying specific limitations that AI agents can overcome (performance feedback, interface rigidity, debuggability).

Conclusion: HLS remains essential in the agentic era as it provides the necessary abstraction layer for AI agents to optimize hardware design, with a symbiotic evolution path from human-driven to AI-autonomous design systems.

Abstract: The rise of large language models has sparked interest in AI-driven hardware design, raising the question: does high-level synthesis (HLS) still matter in the agentic era? We argue that HLS remains essential. While we expect mature agentic hardware systems to leverage both HLS and RTL, this paper focuses on HLS and its role in enabling agentic optimization. HLS offers faster iteration cycles, portability, and design permutability that make it a natural layer for agentic optimization. This position paper makes three contributions. First, we explain why HLS serves as a practical abstraction layer and a golden reference for agentic hardware design. Second, we identify key limitations of current HLS tools, namely inadequate performance feedback, rigid interfaces, and limited debuggability that agents are uniquely positioned to address. Third, we propose a taxonomy for the symbiotic evolution of agentic HLS, clarifying how responsibility shifts from human designers to AI agents as systems advance from copilots to autonomous design partners.

Method: Introduces Wiki Live Challenge (WLC) using newest Wikipedia Good Articles as expert-level references. Curates 100 recent Good Articles and proposes Wiki Eval framework with fine-grained evaluation method (39 criteria for writing quality) and rigorous metrics for factual verifiability.

Result: Experiments on various DRA systems show significant gap between current DRAs and human expert-level Wikipedia articles, validating WLC’s effectiveness in advancing agent research.

Conclusion: WLC provides a reliable benchmark with expert-verified references for evaluating Deep Research Agents, addressing limitations of current LLM-based evaluation approaches and enabling more objective assessment of agent capabilities.

Abstract: Deep Research Agents (DRAs) have demonstrated remarkable capabilities in autonomous information retrieval and report generation, showing great potential to assist humans in complex research tasks. Current evaluation frameworks primarily rely on LLM-generated references or LLM-derived evaluation dimensions. While these approaches offer scalability, they often lack the reliability of expert-verified content and struggle to provide objective, fine-grained assessments of critical dimensions. To bridge this gap, we introduce Wiki Live Challenge (WLC), a live benchmark that leverages the newest Wikipedia Good Articles (GAs) as expert-level references. Wikipedia’s strict standards for neutrality, comprehensiveness, and verifiability serve as a great challenge for DRAs, with GAs representing the pinnacle of which. We curate a dataset of 100 recent Good Articles and propose Wiki Eval, a comprehensive evaluation framework comprising a fine-grained evaluation method with 39 criteria for writing quality and rigorous metrics for factual verifiability. Extensive experiments on various DRA systems demonstrate a significant gap between current DRAs and human expert-level Wikipedia articles, validating the effectiveness of WLC in advancing agent research. We release our benchmark at https://github.com/WangShao2000/Wiki_Live_Challenge

Method: ARTIS decouples exploration from commitment by enabling test-time exploration through simulated interactions before real-world execution. It uses iterative simulation to extend inference-time computation. The paper also introduces a risk-aware tool simulator that emphasizes fidelity on failure-inducing actions through targeted data generation and rebalanced training.

Result: Experiments on multi-turn and multi-step agentic benchmarks show that iterative simulation substantially improves agent reliability, and risk-aware simulation is essential for consistently realizing these gains across models and tasks.

Conclusion: ARTIS provides a framework for improving LLM agent reliability through safe test-time exploration in simulations, with risk-aware simulation being crucial for capturing rare but high-impact failure modes in agentic decision making.

Abstract: Current test-time scaling (TTS) techniques enhance large language model (LLM) performance by allocating additional computation at inference time, yet they remain insufficient for agentic settings, where actions directly interact with external environments and their effects can be irreversible and costly. We propose ARTIS, Agentic Risk-Aware Test-Time Scaling via Iterative Simulation, a framework that decouples exploration from commitment by enabling test-time exploration through simulated interactions prior to real-world execution. This design allows extending inference-time computation to improve action-level reliability and robustness without incurring environmental risk. We further show that naive LLM-based simulators struggle to capture rare but high-impact failure modes, substantially limiting their effectiveness for agentic decision making. To address this limitation, we introduce a risk-aware tool simulator that emphasizes fidelity on failure-inducing actions via targeted data generation and rebalanced training. Experiments on multi-turn and multi-step agentic benchmarks demonstrate that iterative simulation substantially improves agent reliability, and that risk-aware simulation is essential for consistently realizing these gains across models and tasks.

Method: Zero2Text uses recursive online alignment that synergizes LLM priors with dynamic ridge regression to iteratively align text generation to target embeddings on-the-fly, without requiring training data or static datasets.

Result: Zero2Text achieves 1.8x higher ROUGE-L and 6.4x higher BLEU-2 scores compared to baselines on MS MARCO against OpenAI models, recovering sentences from unknown domains without any leaked data pairs, and shows that standard defenses like differential privacy fail against this adaptive threat.

Conclusion: Zero2Text demonstrates a significant advancement in embedding inversion attacks by overcoming limitations of existing methods, highlighting serious privacy vulnerabilities in vector databases that current defenses cannot adequately address.

Abstract: The proliferation of retrieval-augmented generation (RAG) has established vector databases as critical infrastructure, yet they introduce severe privacy risks via embedding inversion attacks. Existing paradigms face a fundamental trade-off: optimization-based methods require computationally prohibitive queries, while alignment-based approaches hinge on the unrealistic assumption of accessible in-domain training data. These constraints render them ineffective in strict black-box and cross-domain settings. To dismantle these barriers, we introduce Zero2Text, a novel training-free framework based on recursive online alignment. Unlike methods relying on static datasets, Zero2Text synergizes LLM priors with a dynamic ridge regression mechanism to iteratively align generation to the target embedding on-the-fly. We further demonstrate that standard defenses, such as differential privacy, fail to effectively mitigate this adaptive threat. Extensive experiments across diverse benchmarks validate Zero2Text; notably, on MS MARCO against the OpenAI victim model, it achieves 1.8x higher ROUGE-L and 6.4x higher BLEU-2 scores compared to baselines, recovering sentences from unknown domains without a single leaked data pair.

Method: Proposes continuous sentence representation called sentence curve (spline curve whose control points affect multiple words), extends diffusion language models to predict sentence curves instead of static word embeddings.

Result: Achieves SOTA performance among diffusion language models on IWSLT14 and WMT14, shows stable training without burdensome knowledge distillation, demonstrates promising potential on LM1B compared to discrete diffusion language models.

Conclusion: Sentence curve representation effectively addresses limitations of static word embeddings by promoting global structure modeling in language models, offering improved performance and training stability.

Abstract: Language models (LMs) are a central component of modern AI systems, and diffusion-based language models (DLMs) have recently emerged as a competitive alternative. Both paradigms rely on word embeddings not only to represent the input sentence, but also to represent the target sentence that backbone models are trained to predict. We argue that such static embedding of the target word is insensitive to neighboring words, encouraging locally accurate word prediction while neglecting global structure across the target sentence. To address this limitation, we propose a continuous sentence representation, termed sentence curve, defined as a spline curve whose control points affect multiple words in the sentence. Based on this representation, we introduce sentence curve language model (SCLM), which extends DLMs to predict sentence curves instead of the static word embeddings. We theoretically show that sentence curve prediction induces a regularization effect that promotes global structure modeling, and characterize how different sentence curve types affect this behavior. Empirically, SCLM achieves SOTA performance among DLMs on IWSLT14 and WMT14, shows stable training without burdensome knowledge distillation, and demonstrates promising potential compared to discrete DLMs on LM1B.

Method: Leverage Wikipedia’s structure where cohesive narratives are grounded in long external reference documents. Sample articles across 12 topics, use external references as retrieval corpus and citation-linked statements as ground truth. Create 1,100 questions across three complexity levels: single-fact QA, multi-fact QA, and section-level summarization.

Result: Experiments show current GraphRAG pipelines help with multi-fact aggregation when evidence comes from moderate number of sources, but the aggregation paradigm may overemphasize high-level statements at the expense of fine-grained details, leading to weaker performance on summarization tasks.

Conclusion: WildGraphBench provides a more realistic evaluation framework for GraphRAG systems, revealing limitations in current approaches for handling fine-grained details in summarization tasks.

Abstract: Graph-based Retrieval-Augmented Generation (GraphRAG) organizes external knowledge as a hierarchical graph, enabling efficient retrieval and aggregation of scattered evidence across multiple documents. However, many existing benchmarks for GraphRAG rely on short, curated passages as external knowledge, failing to adequately evaluate systems in realistic settings involving long contexts and large-scale heterogeneous documents. To bridge this gap, we introduce WildGraphBench, a benchmark designed to assess GraphRAG performance in the wild. We leverage Wikipedia’s unique structure, where cohesive narratives are grounded in long and heterogeneous external reference documents, to construct a benchmark reflecting real-word scenarios. Specifically, we sample articles across 12 top-level topics, using their external references as the retrieval corpus and citation-linked statements as ground truth, resulting in 1,100 questions spanning three levels of complexity: single-fact QA, multi-fact QA, and section-level summarization. Experiments across multiple baselines reveal that current GraphRAG pipelines help on multi-fact aggregation when evidence comes from a moderate number of sources, but this aggregation paradigm may overemphasize high-level statements at the expense of fine-grained details, leading to weaker performance on summarization tasks. Project page:https://github.com/BstWPY/WildGraphBench.

Method: Proposes RPG-Encoder that: (1) Encodes raw code into Repository Planning Graphs combining semantic features with dependencies, (2) Evolves topology incrementally to reduce maintenance costs by 95.7%, and (3) Operates as unified interface for structure-aware navigation.

Result: Achieves state-of-the-art localization performance: 93.7% Acc@5 on SWE-bench Verified, exceeds best baseline by over 10% on SWE-bench Live Lite, and achieves 98.5% reconstruction coverage on RepoCraft.

Conclusion: RPG-Encoder successfully closes the reasoning loop between intent and implementation by creating high-fidelity repository representations that unify comprehension and generation processes.

Abstract: Current repository agents encounter a reasoning disconnect due to fragmented representations, as existing methods rely on isolated API documentation or dependency graphs that lack semantic depth. We consider repository comprehension and generation to be inverse processes within a unified cycle: generation expands intent into implementation, while comprehension compresses implementation back into intent. To address this, we propose RPG-Encoder, a framework that generalizes the Repository Planning Graph (RPG) from a static generative blueprint into a unified, high-fidelity representation. RPG-Encoder closes the reasoning loop through three mechanisms: (1) Encoding raw code into the RPG that combines lifted semantic features with code dependencies; (2) Evolving the topology incrementally to decouple maintenance costs from repository scale, reducing overhead by 95.7%; and (3) Operating as a unified interface for structure-aware navigation. In evaluations, RPG-Encoder establishes state-of-the-art localization performance on SWE-bench Verified with 93.7% Acc@5 and exceeds the best baseline by over 10% in localization accuracy on SWE-bench Live Lite. These results highlight our superior fine-grained precision in complex codebases. Furthermore, it achieves 98.5% reconstruction coverage on RepoCraft, confirming RPG’s high-fidelity capacity to mirror the original codebase and closing the loop between intent and implementation.

Method: Proposes AR-MAP framework that uses preference-aligned autoregressive LLMs as implicit teachers for DLLM alignment through simple weight scaling, exploiting shared architectural structure.

Result: Achieves 69.08% average score across diverse preference alignment tasks, competitive or superior to existing DLLM-specific alignment methods.

Conclusion: AR-MAP enables effective preference alignment for diffusion LLMs by transferring knowledge from autoregressive models, circumventing computational challenges of direct alignment.

Abstract: Diffusion Large Language Models (DLLMs) have emerged as a powerful alternative to autoregressive models, enabling parallel token generation across multiple positions. However, preference alignment of DLLMs remains challenging due to high variance introduced by Evidence Lower Bound (ELBO)-based likelihood estimation. In this work, we propose AR-MAP, a novel transfer learning framework that leverages preference-aligned autoregressive LLMs (AR-LLMs) as implicit teachers for DLLM alignment. We reveal that DLLMs can effectively absorb alignment knowledge from AR-LLMs through simple weight scaling, exploiting the shared architectural structure between these divergent generation paradigms. Crucially, our approach circumvents the high variance and computational overhead of direct DLLM alignment and comprehensive experiments across diverse preference alignment tasks demonstrate that AR-MAP achieves competitive or superior performance compared to existing DLLM-specific alignment methods, achieving 69.08% average score across all tasks and models. Our Code is available at https://github.com/AMAP-ML/AR-MAP.

Method: WorldVQA decouples knowledge retrieval from reasoning by designing questions that test the atomic capability of grounding and naming visual entities. It uses a stratified taxonomy spanning from common head-class objects to long-tail rarities to comprehensively assess visual knowledge breadth.

Result: The benchmark provides a rigorous test for visual factuality in MLLMs, establishing a standard for assessing encyclopedic breadth and hallucination rates across different model generations.

Conclusion: WorldVQA serves as an important evaluation tool for measuring what MLLMs actually memorize about the visual world, helping to establish standards for assessing visual knowledge and reducing hallucinations in multimodal AI systems.

Abstract: We introduce WorldVQA, a benchmark designed to evaluate the atomic visual world knowledge of Multimodal Large Language Models (MLLMs). Unlike current evaluations, which often conflate visual knowledge retrieval with reasoning, WorldVQA decouples these capabilities to strictly measure “what the model memorizes.” The benchmark assesses the atomic capability of grounding and naming visual entities across a stratified taxonomy, spanning from common head-class objects to long-tail rarities. We expect WorldVQA to serve as a rigorous test for visual factuality, thereby establishing a standard for assessing the encyclopedic breadth and hallucination rates of current and next-generation frontier models.

Method: 1) Product-aware adaptive grouping to dynamically organize users based on attributes and product characteristics; 2) Preference-conditioned image generation using a Group-aware Multimodal Large Language Model (G-MLLM) pre-trained to comprehend group features and generate images; 3) Fine-tuning G-MLLM with Group-DPO for group-wise preference alignment.

Result: The framework achieves state-of-the-art performance in both offline and online settings, and introduces GAIP - the first large-scale public dataset of group-wise image preferences with around 600K groups from 40M users.

Conclusion: OSMF successfully bridges the gap in targeted advertising by aligning diverse group-wise click preferences through adaptive grouping and multimodal LLM-based image generation, significantly improving CTR for specific user groups.

Abstract: Advertising image generation has increasingly focused on online metrics like Click-Through Rate (CTR), yet existing approaches adopt a ``one-size-fits-all" strategy that optimizes for overall CTR while neglecting preference diversity among user groups. This leads to suboptimal performance for specific groups, limiting targeted marketing effectiveness. To bridge this gap, we present \textit{One Size, Many Fits} (OSMF), a unified framework that aligns diverse group-wise click preferences in large-scale advertising image generation. OSMF begins with product-aware adaptive grouping, which dynamically organizes users based on their attributes and product characteristics, representing each group with rich collective preference features. Building on these groups, preference-conditioned image generation employs a Group-aware Multimodal Large Language Model (G-MLLM) to generate tailored images for each group. The G-MLLM is pre-trained to simultaneously comprehend group features and generate advertising images. Subsequently, we fine-tune the G-MLLM using our proposed Group-DPO for group-wise preference alignment, which effectively enhances each group’s CTR on the generated images. To further advance this field, we introduce the Grouped Advertising Image Preference Dataset (GAIP), the first large-scale public dataset of group-wise image preferences, including around 600K groups built from 40M users. Extensive experiments demonstrate that our framework achieves the state-of-the-art performance in both offline and online settings. Our code and datasets will be released at https://github.com/JD-GenX/OSMF.

Method: Proposes AdaptMMBench benchmark across five domains (real-world, OCR, GUI, knowledge, math) with Matthews Correlation Coefficient metric to evaluate selection rationality of reasoning modes, dynamically identifying task difficulties based on models’ capability boundaries, and facilitating multi-dimensional process evaluation across key step coverage, tool effectiveness, and computational efficiency.

Result: Evaluation reveals that adaptive mode selection scales with model capacity but decouples from final accuracy, while key step coverage aligns with performance, and tool effectiveness remains highly inconsistent across model architectures.

Conclusion: AdaptMMBench provides a comprehensive framework for evaluating adaptive multimodal reasoning that isolates meta-cognition abilities and enables fine-grained analysis of reasoning processes beyond simple accuracy metrics.

Abstract: Adaptive multimodal reasoning has emerged as a promising frontier in Vision-Language Models (VLMs), aiming to dynamically modulate between tool-augmented visual reasoning and text reasoning to enhance both effectiveness and efficiency. However, existing evaluations rely on static difficulty labels and simplistic metrics, which fail to capture the dynamic nature of difficulty relative to varying model capacities. Consequently, they obscure the distinction between adaptive mode selection and general performance while neglecting fine-grained process analyses. In this paper, we propose AdaptMMBench, a comprehensive benchmark for adaptive multimodal reasoning across five domains: real-world, OCR, GUI, knowledge, and math, encompassing both direct perception and complex reasoning tasks. AdaptMMBench utilizes a Matthews Correlation Coefficient (MCC) metric to evaluate the selection rationality of different reasoning modes, isolating this meta-cognition ability by dynamically identifying task difficulties based on models’ capability boundaries. Moreover, AdaptMMBench facilitates multi-dimensional process evaluation across key step coverage, tool effectiveness, and computational efficiency. Our evaluation reveals that while adaptive mode selection scales with model capacity, it notably decouples from final accuracy. Conversely, key step coverage aligns with performance, though tool effectiveness remains highly inconsistent across model architectures.

Method: Regularized end-to-end deep learning framework trained with Monte Carlo-simulated ground truth, incorporating a physics-based OCT forward model that generates predicted signals from estimated parameters for physics-consistent supervision.

Result: Experiments on synthetic corneal OCT dataset demonstrate robust optical map recovery under noise, improved resolution, and enhanced structural fidelity compared to conventional methods.

Conclusion: The approach enables quantitative multi-parameter tissue characterization and shows the benefit of combining physics-informed modeling with deep learning for computational OCT applications.

Abstract: Inverse scattering in optical coherence tomography (OCT) seeks to recover both structural images and intrinsic tissue optical properties, including refractive index, scattering coefficient, and anisotropy. This inverse problem is challenging due to attenuation, speckle noise, and strong coupling among parameters. We propose a regularized end-to-end deep learning framework that jointly reconstructs optical parameter maps and speckle-reduced OCT structural intensity for layer visualization. Trained with Monte Carlo-simulated ground truth, our network incorporates a physics-based OCT forward model that generates predicted signals from the estimated parameters, providing physics-consistent supervision for parameter recovery and artifact suppression. Experiments on the synthetic corneal OCT dataset demonstrate robust optical map recovery under noise, improved resolution, and enhanced structural fidelity. This approach enables quantitative multi-parameter tissue characterization and highlights the benefit of combining physics-informed modeling with deep learning for computational OCT.

Method: Proposes SVD-ViT with three components: SPC module, SSVA, and ID-RSVD, which use singular value decomposition to extract and aggregate singular vectors capturing object foreground information while suppressing background noise.

Result: Experimental results show improved classification accuracy and effective learning of informative foreground representations while reducing background noise impact.

Conclusion: SVD-ViT successfully addresses ViT’s limitation in foreground-background distinction through SVD-based feature prioritization, enhancing classification performance.

Abstract: Vision Transformers (ViT) have been established as large-scale foundation models. However, because self-attention operates globally, they lack an explicit mechanism to distinguish foreground from background. As a result, ViT may learn unnecessary background features and artifacts, leading to degraded classification performance. To address this issue, we propose SVD-ViT, which leverages singular value decomposition (SVD) to prioritize the learning of foreground features. SVD-ViT consists of three components-\textbf{SPC module}, \textbf{SSVA}, and \textbf{ID-RSVD}-and suppresses task-irrelevant factors such as background noise and artifacts by extracting and aggregating singular vectors that capture object foreground information. Experimental results demonstrate that our method improves classification accuracy and effectively learns informative foreground representations while reducing the impact of background noise.

Method: Proposes Landmark Point Transformer (LmPT) that represents anatomical surfaces as point clouds and uses a transformer architecture with a conditioning mechanism to adapt to different input types, enabling cross-species learning between human and dog femurs.

Result: Demonstrates effective generalization across species (human and dog femurs) with publicly available code and dog femur dataset.

Conclusion: LmPT provides an effective automated solution for anatomical landmark detection that can leverage cross-species data for translational research.

Abstract: Accurate identification of anatomical landmarks is crucial for various medical applications. Traditional manual landmarking is time-consuming and prone to inter-observer variability, while rule-based methods are often tailored to specific geometries or limited sets of landmarks. In recent years, anatomical surfaces have been effectively represented as point clouds, which are lightweight structures composed of spatial coordinates. Following this strategy and to overcome the limitations of existing landmarking techniques, we propose Landmark Point Transformer (LmPT), a method for automatic anatomical landmark detection on point clouds that can leverage homologous bones from different species for translational research. The LmPT model incorporates a conditioning mechanism that enables adaptability to different input types to conduct cross-species learning. We focus the evaluation of our approach on femoral landmarking using both human and newly annotated dog femurs, demonstrating its generalization and effectiveness across species. The code and dog femur dataset will be publicly available at: https://github.com/Pierreoo/LandmarkPointTransformer.

Method: ELIQ automatically constructs positive and aspect-specific negative pairs covering both conventional distortions and AIGC-specific distortion modes. It then adapts a pre-trained multimodal model via instruction tuning and uses lightweight gated fusion with a Quality Query Transformer to predict two-dimensional quality (visual quality and prompt-image alignment).

Result: ELIQ consistently outperforms existing label-free methods across multiple benchmarks, generalizes from AI-generated content to user-generated content scenarios without modification, and demonstrates transferable supervision without human annotations.

Conclusion: ELIQ provides a scalable, label-free framework for quality assessment that can adapt to continuously evolving generative models, paving the way for more efficient evaluation of AI-generated content quality.

Abstract: Generative text-to-image models are advancing at an unprecedented pace, continuously shifting the perceptual quality ceiling and rendering previously collected labels unreliable for newer generations. To address this, we present ELIQ, a Label-free Framework for Quality Assessment of Evolving AI-generated Images. Specifically, ELIQ focuses on visual quality and prompt-image alignment, automatically constructs positive and aspect-specific negative pairs to cover both conventional distortions and AIGC-specific distortion modes, enabling transferable supervision without human annotations. Building on these pairs, ELIQ adapts a pre-trained multimodal model into a quality-aware critic via instruction tuning and predicts two-dimensional quality using lightweight gated fusion and a Quality Query Transformer. Experiments across multiple benchmarks demonstrate that ELIQ consistently outperforms existing label-free methods, generalizes from AI-generated content (AIGC) to user-generated content (UGC) scenarios without modification, and paves the way for scalable and label-free quality assessment under continuously evolving generative models. The code will be released upon publication.

Method: Uses off-the-shelf whole-body person detector with low threshold, retrieves false negatives via temporal tracking and self-supervised multi-view association, uses recovered detections as pseudo labels to iteratively fine-tune detector, then applies whole-body pose estimation and fine-tunes pose model with its own high-score predictions.

Result: Achieves over 97% recall on 4D-OR dataset and real surgery dataset, trains real-time whole-body detector using pseudo labels with comparable performance.

Conclusion: Proposed framework effectively anonymizes OR videos without manual annotation or camera calibration, enabling scalable privacy preservation for medical video research.

Abstract: Privacy preservation is a prerequisite for using video data in Operating Room (OR) research. Effective anonymization relies on the exhaustive localization of every individual; even a single missed detection necessitates extensive manual correction. However, existing approaches face two critical scalability bottlenecks: (1) they usually require manual annotations of each new clinical site for high accuracy; (2) while multi-camera setups have been widely adopted to address single-view ambiguity, camera calibration is typically required whenever cameras are repositioned. To address these problems, we propose a novel self-supervised multi-view video anonymization framework consisting of whole-body person detection and whole-body pose estimation, without annotation or camera calibration. Our core strategy is to enhance the single-view detector by “retrieving” false negatives using temporal and multi-view context, and conducting self-supervised domain adaptation. We first run an off-the-shelf whole-body person detector in each view with a low-score threshold to gather candidate detections. Then, we retrieve the low-score false negatives that exhibit consistency with the high-score detections via tracking and self-supervised uncalibrated multi-view association. These recovered detections serve as pseudo labels to iteratively fine-tune the whole-body detector. Finally, we apply whole-body pose estimation on each detected person, and fine-tune the pose model using its own high-score predictions. Experiments on the 4D-OR dataset of simulated surgeries and our dataset of real surgeries show the effectiveness of our approach achieving over 97% recall. Moreover, we train a real-time whole-body detector using our pseudo labels, achieving comparable performance and highlighting our method’s practical applicability. Code is available at https://github.com/CAMMA-public/OR_anonymization.

Method: 1) Latent-guided token compression module reduces redundancy in visual sequences while preserving semantics. 2) Semi-autoregressive decoding scheme leverages co-occurrence and local correlations of visual entities to predict multiple tokens in one forward pass.

Result: Achieves up to 2.68× speed-up on video captioning and 2.55× on visual instruction tuning compared to original LMMs, consistently outperforming prior speculative decoding baselines.

Conclusion: SpecFLASH effectively accelerates multimodal model inference by exploiting visual structure characteristics, making it a practical solution for efficient multimodal generation tasks.

Abstract: Large language models and large multimodal models (LLMs and LMMs) deliver strong generative performance but suffer from slow decoding, a problem that becomes more severe when handling visual inputs, whose sequences typically contain many more tokens with lower information density than text. Speculative decoding accelerates LLM inference by letting a compact draft model propose candidate tokens that are selectively accepted by a larger target model, achieving speed-up without degrading quality. However, existing multimodal speculative decoding approaches largely ignore the structural characteristics of visual representations and usually rely on text-only draft models. In this paper, we introduce SpecFLASH, a speculative decoding framework tailored to LMMs that explicitly exploits multimodal structure when designing the draft model. We first mitigate redundancy in visual token sequences with a lightweight, latent-guided token compression module that compacts visual features while preserving semantics, and then leverage the co-occurrence and local correlations of visual entities via a semi-autoregressive decoding scheme that predicts multiple tokens in a single forward pass. Extensive experiments demonstrate that SpecFLASH consistently surpasses prior speculative decoding baselines, achieving up to $2.68\times$ speed-up on video captioning and $2.55\times$ on visual instruction tuning, relative to the original LMM. Our code is available here: https://github.com/ZihuaEvan/FlashSD/.

Method: ViThinker enables models to autonomously generate decision (query) tokens that trigger synthesis of expert-aligned visual features on demand. It uses a two-stage curriculum: first distilling frozen expert capabilities into model parameters, then learning task-driven querying via sparsity penalties to discover minimal sufficient perception for each reasoning step.

Result: Evaluations across vision-centric benchmarks demonstrate consistent improvements, showing that active query generation outperforms passive approaches in both perceptual grounding and reasoning accuracy.

Conclusion: ViThinker successfully enables active perception in vision-language models, allowing them to generate queries for on-demand visual feature synthesis, leading to better reasoning performance compared to passive approaches.

Abstract: Chain-of-Thought (CoT) reasoning excels in language models but struggles in vision-language models due to premature visual-to-text conversion that discards continuous information such as geometry and spatial layout. While recent methods enhance CoT through static enumeration or attention-based selection, they remain passive, i.e., processing pre-computed inputs rather than actively seeking task-relevant details. Inspired by human active perception, we introduce ViThinker, a framework that enables vision-language models to autonomously generate decision (query) tokens triggering the synthesis of expert-aligned visual features on demand. ViThinker internalizes vision-expert capabilities during training, performing generative mental simulation during inference without external tool calls. Through a two-stage curriculum: first distilling frozen experts into model parameters, then learning task-driven querying via sparsity penalties, i.e., ViThinker discovers minimal sufficient perception for each reasoning step. Evaluations across vision-centric benchmarks demonstrate consistent improvements, validating that active query generation outperforms passive approaches in both perceptual grounding and reasoning accuracy.

Method: TP-Blend uses two attention processors: Cross-Attention Object Fusion (CAOF) for object blending via optimal transport on attention maps, and Self-Attention Style Fusion (SASF) for style injection using Detail-Sensitive Instance Normalization and Key/Value matrix swapping.

Result: The method produces high-resolution, photo-realistic edits with precise control over both content and appearance, surpassing recent baselines in quantitative fidelity, perceptual quality, and inference speed.

Conclusion: TP-Blend enables simultaneous object replacement and style transfer in diffusion models without additional training, providing a lightweight framework for complex multimodal editing tasks.

Abstract: Current text-conditioned diffusion editors handle single object replacement well but struggle when a new object and a new style must be introduced simultaneously. We present Twin-Prompt Attention Blend (TP-Blend), a lightweight training-free framework that receives two separate textual prompts, one specifying a blend object and the other defining a target style, and injects both into a single denoising trajectory. TP-Blend is driven by two complementary attention processors. Cross-Attention Object Fusion (CAOF) first averages head-wise attention to locate spatial tokens that respond strongly to either prompt, then solves an entropy-regularised optimal transport problem that reassigns complete multi-head feature vectors to those positions. CAOF updates feature vectors at the full combined dimensionality of all heads (e.g., 640 dimensions in SD-XL), preserving rich cross-head correlations while keeping memory low. Self-Attention Style Fusion (SASF) injects style at every self-attention layer through Detail-Sensitive Instance Normalization. A lightweight one-dimensional Gaussian filter separates low- and high-frequency components; only the high-frequency residual is blended back, imprinting brush-stroke-level texture without disrupting global geometry. SASF further swaps the Key and Value matrices with those derived from the style prompt, enforcing context-aware texture modulation that remains independent of object fusion. Extensive experiments show that TP-Blend produces high-resolution, photo-realistic edits with precise control over both content and appearance, surpassing recent baselines in quantitative fidelity, perceptual quality, and inference speed.

Method: Introduces contrastive document-aware reference selection using ROCO embeddings and metadata to balance visual relevance, embedding diversity, and source-level provenance. Proposes Counterfactual-Contrastive Inference with structured pairwise visual comparisons and margin-based decision rules with faithful abstention.

Result: Achieves state-of-the-art performance on MediConfusion benchmark, improving set-level accuracy by nearly 15% relative to prior methods while reducing confusion and improving individual accuracy.

Conclusion: The contrastive evidence selection framework enables more accurate discrimination between confusable medical conditions by optimizing for discrimination rather than similarity, addressing limitations of traditional retrieval methods.

Abstract: Accurate decision making in medical imaging requires reasoning over subtle visual differences between confusable conditions, yet most existing approaches rely on nearest neighbor retrieval that returns redundant evidence and reinforces a single hypothesis. We introduce a contrastive, document-aware reference selection framework that constructs compact evidence sets optimized for discrimination rather than similarity by explicitly balancing visual relevance, embedding diversity, and source-level provenance using ROCO embeddings and metadata. While ROCO provides large-scale image-caption pairs, it does not specify how references should be selected for contrastive reasoning, and naive retrieval frequently yields near-duplicate figures from the same document. To address this gap, we release a reproducible reference selection protocol and curated reference bank that enable a systematic study of contrastive retrieval in medical image reasoning. Building on these contrastive evidence sets, we propose Counterfactual-Contrastive Inference, a confidence-aware reasoning framework that performs structured pairwise visual comparisons and aggregates evidence using margin-based decision rules with faithful abstention. On the MediConfusion benchmark, our approach achieves state-of-the-art performance, improving set-level accuracy by nearly 15% relative to prior methods while reducing confusion and improving individual accuracy.

Method: Developed FaceLinkGen attack that performs identity linkage/matching and face regeneration directly from protected templates without recovering original pixels. Tested on three recent PPFR systems in both standard and near zero knowledge settings.

Result: FaceLinkGen achieved over 98.5% matching accuracy and above 96% regeneration success on three PPFR systems. Even in near zero knowledge settings, it exceeded 92% matching and 94% regeneration, exposing structural gaps in current privacy evaluation metrics.

Conclusion: Pixel distortion metrics (PSNR/SSIM) widely used in PPFR evaluation fail to measure real privacy. Visual obfuscation leaves identity information broadly exposed to both external intruders and untrusted service providers, requiring new privacy evaluation frameworks.

Abstract: Transformation-based privacy-preserving face recognition (PPFR) aims to verify identities while hiding facial data from attackers and malicious service providers. Existing evaluations mostly treat privacy as resistance to pixel-level reconstruction, measured by PSNR and SSIM. We show that this reconstruction-centric view fails. We present FaceLinkGen, an identity extraction attack that performs linkage/matching and face regeneration directly from protected templates without recovering original pixels. On three recent PPFR systems, FaceLinkGen reaches over 98.5% matching accuracy and above 96% regeneration success, and still exceeds 92% matching and 94% regeneration in a near zero knowledge setting. These results expose a structural gap between pixel distortion metrics, which are widely used in PPFR evaluation, and real privacy. We show that visual obfuscation leaves identity information broadly exposed to both external intruders and untrusted service providers.

Method: Proposes hyperbolic representation learning for medical image analysis, exploiting hyperbolic manifolds to model complex data characteristics. Introduces an unsupervised, domain-invariant hyperbolic cross-branch consistency constraint to promote domain-invariant features.

Result: Demonstrates statistically significant gains across 11 in-distribution datasets and 3 ViT models. Outperforms state-of-the-art Euclidean methods by average +2.1% AUC on three domain generalization benchmarks: Fitzpatrick17k, Camelyon17-WILDS, and cross-dataset retinal imaging setup.

Conclusion: Hyperbolic representation learning effectively captures complex hierarchical structures in medical images, enabling better domain generalization across different imaging modalities, data sizes, and label granularities, confirming robust generalization capabilities across substantially different conditions.

Abstract: Robust generalization beyond training distributions remains a critical challenge for deep neural networks. This is especially pronounced in medical image analysis, where data is often scarce and covariate shifts arise from different hardware devices, imaging protocols, and heterogeneous patient populations. These factors collectively hinder reliable performance and slow down clinical adoption. Despite recent progress, existing learning paradigms primarily rely on the Euclidean manifold, whose flat geometry fails to capture the complex, hierarchical structures present in clinical data. In this work, we exploit the advantages of hyperbolic manifolds to model complex data characteristics. We present the first comprehensive validation of hyperbolic representation learning for medical image analysis and demonstrate statistically significant gains across eleven in-distribution datasets and three ViT models. We further propose an unsupervised, domain-invariant hyperbolic cross-branch consistency constraint. Extensive experiments confirm that our proposed method promotes domain-invariant features and outperforms state-of-the-art Euclidean methods by an average of $+2.1%$ AUC on three domain generalization benchmarks: Fitzpatrick17k, Camelyon17-WILDS, and a cross-dataset setup for retinal imaging. These datasets span different imaging modalities, data sizes, and label granularities, confirming generalization capabilities across substantially different conditions. The code is available at https://github.com/francescodisalvo05/hyperbolic-cross-branch-consistency .

Method: MARBLE uses a purely Mamba-based architecture with parallel multi-scale processing and linear-time state-space models. It integrates coarse-to-fine reasoning by processing multiple magnification levels simultaneously, capturing cross-scale dependencies efficiently with minimal parameter overhead.

Result: Experiments on five public datasets show improvements of up to 6.9% in AUC, 20.3% in accuracy, and 2.3% in C-index compared to existing methods, establishing MARBLE as an efficient and generalizable framework for multi-scale WSI analysis.

Conclusion: MARBLE provides a scalable and modular alternative to attention-based architectures for whole-slide image analysis, demonstrating that Mamba-based state-space models can effectively handle multi-scale visual reasoning with linear-time complexity.

Abstract: We introduce Multi-scale Adaptive Recurrent Biomedical Linear-time Encoder (MARBLE), the first \textit{purely Mamba-based} multi-state multiple instance learning (MIL) framework for whole-slide image (WSI) analysis. MARBLE processes multiple magnification levels in parallel and integrates coarse-to-fine reasoning within a linear-time state-space model, efficiently capturing cross-scale dependencies with minimal parameter overhead. WSI analysis remains challenging due to gigapixel resolutions and hierarchical magnifications, while existing MIL methods typically operate at a single scale and transformer-based approaches suffer from quadratic attention costs. By coupling parallel multi-scale processing with linear-time sequence modeling, MARBLE provides a scalable and modular alternative to attention-based architectures. Experiments on five public datasets show improvements of up to \textbf{6.9%} in AUC, \textbf{20.3%} in accuracy, and \textbf{2.3%} in C-index, establishing MARBLE as an efficient and generalizable framework for multi-scale WSI analysis.

Method: Proposes SRA-Seg with similarity-alignment loss using frozen DINOv2 embeddings to pull synthetic representations toward nearest real counterparts, soft edge blending for smooth anatomical transitions, and pseudo-labeling via EMA teacher model with soft-segmentation losses.

Result: Using only 10% labeled real data and 90% synthetic unlabeled data, achieves 89.34% Dice on ACDC and 84.42% on FIVES, outperforming existing semi-supervised methods and matching performance of methods using real unlabeled data.

Conclusion: SRA-Seg effectively bridges the synthetic-real domain gap for medical image segmentation, enabling effective use of synthetic data with minimal real annotations.

Abstract: Synthetic data, an appealing alternative to extensive expert-annotated data for medical image segmentation, consistently fails to improve segmentation performance despite its visual realism. The reason being that synthetic and real medical images exist in different semantic feature spaces, creating a domain gap that current semi-supervised learning methods cannot bridge. We propose SRA-Seg, a framework explicitly designed to align synthetic and real feature distributions for medical image segmentation. SRA-Seg introduces a similarity-alignment (SA) loss using frozen DINOv2 embeddings to pull synthetic representations toward their nearest real counterparts in semantic space. We employ soft edge blending to create smooth anatomical transitions and continuous labels, eliminating the hard boundaries from traditional copy-paste augmentation. The framework generates pseudo-labels for synthetic images via an EMA teacher model and applies soft-segmentation losses that respect uncertainty in mixed regions. Our experiments demonstrate strong results: using only 10% labeled real data and 90% synthetic unlabeled data, SRA-Seg achieves 89.34% Dice on ACDC and 84.42% on FIVES, significantly outperforming existing semi-supervised methods and matching the performance of methods using real unlabeled data.

Method: Hardware-software co-optimized VIO pipeline using ORB feature tracking and bundle adjustment, with emphasis on computational efficiency through parallelization, low memory usage, and optimization for embedded microcontrollers and low-power SoCs.

Result: Achieves 20 FPS real-time performance while consuming less than 100 mW on a parallel-processing ultra-low-power RISC-V SoC, with competitive accuracy on public VIO datasets.

Conclusion: LEVIO demonstrates that efficient VIO is feasible on ultra-low-power hardware, enabling real-time motion tracking for resource-constrained applications like micro-drones and AR glasses.

Abstract: Accurate, infrastructure-less sensor systems for motion tracking are essential for mobile robotics and augmented reality (AR) applications. The most popular state-of-the-art visual-inertial odometry (VIO) systems, however, are too computationally demanding for resource-constrained hardware, such as micro-drones and smart glasses. This work presents LEVIO, a fully featured VIO pipeline optimized for ultra-low-power compute platforms, allowing six-degrees-of-freedom (DoF) real-time sensing. LEVIO incorporates established VIO components such as Oriented FAST and Rotated BRIEF (ORB) feature tracking and bundle adjustment, while emphasizing a computationally efficient architecture with parallelization and low memory usage to suit embedded microcontrollers and low-power systems-on-chip (SoCs). The paper proposes and details the algorithmic design choices and the hardware-software co-optimization approach, and presents real-time performance on resource-constrained hardware. LEVIO is validated on a parallel-processing ultra-low-power RISC-V SoC, achieving 20 FPS while consuming less than 100 mW, and benchmarked against public VIO datasets, offering a compelling balance between efficiency and accuracy. To facilitate reproducibility and adoption, the complete implementation is released as open-source.

Method: Two-stage framework: 1) After vision encoder, apply separation, alignment, and aggregation operations inspired by swarm intelligence to retain information-rich global spatial anchors; 2) Within LLM, perform text-guided pruning to retain task-relevant visual tokens.

Result: Achieves SOTA performance on multiple VQA benchmarks (94% to 95%) and substantial improvements on visual grounding tasks (7% to 47%).

Conclusion: Nüwa enables efficient feature aggregation while maintaining spatial integrity, addressing limitations of existing pruning methods for multimodal tasks requiring spatial understanding.

Abstract: Vision token pruning has proven to be an effective acceleration technique for the efficient Vision Language Model (VLM). However, existing pruning methods demonstrate excellent performance preservation in visual question answering (VQA) and suffer substantial degradation on visual grounding (VG) tasks. Our analysis of the VLM’s processing pipeline reveals that strategies utilizing global semantic similarity and attention scores lose the global spatial reference frame, which is derived from the interactions of tokens’ positional information. Motivated by these findings, we propose $\text{Nüwa}$, a two-stage token pruning framework that enables efficient feature aggregation while maintaining spatial integrity. In the first stage, after the vision encoder, we apply three operations, namely separation, alignment, and aggregation, which are inspired by swarm intelligence algorithms to retain information-rich global spatial anchors. In the second stage, within the LLM, we perform text-guided pruning to retain task-relevant visual tokens. Extensive experiments demonstrate that $\text{Nüwa}$ achieves SOTA performance on multiple VQA benchmarks (from 94% to 95%) and yields substantial improvements on visual grounding tasks (from 7% to 47%).

Method: Proposes a sequential neural network model with spatial-temporal attention mechanism built on encoder-decoder structure with common neural network backbones. The attention mechanism focuses on key lane features and exploits salient spatial-temporal correlations among continuous image frames.

Result: Outperforms state-of-the-art methods on three large-scale datasets, demonstrates robustness in various testing scenarios, and achieves computational efficiency with fewer parameters and reduced MACs compared to baseline sequential models.

Conclusion: The proposed spatial-temporal attention mechanism effectively enhances lane detection performance by focusing on critical features and exploiting temporal correlations, offering both accuracy and computational efficiency for real-time applications.

Abstract: Lane detection is a crucial perception task for all levels of automated vehicles (AVs) and Advanced Driver Assistance Systems, particularly in mixed-traffic environments where AVs must interact with human-driven vehicles (HDVs) and challenging traffic scenarios. Current methods lack versatility in delivering accurate, robust, and real-time compatible lane detection, especially vision-based methods often neglect critical regions of the image and their spatial-temporal (ST) salience, leading to poor performance in difficult circumstances such as serious occlusion and dazzle lighting. This study introduces a novel sequential neural network model with a spatial-temporal attention mechanism to focus on key features of lane lines and exploit salient ST correlations among continuous image frames. The proposed model, built on a standard encoder-decoder structure and common neural network backbones, is trained and evaluated on three large-scale open-source datasets. Extensive experiments demonstrate the strength and robustness of the proposed model, outperforming state-of-the-art methods in various testing scenarios. Furthermore, with the ST attention mechanism, the developed sequential neural network models exhibit fewer parameters and reduced Multiply-Accumulate Operations (MACs) compared to baseline sequential models, highlighting their computational efficiency. Relevant data, code, and models are released at https://doi.org/10.4121/4619cab6-ae4a-40d5-af77-582a77f3d821.

Method: Introduces TRACE, a model that takes prior and current chest X-rays as input and jointly learns temporal comparison, change classification (worsened/improved/stable), and spatial localization via bounding box coordinates for each finding.

Result: TRACE achieves over 90% grounding accuracy and demonstrates that change detection emerges only when temporal comparison and spatial grounding are jointly learned, suggesting grounding provides essential spatial attention for temporal reasoning.

Conclusion: TRACE establishes a foundation for temporal radiology analysis by combining change detection with spatial grounding, revealing that joint learning of these capabilities is essential for meaningful temporal reasoning in medical imaging.

Abstract: Temporal comparison of chest X-rays is fundamental to clinical radiology, enabling detection of disease progression, treatment response, and new findings. While vision-language models have advanced single-image report generation and visual grounding, no existing method combines these capabilities for temporal change detection. We introduce Temporal Radiology with Anatomical Change Explanation (TRACE), the first model that jointly performs temporal comparison, change classification, and spatial localization. Given a prior and current chest X-ray, TRACE generates natural language descriptions of interval changes (worsened, improved, stable) while grounding each finding with bounding box coordinates. TRACE demonstrates effective spatial localization with over 90% grounding accuracy, establishing a foundation for this challenging new task. Our ablation study uncovers an emergent capability: change detection arises only when temporal comparison and spatial grounding are jointly learned, as neither alone enables meaningful change detection. This finding suggests that grounding provides a spatial attention mechanism essential for temporal reasoning.

Method: Proposes DHiF that translates discriminative modeling into generation of dynamic local filter banks. DHiF is sensitive to HFCs due to dynamic parameters adjusted within zero-centered range according to Fourier transformation properties. Works as drop-in replacement for standard convolution in any SIRST detection network.

Result: Extensive experiments on real-scene datasets show DHiF exhibits superior detection performance compared to other state-of-the-art convolution operations with promising improvement.

Conclusion: DHiF effectively addresses the challenge of distinguishing infrared small targets from other high-frequency components through dynamic filter generation and discriminative representation learning.

Abstract: Infrared small targets are typically tiny and locally salient, which belong to high-frequency components (HFCs) in images. Single-frame infrared small target (SIRST) detection is challenging, since there are many HFCs along with targets, such as bright corners, broken clouds, and other clutters. Current learning-based methods rely on the powerful capabilities of deep networks, but neglect explicit modeling and discriminative representation learning of various HFCs, which is important to distinguish targets from other HFCs. To address the aforementioned issues, we propose a dynamic high-frequency convolution (DHiF) to translate the discriminative modeling process into the generation of a dynamic local filter bank. Especially, DHiF is sensitive to HFCs, owing to the dynamic parameters of its generated filters being symmetrically adjusted within a zero-centered range according to Fourier transformation properties. Combining with standard convolution operations, DHiF can adaptively and dynamically process different HFC regions and capture their distinctive grayscale variation characteristics for discriminative representation learning. DHiF functions as a drop-in replacement for standard convolution and can be used in arbitrary SIRST detection networks without significant decrease in computational efficiency. To validate the effectiveness of our DHiF, we conducted extensive experiments across different SIRST detection networks on real-scene datasets. Compared to other state-of-the-art convolution operations, DHiF exhibits superior detection performance with promising improvement. Codes are available at https://github.com/TinaLRJ/DHiF.

Method: The study derives analytical mathematical expressions for depth and range error in fisheye stereo vision systems as a function of object distance, specifically modeling the error behavior at large incidence angles where traditional pinhole camera models break down.

Result: The derived expressions provide a theoretical framework for understanding how depth estimation accuracy degrades with increasing object distance and incidence angle in fisheye stereo systems, enabling better error prediction and system design.

Conclusion: The analytical error expressions for fisheye stereo vision provide valuable tools for system design and performance evaluation, particularly for applications requiring accurate depth estimation across wide fields of view.

Abstract: This study derives analytical expressions for the depth and range error of fisheye stereo vision systems as a function of object distance, specifically accounting for accuracy at large angles.

Method: Proposes a graph network with cross-check feature attention for scene graph prediction, and constructs a graph-variational autoencoder (graph-VAE) with a joint shape and layout block for 3D scene generation from semantic scene graphs.

Result: Outperforms state-of-the-art methods on 3RScan/3DSSG and SG-FRONT datasets in both quantitative and qualitative evaluations, even in complex indoor environments and under challenging scene graph constraints.

Conclusion: SceneLinker enables users to generate consistent 3D spaces from physical environments via scene graphs, facilitating the creation of spatial Mixed Reality content that adapts to individual user spaces.

Abstract: We introduce SceneLinker, a novel framework that generates compositional 3D scenes via semantic scene graph from RGB sequences. To adaptively experience Mixed Reality (MR) content based on each user’s space, it is essential to generate a 3D scene that reflects the real-world layout by compactly capturing the semantic cues of the surroundings. Prior works struggled to fully capture the contextual relationship between objects or mainly focused on synthesizing diverse shapes, making it challenging to generate 3D scenes aligned with object arrangements. We address these challenges by designing a graph network with cross-check feature attention for scene graph prediction and constructing a graph-variational autoencoder (graph-VAE), which consists of a joint shape and layout block for 3D scene generation. Experiments on the 3RScan/3DSSG and SG-FRONT datasets demonstrate that our approach outperforms state-of-the-art methods in both quantitative and qualitative evaluations, even in complex indoor environments and under challenging scene graph constraints. Our work enables users to generate consistent 3D spaces from their physical environments via scene graphs, allowing them to create spatial MR content. Project page is https://scenelinker2026.github.io.

Method: Proposes CAFT with a fine-to-coarse visual encoder and hierarchical text transformer, using hierarchical alignment loss that matches whole images with whole captions while biasing region-sentence correspondences.

Result: Achieves state-of-the-art performance on six long-text retrieval benchmarks, exhibits strong scaling behavior, and enables fine-grained visually grounded representations to emerge without explicit region-level supervision.

Conclusion: Hierarchical cross-domain alignment enables effective fine-grained vision-language understanding for long captions without requiring pixel-level supervision.

Abstract: Large vision-language models such as CLIP struggle with long captions because they align images and texts as undifferentiated wholes. Fine-grained vision-language understanding requires hierarchical semantics capturing both global context and localized details across visual and textual domains. Yet linguistic hierarchies from syntax or semantics rarely match visual organization, and purely visual hierarchies tend to fragment scenes into appearance-driven parts without semantic focus. We propose CAFT (Cross-domain Alignment of Forests and Trees), a hierarchical image-text representation learning framework that aligns global and local semantics across images and long captions without pixel-level supervision. Coupling a fine-to-coarse visual encoder with a hierarchical text transformer, it uses a hierarchical alignment loss that matches whole images with whole captions while biasing region-sentence correspondences, so that coarse semantics are built from fine-grained evidence rather than from aggregation untethered to part-level grounding. Trained on 30M image-text pairs, CAFT achieves state-of-the-art performance on six long-text retrieval benchmarks and exhibits strong scaling behavior. Experiments show that hierarchical cross-domain alignment enables fine-grained, visually grounded image-text representations to emerge without explicit region-level supervision.

Method: Proposes a lifespan-aware 4D Gaussian framework with learnable lifespan parameters that reformulate temporal visibility from Gaussian decay to flat-top profiles, modulate motion based on lifespan, and use lifespan-velocity-aware densification for balanced optimization.

Result: Achieves state-of-the-art performance on multiple benchmarks while supporting real-time rendering up to 4K resolution at 100 FPS on one RTX 4090.

Conclusion: SharpTimeGS effectively balances static and dynamic region modeling through lifespan-aware optimization, enabling high-quality real-time 4D reconstruction of dynamic scenes.

Abstract: Novel view synthesis of dynamic scenes is fundamental to achieving photorealistic 4D reconstruction and immersive visual experiences. Recent progress in Gaussian-based representations has significantly improved real-time rendering quality, yet existing methods still struggle to maintain a balance between long-term static and short-term dynamic regions in both representation and optimization. To address this, we present SharpTimeGS, a lifespan-aware 4D Gaussian framework that achieves temporally adaptive modeling of both static and dynamic regions under a unified representation. Specifically, we introduce a learnable lifespan parameter that reformulates temporal visibility from a Gaussian-shaped decay into a flat-top profile, allowing primitives to remain consistently active over their intended duration and avoiding redundant densification. In addition, the learned lifespan modulates each primitives’ motion, reducing drift in long-lived static points while retaining unrestricted motion for short-lived dynamic ones. This effectively decouples motion magnitude from temporal duration, improving long-term stability without compromising dynamic fidelity. Moreover, we design a lifespan-velocity-aware densification strategy that mitigates optimization imbalance between static and dynamic regions by allocating more capacity to regions with pronounced motion while keeping static areas compact and stable. Extensive experiments on multiple benchmarks demonstrate that our method achieves state-of-the-art performance while supporting real-time rendering up to 4K resolution at 100 FPS on one RTX 4090.

Method: Proposes Video-OPD framework that optimizes trajectories sampled from current policy using reverse KL divergence with dense token-level supervision from a frontier teacher. Also introduces Teacher-Validated Disagreement Focusing (TVDF) curriculum that prioritizes teacher-reliable and informative trajectories.

Result: Video-OPD consistently outperforms GRPO-based RL methods while achieving substantially faster convergence and lower computational cost across empirical evaluations.

Conclusion: On-policy distillation is an effective alternative to conventional reinforcement learning for Temporal Video Grounding, offering better performance with improved efficiency through dense supervision and curriculum learning.

Abstract: Reinforcement learning has emerged as a principled post-training paradigm for Temporal Video Grounding (TVG) due to its on-policy optimization, yet existing GRPO-based methods remain fundamentally constrained by sparse reward signals and substantial computational overhead. We propose Video-OPD, an efficient post-training framework for TVG inspired by recent advances in on-policy distillation. Video-OPD optimizes trajectories sampled directly from the current policy, thereby preserving alignment between training and inference distributions, while a frontier teacher supplies dense, token-level supervision via a reverse KL divergence objective. This formulation preserves the on-policy property critical for mitigating distributional shift, while converting sparse, episode-level feedback into fine-grained, step-wise learning signals. Building on Video-OPD, we introduce Teacher-Validated Disagreement Focusing (TVDF), a lightweight training curriculum that iteratively prioritizes trajectories that are both teacher-reliable and maximally informative for the student, thereby improving training efficiency. Empirical results demonstrate that Video-OPD consistently outperforms GRPO while achieving substantially faster convergence and lower computational cost, establishing on-policy distillation as an effective alternative to conventional reinforcement learning for TVG.

Method: Two-stage pipeline: 1) gradient-boosted regressor estimates correctness likelihood at each fidelity from question features alone, 2) isotonic calibrator refines probabilities for reliable decision-making. System selects minimum-cost fidelity maximizing expected utility given predicted accuracy and retrieval costs.

Result: Evaluated across three deployment scenarios using five datasets (VQA-v2, GQA, TextVQA, LoCoMo, FloodNet) and six VLMs with 7B-235B parameters. Consistently achieves 50-60% cost reductions while retaining 90-95% of full-resolution accuracy across diverse query types and model architectures.

Conclusion: Pre-retrieval fidelity selection is vital to optimize multimodal inference under resource constraints. VOILA demonstrates that adaptive fidelity selection can significantly reduce costs while maintaining high accuracy in vision-language tasks.

Abstract: Despite significant costs from retrieving and processing high-fidelity visual inputs, most multimodal vision-language systems operate at fixed fidelity levels. We introduce VOILA, a framework for Value-Of-Information-driven adaptive fidelity selection in Visual Question Answering (VQA) that optimizes what information to retrieve before model execution. Given a query, VOILA uses a two-stage pipeline: a gradient-boosted regressor estimates correctness likelihood at each fidelity from question features alone, then an isotonic calibrator refines these probabilities for reliable decision-making. The system selects the minimum-cost fidelity maximizing expected utility given predicted accuracy and retrieval costs. We evaluate VOILA across three deployment scenarios using five datasets (VQA-v2, GQA, TextVQA, LoCoMo, FloodNet) and six Vision-Language Models (VLMs) with 7B-235B parameters. VOILA consistently achieves 50-60% cost reductions while retaining 90-95% of full-resolution accuracy across diverse query types and model architectures, demonstrating that pre-retrieval fidelity selection is vital to optimize multimodal inference under resource constraints.

Method: Proposes using statistical normalization and denormalization strategy for reconstruction guidance during convolutional downsampling. Given structure and texture feature maps, the method leverages their mutual relationships to preserve information through the downsampling process.

Result: Extensive experiments show advantages over state-of-the-art methods on images from low-to-high resolutions (256×256 and 512×512). The method is particularly effective when substituting all encoders with the proposed approach.

Conclusion: The proposed statistical normalization/denormalization strategy effectively addresses information loss in convolutional downsampling for image inpainting, improving reconstruction quality across various resolutions.

Abstract: Image inpainting has earned substantial progress, owing to the encoder-and-decoder pipeline, which is benefited from the Convolutional Neural Networks (CNNs) with convolutional downsampling to inpaint the masked regions semantically from the known regions within the encoder, coupled with an upsampling process from the decoder for final inpainting output. Recent studies intuitively identify the high-frequency structure and low-frequency texture to be extracted by CNNs from the encoder, and subsequently for a desirable upsampling recovery. However, the existing arts inevitably overlook the information loss for both structure and texture feature maps during the convolutional downsampling process, hence suffer from a non-ideal upsampling output. In this paper, we systematically answer whether and how the structure and texture feature map can mutually help to alleviate the information loss during the convolutional downsampling. Given the structure and texture feature maps, we adopt the statistical normalization and denormalization strategy for the reconstruction guidance during the convolutional downsampling process. The extensive experimental results validate its advantages to the state-of-the-arts over the images from low-to-high resolutions including 256256 and 512512, especially holds by substituting all the encoders by ours. Our code is available at https://github.com/htyjers/ConvInpaint-TSGL

Method: Three-step approach: 1) Establish reliable plane correspondence, 2) Build pseudo-plane-constrained Gauss-Helmert model, 3) Adaptively estimate vertical translation to handle model uncertainty.

Result: Extensive experiments on five real-world datasets show L2M-Reg is more accurate and computationally efficient than current leading ICP-based and plane-based methods.

Conclusion: L2M-Reg provides a novel building-level solution for LiDAR-to-Model registration when model uncertainty is present, with open-sourced datasets and code.

Abstract: Accurate registration between LiDAR (Light Detection and Ranging) point clouds and semantic 3D city models is a fundamental topic in urban digital twinning and a prerequisite for downstream tasks, such as digital construction, change detection, and model refinement. However, achieving accurate LiDAR-to-Model registration at the individual building level remains challenging, particularly due to the generalization uncertainty in semantic 3D city models at the Level of Detail 2 (LoD2). This paper addresses this gap by proposing L2M-Reg, a plane-based fine registration method that explicitly accounts for model uncertainty. L2M-Reg consists of three key steps: establishing reliable plane correspondence, building a pseudo-plane-constrained Gauss-Helmert model, and adaptively estimating vertical translation. Overall, extensive experiments on five real-world datasets demonstrate that L2M-Reg is both more accurate and computationally efficient than current leading ICP-based and plane-based methods. Therefore, L2M-Reg provides a novel building-level solution regarding LiDAR-to-Model registration when model uncertainty is present. The datasets and code for L2M-Reg can be found: https://github.com/Ziyang-Geodesy/L2M-Reg.

Method: Developed a computer vision pipeline to process footage from over 900 traffic cameras throughout Manhattan and NYC, comparing traffic patterns from November 2024 (pre-implementation baseline) through the program’s implementation in January 2025 until January 2026.

Result: Identified systematic changes in vehicle density across the monitored region following congestion pricing implementation, establishing clear before-and-after patterns through automated analysis.

Conclusion: Automated computer vision analysis of traffic camera data provides an effective method for measuring the real-world impact of urban transportation policies like congestion pricing.

Abstract: We examine the impact of New York City’s congestion pricing program through automated analysis of traffic camera data. Our computer vision pipeline processes footage from over 900 cameras distributed throughout Manhattan and New York, comparing traffic patterns from November 2024 through the program’s implementation in January 2025 until January 2026. We establish baseline traffic patterns and identify systematic changes in vehicle density across the monitored region.

Method: MUSE formulates storytelling as a closed-loop constraint enforcement problem using a multi-agent framework with iterative plan-execute-verify-revise loop. It translates narrative intent into explicit machine-executable controls over identity, spatial composition, and temporal continuity, applying targeted multimodal feedback to correct violations during generation.

Result: Experiments show MUSE substantially improves long-horizon narrative coherence, cross-modal identity consistency, and cinematic quality compared to representative baselines. The paper also introduces MUSEBench, a reference-free evaluation protocol validated by human judgments.

Conclusion: MUSE effectively addresses the intent-execution gap in long-form audio-visual storytelling through closed-loop constraint enforcement and multimodal feedback, demonstrating superior coherence and consistency over existing approaches.

Abstract: Generating long-form audio-visual stories from a short user prompt remains challenging due to an intent-execution gap, where high-level narrative intent must be preserved across coherent, shot-level multimodal generation over long horizons. Existing approaches typically rely on feed-forward pipelines or prompt-only refinement, which often leads to semantic drift and identity inconsistency as sequences grow longer. We address this challenge by formulating storytelling as a closed-loop constraint enforcement problem and propose MUSE, a multi-agent framework that coordinates generation through an iterative plan-execute-verify-revise loop. MUSE translates narrative intent into explicit, machine-executable controls over identity, spatial composition, and temporal continuity, and applies targeted multimodal feedback to correct violations during generation. To evaluate open-ended storytelling without ground-truth references, we introduce MUSEBench, a reference-free evaluation protocol validated by human judgments. Experiments demonstrate that MUSE substantially improves long-horizon narrative coherence, cross-modal identity consistency, and cinematic quality compared with representative baselines.

Method: Neurosymbolic approach: use LLMs to generate parameterized programmatic representations for hypothesized solution rules, then perform parameter fitting using Bayesian optimization

Result: Method evaluated on classifying Bongard problem images given ground truth rules, as well as solving problems from scratch

Conclusion: The approach demonstrates a way to tackle challenging visual reasoning problems that require generalization to novel situations beyond typical VLM capabilities

Abstract: Vision-Language Models (VLMs) have made great strides in everyday visual tasks, such as captioning a natural image, or answering commonsense questions about such images. But humans possess the puzzling ability to deploy their visual reasoning abilities in radically new situations, a skill rigorously tested by the classic set of visual reasoning challenges known as the Bongard problems. We present a neurosymbolic approach to solving these problems: given a hypothesized solution rule for a Bongard problem, we leverage LLMs to generate parameterized programmatic representations for the rule and perform parameter fitting using Bayesian optimization. We evaluate our method on classifying Bongard problem images given the ground truth rule, as well as on solving the problems from scratch.

Method: Two main strategies: 1) FakeTwins uses pretrained networks as encoders to compute self-supervised loss applied through generated images to train the generator. 2) Discriminator consistency enforces alignment between discriminators that evaluate feature maps from CNN and ViT networks to promote coherent learning and training robustness.

Result: Extensive evaluation across 17 datasets (large, small, limited data, various domains) shows HP-GAN consistently outperforms state-of-the-art methods in FID scores, achieving significant improvements in image diversity and quality.

Conclusion: HP-GAN effectively exploits neural network priors through self-supervised learning and discriminator consistency, leading to superior image generation performance across diverse datasets and conditions.

Abstract: Generative Adversarial Networks (GANs) have made significant progress in enhancing the quality of image synthesis. Recent methods frequently leverage pretrained networks to calculate perceptual losses or utilize pretrained feature spaces. In this paper, we extend the capabilities of pretrained networks by incorporating innovative self-supervised learning techniques and enforcing consistency between discriminators during GAN training. Our proposed method, named HP-GAN, effectively exploits neural network priors through two primary strategies: FakeTwins and discriminator consistency. FakeTwins leverages pretrained networks as encoders to compute a self-supervised loss and applies this through the generated images to train the generator, thereby enabling the generation of more diverse and high quality images. Additionally, we introduce a consistency mechanism between discriminators that evaluate feature maps extracted from Convolutional Neural Network (CNN) and Vision Transformer (ViT) feature networks. Discriminator consistency promotes coherent learning among discriminators and enhances training robustness by aligning their assessments of image quality. Our extensive evaluation across seventeen datasets-including scenarios with large, small, and limited data, and covering a variety of image domains-demonstrates that HP-GAN consistently outperforms current state-of-the-art methods in terms of Fréchet Inception Distance (FID), achieving significant improvements in image diversity and quality. Code is available at: https://github.com/higun2/HP-GAN.

Method: IVC-Prune identifies implicit visual coordinate (IVC) tokens by analyzing RoPE’s mathematical properties (positions where rotation matrices approximate identity or 90° rotation). Foreground tokens are identified via two-stage process: semantic seed discovery followed by contextual refinement using value-vector similarity. The method is training-free and prompt-aware.

Result: Extensive evaluations across 4 LVLMs and 20 benchmarks show IVC-Prune reduces visual tokens by ~50% while maintaining ≥99% of original performance, with improvements on several benchmarks.

Conclusion: The paper introduces a novel insight about LVLMs’ implicit visual coordinate systems through RoPE and proposes an effective pruning strategy that preserves both spatial reasoning capabilities and semantic relevance, significantly reducing computational costs.

Abstract: Large Vision-Language Models (LVLMs) achieve impressive performance across multiple tasks. A significant challenge, however, is their prohibitive inference cost when processing high-resolution visual inputs. While visual token pruning has emerged as a promising solution, existing methods that primarily focus on semantic relevance often discard tokens that are crucial for spatial reasoning. We address this gap through a novel insight into \emph{how LVLMs process spatial reasoning}. Specifically, we reveal that LVLMs implicitly establish visual coordinate systems through Rotary Position Embeddings (RoPE), where specific token positions serve as \textbf{implicit visual coordinates} (IVC tokens) that are essential for spatial reasoning. Based on this insight, we propose \textbf{IVC-Prune}, a training-free, prompt-aware pruning strategy that retains both IVC tokens and semantically relevant foreground tokens. IVC tokens are identified by theoretically analyzing the mathematical properties of RoPE, targeting positions at which its rotation matrices approximate identity matrix or the $90^\circ$ rotation matrix. Foreground tokens are identified through a robust two-stage process: semantic seed discovery followed by contextual refinement via value-vector similarity. Extensive evaluations across four representative LVLMs and twenty diverse benchmarks show that IVC-Prune reduces visual tokens by approximately 50% while maintaining $\geq$ 99% of the original performance and even achieving improvements on several benchmarks. Source codes are available at https://github.com/FireRedTeam/IVC-Prune.

Method: Introduces JRDB-Pose3D dataset captured from mobile robotic platforms in complex indoor/outdoor environments, providing SMPL-based 3D pose annotations with consistent body-shape parameters, track IDs, and inheriting all JRDB dataset annotations including 2D poses, social grouping, activities, and demographic information.

Result: Dataset contains 5-10 human poses per frame on average, with some scenes featuring up to 35 individuals simultaneously, presenting challenges like occlusions, truncated bodies, and out-of-frame body parts that reflect real-world complexity.

Conclusion: JRDB-Pose3D bridges the gap between controlled lab datasets and real-world applications by providing comprehensive 3D human pose data in crowded environments, enabling research in multi-human perception and human-centric understanding tasks.

Abstract: Real-world scenes are inherently crowded. Hence, estimating 3D poses of all nearby humans, tracking their movements over time, and understanding their activities within social and environmental contexts are essential for many applications, such as autonomous driving, robot perception, robot navigation, and human-robot interaction. However, most existing 3D human pose estimation datasets primarily focus on single-person scenes or are collected in controlled laboratory environments, which restricts their relevance to real-world applications. To bridge this gap, we introduce JRDB-Pose3D, which captures multi-human indoor and outdoor environments from a mobile robotic platform. JRDB-Pose3D provides rich 3D human pose annotations for such complex and dynamic scenes, including SMPL-based pose annotations with consistent body-shape parameters and track IDs for each individual over time. JRDB-Pose3D contains, on average, 5-10 human poses per frame, with some scenes featuring up to 35 individuals simultaneously. The proposed dataset presents unique challenges, including frequent occlusions, truncated bodies, and out-of-frame body parts, which closely reflect real-world environments. Moreover, JRDB-Pose3D inherits all available annotations from the JRDB dataset, such as 2D pose, information about social grouping, activities, and interactions, full-scene semantic masks with consistent human- and object-level tracking, and detailed annotations for each individual, such as age, gender, and race, making it a holistic dataset for a wide range of downstream perception and human-centric understanding tasks.

Method: Proposes Gaussian Boundary Optimization (GBO), a training-free inference framework that predicts segment boundaries by solving an optimization problem that balances proposal coverage and segment compactness. Derives closed-form solution and analyzes optimality conditions under varying penalty regimes.

Result: GBO significantly improves localization performance, achieving state-of-the-art results across standard benchmarks. It’s efficient, generalizable across various proposal schemes, and compatible with both single-Gaussian and mixture-based architectures.

Conclusion: GBO provides a principled optimization-based inference framework for weakly supervised temporal video grounding that overcomes limitations of heuristic boundary mapping, offering both theoretical foundations and practical advantages for improved video segment localization.

Abstract: Weakly supervised temporal video grounding aims to localize query-relevant segments in untrimmed videos using only video-sentence pairs, without requiring ground-truth segment annotations that specify exact temporal boundaries. Recent approaches tackle this task by utilizing Gaussian-based temporal proposals to represent query-relevant segments. However, their inference strategies rely on heuristic mappings from Gaussian parameters to segment boundaries, resulting in suboptimal localization performance. To address this issue, we propose Gaussian Boundary Optimization (GBO), a novel inference framework that predicts segment boundaries by solving a principled optimization problem that balances proposal coverage and segment compactness. We derive a closed-form solution for this problem and rigorously analyze the optimality conditions under varying penalty regimes. Beyond its theoretical foundations, GBO offers several practical advantages: it is training-free and compatible with both single-Gaussian and mixture-based proposal architectures. Our experiments show that GBO significantly improves localization, achieving state-of-the-art results across standard benchmarks. Extensive experiments demonstrate the efficiency and generalizability of GBO across various proposal schemes. The code is available at \href{https://github.com/sunoh-kim/gbo}{https://github.com/sunoh-kim/gbo}.

Method: Developed SKELEX using self-supervised learning on 1.2 million diverse musculoskeletal radiographs. The model was evaluated on 12 downstream diagnostic tasks including fracture detection, osteoarthritis grading, and bone tumor classification. Also demonstrated zero-shot abnormality localization and developed an interpretable, region-guided model for bone tumor prediction.

Result: SKELEX generally outperformed baselines on diagnostic tasks and demonstrated zero-shot abnormality localization without task-specific training. The bone tumor model maintained robust performance on independent external datasets and was deployed as a publicly accessible web application.

Conclusion: SKELEX provides a scalable, label-efficient, and generalizable AI framework for musculoskeletal imaging, establishing a foundation for clinical translation and data-efficient research in musculoskeletal radiology.

Abstract: Artificial intelligence (AI) has shown promise in detecting and characterizing musculoskeletal diseases from radiographs. However, most existing models remain task-specific, annotation-dependent, and limited in generalizability across diseases and anatomical regions. Although a generalizable foundation model trained on large-scale musculoskeletal radiographs is clinically needed, publicly available datasets remain limited in size and lack sufficient diversity to enable training across a wide range of musculoskeletal conditions and anatomical sites. Here, we present SKELEX, a large-scale foundation model for musculoskeletal radiographs, trained using self-supervised learning on 1.2 million diverse, condition-rich images. The model was evaluated on 12 downstream diagnostic tasks and generally outperformed baselines in fracture detection, osteoarthritis grading, and bone tumor classification. Furthermore, SKELEX demonstrated zero-shot abnormality localization, producing error maps that identified pathologic regions without task-specific training. Building on this capability, we developed an interpretable, region-guided model for predicting bone tumors, which maintained robust performance on independent external datasets and was deployed as a publicly accessible web application. Overall, SKELEX provides a scalable, label-efficient, and generalizable AI framework for musculoskeletal imaging, establishing a foundation for both clinical translation and data-efficient research in musculoskeletal radiology.

Method: Replace Stable Diffusion features with an optimal transport algorithm that includes a Gromov-Wasserstein spatial smoothness prior, combined with DINOv2 features for semantic matching.

Result: Significantly boosts DINOv2 baseline performance, achieves competitive or superior results compared to Stable Diffusion ensemble methods, while being 5-10x more computationally efficient.

Conclusion: Optimal transport with spatial priors can effectively replace Stable Diffusion features for semantic correspondence, providing better efficiency without sacrificing performance.

Abstract: Establishing correspondences between image pairs is a long studied problem in computer vision. With recent large-scale foundation models showing strong zero-shot performance on downstream tasks including classification and segmentation, there has been interest in using the internal feature maps of these models for the semantic correspondence task. Recent works observe that features from DINOv2 and Stable Diffusion (SD) are complementary, the former producing accurate but sparse correspondences, while the latter produces spatially consistent correspondences. As a result, current state-of-the-art methods for semantic correspondence involve combining features from both models in an ensemble. While the performance of these methods is impressive, they are computationally expensive, requiring evaluating feature maps from large-scale foundation models. In this work we take a different approach, instead replacing SD features with a superior matching algorithm which is imbued with the desirable spatial consistency property. Specifically, we replace the standard nearest neighbours matching with an optimal transport algorithm that includes a Gromov Wasserstein spatial smoothness prior. We show that we can significantly boost the performance of the DINOv2 baseline, and be competitive and sometimes surpassing state-of-the-art methods using Stable Diffusion features, while being 5–10x more efficient. We make code available at https://github.com/fsnelgar/semantic_matching_gwot .

Method: EvoAug combines generative models (conditional diffusion, few-shot NeRFs) with efficient evolutionary algorithm to learn stochastic augmentation trees that hierarchically compose augmentations, enabling structured and adaptive transformations.

Result: Strong performance across fine-grained classification and few-shot learning tasks. Pipeline discovers augmentations that align with domain knowledge even in low-data settings.

Conclusion: Learned generative augmentations have significant potential for robust model training, unlocking new possibilities beyond traditional augmentation methods.

Abstract: Data augmentation has long been a cornerstone for reducing overfitting in vision models, with methods like AutoAugment automating the design of task-specific augmentations. Recent advances in generative models, such as conditional diffusion and few-shot NeRFs, offer a new paradigm for data augmentation by synthesizing data with significantly greater diversity and realism. However, unlike traditional augmentations like cropping or rotation, these methods introduce substantial changes that enhance robustness but also risk degrading performance if the augmentations are poorly matched to the task. In this work, we present EvoAug, an automated augmentation learning pipeline, which leverages these generative models alongside an efficient evolutionary algorithm to learn optimal task-specific augmentations. Our pipeline introduces a novel approach to image augmentation that learns stochastic augmentation trees that hierarchically compose augmentations, enabling more structured and adaptive transformations. We demonstrate strong performance across fine-grained classification and few-shot learning tasks. Notably, our pipeline discovers augmentations that align with domain knowledge, even in low-data settings. These results highlight the potential of learned generative augmentations, unlocking new possibilities for robust model training.

Method: Species-fair design comparing children and CNNs on few-shot semi-supervised category learning task with novel object categories. Both exposed to mixtures of labeled and unlabeled exemplars while varying supervision (1/3/6 labels), target feature (size, shape, pattern), and perceptual alignment (high/low).

Result: Children generalize rapidly from minimal labels but show strong feature-specific biases and sensitivity to alignment. CNNs show different interaction profile: added supervision improves performance, but both alignment and feature structure moderate the impact additional supervision has on learning.

Conclusion: Human-model comparisons must be drawn under the right conditions, emphasizing interactions among supervision, feature structure, and alignment rather than overall accuracy. The study reveals fundamental differences in how biological and artificial systems learn from sparse data.

Abstract: Understanding how humans and machines learn from sparse data is central to cognitive science and machine learning. Using a species-fair design, we compare children and convolutional neural networks (CNNs) in a few-shot semi-supervised category learning task. Both learners are exposed to novel object categories under identical conditions. Learners receive mixtures of labeled and unlabeled exemplars while we vary supervision (1/3/6 labels), target feature (size, shape, pattern), and perceptual alignment (high/low). We find that children generalize rapidly from minimal labels but show strong feature-specific biases and sensitivity to alignment. CNNs show a different interaction profile: added supervision improves performance, but both alignment and feature structure moderate the impact additional supervision has on learning. These results show that human-model comparisons must be drawn under the right conditions, emphasizing interactions among supervision, feature structure, and alignment rather than overall accuracy.

Method: Uses diffusion models with guidance framework: unconditional diffusion model trained only on 3D pose data, guided by gradients from 2D keypoint detector heatmaps to generate poses consistent with 2D images

Result: State-of-the-art on Human 3.6M under best-of-m evaluation for methods without paired data; competitive on MPI-INF-3DHP and 3DPW; demonstrates flexibility for pose generation and completion tasks

Conclusion: Diffusion models provide effective probabilistic framework for 3D pose estimation without paired data, handling ambiguity and enabling novel applications like pose generation/completion

Abstract: 3D human pose estimation from 2D images is a challenging problem due to depth ambiguity and occlusion. Because of these challenges the task is underdetermined, where there exists multiple – possibly infinite – poses that are plausible given the image. Despite this, many prior works assume the existence of a deterministic mapping and estimate a single pose given an image. Furthermore, methods based on machine learning require a large amount of paired 2D-3D data to train and suffer from generalization issues to unseen scenarios. To address both of these issues, we propose a framework for pose estimation using diffusion models, which enables sampling from a probability distribution over plausible poses which are consistent with a 2D image. Our approach falls under the guidance framework for conditional generation, and guides samples from an unconditional diffusion model, trained only on 3D data, using the gradients of the heatmaps from a 2D keypoint detector. We evaluate our method on the Human 3.6M dataset under best-of-$m$ multiple hypothesis evaluation, showing state-of-the-art performance among methods which do not require paired 2D-3D data for training. We additionally evaluate the generalization ability using the MPI-INF-3DHP and 3DPW datasets and demonstrate competitive performance. Finally, we demonstrate the flexibility of our framework by using it for novel tasks including pose generation and pose completion, without the need to train bespoke conditional models. We make code available at https://github.com/fsnelgar/diffusion_pose .

Method: Created FinMTM benchmark with 11,133 bilingual QA pairs grounded in financial visuals (candlestick charts, statistical plots, report figures). Covers multiple task types: single/multiple-choice questions, multi-turn dialogues, and agent-based tasks. Developed specialized evaluation protocols for each task type.

Result: Evaluation of 22 VLMs revealed significant limitations in fine-grained visual perception, long-context reasoning, and complex agent workflows, demonstrating the need for improved multimodal reasoning capabilities in financial domains.

Conclusion: FinMTM provides a comprehensive benchmark for evaluating VLMs on realistic financial tasks, highlighting current model limitations and establishing a foundation for future improvements in multimodal financial reasoning.

Abstract: The financial domain poses substantial challenges for vision-language models (VLMs) due to specialized chart formats and knowledge-intensive reasoning requirements. However, existing financial benchmarks are largely single-turn and rely on a narrow set of question formats, limiting comprehensive evaluation in realistic application scenarios. To address this gap, we propose FinMTM, a multi-turn multimodal benchmark that expands diversity along both data and task dimensions. On the data side, we curate and annotate 11{,}133 bilingual (Chinese and English) financial QA pairs grounded in financial visuals, including candlestick charts, statistical plots, and report figures. On the task side, FinMTM covers single- and multiple-choice questions, multi-turn open-ended dialogues, and agent-based tasks. We further design task-specific evaluation protocols, including a set-overlap scoring rule for multiple-choice questions, a weighted combination of turn-level and session-level scores for multi-turn dialogues, and a composite metric that integrates planning quality with final outcomes for agent tasks. Extensive experimental evaluation of 22 VLMs reveal their limitations in fine-grained visual perception, long-context reasoning, and complex agent workflows.

Method: SwiftVLM introduces a bypass pruning paradigm where unselected visual tokens are preserved and forwarded to subsequent pruning stages for re-evaluation. It performs pruning at model-specific layers with strong visual token selection capability, enabling independent pruning decisions across layers without requiring training.

Result: Experiments across multiple VLMs and benchmarks show SwiftVLM consistently outperforms existing pruning strategies, achieving superior accuracy-efficiency trade-offs and more faithful visual token selection behavior compared to methods that make irreversible early pruning decisions.

Conclusion: The bypass paradigm effectively addresses the limitations of early pruning by preserving potentially important visual tokens for later re-evaluation, enabling VLMs to maintain fine-grained visual understanding while reducing computational costs.

Abstract: Visual token pruning is a promising approach for reducing the computational cost of vision-language models (VLMs), and existing methods often rely on early pruning decisions to improve efficiency. While effective on coarse-grained reasoning tasks, they suffer from significant performance degradation on tasks requiring fine-grained visual details. Through layer-wise analysis, we reveal substantial discrepancies in visual token importance across layers, showing that tokens deemed unimportant at shallow layers can later become highly relevant for text-conditioned reasoning. To avoid irreversible critical information loss caused by premature pruning, we introduce a new pruning paradigm, termed bypass, which preserves unselected visual tokens and forwards them to subsequent pruning stages for re-evaluation. Building on this paradigm, we propose SwiftVLM, a simple and training-free method that performs pruning at model-specific layers with strong visual token selection capability, while enabling independent pruning decisions across layers. Experiments across multiple VLMs and benchmarks demonstrate that SwiftVLM consistently outperforms existing pruning strategies, achieving superior accuracy-efficiency trade-offs and more faithful visual token selection behavior.

Method: Integrates universal proposal network (UPN) for category-agnostic boxes, SAM2 for mask extraction, and DINOv2 features. Introduces graph-based confidence reweighting where bounding boxes are nodes in a directed graph, using graph diffusion to propagate confidence scores and refine proposals.

Result: Substantially outperforms existing approaches on Pascal-5^i, COCO-20^i, and CD-FSOD datasets. On challenging CD-FSOD dataset, achieves 31.6 AP in 10-shot setting vs. 21.4 AP for previous training-free methods.

Conclusion: FSOD-VFM effectively addresses fragmentation issues in foundation model proposals through graph-based confidence reweighting, achieving state-of-the-art few-shot object detection without additional training.

Abstract: In this paper, we present FSOD-VFM: Few-Shot Object Detectors with Vision Foundation Models, a framework that leverages vision foundation models to tackle the challenge of few-shot object detection. FSOD-VFM integrates three key components: a universal proposal network (UPN) for category-agnostic bounding box generation, SAM2 for accurate mask extraction, and DINOv2 features for efficient adaptation to new object categories. Despite the strong generalization capabilities of foundation models, the bounding boxes generated by UPN often suffer from overfragmentation, covering only partial object regions and leading to numerous small, false-positive proposals rather than accurate, complete object detections. To address this issue, we introduce a novel graph-based confidence reweighting method. In our approach, predicted bounding boxes are modeled as nodes in a directed graph, with graph diffusion operations applied to propagate confidence scores across the network. This reweighting process refines the scores of proposals, assigning higher confidence to whole objects and lower confidence to local, fragmented parts. This strategy improves detection granularity and effectively reduces the occurrence of false-positive bounding box proposals. Through extensive experiments on Pascal-5$^i$, COCO-20$^i$, and CD-FSOD datasets, we demonstrate that our method substantially outperforms existing approaches, achieving superior performance without requiring additional training. Notably, on the challenging CD-FSOD dataset, which spans multiple datasets and domains, our FSOD-VFM achieves 31.6 AP in the 10-shot setting, substantially outperforming previous training-free methods that reach only 21.4 AP. Code is available at: https://intellindust-ai-lab.github.io/projects/FSOD-VFM.

Method: Proposes Diversity-Preserved DMD (DP-DMD) with role-separated distillation: first step uses target-prediction objective (e.g., v-prediction) to preserve diversity, subsequent steps use standard DMD loss for quality refinement, with gradients from DMD blocked at first step.

Result: Preserves sample diversity while maintaining visual quality on par with state-of-the-art methods in extensive text-to-image experiments, without requiring perceptual backbones, discriminators, auxiliary networks, or additional ground-truth images.

Conclusion: DP-DMD provides a simple yet effective solution to mode collapse in diffusion model distillation, achieving diversity preservation and quality maintenance without complex regularization schemes.

Abstract: Distribution matching distillation (DMD) aligns a multi-step generator with its few-step counterpart to enable high-quality generation under low inference cost. However, DMD tends to suffer from mode collapse, as its reverse-KL formulation inherently encourages mode-seeking behavior, for which existing remedies typically rely on perceptual or adversarial regularization, thereby incurring substantial computational overhead and training instability. In this work, we propose a role-separated distillation framework that explicitly disentangles the roles of distilled steps: the first step is dedicated to preserving sample diversity via a target-prediction (e.g., v-prediction) objective, while subsequent steps focus on quality refinement under the standard DMD loss, with gradients from the DMD objective blocked at the first step. We term this approach Diversity-Preserved DMD (DP-DMD), which, despite its simplicity – no perceptual backbone, no discriminator, no auxiliary networks, and no additional ground-truth images – preserves sample diversity while maintaining visual quality on par with state-of-the-art methods in extensive text-to-image experiments.

Method: Three key innovations: (1) Share-activation KAN (SaKAN) reformulates Sprecher’s variant of Kolmogorov-Arnold representation theorem for better optimization; (2) Grad-Free Spline eliminates spline gradients that consume huge GPU memory while contributing negligibly to training; (3) ALL U-KAN implements the first fully KA-based deep model with KA and KAonv layers replacing FC and Conv layers.

Result: Extensive evaluations on three medical image segmentation tasks show superiority over partial KA-based and traditional architectures. Compared to directly deeply stacked KAN, ALL U-KAN achieves 10× parameter reduction and >20× memory consumption reduction while maintaining higher segmentation accuracy.

Conclusion: The paper successfully demonstrates that KA-based layers can entirely replace traditional architectures in deep learning, achieving superior learning capacity while overcoming previous limitations, unlocking new explorations into deep KAN architectures.

Abstract: Deeply stacked KANs are practically impossible due to high training difficulties and substantial memory requirements. Consequently, existing studies can only incorporate few KAN layers, hindering the comprehensive exploration of KANs. This study overcomes these limitations and introduces the first fully KA-based deep model, demonstrating that KA-based layers can entirely replace traditional architectures in deep learning and achieve superior learning capacity. Specifically, (1) the proposed Share-activation KAN (SaKAN) reformulates Sprecher’s variant of Kolmogorov-Arnold representation theorem, which achieves better optimization due to its simplified parameterization and denser training samples, to ease training difficulty, (2) this paper indicates that spline gradients contribute negligibly to training while consuming huge GPU memory, thus proposes the Grad-Free Spline to significantly reduce memory usage and computational overhead. (3) Building on these two innovations, our ALL U-KAN is the first representative implementation of fully KA-based deep model, where the proposed KA and KAonv layers completely replace FC and Conv layers. Extensive evaluations on three medical image segmentation tasks confirm the superiority of the full KA-based architecture compared to partial KA-based and traditional architectures, achieving all higher segmentation accuracy. Compared to directly deeply stacked KAN, ALL U-KAN achieves 10 times reduction in parameter count and reduces memory consumption by more than 20 times, unlocking the new explorations into deep KAN architectures.

Method: 1) Self-supervised pre-training of Group Activity Feature (GAF) space based on activity similarity. 2) Human-in-the-loop interactive fine-tuning where users label selected videos as positive/negative. 3) Data-efficient video selection process to choose informative videos for labeling. 4) Contrastive learning to update GAF space by moving positive videos closer and negative videos farther from queries.

Result: Significant improvement in retrieval performance on two team sports datasets. Ablation studies show that various components of the human-in-the-loop adaptation contribute to performance gains.

Conclusion: The proposed human-in-the-loop adaptation enables effective group activity video retrieval without requiring group activity annotations, offering a flexible and adaptive approach that can be tailored to user preferences.

Abstract: This paper proposes human-in-the-loop adaptation for Group Activity Feature Learning (GAFL) without group activity annotations. This human-in-the-loop adaptation is employed in a group-activity video retrieval framework to improve its retrieval performance. Our method initially pre-trains the GAF space based on the similarity of group activities in a self-supervised manner, unlike prior work that classifies videos into pre-defined group activity classes in a supervised learning manner. Our interactive fine-tuning process updates the GAF space to allow a user to better retrieve videos similar to query videos given by the user. In this fine-tuning, our proposed data-efficient video selection process provides several videos, which are selected from a video database, to the user in order to manually label these videos as positive or negative. These labeled videos are used to update (i.e., fine-tune) the GAF space, so that the positive and negative videos move closer to and farther away from the query videos through contrastive learning. Our comprehensive experimental results on two team sports datasets validate that our method significantly improves the retrieval performance. Ablation studies also demonstrate that several components in our human-in-the-loop adaptation contribute to the improvement of the retrieval performance. Code: https://github.com/chihina/GAFL-FINE-CVIU.

Method: Proposes BinaryDemoire with two key components: 1) Moiré-aware binary gate (MABG) extracts lightweight frequency descriptors and activation statistics to predict channel-wise gating coefficients for binary convolution responses. 2) Shuffle-grouped residual adapter (SGRA) performs structured sparse shortcut alignment with interleaved mixing for cross-channel information exchange.

Result: Extensive experiments on four benchmarks demonstrate BinaryDemoire surpasses current binarization methods for demoiréing tasks.

Conclusion: BinaryDemoire effectively addresses the challenge of binarized demoiréing by explicitly accommodating the frequency structure of moiré degradations, offering an efficient solution for deployment.

Abstract: Image demoiréing aims to remove structured moiré artifacts in recaptured imagery, where degradations are highly frequency-dependent and vary across scales and directions. While recent deep networks achieve high-quality restoration, their full-precision designs remain costly for deployment. Binarization offers an extreme compression regime by quantizing both activations and weights to 1-bit. Yet, it has been rarely studied for demoiréing and performs poorly when naively applied. In this work, we propose BinaryDemoire, a binarized demoiréing framework that explicitly accommodates the frequency structure of moiré degradations. First, we introduce a moiré-aware binary gate (MABG) that extracts lightweight frequency descriptors together with activation statistics. It predicts channel-wise gating coefficients to condition the aggregation of binary convolution responses. Second, we design a shuffle-grouped residual adapter (SGRA) that performs structured sparse shortcut alignment. It further integrates interleaved mixing to promote information exchange across different channel partitions. Extensive experiments on four benchmarks demonstrate that the proposed BinaryDemoire surpasses current binarization methods. Code: https://github.com/zhengchen1999/BinaryDemoire.

Method: Proposes LSGQuant with three key components: 1) Dynamic Range Adaptive Quantizer (DRAQ) to fit video token activations, 2) Variance-Oriented Layer Training Strategy (VOLTS) based on layer sensitivity analysis, and 3) Quantization-Aware Optimization (QAO) to jointly refine quantized and high-precision branches.

Result: Extensive experiments show the method achieves nearly the same performance as the original full-precision model and significantly outperforms existing quantization techniques for video super-resolution.

Conclusion: LSGQuant effectively compresses one-step diffusion models for video super-resolution while maintaining performance, addressing key challenges in quantizing diffusion transformers through adaptive quantization and layer-aware optimization.

Abstract: One-Step Diffusion Models have demonstrated promising capability and fast inference in video super-resolution (VSR) for real-world. Nevertheless, the substantial model size and high computational cost of Diffusion Transformers (DiTs) limit downstream applications. While low-bit quantization is a common approach for model compression, the effectiveness of quantized models is challenged by the high dynamic range of input latent and diverse layer behaviors. To deal with these challenges, we introduce LSGQuant, a layer-sensitivity guided quantizing approach for one-step diffusion-based real-world VSR. Our method incorporates a Dynamic Range Adaptive Quantizer (DRAQ) to fit video token activations. Furthermore, we estimate layer sensitivity and implement a Variance-Oriented Layer Training Strategy (VOLTS) by analyzing layer-wise statistics in calibration. We also introduce Quantization-Aware Optimization (QAO) to jointly refine the quantized branch and a retained high-precision branch. Extensive experiments demonstrate that our method has nearly performance to origin model with full-precision and significantly exceeds existing quantization techniques. Code is available at: https://github.com/zhengchen1999/LSGQuant.

Method: Three-module approach: 1) Global Coordinate Estimation predicts point-wise global coordinates with uncertainties for each scan, 2) Prior Coordinate Generation estimates inter-frame point correspondences using attention mechanism, 3) Uncertainty-Guided Coordinate Fusion integrates both predictions end-to-end for temporally consistent 6-DoF pose estimation.

Result: TempLoc outperforms state-of-the-art methods by a large margin on NCLT and Oxford Robot-Car benchmarks, demonstrating effectiveness of temporal-aware correspondence modeling.

Conclusion: Modeling sequential consistency through attention-based correspondence estimation and uncertainty-guided fusion significantly improves LiDAR relocalization robustness in challenging scenarios.

Abstract: LiDAR relocalization aims to estimate the global 6-DoF pose of a sensor in the environment. However, existing regression-based approaches are prone to dynamic or ambiguous scenarios, as they either solely rely on single-frame inference or neglect the spatio-temporal consistency across scans. In this paper, we propose TempLoc, a new LiDAR relocalization framework that enhances the robustness of localization by effectively modeling sequential consistency. Specifically, a Global Coordinate Estimation module is first introduced to predict point-wise global coordinates and associated uncertainties for each LiDAR scan. A Prior Coordinate Generation module is then presented to estimate inter-frame point correspondences by the attention mechanism. Lastly, an Uncertainty-Guided Coordinate Fusion module is deployed to integrate both predictions of point correspondence in an end-to-end fashion, yielding a more temporally consistent and accurate global 6-DoF pose. Experimental results on the NCLT and Oxford Robot-Car benchmarks show that our TempLoc outperforms stateof-the-art methods by a large margin, demonstrating the effectiveness of temporal-aware correspondence modeling in LiDAR relocalization. Our code will be released soon.

Method: Synergizes pre-trained hand expert with 4D scene foundation model via scene-aware visual prompting mechanism; injects high-fidelity hand priors into persistent scene memory for simultaneous reconstruction in single forward pass.

Result: Bypasses reliance on offline optimization, delivers competitive performance in both local hand reconstruction and global positioning, enabling accurate hand meshes and dense metric-scale scene geometry.

Conclusion: Hand3R enables joint 4D hand-scene reconstruction from monocular video, addressing the gap between isolated hand reconstruction and full scene understanding for embodied AI applications.

Abstract: For Embodied AI, jointly reconstructing dynamic hands and the dense scene context is crucial for understanding physical interaction. However, most existing methods recover isolated hands in local coordinates, overlooking the surrounding 3D environment. To address this, we present Hand3R, the first online framework for joint 4D hand-scene reconstruction from monocular video. Hand3R synergizes a pre-trained hand expert with a 4D scene foundation model via a scene-aware visual prompting mechanism. By injecting high-fidelity hand priors into a persistent scene memory, our approach enables simultaneous reconstruction of accurate hand meshes and dense metric-scale scene geometry in a single forward pass. Experiments demonstrate that Hand3R bypasses the reliance on offline optimization and delivers competitive performance in both local hand reconstruction and global positioning.

Method: Formulates ICL as conditional generation via visual analogy (x_s:x_t::x_q:y_q). Uses a frozen Diffusion Transformer (DiT) with role-aware multi-image conditioning and introduces Mixture-of-Experts LoRA to mitigate gradient interference across tasks. Also curates a large-scale dataset spanning perception, restoration, and editing tasks.

Result: VIRAL outperforms existing methods and demonstrates that a unified visual ICL paradigm can handle the majority of visual tasks, including open-domain editing.

Conclusion: The proposed VIRAL framework successfully enables in-context learning for diverse visual tasks through visual analogies, showing that a unified approach can effectively handle perception, restoration, and editing tasks.

Abstract: Replicating In-Context Learning (ICL) in computer vision remains challenging due to task heterogeneity. We propose \textbf{VIRAL}, a framework that elicits visual reasoning from a pre-trained image editing model by formulating ICL as conditional generation via visual analogy ($x_s : x_t :: x_q : y_q$). We adapt a frozen Diffusion Transformer (DiT) using role-aware multi-image conditioning and introduce a Mixture-of-Experts LoRA to mitigate gradient interference across diverse tasks. Additionally, to bridge the gaps in current visual context datasets, we curate a large-scale dataset spanning perception, restoration, and editing. Experiments demonstrate that VIRAL outperforms existing methods, validating that a unified V-ICL paradigm can handle the majority of visual tasks, including open-domain editing. Our code is available at https://anonymous.4open.science/r/VIRAL-744A

Method: Two key components: (1) Instance-Masked Attention applies instance identity masks and trajectory masks in attention blocks to ensure visual tokens interact only with corresponding instance features across space and time; (2) Instance-Masked Loss adaptively emphasizes foreground regions with probabilistic instance masking to reduce background noise while maintaining scene fidelity.

Result: Achieves state-of-the-art driving video generation quality and demonstrates significant improvements in downstream autonomous driving tasks on the nuScenes dataset.

Conclusion: ConsisDrive effectively addresses identity drift in driving world models through instance-level temporal consistency mechanisms, improving both generation quality and downstream task performance.

Abstract: Autonomous driving relies on robust models trained on large-scale, high-quality multi-view driving videos. Although world models provide a cost-effective solution for generating realistic driving data, they often suffer from identity drift, where the same object changes its appearance or category across frames due to the absence of instance-level temporal constraints. We introduce ConsisDrive, an identity-preserving driving world model designed to enforce temporal consistency at the instance level. Our framework incorporates two key components: (1) Instance-Masked Attention, which applies instance identity masks and trajectory masks within attention blocks to ensure that visual tokens interact only with their corresponding instance features across spatial and temporal dimensions, thereby preserving object identity consistency; and (2) Instance-Masked Loss, which adaptively emphasizes foreground regions with probabilistic instance masking, reducing background noise while maintaining overall scene fidelity. By integrating these mechanisms, ConsisDrive achieves state-of-the-art driving video generation quality and demonstrates significant improvements in downstream autonomous driving tasks on the nuScenes dataset. Our project page is https://shanpoyang654.github.io/ConsisDrive/page.html.

Method: FARTrack introduces two key techniques: 1) Task-Specific Self-Distillation - compresses the model by distilling task-specific tokens layer by layer, avoiding suboptimal manual teacher-student layer assignments and enhancing inference speed; 2) Inter-frame Autoregressive Sparsification - sequentially condenses multiple templates to learn a temporally-global optimal sparsification strategy without additional runtime overhead.

Result: FARTrack achieves an AO of 70.6% on GOT-10k benchmark in real-time. The fastest model reaches 343 FPS on GPU and 121 FPS on CPU, demonstrating outstanding speed while maintaining competitive tracking performance.

Conclusion: The proposed FARTrack framework successfully addresses the speed-performance trade-off in visual tracking through autoregressive design and novel optimization techniques, making high-performance tracking practical for deployment on resource-constrained devices.

Abstract: Inference speed and tracking performance are two critical evaluation metrics in the field of visual tracking. However, high-performance trackers often suffer from slow processing speeds, making them impractical for deployment on resource-constrained devices. To alleviate this issue, we propose FARTrack, a Fast Auto-Regressive Tracking framework. Since autoregression emphasizes the temporal nature of the trajectory sequence, it can maintain high performance while achieving efficient execution across various devices. FARTrack introduces Task-Specific Self-Distillation and Inter-frame Autoregressive Sparsification, designed from the perspectives of shallow-yet-accurate distillation and redundant-to-essential token optimization, respectively. Task-Specific Self-Distillation achieves model compression by distilling task-specific tokens layer by layer, enhancing the model’s inference speed while avoiding suboptimal manual teacher-student layer pairs assignments. Meanwhile, Inter-frame Autoregressive Sparsification sequentially condenses multiple templates, avoiding additional runtime overhead while learning a temporally-global optimal sparsification strategy. FARTrack demonstrates outstanding speed and competitive performance. It delivers an AO of 70.6% on GOT-10k in real-time. Beyond, our fastest model achieves a speed of 343 FPS on the GPU and 121 FPS on the CPU.

Method: Proposes PokeFusion Attention, a decoder-level cross-attention mechanism that fuses textual semantics with learned style embeddings directly inside the diffusion decoder. It decouples text and style conditioning at the attention level, keeping the pretrained diffusion backbone frozen while training only decoder cross-attention layers and a compact style projection module.

Result: Experiments on a Pokemon-style character generation benchmark show consistent improvements in style fidelity, semantic alignment, and character shape consistency compared to adapter-based baselines, while maintaining low parameter overhead and inference-time simplicity.

Conclusion: PokeFusion Attention provides a parameter-efficient, plug-and-play control component for reference-free stylized generation that can be easily integrated into existing diffusion pipelines and transferred across different backbones.

Abstract: This paper studies reference-free style-conditioned character generation in text-to-image diffusion models, where high-quality synthesis requires both stable character structure and consistent, fine-grained style expression across diverse prompts. Existing approaches primarily rely on text-only prompting, which is often under-specified for visual style and tends to produce noticeable style drift and geometric inconsistency, or introduce reference-based adapters that depend on external images at inference time, increasing architectural complexity and limiting deployment flexibility.We propose PokeFusion Attention, a lightweight decoder-level cross-attention mechanism that fuses textual semantics with learned style embeddings directly inside the diffusion decoder. By decoupling text and style conditioning at the attention level, our method enables effective reference-free stylized generation while keeping the pretrained diffusion backbone fully frozen.PokeFusion Attention trains only decoder cross-attention layers together with a compact style projection module, resulting in a parameter-efficient and plug-and-play control component that can be easily integrated into existing diffusion pipelines and transferred across different backbones.Experiments on a stylized character generation benchmark (Pokemon-style) demonstrate that our method consistently improves style fidelity, semantic alignment, and character shape consistency compared with representative adapter-based baselines, while maintaining low parameter overhead and inference-time simplicity.

Method: Proposes Spiral RoPE which partitions embedding channels into multiple groups associated with uniformly distributed directions. Each group is rotated according to the projection of patch position onto its corresponding direction, enabling spatial relationships beyond horizontal/vertical axes.

Result: Consistent performance improvements across vision tasks including classification, segmentation, and generation. Qualitative analysis shows more concentrated attention activations on semantically relevant objects and better respect for local object boundaries.

Conclusion: Spiral RoPE demonstrates the importance of multi-directional positional encoding in vision transformers, overcoming fundamental limitations of standard axial 2D RoPE for better modeling of spatial relationships in images.

Abstract: Rotary Position Embedding (RoPE) is the de facto positional encoding in large language models due to its ability to encode relative positions and support length extrapolation. When adapted to vision transformers, the standard axial formulation decomposes two-dimensional spatial positions into horizontal and vertical components, implicitly restricting positional encoding to axis-aligned directions. We identify this directional constraint as a fundamental limitation of the standard axial 2D RoPE, which hinders the modeling of oblique spatial relationships that naturally exist in natural images. To overcome this limitation, we propose Spiral RoPE, a simple yet effective extension that enables multi-directional positional encoding by partitioning embedding channels into multiple groups associated with uniformly distributed directions. Each group is rotated according to the projection of the patch position onto its corresponding direction, allowing spatial relationships to be encoded beyond the horizontal and vertical axes. Across a wide range of vision tasks including classification, segmentation, and generation, Spiral RoPE consistently improves performance. Qualitative analysis of attention maps further show that Spiral RoPE exhibits more concentrated activations on semantically relevant objects and better respects local object boundaries, highlighting the importance of multi-directional positional encoding in vision transformers.

Method: 1) Created EventMind dataset with 500k+ instruction sets for curriculum training; 2) Adaptive temporal window aggregation module to compress temporal tokens while preserving key cues; 3) Sparse density-guided attention module to select informative spatial regions and suppress sparse areas.

Result: Achieves 12.4× throughput improvement over baseline (EventFlash-Zero) while maintaining comparable performance. Supports long-range event stream processing with up to 1,000 bins, significantly outperforming EventGPT’s 5-bin limit.

Conclusion: EventFlash serves as an efficient foundation model for event-based vision by effectively leveraging spatiotemporal sparsity to reduce redundancy and accelerate inference while maintaining robust perception capabilities.

Abstract: Event-based multimodal large language models (MLLMs) enable robust perception in high-speed and low-light scenarios, addressing key limitations of frame-based MLLMs. However, current event-based MLLMs often rely on dense image-like processing paradigms, overlooking the spatiotemporal sparsity of event streams and resulting in high computational cost. In this paper, we propose EventFlash, a novel and efficient MLLM to explore spatiotemporal token sparsification for reducing data redundancy and accelerating inference. Technically, we build EventMind, a large-scale and scene-diverse dataset with over 500k instruction sets, providing both short and long event stream sequences to support our curriculum training strategy. We then present an adaptive temporal window aggregation module for efficient temporal sampling, which adaptively compresses temporal tokens while retaining key temporal cues. Finally, a sparse density-guided attention module is designed to improve spatial token efficiency by selecting informative regions and suppressing empty or sparse areas. Experimental results show that EventFlash achieves a $12.4\times$ throughput improvement over the baseline (EventFlash-Zero) while maintaining comparable performance. It supports long-range event stream processing with up to 1,000 bins, significantly outperforming the 5-bin limit of EventGPT. We believe EventFlash serves as an efficient foundation model for event-based vision.

Method: Two key components: (1) Instance Flow Guider extracts and propagates instance features across frames to enforce temporal consistency, and (2) Spatial Geometric Aligner improves spatial reasoning, ensures precise instance positioning, and explicitly models occlusion hierarchies. Uses CARLA’s autopilot for simulating rare safety-critical scenarios.

Result: Achieves state-of-the-art video generation quality on nuScenes dataset and enhances downstream autonomous driving tasks. Enables rigorous safety evaluation through procedural simulation of rare driving scenarios.

Conclusion: InstaDrive addresses key limitations in driving video generation by introducing instance-aware mechanisms that improve both temporal consistency and spatial geometric fidelity, leading to more realistic training data and better safety evaluation capabilities.

Abstract: Autonomous driving relies on robust models trained on high-quality, large-scale multi-view driving videos. While world models offer a cost-effective solution for generating realistic driving videos, they struggle to maintain instance-level temporal consistency and spatial geometric fidelity. To address these challenges, we propose InstaDrive, a novel framework that enhances driving video realism through two key advancements: (1) Instance Flow Guider, which extracts and propagates instance features across frames to enforce temporal consistency, preserving instance identity over time. (2) Spatial Geometric Aligner, which improves spatial reasoning, ensures precise instance positioning, and explicitly models occlusion hierarchies. By incorporating these instance-aware mechanisms, InstaDrive achieves state-of-the-art video generation quality and enhances downstream autonomous driving tasks on the nuScenes dataset. Additionally, we utilize CARLA’s autopilot to procedurally and stochastically simulate rare but safety-critical driving scenarios across diverse maps and regions, enabling rigorous safety evaluation for autonomous systems. Our project page is https://shanpoyang654.github.io/InstaDrive/page.html.

Method: Extends existing VPR datasets with natural language descriptions, investigates multi-modal fusion for robustness and cross-modal retrieval using Low-Rank Adaptation (LoRA) and Multi-Similarity loss.

Result: Language descriptions yield consistent gains in visually degraded conditions, with most significant impact on smaller backbones. Compact models with language can rival larger vision-only architectures. Cross-modal retrieval baseline substantially outperforms standard contrastive methods.

Conclusion: LaVPR enables a new class of localization systems that are resilient to real-world stochasticity and practical for resource-constrained deployment.

Abstract: Visual Place Recognition (VPR) often fails under extreme environmental changes and perceptual aliasing. Furthermore, standard systems cannot perform “blind” localization from verbal descriptions alone, a capability needed for applications such as emergency response. To address these challenges, we introduce LaVPR, a large-scale benchmark that extends existing VPR datasets with over 650,000 rich natural-language descriptions. Using LaVPR, we investigate two paradigms: Multi-Modal Fusion for enhanced robustness and Cross-Modal Retrieval for language-based localization. Our results show that language descriptions yield consistent gains in visually degraded conditions, with the most significant impact on smaller backbones. Notably, adding language allows compact models to rival the performance of much larger vision-only architectures. For cross-modal retrieval, we establish a baseline using Low-Rank Adaptation (LoRA) and Multi-Similarity loss, which substantially outperforms standard contrastive methods across vision-language models. Ultimately, LaVPR enables a new class of localization systems that are both resilient to real-world stochasticity and practical for resource-constrained deployment. Our dataset and code are available at https://github.com/oferidan1/LaVPR.

Method: Tested 21 vision encoders to compare geometric metrics against compositional binding capabilities. Used standard geometry-based statistics and functional sensitivity measured by input-output Jacobian. Provided analytic analysis showing how objective designs constrain embedding geometry but leave local input-output mapping unconstrained.

Result: Found near-zero correlation between standard geometry-based statistics and compositional binding. In contrast, functional sensitivity reliably tracked compositional binding capability. Analytic analysis confirmed this disparity stems from how existing losses explicitly constrain embedding geometry while leaving functional sensitivity unconstrained.

Conclusion: Global embedding geometry captures only a partial view of representational competence. Functional sensitivity serves as a critical complementary axis for modeling composite structure, suggesting the need to move beyond geometry-focused evaluation protocols.

Abstract: A common assumption in representation learning is that globally well-distributed embeddings support robust and generalizable representations. This focus has shaped both training objectives and evaluation protocols, implicitly treating global geometry as a proxy for representational competence. While global geometry effectively encodes which elements are present, it is often insensitive to how they are composed. We investigate this limitation by testing the ability of geometric metrics to predict compositional binding across 21 vision encoders. We find that standard geometry-based statistics exhibit near-zero correlation with compositional binding. In contrast, functional sensitivity, as measured by the input-output Jacobian, reliably tracks this capability. We further provide an analytic account showing that this disparity arises from objective design, as existing losses explicitly constrain embedding geometry but leave the local input-output mapping unconstrained. These results suggest that global embedding geometry captures only a partial view of representational competence and establish functional sensitivity as a critical complementary axis for modeling composite structure.

Method: Proposes A3-TTA framework that identifies well-predicted target domain images using class compact density metric as anchors, uses them to guide pseudo-label generation, and regularizes via semantic consistency and boundary-aware entropy minimization with self-adaptive exponential moving average for noise mitigation.

Result: Significantly improves average Dice scores by 10.40 to 17.68 percentage points compared to source model, outperforms state-of-the-art TTA methods across medical and natural images, and excels in continual TTA with strong anti-forgetting ability.

Conclusion: A3-TTA provides an effective anchor-guided pseudo-labeling approach for test-time adaptation in image segmentation that stabilizes adaptation, prevents catastrophic forgetting, and works well across different segmentation architectures and domains.

Abstract: Test-Time Adaptation (TTA) offers a practical solution for deploying image segmentation models under domain shift without accessing source data or retraining. Among existing TTA strategies, pseudo-label-based methods have shown promising performance. However, they often rely on perturbation-ensemble heuristics (e.g., dropout sampling, test-time augmentation, Gaussian noise), which lack distributional grounding and yield unstable training signals. This can trigger error accumulation and catastrophic forgetting during adaptation. To address this, we propose \textbf{A3-TTA}, a TTA framework that constructs reliable pseudo-labels through anchor-guided supervision. Specifically, we identify well-predicted target domain images using a class compact density metric, under the assumption that confident predictions imply distributional proximity to the source domain. These anchors serve as stable references to guide pseudo-label generation, which is further regularized via semantic consistency and boundary-aware entropy minimization. Additionally, we introduce a self-adaptive exponential moving average strategy to mitigate label noise and stabilize model update during adaptation. Evaluated on both multi-domain medical images (heart structure and prostate segmentation) and natural images, A3-TTA significantly improves average Dice scores by 10.40 to 17.68 percentage points compared to the source model, outperforming several state-of-the-art TTA methods under different segmentation model architectures. A3-TTA also excels in continual TTA, maintaining high performance across sequential target domains with strong anti-forgetting ability. The code will be made publicly available at https://github.com/HiLab-git/A3-TTA.

Method: Uses EfficientNetV2-S for image quality assessment, fine-tuned Vision Foundation Model for abnormality detection and multi-disease classification, with unified adaptive aggregation to integrate 2D slice predictions into 3D patient-level diagnosis.

Result: Achieved high F1 scores: 99.01% for quality assessment, 97.46% for abnormality detection, 94.39% for patient-level diagnosis; matched expert performance in human-machine comparisons with better efficiency.

Conclusion: FOCUS automates the image-to-diagnosis pipeline, advancing unmanned ophthalmology with a validated blueprint for autonomous screening to enhance retinal care accessibility and efficiency at population scale.

Abstract: Optical coherence tomography (OCT) has revolutionized retinal disease diagnosis with its high-resolution and three-dimensional imaging nature, yet its full diagnostic automation in clinical practices remains constrained by multi-stage workflows and conventional single-slice single-task AI models. We present Full-process OCT-based Clinical Utility System (FOCUS), a foundation model-driven framework enabling end-to-end automation of 3D OCT retinal disease diagnosis. FOCUS sequentially performs image quality assessment with EfficientNetV2-S, followed by abnormality detection and multi-disease classification using a fine-tuned Vision Foundation Model. Crucially, FOCUS leverages a unified adaptive aggregation method to intelligently integrate 2D slices-level predictions into comprehensive 3D patient-level diagnosis. Trained and tested on 3,300 patients (40,672 slices), and externally validated on 1,345 patients (18,498 slices) across four different-tier centers and diverse OCT devices, FOCUS achieved high F1 scores for quality assessment (99.01%), abnormally detection (97.46%), and patient-level diagnosis (94.39%). Real-world validation across centers also showed stable performance (F1: 90.22%-95.24%). In human-machine comparisons, FOCUS matched expert performance in abnormality detection (F1: 95.47% vs 90.91%) and multi-disease diagnosis (F1: 93.49% vs 91.35%), while demonstrating better efficiency. FOCUS automates the image-to-diagnosis pipeline, representing a critical advance towards unmanned ophthalmology with a validated blueprint for autonomous screening to enhance population scale retinal care accessibility and efficiency.

Method: Proposes Pixel-wise Quantitative Thermography Neural Network (PQT-Net) with novel data augmentation converting thermal sequences to 2D stripe images preserving temporal heat diffusion. Uses pre-trained EfficientNetV2-S backbone with custom Residual Regression Head for output refinement.

Result: Achieves minimum Mean Absolute Error of 0.0094 mm and coefficient of determination exceeding 99%, outperforming other deep learning models for defect depth quantification.

Conclusion: PQT-Net demonstrates high precision for robust quantitative defect characterization in additive manufacturing, showing potential for industrial non-destructive testing applications.

Abstract: Defect depth quantification in additively manufactured (AM) components remains a significant challenge for non-destructive testing (NDT). This study proposes a Pixel-wise Quantitative Thermography Neural Network (PQT-Net) to address this challenge for polylactic acid (PLA) parts. A key innovation is a novel data augmentation strategy that reconstructs thermal sequence data into two-dimensional stripe images, preserving the complete temporal evolution of heat diffusion for each pixel. The PQT-Net architecture incorporates a pre-trained EfficientNetV2-S backbone and a custom Residual Regression Head (RRH) with learnable parameters to refine outputs. Comparative experiments demonstrate the superiority of PQT-Net over other deep learning models, achieving a minimum Mean Absolute Error (MAE) of 0.0094 mm and a coefficient of determination (R) exceeding 99%. The high precision of PQT-Net underscores its potential for robust quantitative defect characterization in AM.

Method: Proposes InvLBA (Invisible Latent Backdoor Attack) that operates at the latent feature level rather than pixel level. The method uses latent perturbation for clean-label backdoor attacks in generative data augmentation settings, with theoretical guarantees on generalization of clean accuracy and attack success rates.

Result: Experiments on multiple datasets show InvLBA improves attack success rate by 46.43% on average compared to existing methods, with almost no reduction in clean accuracy and high robustness against state-of-the-art defense methods.

Conclusion: Latent-level attacks are more effective than pixel-level attacks for backdoor attacks in generative data augmentation scenarios, and InvLBA provides a theoretically-grounded, practical approach with strong performance and robustness.

Abstract: With the rapid advancement of image generative models, generative data augmentation has become an effective way to enrich training images, especially when only small-scale datasets are available. At the same time, in practical applications, generative data augmentation can be vulnerable to clean-label backdoor attacks, which aim to bypass human inspection. However, based on theoretical analysis and preliminary experiments, we observe that directly applying existing pixel-level clean-label backdoor attack methods (e.g., COMBAT) to generated images results in low attack success rates. This motivates us to move beyond pixel-level triggers and focus instead on the latent feature level. To this end, we propose InvLBA, an invisible clean-label backdoor attack method for generative data augmentation by latent perturbation. We theoretically prove that the generalization of the clean accuracy and attack success rates of InvLBA can be guaranteed. Experiments on multiple datasets show that our method improves the attack success rate by 46.43% on average, with almost no reduction in clean accuracy and high robustness against SOTA defense methods.

Method: Proposes MedSAM-Agent framework with: 1) Hybrid prompting strategy for expert-curated trajectory generation to internalize human-like decision heuristics, and 2) Two-stage training pipeline integrating multi-turn end-to-end outcome verification with clinical-fidelity process reward design.

Result: Extensive experiments across 6 medical modalities and 21 datasets demonstrate state-of-the-art performance, effectively unifying autonomous medical reasoning with robust, iterative optimization.

Conclusion: MedSAM-Agent successfully bridges the gap in interactive medical segmentation by enabling multi-step autonomous decision-making with efficient interaction strategies and process-level supervision.

Abstract: Medical image segmentation is evolving from task-specific models toward generalizable frameworks. Recent research leverages Multi-modal Large Language Models (MLLMs) as autonomous agents, employing reinforcement learning with verifiable reward (RLVR) to orchestrate specialized tools like the Segment Anything Model (SAM). However, these approaches often rely on single-turn, rigid interaction strategies and lack process-level supervision during training, which hinders their ability to fully exploit the dynamic potential of interactive tools and leads to redundant actions. To bridge this gap, we propose MedSAM-Agent, a framework that reformulates interactive segmentation as a multi-step autonomous decision-making process. First, we introduce a hybrid prompting strategy for expert-curated trajectory generation, enabling the model to internalize human-like decision heuristics and adaptive refinement strategies. Furthermore, we develop a two-stage training pipeline that integrates multi-turn, end-to-end outcome verification with a clinical-fidelity process reward design to promote interaction parsimony and decision efficiency. Extensive experiments across 6 medical modalities and 21 datasets demonstrate that MedSAM-Agent achieves state-of-the-art performance, effectively unifying autonomous medical reasoning with robust, iterative optimization. Code is available \href{https://github.com/CUHK-AIM-Group/MedSAM-Agent}{here}.

Method: Proposes PWAVEP purification framework that: 1) Computes spectral graph wavelet domain saliency and local sparsity scores for each point; 2) Hierarchically eliminates the most salient points (hard-to-recover adversarial outliers); 3) Applies spectral filtering to moderately salient points using graph wavelet transform to attenuate high-frequency coefficients associated with adversarial noise.

Result: Extensive evaluations show PWAVEP achieves superior accuracy and robustness compared to existing approaches, advancing state-of-the-art in 3D point cloud purification. The method effectively suppresses adversarial noise while maintaining point cloud integrity.

Conclusion: PWAVEP provides an effective plug-and-play defense against adversarial attacks on 3D point clouds by leveraging spectral domain analysis and graph wavelet transforms, offering a non-invasive solution that doesn’t require model modifications or expensive training procedures.

Abstract: Recent progress in adversarial attacks on 3D point clouds, particularly in achieving spatial imperceptibility and high attack performance, presents significant challenges for defenders. Current defensive approaches remain cumbersome, often requiring invasive model modifications, expensive training procedures or auxiliary data access. To address these threats, in this paper, we propose a plug-and-play and non-invasive defense mechanism in the spectral domain, grounded in a theoretical and empirical analysis of the relationship between imperceptible perturbations and high-frequency spectral components. Building upon these insights, we introduce a novel purification framework, termed PWAVEP, which begins by computing a spectral graph wavelet domain saliency score and local sparsity score for each point. Guided by these values, PWAVEP adopts a hierarchical strategy, it eliminates the most salient points, which are identified as hardly recoverable adversarial outliers. Simultaneously, it applies a spectral filtering process to a broader set of moderately salient points. This process leverages a graph wavelet transform to attenuate high-frequency coefficients associated with the targeted points, thereby effectively suppressing adversarial noise. Extensive evaluations demonstrate that the proposed PWAVEP achieves superior accuracy and robustness compared to existing approaches, advancing the state-of-the-art in 3D point cloud purification. Code and datasets are available at https://github.com/a772316182/pwavep

Method: Uses token-conditioned diffusion decoder with InfoGAN-style objective (training recognition model to predict tokens from decoded images). Introduces token swapping between images during training to promote compositionality. Applies adversarial flow regularizer to keep unpaired swap generations on natural-image distribution. Proposes two metrics to measure token space compositionality and learnability.

Result: Achieves state-of-the-art performance on image class-conditioned generation. Enables high-level semantic editing through token swapping between images. Improves on proposed compositionality metrics and supports state-of-the-art generators.

Conclusion: CompTok successfully creates compositional visual tokenizers that enable better control over image generation through token manipulation, with applications in semantic editing and improved generation quality.

Abstract: We introduce CompTok, a training framework for learning visual tokenizers whose tokens are enhanced for compositionality. CompTok uses a token-conditioned diffusion decoder. By employing an InfoGAN-style objective, where we train a recognition model to predict the tokens used to condition the diffusion decoder using the decoded images, we enforce the decoder to not ignore any of the tokens. To promote compositional control, besides the original images, CompTok also trains on tokens formed by swapping token subsets between images, enabling more compositional control of the token over the decoder. As the swapped tokens between images do not have ground truth image targets, we apply a manifold constraint via an adversarial flow regularizer to keep unpaired swap generations on the natural-image distribution. The resulting tokenizer not only achieves state-of-the-art performance on image class-conditioned generation, but also demonstrates properties such as swapping tokens between images to achieve high level semantic editing of an image. Additionally, we propose two metrics that measures the landscape of the token space that can be useful to describe not only the compositionality of the tokens, but also how easy to learn the landscape is for a generator to be trained on this space. We show in experiments that CompTok can improve on both of the metrics as well as supporting state-of-the-art generators for class conditioned generation.

Method: Proposes Tiled Prompts framework that generates tile-specific prompts for each latent tile and performs super-resolution under locally text-conditioned posteriors, providing high-information guidance with minimal overhead.

Result: Experiments on high-resolution real-world images and videos show consistent gains in perceptual quality and text alignment, while reducing hallucinations and tile-level artifacts compared to global-prompt baselines.

Conclusion: Tiled Prompts effectively resolves prompt underspecification in diffusion-based super-resolution by providing localized text guidance, improving both image and video super-resolution quality.

Abstract: Text-conditioned diffusion models have advanced image and video super-resolution by using prompts as semantic priors, but modern super-resolution pipelines typically rely on latent tiling to scale to high resolutions, where a single global caption causes prompt underspecification. A coarse global prompt often misses localized details (prompt sparsity) and provides locally irrelevant guidance (prompt misguidance) that can be amplified by classifier-free guidance. We propose Tiled Prompts, a unified framework for image and video super-resolution that generates a tile-specific prompt for each latent tile and performs super-resolution under locally text-conditioned posteriors, providing high-information guidance that resolves prompt underspecification with minimal overhead. Experiments on high resolution real-world images and videos show consistent gains in perceptual quality and text alignment, while reducing hallucinations and tile-level artifacts relative to global-prompt baselines.

Method: Uses multi-view images with optional camera poses/depth, employs state-of-the-art zero-shot 3D instance segmentation for bounding box proposals, and leverages prompt-based segmentation with modern VLMs for reasoning.

Result: Achieves state-of-the-art performance among zero-shot methods on ScanRefer and Nr3D benchmarks.

Conclusion: Demonstrates effective zero-shot 3D visual grounding using only multi-view images, advancing the field toward more practical 3D scene understanding.

Abstract: 3D visual grounding (3DVG) aims to localize objects in a 3D scene based on natural language queries. In this work, we explore zero-shot 3DVG from multi-view images alone, without requiring any geometric supervision or object priors. We introduce Z3D, a universal grounding pipeline that flexibly operates on multi-view images while optionally incorporating camera poses and depth maps. We identify key bottlenecks in prior zero-shot methods causing significant performance degradation and address them with (i) a state-of-the-art zero-shot 3D instance segmentation method to generate high-quality 3D bounding box proposals and (ii) advanced reasoning via prompt-based segmentation, which utilizes full capabilities of modern VLMs. Extensive experiments on the ScanRefer and Nr3D benchmarks demonstrate that our approach achieves state-of-the-art performance among zero-shot methods. Code is available at https://github.com/col14m/z3d .

Method: Proposes a discrete diffusion framework that treats HMER as iterative symbolic refinement through multi-step remasking. Uses symbol-aware tokenization and Random-Masking Mutual Learning to enhance syntactic alignment and robustness to handwriting diversity.

Result: Achieves 5.56% CER and 60.42% EM on MathWriting benchmark, outperforming strong Transformer and commercial baselines. Shows consistent gains on CROHME 2014-2023 datasets.

Conclusion: Discrete diffusion provides a new paradigm for structure-aware visual recognition beyond generative modeling, offering better handling of 2D structural layouts in mathematical expressions without autoregressive limitations.

Abstract: Handwritten Mathematical Expression Recognition (HMER) requires reasoning over diverse symbols and 2D structural layouts, yet autoregressive models struggle with exposure bias and syntactic inconsistency. We present a discrete diffusion framework that reformulates HMER as iterative symbolic refinement instead of sequential generation. Through multi-step remasking, the proposal progressively refines both symbols and structural relations, removing causal dependencies and improving structural consistency. A symbol-aware tokenization and Random-Masking Mutual Learning further enhance syntactic alignment and robustness to handwriting diversity. On the MathWriting benchmark, the proposal achieves 5.56% CER and 60.42% EM, outperforming strong Transformer and commercial baselines. Consistent gains on CROHME 2014–2023 demonstrate that discrete diffusion provides a new paradigm for structure-aware visual recognition beyond generative modeling.

Method: Proposes Multi-Resolution Alignment (MRA) with three modules: 1) Multi-resolution View Transformer projects 2D image features to multi-resolution 3D features with scene-level alignment, 2) Cubic Semantic Anisotropy identifies instance-level semantic significance of voxels, and 3) Critical Distribution Alignment selects critical voxels as anchors and applies circulated loss for feature distribution consistency across resolutions.

Result: The approach mitigates voxel sparsity in camera-based 3D semantic scene completion and improves model performance through auxiliary supervision from multi-resolution alignment.

Conclusion: MRA effectively addresses voxel sparsity in camera-based 3D SSC by exploiting multi-resolution alignment at both scene and instance levels, providing better optimization and performance for autonomous driving perception systems.

Abstract: Camera-based 3D semantic scene completion (SSC) offers a cost-effective solution for assessing the geometric occupancy and semantic labels of each voxel in the surrounding 3D scene with image inputs, providing a voxel-level scene perception foundation for the perception-prediction-planning autonomous driving systems. Although significant progress has been made in existing methods, their optimization rely solely on the supervision from voxel labels and face the challenge of voxel sparsity as a large portion of voxels in autonomous driving scenarios are empty, which limits both optimization efficiency and model performance. To address this issue, we propose a \textit{Multi-Resolution Alignment (MRA)} approach to mitigate voxel sparsity in camera-based 3D semantic scene completion, which exploits the scene and instance level alignment across multi-resolution 3D features as auxiliary supervision. Specifically, we first propose the Multi-resolution View Transformer module, which projects 2D image features into multi-resolution 3D features and aligns them at the scene level through fusing discriminative seed features. Furthermore, we design the Cubic Semantic Anisotropy module to identify the instance-level semantic significance of each voxel, accounting for the semantic differences of a specific voxel against its neighboring voxels within a cubic area. Finally, we devise a Critical Distribution Alignment module, which selects critical voxels as instance-level anchors with the guidance of cubic semantic anisotropy, and applies a circulated loss for auxiliary supervision on the critical feature distribution consistency across different resolutions. The code is available at https://github.com/PKU-ICST-MIPL/MRA_TIP.

Method: Proposes SLIM-Diff with: (1) single shared-bottleneck U-Net enforcing tight coupling between anatomy and lesion geometry using 2-channel image+mask representation, (2) loss-geometry tuning via tunable Lp objective, comparing x0-prediction vs ε-prediction with different Lp norms.

Result: x0-prediction is consistently strongest for joint synthesis; fractional sub-quadratic penalties (L1.5) improve image fidelity while L2 better preserves lesion mask morphology.

Conclusion: SLIM-Diff provides stable joint synthesis of medical images with subtle lesions, with optimal configurations identified for balancing image fidelity and lesion morphology preservation.

Abstract: Focal cortical dysplasia (FCD) lesions in epilepsy FLAIR MRI are subtle and scarce, making joint image–mask generative modeling prone to instability and memorization. We propose SLIM-Diff, a compact joint diffusion model whose main contributions are (i) a single shared-bottleneck U-Net that enforces tight coupling between anatomy and lesion geometry from a 2-channel image+mask representation, and (ii) loss-geometry tuning via a tunable $L_p$ objective. As an internal baseline, we include the canonical DDPM-style objective ($ε$-prediction with $L_2$ loss) and isolate the effect of prediction parameterization and $L_p$ geometry under a matched setup. Experiments show that $x_0$-prediction is consistently the strongest choice for joint synthesis, and that fractional sub-quadratic penalties ($L_{1.5}$) improve image fidelity while $L_2$ better preserves lesion mask morphology. Our code and model weights are available in https://github.com/MarioPasc/slim-diff

Method: DiM formulates watermarking as a dimension-aware mapping problem where watermark information is modeled as payloads of different dimensionalities. The framework analyzes how dimensional configuration of embedding and extraction determines watermarking behavior, with same-dimensional mappings preserving payload structure and cross-dimensional mappings enabling spatial/spatiotemporal localization.

Result: Experiments in the video domain demonstrate that varying only embedding and extraction dimensions (without architectural changes) leads to different watermarking capabilities including spatiotemporal tamper localization, local embedding control, and recovery of temporal order under frame disruptions.

Conclusion: DiM provides a unified framework for understanding and designing watermarking methods through dimension-aware mapping, revealing that dimensional configuration is a key determinant of watermarking behavior and capabilities.

Abstract: Deep watermarking methods often share similar encoder-decoder architectures, yet differ substantially in their functional behaviors. We propose DiM, a new multi-dimensional watermarking framework that formulates watermarking as a dimension-aware mapping problem, thereby unifying existing watermarking methods at the functional level. Under DiM, watermark information is modeled as payloads of different dimensionalities, including one-dimensional binary messages, two-dimensional spatial masks, and three-dimensional spatiotemporal structures. We find that the dimensional configuration of embedding and extraction largely determines the resulting watermarking behavior. Same-dimensional mappings preserve payload structure and support fine-grained control, while cross-dimensional mappings enable spatial or spatiotemporal localization. We instantiate DiM in the video domain, where spatiotemporal representations enable a broader set of dimension mappings. Experiments demonstrate that varying only the embedding and extraction dimensions, without architectural changes, leads to different watermarking capabilities, including spatiotemporal tamper localization, local embedding control, and recovery of temporal order under frame disruptions.

Method: Proposes C3PO with two components: 1) Chain-of-Thought Compression - selectively filters redundant thinking tokens for compact, signal-efficient CoT representations; 2) Contrastive Preference Optimization - uses reasoning-enhanced preference tuning with AI feedback and a hallucination-inducing mechanism to create informative negative signals for contrastive correction.

Result: Demonstrates consistent hallucination reduction across diverse multimodal reasoning models and benchmarks, with theoretical justification for effectiveness.

Conclusion: C3PO effectively addresses hallucination issues in multimodal reasoning models by improving reasoning trace quality and reducing reliance on language priors through compression and contrastive optimization techniques.

Abstract: While multimodal reasoning models (MLRMs) have exhibited impressive capabilities, they remain prone to hallucinations, and effective solutions are still underexplored. In this paper, we experimentally analyze the hallucination cause and propose C3PO, a training-based mitigation framework comprising \textbf{C}hain-of-Thought \textbf{C}ompression and \textbf{C}ontrastive \textbf{P}reference \textbf{O}ptimization. Firstly, we identify that introducing reasoning mechanisms exacerbates models’ reliance on language priors while overlooking visual inputs, which can produce CoTs with reduced visual cues but redundant text tokens. To this end, we propose to selectively filter redundant thinking tokens for a more compact and signal-efficient CoT representation that preserves task-relevant information while suppressing noise. In addition, we observe that the quality of the reasoning trace largely determines whether hallucination emerges in subsequent responses. To leverage this insight, we introduce a reasoning-enhanced preference tuning scheme that constructs training pairs using high-quality AI feedback. We further design a multimodal hallucination-inducing mechanism that elicits models’ inherent hallucination patterns via carefully crafted inducers, yielding informative negative signals for contrastive correction. We provide theoretical justification for the effectiveness and demonstrate consistent hallucination reduction across diverse MLRMs and benchmarks.

Method: Synergistic Representation Learning (SRL) establishes a virtuous cycle where encoder and decoder mutually refine each other - encoder’s sharpness deblurs semantic boundaries in decoder output, while decoder’s spatial consistency denoises encoder’s features. Includes warm-up phase with slot regularization to initially allocate distinct entities per slot.

Result: SRL achieves state-of-the-art results on video object-centric learning benchmarks by bridging the representational gap between encoder and decoder.

Conclusion: The proposed synergistic learning approach successfully breaks the vicious cycle in unsupervised object-centric models, enabling mutual refinement between encoder and decoder representations for improved scene decomposition.

Abstract: Unsupervised object-centric learning models, particularly slot-based architectures, have shown great promise in decomposing complex scenes. However, their reliance on reconstruction-based training creates a fundamental conflict between the sharp, high-frequency attention maps of the encoder and the spatially consistent but blurry reconstruction maps of the decoder. We identify that this discrepancy gives rise to a vicious cycle: the noisy feature map from the encoder forces the decoder to average over possibilities and produce even blurrier outputs, while the gradient computed from blurry reconstruction maps lacks high-frequency details necessary to supervise encoder features. To break this cycle, we introduce Synergistic Representation Learning (SRL) that establishes a virtuous cycle where the encoder and decoder mutually refine one another. SRL leverages the encoder’s sharpness to deblur the semantic boundary within the decoder output, while exploiting the decoder’s spatial consistency to denoise the encoder’s features. This mutual refinement process is stabilized by a warm-up phase with a slot regularization objective that initially allocates distinct entities per slot. By bridging the representational gap between the encoder and decoder, SRL achieves state-of-the-art results on video object-centric learning benchmarks. Codes are available at https://github.com/hynnsk/SRL.

Method: Incorporates hypernetworks into single- and multi-concept LoRA training; hypernetwork dynamically generates adaptive LoRA weights based on CLIP embeddings during inference for context-aware unlearning; compatible with Stable Diffusion and flow-based text-to-image models.

Result: Demonstrates stable training behavior and effective concept control across challenging tasks including object erasure, celebrity erasure, and explicit content removal; shows effectiveness and versatility in concept removal.

Conclusion: UnHype provides a scalable and adaptive framework for machine unlearning in diffusion models, addressing key limitations of existing LoRA-based approaches through hypernetwork integration.

Abstract: Recent advances in large-scale diffusion models have intensified concerns about their potential misuse, particularly in generating realistic yet harmful or socially disruptive content. This challenge has spurred growing interest in effective machine unlearning, the process of selectively removing specific knowledge or concepts from a model without compromising its overall generative capabilities. Among various approaches, Low-Rank Adaptation (LoRA) has emerged as an effective and efficient method for fine-tuning models toward targeted unlearning. However, LoRA-based methods often exhibit limited adaptability to concept semantics and struggle to balance removing closely related concepts with maintaining generalization across broader meanings. Moreover, these methods face scalability challenges when multiple concepts must be erased simultaneously. To address these limitations, we introduce UnHype, a framework that incorporates hypernetworks into single- and multi-concept LoRA training. The proposed architecture can be directly plugged into Stable Diffusion as well as modern flow-based text-to-image models, where it demonstrates stable training behavior and effective concept control. During inference, the hypernetwork dynamically generates adaptive LoRA weights based on the CLIP embedding, enabling more context-aware, scalable unlearning. We evaluate UnHype across several challenging tasks, including object erasure, celebrity erasure, and explicit content removal, demonstrating its effectiveness and versatility. Repository: https://github.com/gmum/UnHype.

Method: Multi-agent framework with Teacher agent generating parameterized Python scripts with reflective feedback, Solver agent optimizing reasoning through preference learning, and Generator learning image generation from accumulated “image-code-instruction” triplets.

Result: Socratic-Solver achieves 49.11 on six benchmarks using only 1/4 of baseline data, surpassing baselines by 2.43 points. Socratic-Generator achieves 42.4% on GenExam, new SOTA for open-source models.

Conclusion: Socratic-Geo demonstrates effective autonomous data synthesis for geometric reasoning in MLLMs, significantly improving both reasoning and generation capabilities with minimal seed data.

Abstract: Multimodal Large Language Models (MLLMs) have significantly advanced vision-language understanding. However, even state-of-the-art models struggle with geometric reasoning, revealing a critical bottleneck: the extreme scarcity of high-quality image-text pairs. Human annotation is prohibitively expensive, while automated methods fail to ensure fidelity and training effectiveness. Existing approaches either passively adapt to available images or employ inefficient random exploration with filtering, decoupling generation from learning needs. We propose Socratic-Geo, a fully autonomous framework that dynamically couples data synthesis with model learning through multi-agent interaction. The Teacher agent generates parameterized Python scripts with reflective feedback (Reflect for solvability, RePI for visual validity), ensuring image-text pair purity. The Solver agent optimizes reasoning through preference learning, with failure paths guiding Teacher’s targeted augmentation. Independently, the Generator learns image generation capabilities on accumulated “image-code-instruction” triplets, distilling programmatic drawing intelligence into visual generation. Starting from only 108 seed problems, Socratic-Solver achieves 49.11 on six benchmarks using one-quarter of baseline data, surpassing strong baselines by 2.43 points. Socratic-Generator achieves 42.4% on GenExam, establishing new state-of-the-art for open-source models, surpassing Seedream-4.0 (39.8%) and approaching Gemini-2.5-Flash-Image (43.1%).

Method: Proposes ConsistentRFT framework with two key components: 1) Dynamic Granularity Rollout (DGR) that balances exploration between global semantics and local details by dynamically scheduling noise sources, and 2) Consistent Policy Gradient Optimization (CPGO) that preserves model consistency by aligning current policy with a stable prior.

Result: Significantly reduces visual hallucinations: 49% reduction for low-level and 38% for high-level perceptual hallucinations. Outperforms other RFT methods on out-of-domain metrics, showing 5.1% improvement over FLUX1.dev (vs baseline’s -0.4% decrease).

Conclusion: ConsistentRFT effectively mitigates visual hallucinations in flow-based models by addressing exploration-exploitation tradeoffs and preserving model consistency, leading to better alignment and generation quality.

Abstract: Reinforcement Fine-Tuning (RFT) on flow-based models is crucial for preference alignment. However, they often introduce visual hallucinations like over-optimized details and semantic misalignment. This work preliminarily explores why visual hallucinations arise and how to reduce them. We first investigate RFT methods from a unified perspective, and reveal the core problems stemming from two aspects, exploration and exploitation: (1) limited exploration during stochastic differential equation (SDE) rollouts, leading to an over-emphasis on local details at the expense of global semantics, and (2) trajectory imitation process inherent in policy gradient methods, distorting the model’s foundational vector field and its cross-step consistency. Building on this, we propose ConsistentRFT, a general framework to mitigate these hallucinations. Specifically, we design a Dynamic Granularity Rollout (DGR) mechanism to balance exploration between global semantics and local details by dynamically scheduling different noise sources. We then introduce a Consistent Policy Gradient Optimization (CPGO) that preserves the model’s consistency by aligning the current policy with a more stable prior. Extensive experiments demonstrate that ConsistentRFT significantly mitigates visual hallucinations, achieving average reductions of 49% for low-level and 38% for high-level perceptual hallucinations. Furthermore, ConsistentRFT outperforms other RFT methods on out-of-domain metrics, showing an improvement of 5.1% (v.s. the baseline’s decrease of -0.4%) over FLUX1.dev. This is \href{https://xiaofeng-tan.github.io/projects/ConsistentRFT}{Project Page}.

Method: Hierarchical Concept-to-Appearance Guidance (CAG) with two key components: 1) VAE dropout training that randomly omits reference VAE features to encourage reliance on robust semantic signals from a Visual Language Model (VLM), and 2) correspondence-aware masked attention module in Diffusion Transformer that restricts each text token to attend only to matched reference regions.

Result: Extensive experiments demonstrate state-of-the-art performance on multi-subject image generation, substantially improving prompt following and subject consistency.

Conclusion: The proposed CAG framework provides explicit, structured supervision from high-level concepts to fine-grained appearances, achieving better identity preservation and compositional control in multi-subject image generation.

Abstract: Multi-subject image generation aims to synthesize images that faithfully preserve the identities of multiple reference subjects while following textual instructions. However, existing methods often suffer from identity inconsistency and limited compositional control, as they rely on diffusion models to implicitly associate text prompts with reference images. In this work, we propose Hierarchical Concept-to-Appearance Guidance (CAG), a framework that provides explicit, structured supervision from high-level concepts to fine-grained appearances. At the conceptual level, we introduce a VAE dropout training strategy that randomly omits reference VAE features, encouraging the model to rely more on robust semantic signals from a Visual Language Model (VLM) and thereby promoting consistent concept-level generation in the absence of complete appearance cues. At the appearance level, we integrate the VLM-derived correspondences into a correspondence-aware masked attention module within the Diffusion Transformer (DiT). This module restricts each text token to attend only to its matched reference regions, ensuring precise attribute binding and reliable multi-subject composition. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the multi-subject image generation, substantially improving prompt following and subject consistency.

Method: Proposes CoViP framework that treats personalized image captioning as a core task. Uses reinforcement-learning-based post-training and caption-augmented generation to improve contextualized visual personalization. Introduces diagnostic evaluations to ensure models truly leverage visual context rather than textual shortcuts.

Result: Extensive experiments show existing open-source and proprietary VLMs have substantial limitations in contextualized visual personalization. CoViP improves personalized image captioning and yields holistic gains across downstream personalization tasks.

Conclusion: CoViP represents a crucial stage for enabling robust and generalizable contextualized visual personalization in vision-language models, addressing the gap in generating truly personalized responses based on users’ visual experiences.

Abstract: Despite recent progress in vision-language models (VLMs), existing approaches often fail to generate personalized responses based on the user’s specific experiences, as they lack the ability to associate visual inputs with a user’s accumulated visual-textual context. We newly formalize this challenge as contextualized visual personalization, which requires the visual recognition and textual retrieval of personalized visual experiences by VLMs when interpreting new images. To address this issue, we propose CoViP, a unified framework that treats personalized image captioning as a core task for contextualized visual personalization and improves this capability through reinforcement-learning-based post-training and caption-augmented generation. We further introduce diagnostic evaluations that explicitly rule out textual shortcut solutions and verify whether VLMs truly leverage visual context. Extensive experiments demonstrate that existing open-source and proprietary VLMs exhibit substantial limitations, while CoViP not only improves personalized image captioning but also yields holistic gains across downstream personalization tasks. These results highlight CoViP as a crucial stage for enabling robust and generalizable contextualized visual personalization.

Method: Proposes InteractAvatar with dual-stream framework: 1) Perception and Interaction Module (PIM) for text-aligned interaction motion generation using detection for environmental perception, 2) Audio-Interaction Aware Generation Module (AIM) for synthesizing vivid talking avatars performing object interactions, and 3) motion-to-video aligner enabling parallel co-generation of motions and videos.

Result: Establishes GroundedInter benchmark for evaluating GHOI video generation. Extensive experiments demonstrate effectiveness in generating grounded human-object interactions for talking avatars, addressing the control-quality dilemma.

Conclusion: InteractAvatar successfully generates talking avatars performing grounded human-object interactions by decoupling perception/planning from video synthesis, enabling environmental perception and coordinated audio-visual generation.

Abstract: Generating talking avatars is a fundamental task in video generation. Although existing methods can generate full-body talking avatars with simple human motion, extending this task to grounded human-object interaction (GHOI) remains an open challenge, requiring the avatar to perform text-aligned interactions with surrounding objects. This challenge stems from the need for environmental perception and the control-quality dilemma in GHOI generation. To address this, we propose a novel dual-stream framework, InteractAvatar, which decouples perception and planning from video synthesis for grounded human-object interaction. Leveraging detection to enhance environmental perception, we introduce a Perception and Interaction Module (PIM) to generate text-aligned interaction motions. Additionally, an Audio-Interaction Aware Generation Module (AIM) is proposed to synthesize vivid talking avatars performing object interactions. With a specially designed motion-to-video aligner, PIM and AIM share a similar network structure and enable parallel co-generation of motions and plausible videos, effectively mitigating the control-quality dilemma. Finally, we establish a benchmark, GroundedInter, for evaluating GHOI video generation. Extensive experiments and comparisons demonstrate the effectiveness of our method in generating grounded human-object interactions for talking avatars. Project page: https://interactavatar.github.io

Method: InlierQ computes gradient-aware volume saliency scores for each activation volume, classifies volumes as inliers or anomalies, and fits a posterior distribution over these scores using Expectation-Maximization (EM) algorithm. This suppresses anomalies while preserving informative features. The method is label-free, drop-in, and requires only 64 calibration samples.

Result: Experiments on COCO and nuScenes benchmarks show consistent reductions in quantization error for camera-based (2D and 3D) and LiDAR-based (3D) object detection across different quantization settings.

Conclusion: InlierQ effectively addresses the challenge of task-irrelevant anomalies in object detection quantization by separating anomalies from informative inliers, enabling more efficient bit allocation and better preservation of critical features for deployment on resource-constrained devices.

Abstract: Object detection is pivotal in computer vision, yet its immense computational demands make deployment slow and power-hungry, motivating quantization. However, task-irrelevant morphologies such as background clutter and sensor noise induce redundant activations (or anomalies). These anomalies expand activation ranges and skew activation distributions toward task-irrelevant responses, complicating bit allocation and weakening the preservation of informative features. Without a clear criterion to distinguish anomalies, suppressing them can inadvertently discard useful information. To address this, we present InlierQ, an inlier-centric post-training quantization approach that separates anomalies from informative inliers. InlierQ computes gradient-aware volume saliency scores, classifies each volume as an inlier or anomaly, and fits a posterior distribution over these scores using the Expectation-Maximization (EM) algorithm. This design suppresses anomalies while preserving informative features. InlierQ is label-free, drop-in, and requires only 64 calibration samples. Experiments on the COCO and nuScenes benchmarks show consistent reductions in quantization error for camera-based (2D and 3D) and LiDAR-based (3D) object detection.

Method: Two-stage framework: 1) DiSCo disentangles structural abstraction from semantic grounding during multimodal alignment; 2) Table-GLS performs global-to-local structure-guided reasoning via structured exploration and evidence-grounded inference.

Result: Extensive experiments across diverse benchmarks show the framework efficiently enhances LVLM’s table understanding and reasoning capabilities, with strong generalization to unseen table structures.

Conclusion: The proposed framework enables LVLMs to adapt to table reasoning with minimal annotation and no external tools, addressing key challenges in table image understanding.

Abstract: Reasoning over table images remains challenging for Large Vision-Language Models (LVLMs) due to complex layouts and tightly coupled structure-content information. Existing solutions often depend on expensive supervised training, reinforcement learning, or external tools, limiting efficiency and scalability. This work addresses a key question: how to adapt LVLMs to table reasoning with minimal annotation and no external tools? Specifically, we first introduce DiSCo, a Disentangled Structure-Content alignment framework that explicitly separates structural abstraction from semantic grounding during multimodal alignment, efficiently adapting LVLMs to tables structures. Building on DiSCo, we further present Table-GLS, a Global-to-Local Structure-guided reasoning framework that performs table reasoning via structured exploration and evidence-grounded inference. Extensive experiments across diverse benchmarks demonstrate that our framework efficiently enhances LVLM’s table understanding and reasoning capabilities, particularly generalizing to unseen table structures.

Method: Introduces a unified normalized convex fusion framework with lightweight gates to systematically organize multi-layer LLM hidden states via three strategies: time-wise fusion (conditioning varies with diffusion timestep), depth-wise fusion (conditioning varies with DiT network depth), and joint fusion (both).

Result: Depth-wise semantic routing emerges as superior, consistently improving text-image alignment and compositional generation (e.g., +9.97 on GenAI-Bench Counting task). Time-wise fusion degrades visual fidelity due to train-inference trajectory mismatch under classifier-free guidance.

Conclusion: Depth-wise routing is an effective baseline for dynamic text conditioning in DiT models. Time-dependent conditioning requires trajectory-aware signals to avoid semantic mistiming during inference.

Abstract: Recent DiT-based text-to-image models increasingly adopt LLMs as text encoders, yet text conditioning remains largely static and often utilizes only a single LLM layer, despite pronounced semantic hierarchy across LLM layers and non-stationary denoising dynamics over both diffusion time and network depth. To better match the dynamic process of DiT generation and thereby enhance the diffusion model’s generative capability, we introduce a unified normalized convex fusion framework equipped with lightweight gates to systematically organize multi-layer LLM hidden states via time-wise, depth-wise, and joint fusion. Experiments establish Depth-wise Semantic Routing as the superior conditioning strategy, consistently improving text-image alignment and compositional generation (e.g., +9.97 on the GenAI-Bench Counting task). Conversely, we find that purely time-wise fusion can paradoxically degrade visual generation fidelity. We attribute this to a train-inference trajectory mismatch: under classifier-free guidance, nominal timesteps fail to track the effective SNR, causing semantically mistimed feature injection during inference. Overall, our results position depth-wise routing as a strong and effective baseline and highlight the critical need for trajectory-aware signals to enable robust time-dependent conditioning.

Method: Proposes LogiCls framework that decomposes complex logical constraints into verifiable subqueries, uses data-centric instruction synthesis with chain-of-thought supervision, couples precise grounding annotations with image-text augmentations, and employs difficulty-aware resampling to emphasize challenging subqueries.

Result: Extensive experiments show LogiCls delivers robust, interpretable, and accurate industrial logical anomaly classification, providing both predicted violation categories and evidence trails.

Conclusion: The paper introduces Logical Anomaly Classification (LAC) as a unified task for anomaly detection and violation classification, with LogiCls demonstrating effective vision-language reasoning for industrial quality assurance.

Abstract: Logical anomalies are violations of predefined constraints on object quantity, spatial layout, and compositional relationships in industrial images. While prior work largely treats anomaly detection as a binary decision, such formulations cannot indicate which logical rule is broken and therefore offer limited value for quality assurance. We introduce Logical Anomaly Classification (LAC), a task that unifies anomaly detection and fine-grained violation classification in a single inference step. To tackle LAC, we propose LogiCls, a vision-language framework that decomposes complex logical constraints into a sequence of verifiable subqueries. We further present a data-centric instruction synthesis pipeline that generates chain-of-thought (CoT) supervision for these subqueries, coupling precise grounding annotations with diverse image-text augmentations to adapt vision language models (VLMs) to logic-sensitive reasoning. Training is stabilized by a difficulty-aware resampling strategy that emphasizes challenging subqueries and long tail constraint types. Extensive experiments demonstrate that LogiCls delivers robust, interpretable, and accurate industrial logical anomaly classification, providing both the predicted violation categories and their evidence trails.

Method: Combines autoregressive next-token prediction for 3D understanding with continuous diffusion for 3D generation. Uses lightweight transformer to bridge feature space of LLMs with conditional space of 3D diffusion models, enabling cross-modal information exchange while preserving standalone model priors.

Result: Achieves state-of-the-art performance across diverse 3D understanding and generation benchmarks, and excels in 3D editing tasks. Demonstrates effective information interaction between generation and understanding with minimal compromise to inherent capabilities.

Conclusion: Unified AR+diffusion models represent a promising direction for building general-purpose 3D intelligence, showing that effective information interaction between understanding and generation can be achieved without forcing a single paradigm.

Abstract: The rapid progress of large multimodal models has inspired efforts toward unified frameworks that couple understanding and generation. While such paradigms have shown remarkable success in 2D, extending them to 3D remains largely underexplored. Existing attempts to unify 3D tasks under a single autoregressive (AR) paradigm lead to significant performance degradation due to forced signal quantization and prohibitive training cost. Our key insight is that the essential challenge lies not in enforcing a unified autoregressive paradigm, but in enabling effective information interaction between generation and understanding while minimally compromising their inherent capabilities and leveraging pretrained models to reduce training cost. Guided by this perspective, we present the first unified framework for 3D understanding and generation that combines autoregression with diffusion. Specifically, we adopt an autoregressive next-token prediction paradigm for 3D understanding, and a continuous diffusion paradigm for 3D generation. A lightweight transformer bridges the feature space of large language models and the conditional space of 3D diffusion models, enabling effective cross-modal information exchange while preserving the priors learned by standalone models. Extensive experiments demonstrate that our framework achieves state-of-the-art performance across diverse 3D understanding and generation benchmarks, while also excelling in 3D editing tasks. These results highlight the potential of unified AR+diffusion models as a promising direction for building more general-purpose 3D intelligence.

Method: Proposes Constrained Dynamic Gaussian Splatting (CDGS) with: 1) Differentiable budget controller using multi-modal unified importance score (geometric, motion, perceptual cues), 2) Decoupled optimization of static/dynamic elements with adaptive capacity allocation, 3) Three-phase training strategy, 4) Dual-mode hybrid compression scheme.

Result: CDGS strictly adheres to hardware constraints (error < 2%), pushes Pareto frontier of rate-distortion performance, achieves over 3x compression compared to state-of-the-art methods, and delivers optimal rendering quality under varying capacity limits.

Conclusion: CDGS successfully addresses the deployment dilemma of dynamic Gaussian splatting by formulating it as a budget-constrained optimization problem, enabling high-quality 4D reconstruction within strict hardware constraints.

Abstract: While Dynamic Gaussian Splatting enables high-fidelity 4D reconstruction, its deployment is severely hindered by a fundamental dilemma: unconstrained densification leads to excessive memory consumption incompatible with edge devices, whereas heuristic pruning fails to achieve optimal rendering quality under preset Gaussian budgets. In this work, we propose Constrained Dynamic Gaussian Splatting (CDGS), a novel framework that formulates dynamic scene reconstruction as a budget-constrained optimization problem to enforce a strict, user-defined Gaussian budget during training. Our key insight is to introduce a differentiable budget controller as the core optimization driver. Guided by a multi-modal unified importance score, this controller fuses geometric, motion, and perceptual cues for precise capacity regulation. To maximize the utility of this fixed budget, we further decouple the optimization of static and dynamic elements, employing an adaptive allocation mechanism that dynamically distributes capacity based on motion complexity. Furthermore, we implement a three-phase training strategy to seamlessly integrate these constraints, ensuring precise adherence to the target count. Coupled with a dual-mode hybrid compression scheme, CDGS not only strictly adheres to hardware constraints (error < 2%}) but also pushes the Pareto frontier of rate-distortion performance. Extensive experiments demonstrate that CDGS delivers optimal rendering quality under varying capacity limits, achieving over 3x compression compared to state-of-the-art methods.

Method: Evaluated four inter-image DA strategies on two organ segmentation datasets: CutMix (randomly cutting and pasting patches), CarveMix (carving and pasting irregular regions), ObjectAug (object-level augmentation), and AnatoMix (anatomy-aware mixing). Compared performance against state-of-the-art nnUNet baseline without DA.

Result: CutMix improved average dice score by 4.9, CarveMix by 2.0, and AnatoMix by 1.9 compared to nnUNet without DA. Performance further improved when combined with traditional DA strategies. CutMix proved robust and effective despite producing intuitively ‘wrong’ images.

Conclusion: Inter-image data augmentation strategies, particularly CutMix, significantly improve multi-organ segmentation performance with limited data. These methods create more diverse training samples by combining content from different individuals, addressing data scarcity in medical imaging.

Abstract: Multi-organ segmentation is a widely applied clinical routine and automated organ segmentation tools dramatically improve the pipeline of the radiologists. Recently, deep learning (DL) based segmentation models have shown the capacity to accomplish such a task. However, the training of the segmentation networks requires large amount of data with manual annotations, which is a major concern due to the data scarcity from clinic. Working with limited data is still common for researches on novel imaging modalities. To enhance the effectiveness of DL models trained with limited data, data augmentation (DA) is a crucial regularization technique. Traditional DA (TDA) strategies focus on basic intra-image operations, i.e. generating images with different orientations and intensity distributions. In contrast, the interimage and object-level DA operations are able to create new images from separate individuals. However, such DA strategies are not well explored on the task of multi-organ segmentation. In this paper, we investigated four possible inter-image DA strategies: CutMix, CarveMix, ObjectAug and AnatoMix, on two organ segmentation datasets. The result shows that CutMix, CarveMix and AnatoMix can improve the average dice score by 4.9, 2.0 and 1.9, compared with the state-of-the-art nnUNet without DA strategies. These results can be further improved by adding TDA strategies. It is revealed in our experiments that Cut-Mix is a robust but simple DA strategy to drive up the segmentation performance for multi-organ segmentation, even when CutMix produces intuitively ‘wrong’ images. Our implementation is publicly available for future benchmarks.

Method: SlowFocus mechanism: 1) Identifies query-related temporal segments, 2) Performs dense sampling on these segments for local high-frequency features, 3) Uses multi-frequency mixing attention to aggregate local high-frequency details with global low-frequency contexts, 4) Includes training strategies to enhance temporal grounding and detailed reasoning.

Result: Superior performance on existing public video understanding benchmarks and the proposed FineAction-CGR benchmark, demonstrating enhanced fine-grained temporal understanding capabilities.

Conclusion: SlowFocus effectively addresses the trade-off between frame-level quality and temporal coverage in Video LLMs, enabling better fine-grained video understanding through adaptive temporal focusing and multi-frequency feature aggregation.

Abstract: Large language models (LLMs) have demonstrated exceptional capabilities in text understanding, which has paved the way for their expansion into video LLMs (Vid-LLMs) to analyze video data. However, current Vid-LLMs struggle to simultaneously retain high-quality frame-level semantic information (i.e., a sufficient number of tokens per frame) and comprehensive video-level temporal information (i.e., an adequate number of sampled frames per video). This limitation hinders the advancement of Vid-LLMs towards fine-grained video understanding. To address this issue, we introduce the SlowFocus mechanism, which significantly enhances the equivalent sampling frequency without compromising the quality of frame-level visual tokens. SlowFocus begins by identifying the query-related temporal segment based on the posed question, then performs dense sampling on this segment to extract local high-frequency features. A multi-frequency mixing attention module is further leveraged to aggregate these local high-frequency details with global low-frequency contexts for enhanced temporal comprehension. Additionally, to tailor Vid-LLMs to this innovative mechanism, we introduce a set of training strategies aimed at bolstering both temporal grounding and detailed temporal reasoning capabilities. Furthermore, we establish FineAction-CGR, a benchmark specifically devised to assess the ability of Vid-LLMs to process fine-grained temporal understanding tasks. Comprehensive experiments demonstrate the superiority of our mechanism across both existing public video understanding benchmarks and our proposed FineAction-CGR.

Method: Proposes DeepTopo-Net with Water-Conditioned Adaptive Perceptor (WCAP) using Riemannian metric tensors to dynamically deform convolutional sampling fields, and Abyssal-Topology Refinement Module (ATRM) to maintain structural connectivity through skeletal priors. Also introduces GBU-UCOD, the first high-resolution (2K) benchmark for marine vertical zonation.

Result: Extensive experiments on MAS3K, RMAS, and the proposed GBU-UCOD datasets demonstrate state-of-the-art performance, particularly in preserving morphological integrity of complex underwater patterns.

Conclusion: DeepTopo-Net effectively addresses challenges in underwater camouflaged object detection through topology-aware modeling and frequency-decoupled perception, with the new GBU-UCOD benchmark filling data gaps for hadal and abyssal zones.

Abstract: Underwater Camouflaged Object Detection (UCOD) is a challenging task due to the extreme visual similarity between targets and backgrounds across varying marine depths. Existing methods often struggle with topological fragmentation of slender creatures in the deep sea and the subtle feature extraction of transparent organisms. In this paper, we propose DeepTopo-Net, a novel framework that integrates topology-aware modeling with frequency-decoupled perception. To address physical degradation, we design the Water-Conditioned Adaptive Perceptor (WCAP), which employs Riemannian metric tensors to dynamically deform convolutional sampling fields. Furthermore, the Abyssal-Topology Refinement Module (ATRM) is developed to maintain the structural connectivity of spindly targets through skeletal priors. Specifically, we first introduce GBU-UCOD, the first high-resolution (2K) benchmark tailored for marine vertical zonation, filling the data gap for hadal and abyssal zones. Extensive experiments on MAS3K, RMAS, and our proposed GBU-UCOD datasets demonstrate that DeepTopo-Net achieves state-of-the-art performance, particularly in preserving the morphological integrity of complex underwater patterns. The datasets and codes will be released at https://github.com/Wuwenji18/GBU-UCOD.

Method: Uses TIPS VLM (trained with spatially aware objectives) as backbone, addresses distributional gap between global and local features with decoupled prompts (fixed for image-level detection, learnable for pixel-level localization), and injects local evidence into global score.

Result: Improves image-level performance by 1.1-3.9% and pixel-level by 1.5-6.9% across seven industrial datasets, delivering strong generalization with lean architecture without CLIP-specific tricks.

Conclusion: TIPS-based pipeline effectively addresses CLIP’s limitations for zero-shot anomaly detection, achieving significant improvements in both detection and localization with simpler architecture.

Abstract: Anomaly detection identifies departures from expected behavior in safety-critical settings. When target-domain normal data are unavailable, zero-shot anomaly detection (ZSAD) leverages vision-language models (VLMs). However, CLIP’s coarse image-text alignment limits both localization and detection due to (i) spatial misalignment and (ii) weak sensitivity to fine-grained anomalies; prior work compensates with complex auxiliary modules yet largely overlooks the choice of backbone. We revisit the backbone and use TIPS-a VLM trained with spatially aware objectives. While TIPS alleviates CLIP’s issues, it exposes a distributional gap between global and local features. We address this with decoupled prompts-fixed for image-level detection and learnable for pixel-level localization-and by injecting local evidence into the global score. Without CLIP-specific tricks, our TIPS-based pipeline improves image-level performance by 1.1-3.9% and pixel-level by 1.5-6.9% across seven industrial datasets, delivering strong generalization with a lean architecture. Code is available at github.com/AlirezaSalehy/Tipsomaly.

Method: Proposes Refer-Agent, a collaborative multi-agent system with alternating reasoning-reflection mechanisms. Features: 1) Coarse-to-Fine frame selection for diversity and textual relevance, 2) Dynamic Focus Layout for adaptive visual attention, 3) Chain-of-Reflection mechanism with Questioner-Responder pair for self-verification and feedback generation.

Result: Extensive experiments on five challenging benchmarks show Refer-Agent significantly outperforms state-of-the-art methods, including both SFT-based models and zero-shot approaches. The system is flexible and enables fast integration of new MLLMs without additional fine-tuning.

Conclusion: Refer-Agent provides an effective zero-shot solution for RVOS that overcomes limitations of supervised fine-tuning while achieving superior performance through collaborative multi-agent reasoning and reflection mechanisms.

Abstract: Referring Video Object Segmentation (RVOS) aims to segment objects in videos based on textual queries. Current methods mainly rely on large-scale supervised fine-tuning (SFT) of Multi-modal Large Language Models (MLLMs). However, this paradigm suffers from heavy data dependence and limited scalability against the rapid evolution of MLLMs. Although recent zero-shot approaches offer a flexible alternative, their performance remains significantly behind SFT-based methods, due to the straightforward workflow designs. To address these limitations, we propose \textbf{Refer-Agent}, a collaborative multi-agent system with alternating reasoning-reflection mechanisms. This system decomposes RVOS into step-by-step reasoning process. During reasoning, we introduce a Coarse-to-Fine frame selection strategy to ensure the frame diversity and textual relevance, along with a Dynamic Focus Layout that adaptively adjusts the agent’s visual focus. Furthermore, we propose a Chain-of-Reflection mechanism, which employs a Questioner-Responder pair to generate a self-reflection chain, enabling the system to verify intermediate results and generates feedback for next-round reasoning refinement. Extensive experiments on five challenging benchmarks demonstrate that Refer-Agent significantly outperforms state-of-the-art methods, including both SFT-based models and zero-shot approaches. Moreover, Refer-Agent is flexible and enables fast integration of new MLLMs without any additional fine-tuning costs. Code will be released.

Method: Developed an open-source library with modular JEPA implementations that learn representations by predicting in embedding space rather than pixel space. Applied to CIFAR-10 for image representation learning, Moving MNIST for video temporal modeling, and Two Rooms navigation task for action-conditioned world models.

Result: Achieved 91% accuracy on CIFAR-10 representation probing, demonstrated multi-step prediction on Moving MNIST, and achieved 97% planning success rate on Two Rooms navigation task. Comprehensive ablations showed critical importance of regularization components for preventing representation collapse.

Conclusion: EB-JEPA successfully demonstrates that JEPA-based representation learning scales effectively from images to video and action-conditioned world models, providing accessible implementations for research and education while achieving strong performance on benchmark tasks.

Abstract: We present EB-JEPA, an open-source library for learning representations and world models using Joint-Embedding Predictive Architectures (JEPAs). JEPAs learn to predict in representation space rather than pixel space, avoiding the pitfalls of generative modeling while capturing semantically meaningful features suitable for downstream tasks. Our library provides modular, self-contained implementations that illustrate how representation learning techniques developed for image-level self-supervised learning can transfer to video, where temporal dynamics add complexity, and ultimately to action-conditioned world models, where the model must additionally learn to predict the effects of control inputs. Each example is designed for single-GPU training within a few hours, making energy-based self-supervised learning accessible for research and education. We provide ablations of JEA components on CIFAR-10. Probing these representations yields 91% accuracy, indicating that the model learns useful features. Extending to video, we include a multi-step prediction example on Moving MNIST that demonstrates how the same principles scale to temporal modeling. Finally, we show how these representations can drive action-conditioned world models, achieving a 97% planning success rate on the Two Rooms navigation task. Comprehensive ablations reveal the critical importance of each regularization component for preventing representation collapse. Code is available at https://github.com/facebookresearch/eb_jepa.

Method: Two-stage framework: 1) Question-agnostic keyframe selection by clustering frame-level visual features to get compact, diverse, representative subset; 2) Key visual token selection by pruning redundant or less informative tokens from each keyframe based on token importance and redundancy.

Result: Outperforms state-of-the-art training-free baselines on Multiple-Choice VideoQA while using significantly fewer visual tokens (e.g., only 504 tokens for 60-min video with 10800 frames). Achieves 44.8% accuracy on MLVU-Test benchmark and exceeds several training-based approaches on certain benchmarks.

Conclusion: KTV provides an efficient and effective training-free video understanding framework that addresses visual redundancy and computational overhead issues while maintaining strong video comprehension performance.

Abstract: Training-free video understanding leverages the strong image comprehension capabilities of pre-trained vision language models (VLMs) by treating a video as a sequence of static frames, thus obviating the need for costly video-specific training. However, this paradigm often suffers from severe visual redundancy and high computational overhead, especially when processing long videos. Crucially, existing keyframe selection strategies, especially those based on CLIP similarity, are prone to biases and may inadvertently overlook critical frames, resulting in suboptimal video comprehension. To address these significant challenges, we propose \textbf{KTV}, a novel two-stage framework for efficient and effective training-free video understanding. In the first stage, KTV performs question-agnostic keyframe selection by clustering frame-level visual features, yielding a compact, diverse, and representative subset of frames that mitigates temporal redundancy. In the second stage, KTV applies key visual token selection, pruning redundant or less informative tokens from each selected keyframe based on token importance and redundancy, which significantly reduces the number of tokens fed into the LLM. Extensive experiments on the Multiple-Choice VideoQA task demonstrate that KTV outperforms state-of-the-art training-free baselines while using significantly fewer visual tokens, \emph{e.g.}, only 504 visual tokens for a 60-min video with 10800 frames, achieving $44.8%$ accuracy on the MLVU-Test benchmark. In particular, KTV also exceeds several training-based approaches on certain benchmarks.

Method: Proposes a unified framework that: 1) synthesizes multimodal data including FFA, MSI, and saliency maps highlighting lesions and optic regions; 2) trains parallel models to learn modality-specific representations capturing cross-pathophysiological signatures; 3) adaptively calibrates features within and across modalities for information pruning and flexible integration based on downstream tasks.

Result: Achieves superior performance on two public datasets: multi-label classification (F1-score: 0.683, AUC: 0.953) and diabetic retinopathy grading (Accuracy: 0.842, Kappa: 0.861). The system is thoroughly interpreted through visualizations in image and feature spaces.

Conclusion: The framework enhances retinal disease screening accuracy and efficiency while providing a scalable approach for data augmentation across medical imaging modalities. It addresses key challenges in multimodal medical diagnosis through integrated synthesis and fusion.

Abstract: Retinal diseases spanning a broad spectrum can be effectively identified and diagnosed using complementary signals from multimodal data. However, multimodal diagnosis in ophthalmic practice is typically challenged in terms of data heterogeneity, potential invasiveness, registration complexity, and so on. As such, a unified framework that integrates multimodal data synthesis and fusion is proposed for retinal disease classification and grading. Specifically, the synthesized multimodal data incorporates fundus fluorescein angiography (FFA), multispectral imaging (MSI), and saliency maps that emphasize latent lesions as well as optic disc/cup regions. Parallel models are independently trained to learn modality-specific representations that capture cross-pathophysiological signatures. These features are then adaptively calibrated within and across modalities to perform information pruning and flexible integration according to downstream tasks. The proposed learning system is thoroughly interpreted through visualizations in both image and feature spaces. Extensive experiments on two public datasets demonstrated the superiority of our approach over state-of-the-art ones in the tasks of multi-label classification (F1-score: 0.683, AUC: 0.953) and diabetic retinopathy grading (Accuracy:0.842, Kappa: 0.861). This work not only enhances the accuracy and efficiency of retinal disease screening but also offers a scalable framework for data augmentation across various medical imaging modalities.

Method: Use multi-objective genetic algorithms to optimize style transfer pipelines, balancing structural coherence and style similarity. Employ paired-image metrics on individual samples during evolution for rapid evaluation, then validate with distributional metrics and segmentation performance on target domains.

Result: Evolutionary algorithms can propose diverse augmentation pipelines adapted to different objectives. Applied to GTA5→Cityscapes/ACDC domain adaptation, the approach demonstrates effective pipeline optimization for synthetic-to-real adaptation.

Conclusion: Style transfer can be formulated as a sequencing problem suitable for evolutionary optimization. Efficient metrics enable feasible search in this space, providing diverse augmentation pipelines for domain adaptation in semantic segmentation.

Abstract: Semantic segmentation networks require large amounts of pixel-level annotated data, which are costly to obtain for real-world images. Computer graphics engines can generate synthetic images alongside their ground-truth annotations. However, models trained on such images can perform poorly on real images due to the domain gap between real and synthetic images. Style transfer methods can reduce this difference by applying a realistic style to synthetic images. Choosing effective data transformations and their sequence is difficult due to the large combinatorial search space of style transfer operators. Using multi-objective genetic algorithms, we optimize pipelines to balance structural coherence and style similarity to target domains. We study the use of paired-image metrics on individual image samples during evolution to enable rapid pipeline evaluation, as opposed to standard distributional metrics that require the generation of many images. After optimization, we evaluate the resulting Pareto front using distributional metrics and segmentation performance. We apply this approach to standard datasets in synthetic-to-real domain adaptation: from the video game GTA5 to real image datasets Cityscapes and ACDC, focusing on adverse conditions. Results demonstrate that evolutionary algorithms can propose diverse augmentation pipelines adapted to different objectives. The contribution of this work is the formulation of style transfer as a sequencing problem suitable for evolutionary optimization and the study of efficient metrics that enable feasible search in this space. The source code is available at: https://github.com/echigot/MOOSS.

Method: Three key innovations: 1) SOS-Student model separates unlabeled objects from background and learns orientation/scale from weak annotations; 2) Multi-level Pseudo-label Filtering uses model prediction distributions; 3) Sparse partitioning ensures equal category treatment.

Result: Extensive experiments on DOTA and DIOR datasets show significant performance gains over traditional oriented object detection methods, offering a highly cost-effective solution.

Conclusion: The SPWOOD framework successfully addresses large-scale labeling challenges in remote sensing by efficiently leveraging sparse weak labels and unlabeled data, achieving strong performance with reduced annotation costs.

Abstract: A consistent trend throughout the research of oriented object detection has been the pursuit of maintaining comparable performance with fewer and weaker annotations. This is particularly crucial in the remote sensing domain, where the dense object distribution and a wide variety of categories contribute to prohibitively high costs. Based on the supervision level, existing oriented object detection algorithms can be broadly grouped into fully supervised, semi-supervised, and weakly supervised methods. Within the scope of this work, we further categorize them to include sparsely supervised and partially weakly-supervised methods. To address the challenges of large-scale labeling, we introduce the first Sparse Partial Weakly-Supervised Oriented Object Detection framework, designed to efficiently leverage only a few sparse weakly-labeled data and plenty of unlabeled data. Our framework incorporates three key innovations: (1) We design a Sparse-annotation-Orientation-and-Scale-aware Student (SOS-Student) model to separate unlabeled objects from the background in a sparsely-labeled setting, and learn orientation and scale information from orientation-agnostic or scale-agnostic weak annotations. (2) We construct a novel Multi-level Pseudo-label Filtering strategy that leverages the distribution of model predictions, which is informed by the model’s multi-layer predictions. (3) We propose a unique sparse partitioning approach, ensuring equal treatment for each category. Extensive experiments on the DOTA and DIOR datasets show that our framework achieves a significant performance gain over traditional oriented object detection methods mentioned above, offering a highly cost-effective solution. Our code is publicly available at https://github.com/VisionXLab/SPWOOD.

Method: Created MM-SCALE dataset with image-scenario pairs annotated with 5-point moral acceptability scores and grounded reasoning labels, enabling listwise preference optimization over ranked scenario sets.

Result: VLMs fine-tuned on MM-SCALE achieve higher ranking fidelity and more stable safety calibration compared to those trained with binary signals.

Conclusion: Scalar supervision provides richer alignment signals and finer calibration for multimodal moral reasoning in VLMs, addressing limitations of discrete supervision methods.

Abstract: Vision-Language Models (VLMs) continue to struggle to make morally salient judgments in multimodal and socially ambiguous contexts. Prior works typically rely on binary or pairwise supervision, which often fail to capture the continuous and pluralistic nature of human moral reasoning. We present MM-SCALE (Multimodal Moral Scale), a large-scale dataset for aligning VLMs with human moral preferences through 5-point scalar ratings and explicit modality grounding. Each image-scenario pair is annotated with moral acceptability scores and grounded reasoning labels by humans using an interface we tailored for data collection, enabling listwise preference optimization over ranked scenario sets. By moving from discrete to scalar supervision, our framework provides richer alignment signals and finer calibration of multimodal moral reasoning. Experiments show that VLMs fine-tuned on MM-SCALE achieve higher ranking fidelity and more stable safety calibration than those trained with binary signals.

Method: Proposes Referring Industrial Anomaly Segmentation (RIAS) paradigm using language to guide detection. Introduces MVTec-Ref dataset with diverse referring expressions focusing on anomaly patterns (95% small anomalies). Presents DQFormer benchmark with Dual Query Token (Anomaly/Background) and Mask Group Transformer, enhanced by Language-Gated Multi-Level Aggregation for multi-scale segmentation.

Result: RIAS generates precise masks from text descriptions without manual thresholds and uses universal prompts to detect diverse anomalies with a single model. Experiments demonstrate effectiveness in advancing IAD toward open-set capabilities.

Conclusion: RIAS represents a significant advancement in industrial anomaly detection by leveraging language guidance to overcome traditional limitations, enabling more flexible and efficient detection with a single model for diverse anomaly types.

Abstract: Industrial Anomaly Detection (IAD) is vital for manufacturing, yet traditional methods face significant challenges: unsupervised approaches yield rough localizations requiring manual thresholds, while supervised methods overfit due to scarce, imbalanced data. Both suffer from the “One Anomaly Class, One Model” limitation. To address this, we propose Referring Industrial Anomaly Segmentation (RIAS), a paradigm leveraging language to guide detection. RIAS generates precise masks from text descriptions without manual thresholds and uses universal prompts to detect diverse anomalies with a single model. We introduce the MVTec-Ref dataset to support this, designed with diverse referring expressions and focusing on anomaly patterns, notably with 95% small anomalies. We also propose the Dual Query Token with Mask Group Transformer (DQFormer) benchmark, enhanced by Language-Gated Multi-Level Aggregation (LMA) to improve multi-scale segmentation. Unlike traditional methods using redundant queries, DQFormer employs only “Anomaly” and “Background” tokens for efficient visual-textual integration. Experiments demonstrate RIAS’s effectiveness in advancing IAD toward open-set capabilities. Code: https://github.com/swagger-coder/RIAS-MVTec-Ref.

Method: Propose RegionReasoner: reinforcement learning framework requiring reasoning traces to explicitly cite reference bounding boxes. Uses global-local consistency reward extracting key objects/nouns from scene and region captions, aligning with reasoning trace. Optimized with structured rewards for grounding fidelity and semantic alignment.

Result: RegionReasoner-7B with new benchmark RegionDial-Bench improves multi-round reasoning accuracy, spatial grounding precision, and global-local consistency on detection and segmentation tasks.

Conclusion: Establishes strong baseline for multi-round visual reasoning with explicit grounding and consistency mechanisms, advancing beyond single-step reasoning approaches.

Abstract: Large vision-language models have achieved remarkable progress in visual reasoning, yet most existing systems rely on single-step or text-only reasoning, limiting their ability to iteratively refine understanding across multiple visual contexts. To address this limitation, we introduce a new multi-round visual reasoning benchmark with training and test sets spanning both detection and segmentation tasks, enabling systematic evaluation under iterative reasoning scenarios. We further propose RegionReasoner, a reinforcement learning framework that enforces grounded reasoning by requiring each reasoning trace to explicitly cite the corresponding reference bounding boxes, while maintaining semantic coherence via a global-local consistency reward. This reward extracts key objects and nouns from both global scene captions and region-level captions, aligning them with the reasoning trace to ensure consistency across reasoning steps. RegionReasoner is optimized with structured rewards combining grounding fidelity and global-local semantic alignment. Experiments on detection and segmentation tasks show that RegionReasoner-7B, together with our newly introduced benchmark RegionDial-Bench, considerably improves multi-round reasoning accuracy, spatial grounding precision, and global-local consistency, establishing a strong baseline for this emerging research direction.

Method: Two-stage pipeline: 1) RAPID-SCAN segmentation model (0.64M parameters) for defect detection, 2) Fine-tuned Phi-3.5 VLM for natural language summarization. Uses post-training quantization and hardware optimization for edge deployment.

Result: Achieves 0.834 F1-score for segmentation, generates domain-specific summaries, reduces model size and latency via quantization, and demonstrates effectiveness on mobile robotic platform in real-world scenarios.

Conclusion: Edge-deployable integrated AI systems can bridge automated defect detection and actionable maintenance insights, enabling scalable autonomous inspection solutions for underground infrastructure.

Abstract: Autonomous inspection of underground infrastructure, such as sewer and culvert systems, is critical to public safety and urban sustainability. Although robotic platforms equipped with visual sensors can efficiently detect structural deficiencies, the automated generation of human-readable summaries from these detections remains a significant challenge, especially on resource-constrained edge devices. This paper presents a novel two-stage pipeline for end-to-end summarization of underground deficiencies, combining our lightweight RAPID-SCAN segmentation model with a fine-tuned Vision-Language Model (VLM) deployed on an edge computing platform. The first stage employs RAPID-SCAN (Resource-Aware Pipeline Inspection and Defect Segmentation using Compact Adaptive Network), achieving 0.834 F1-score with only 0.64M parameters for efficient defect segmentation. The second stage utilizes a fine-tuned Phi-3.5 VLM that generates concise, domain-specific summaries in natural language from the segmentation outputs. We introduce a curated dataset of inspection images with manually verified descriptions for VLM fine-tuning and evaluation. To enable real-time performance, we employ post-training quantization with hardware-specific optimization, achieving significant reductions in model size and inference latency without compromising summarization quality. We deploy and evaluate our complete pipeline on a mobile robotic platform, demonstrating its effectiveness in real-world inspection scenarios. Our results show the potential of edge-deployable integrated AI systems to bridge the gap between automated defect detection and actionable insights for infrastructure maintenance, paving the way for more scalable and autonomous inspection solutions.

Method: LIVE enforces bounded error accumulation via a cycle-consistency objective: forward rollout from ground-truth frames followed by reverse generation to reconstruct initial state, with diffusion loss computed on reconstructed terminal state. Also includes progressive training curriculum.

Result: Achieves state-of-the-art performance on long-horizon benchmarks, generating stable, high-quality videos far beyond training rollout lengths.

Conclusion: LIVE provides an effective approach for long-horizon video generation with bounded error accumulation, outperforming previous methods without requiring teacher models.

Abstract: Autoregressive video world models predict future visual observations conditioned on actions. While effective over short horizons, these models often struggle with long-horizon generation, as small prediction errors accumulate over time. Prior methods alleviate this by introducing pre-trained teacher models and sequence-level distribution matching, which incur additional computational cost and fail to prevent error propagation beyond the training horizon. In this work, we propose LIVE, a Long-horizon Interactive Video world modEl that enforces bounded error accumulation via a novel cycle-consistency objective, thereby eliminating the need for teacher-based distillation. Specifically, LIVE first performs a forward rollout from ground-truth frames and then applies a reverse generation process to reconstruct the initial state. The diffusion loss is subsequently computed on the reconstructed terminal state, providing an explicit constraint on long-horizon error propagation. Moreover, we provide an unified view that encompasses different approaches and introduce progressive training curriculum to stabilize training. Experiments demonstrate that LIVE achieves state-of-the-art performance on long-horizon benchmarks, generating stable, high-quality videos far beyond training rollout lengths.

Method: Decomposes single images into fully inpainted, semantically distinct layers with inferred drawing orders. Uses a scalable engine that bootstraps supervision from commercial Live2D models. Combines diffusion-based Body Part Consistency Module for global geometric coherence with pixel-level pseudo-depth inference for intricate stratification resolution.

Result: Approach yields high-fidelity, manipulatable models suitable for professional, real-time animation applications, effectively resolving complex layer structures like interleaving hair strands.

Conclusion: The framework successfully automates the transformation of static anime illustrations into dynamic 2.5D models, overcoming data scarcity through bootstrapped supervision and achieving professional-quality results.

Abstract: We introduce a framework that automates the transformation of static anime illustrations into manipulatable 2.5D models. Current professional workflows require tedious manual segmentation and the artistic ``hallucination’’ of occluded regions to enable motion. Our approach overcomes this by decomposing a single image into fully inpainted, semantically distinct layers with inferred drawing orders. To address the scarcity of training data, we introduce a scalable engine that bootstraps high-quality supervision from commercial Live2D models, capturing pixel-perfect semantics and hidden geometry. Our methodology couples a diffusion-based Body Part Consistency Module, which enforces global geometric coherence, with a pixel-level pseudo-depth inference mechanism. This combination resolves the intricate stratification of anime characters, e.g., interleaving hair strands, allowing for dynamic layer reconstruction. We demonstrate that our approach yields high-fidelity, manipulatable models suitable for professional, real-time animation applications.

Method: Zero-shot prompting strategy using a state-of-the-art Large Vision Language Model (LVLM). Pipeline converts raw DICOM files to bone-windowed PNGs, submits them to LVLM with carefully engineered prompts, and receives structured JSON outputs that are extracted and formatted for validation.

Result: On 100 expert-reviewed images: 92% main bone accuracy, 80% projection view accuracy, and 100% laterality accuracy, with confidence flags for ambiguous cases. Demonstrates LVLMs can substantially accelerate code word development for paleoradiology datasets.

Conclusion: LVLMs can effectively automate identification tasks in heterogeneous medical imaging datasets, enabling efficient content navigation and accelerating anthropological workflows despite challenging conditions like disarticulated bones and variable positioning.

Abstract: Paleoradiology, the use of modern imaging technologies to study archaeological and anthropological remains, offers new windows on millennial scale patterns of human health. Unfortunately, the radiographs collected during field campaigns are heterogeneous: bones are disarticulated, positioning is ad hoc, and laterality markers are often absent. Additionally, factors such as age at death, age of bone, sex, and imaging equipment introduce high variability. Thus, content navigation, such as identifying a subset of images with a specific projection view, can be time consuming and difficult, making efficient triaging a bottleneck for expert analysis. We report a zero shot prompting strategy that leverages a state of the art Large Vision Language Model (LVLM) to automatically identify the main bone, projection view, and laterality in such images. Our pipeline converts raw DICOM files to bone windowed PNGs, submits them to the LVLM with a carefully engineered prompt, and receives structured JSON outputs, which are extracted and formatted onto a spreadsheet in preparation for validation. On a random sample of 100 images reviewed by an expert board certified paleoradiologist, the system achieved 92% main bone accuracy, 80% projection view accuracy, and 100% laterality accuracy, with low or medium confidence flags for ambiguous cases. These results suggest that LVLMs can substantially accelerate code word development for large paleoradiology datasets, allowing for efficient content navigation in future anthropology workflows.

Method: REPA-G leverages aligned representations from pre-trained feature extractors to guide denoising. It optimizes a similarity objective (potential) at inference time to steer generation toward conditioned representations, enabling multi-scale control from fine textures to global semantics.

Result: Quantitative results on ImageNet and COCO show high-quality, diverse generations. The method enables versatile control including texture matching, semantic guidance, and multi-concept composition.

Conclusion: REPA-G provides flexible, precise inference-time conditioning for diffusion models using aligned representations, offering an alternative to text prompts and class labels with theoretical justification.

Abstract: While representation alignment with self-supervised models has been shown to improve diffusion model training, its potential for enhancing inference-time conditioning remains largely unexplored. We introduce Representation-Aligned Guidance (REPA-G), a framework that leverages these aligned representations, with rich semantic properties, to enable test-time conditioning from features in generation. By optimizing a similarity objective (the potential) at inference, we steer the denoising process toward a conditioned representation extracted from a pre-trained feature extractor. Our method provides versatile control at multiple scales, ranging from fine-grained texture matching via single patches to broad semantic guidance using global image feature tokens. We further extend this to multi-concept composition, allowing for the faithful combination of distinct concepts. REPA-G operates entirely at inference time, offering a flexible and precise alternative to often ambiguous text prompts or coarse class labels. We theoretically justify how this guidance enables sampling from the potential-induced tilted distribution. Quantitative results on ImageNet and COCO demonstrate that our approach achieves high-quality, diverse generations. Code is available at https://github.com/valeoai/REPA-G.

Method: Introduces RAWDet-7, a dataset of ~25k training and 7.6k test RAW images collected across diverse cameras, lighting conditions, and environments. Images are densely annotated for seven object categories following MS-COCO and LVIS conventions, with object-level descriptions derived from corresponding high-resolution sRGB images.

Result: Provides a benchmark dataset for evaluating object detection and description under simulated 4-bit, 6-bit, and 8-bit quantization, reflecting realistic sensor constraints. Enables study of detection performance, description quality & detail, and generalization in low-bit RAW image processing.

Conclusion: RAWDet-7 supports research on leveraging RAW image data for machine vision tasks, addressing limitations of human-optimized ISP processing and enabling exploration of sensor-level information preservation under quantization constraints.

Abstract: Most vision models are trained on RGB images processed through ISP pipelines optimized for human perception, which can discard sensor-level information useful for machine reasoning. RAW images preserve unprocessed scene data, enabling models to leverage richer cues for both object detection and object description, capturing fine-grained details, spatial relationships, and contextual information often lost in processed images. To support research in this domain, we introduce RAWDet-7, a large-scale dataset of ~25k training and 7.6k test RAW images collected across diverse cameras, lighting conditions, and environments, densely annotated for seven object categories following MS-COCO and LVIS conventions. In addition, we provide object-level descriptions derived from the corresponding high-resolution sRGB images, facilitating the study of object-level information preservation under RAW image processing and low-bit quantization. The dataset allows evaluation under simulated 4-bit, 6-bit, and 8-bit quantization, reflecting realistic sensor constraints, and provides a benchmark for studying detection performance, description quality & detail, and generalization in low-bit RAW image processing. Dataset & code upon acceptance.

Method: Proposes FOVI that reformats retina-like variable-resolution sensor array into uniform V1-like manifold, uses k-nearest-neighborhood receptive fields with novel kernel mapping for kNN-convolution, and adapts DINOv3 ViT with LoRA.

Result: Models achieve competitive performance at fraction of computational cost compared to non-foveated baselines, enabling efficient high-resolution egocentric vision.

Conclusion: FOVI opens pathways for efficient and scalable active sensing for high-resolution vision tasks by mimicking biological foveation principles.

Abstract: Human vision is foveated, with variable resolution peaking at the center of a large field of view; this reflects an efficient trade-off for active sensing, allowing eye-movements to bring different parts of the world into focus with other parts of the world in context. In contrast, most computer vision systems encode the visual world at a uniform resolution, raising challenges for processing full-field high-resolution images efficiently. We propose a foveated vision interface (FOVI) based on the human retina and primary visual cortex, that reformats a variable-resolution retina-like sensor array into a uniformly dense, V1-like sensor manifold. Receptive fields are defined as k-nearest-neighborhoods (kNNs) on the sensor manifold, enabling kNN-convolution via a novel kernel mapping technique. We demonstrate two use cases: (1) an end-to-end kNN-convolutional architecture, and (2) a foveated adaptation of the foundational DINOv3 ViT model, leveraging low-rank adaptation (LoRA). These models provide competitive performance at a fraction of the computational cost of non-foveated baselines, opening pathways for efficient and scalable active sensing for high-resolution egocentric vision. Code and pre-trained models are available at https://github.com/nblauch/fovi and https://huggingface.co/fovi-pytorch.

Method: QVLA introduces a granular channel-wise bit allocation strategy that measures action-space sensitivity when quantizing each channel to various bit-widths. This creates per-channel importance metrics guiding global optimization that unifies quantization and pruning into a single framework.

Result: On LIBERO benchmark, QVLA reduces OpenVLA-OFT to 29.2% of original VRAM while maintaining 98.9% performance and achieving 1.49x speedup, outperforming LLM-derived SmoothQuant by 22.6% performance improvement.

Conclusion: QVLA establishes a new principled foundation for compressing VLA models in robotics, enabling deployment of powerful large-scale models on real-world hardware through action-centric quantization.

Abstract: The advent of Vision-Language-Action (VLA) models represents a significant leap for embodied intelligence, yet their immense computational demands critically hinder deployment on resource-constrained robotic platforms. Intuitively, low-bit quantization is a prevalent and preferred technique for large-scale model compression. However, we find that a systematic analysis of VLA model’s quantization is fundamentally lacking. We argue that naively applying uniform-bit quantization from Large Language Models (LLMs) to robotics is flawed, as these methods prioritize passive data fidelity while ignoring how minor action deviations compound into catastrophic task failures. To bridge this gap, we introduce QVLA, the first action-centric quantization framework specifically designed for embodied control. In a sharp departure from the rigid, uniform-bit quantization of LLM-based methods, QVLA introduces a highly granular, channel-wise bit allocation strategy. Its core mechanism is to directly measure the final action-space sensitivity when quantizing each individual channel to various bit-widths. This process yields a precise, per-channel importance metric that guides a global optimization, which elegantly unifies quantization and pruning (0-bit) into a single, cohesive framework. Extensive evaluations on different baselines demonstrate the superiority of our approach. In the LIBERO, the quantization version of OpenVLA-OFT with our method requires only 29.2% of the original model’s VRAM while maintaining 98.9% of its original performance and achieving a 1.49x speedup. This translates to a 22.6% performance improvement over the LLM-derived method SmoothQuant. Our work establishes a new, principled foundation for compressing VLA models in robotics, paving the way for deploying powerful, large-scale models on real-world hardware. Code will be released.

Method: U-Net architecture trained to predict brain shift entirely from preoperative MRI, evaluated using Target Registration Errors (TREs) on anatomical landmarks and DICE scores comparing predicted vs. actual intraoperative masks.

Result: Achieved DICE score of 0.97 for global brain deformation prediction and landmark TRE as low as 1.12 mm, effectively compensating for large brain shifts during temporal lobe removal.

Conclusion: The model successfully predicts brain deformation during temporal lobe resection using only preoperative images, potentially increasing surgical safety and efficiency while improving patient outcomes.

Abstract: Introduction: In neurosurgery, image-guided Neurosurgery Systems (IGNS) highly rely on preoperative brain magnetic resonance images (MRI) to assist surgeons in locating surgical targets and determining surgical paths. However, brain shift invalidates the preoperative MRI after dural opening. Updated intraoperative brain MRI with brain shift compensation is crucial for enhancing the precision of neuronavigation systems and ensuring the optimal outcome of surgical interventions. Methodology: We propose NeuralShift, a U-Net-based model that predicts brain shift entirely from pre-operative MRI for patients undergoing temporal lobe resection. We evaluated our results using Target Registration Errors (TREs) computed on anatomical landmarks located on the resection side and along the midline, and DICE scores comparing predicted intraoperative masks with masks derived from intraoperative MRI. Results: Our experimental results show that our model can predict the global deformation of the brain (DICE of 0.97) with accurate local displacements (achieve landmark TRE as low as 1.12 mm), compensating for large brain shifts during temporal lobe removal neurosurgery. Conclusion: Our proposed model is capable of predicting the global deformation of the brain during temporal lobe resection using only preoperative images, providing potential opportunities to the surgical team to increase safety and efficiency of neurosurgery and better outcomes to patients. Our contributions will be publicly available after acceptance in https://github.com/SurgicalDataScienceKCL/NeuralShift.

Method: Jointly trains a motion encoder with a pretrained video generator to distill driving frames into compact, view-agnostic motion tokens injected via cross-attention. Uses view-rich supervision (single-view, multi-view, moving-camera videos) for 3D awareness, and auxiliary geometric supervision with SMPL that anneals to zero.

Result: 3DiMo faithfully reproduces driving motions with flexible, text-driven camera control, significantly surpassing existing methods in both motion fidelity and visual quality.

Conclusion: Implicit, view-agnostic motion representation better aligns with video generators’ spatial priors than explicit 3D constraints, enabling superior motion control with flexible viewpoint manipulation.

Abstract: Existing methods for human motion control in video generation typically rely on either 2D poses or explicit 3D parametric models (e.g., SMPL) as control signals. However, 2D poses rigidly bind motion to the driving viewpoint, precluding novel-view synthesis. Explicit 3D models, though structurally informative, suffer from inherent inaccuracies (e.g., depth ambiguity and inaccurate dynamics) which, when used as a strong constraint, override the powerful intrinsic 3D awareness of large-scale video generators. In this work, we revisit motion control from a 3D-aware perspective, advocating for an implicit, view-agnostic motion representation that naturally aligns with the generator’s spatial priors rather than depending on externally reconstructed constraints. We introduce 3DiMo, which jointly trains a motion encoder with a pretrained video generator to distill driving frames into compact, view-agnostic motion tokens, injected semantically via cross-attention. To foster 3D awareness, we train with view-rich supervision (i.e., single-view, multi-view, and moving-camera videos), forcing motion consistency across diverse viewpoints. Additionally, we use auxiliary geometric supervision that leverages SMPL only for early initialization and is annealed to zero, enabling the model to transition from external 3D guidance to learning genuine 3D spatial motion understanding from the data and the generator’s priors. Experiments confirm that 3DiMo faithfully reproduces driving motions with flexible, text-driven camera control, significantly surpassing existing methods in both motion fidelity and visual quality.

Method: Uses a flexible, fixed ordering based on progressive checkerboards for multiscale autoregressive image generation. The ordering draws samples in parallel from evenly spaced regions at each scale, maintaining full balance in all levels of a quadtree subdivision.

Result: Achieves competitive performance compared to recent state-of-the-art autoregressive systems with similar model capacity on class-conditional ImageNet, using fewer sampling steps.

Conclusion: Progressive checkerboard ordering enables effective conditioning both between and within scales, and a wide range of scale-up factors lead to similar results when total serial steps are constant.

Abstract: A key challenge in autoregressive image generation is to efficiently sample independent locations in parallel, while still modeling mutual dependencies with serial conditioning. Some recent works have addressed this by conditioning between scales in a multiscale pyramid. Others have looked at parallelizing samples in a single image using regular partitions or randomized orders. In this work we examine a flexible, fixed ordering based on progressive checkerboards for multiscale autoregressive image generation. Our ordering draws samples in parallel from evenly spaced regions at each scale, maintaining full balance in all levels of a quadtree subdivision at each step. This enables effective conditioning both between and within scales. Intriguingly, we find evidence that in our balanced setting, a wide range of scale-up factors lead to similar results, so long as the total number of serial steps is constant. On class-conditional ImageNet, our method achieves competitive performance compared to recent state-of-the-art autoregressive systems with like model capacity, using fewer sampling steps.

Method: Proposes DualSpeed with two modes: fast-mode (primary) uses VTP to reduce visual tokens with mode isolator; slow-mode (auxiliary) trains on full visual sequences for consistency, enhanced by self-distillation from fast-mode.

Result: Accelerates training of LLaVA-1.5 by 2.1× and LLaVA-NeXT by 4.0× while retaining over 99% performance, achieving both training efficiency and non-degraded performance.

Conclusion: DualSpeed effectively addresses training inefficiency in MLLMs through a dual-mode framework that combines visual token pruning with training-inference consistency, offering significant speedups without performance degradation.

Abstract: Multimodal Large Language Models (MLLMs) suffer from severe training inefficiency issue, which is associated with their massive model sizes and visual token numbers. Existing efforts in efficient training focus on reducing model sizes or trainable parameters. Inspired by the success of Visual Token Pruning (VTP) in improving inference efficiency, we are exploring another substantial research direction for efficient training by reducing visual tokens. However, applying VTP at the training stage results in a training-inference mismatch: pruning-trained models perform poorly when inferring on non-pruned full visual token sequences. To close this gap, we propose DualSpeed, a fast-slow framework for efficient training of MLLMs. The fast-mode is the primary mode, which incorporates existing VTP methods as plugins to reduce visual tokens, along with a mode isolator to isolate the model’s behaviors. The slow-mode is the auxiliary mode, where the model is trained on full visual sequences to retain training-inference consistency. To boost its training, it further leverages self-distillation to learn from the sufficiently trained fast-mode. Together, DualSpeed can achieve both training efficiency and non-degraded performance. Experiments show DualSpeed accelerates the training of LLaVA-1.5 by 2.1$\times$ and LLaVA-NeXT by 4.0$\times$, retaining over 99% performance. Code: https://github.com/dingkun-zhang/DualSpeed

Method: Proposes Adaptive-Origin Guidance (AdaOr) that adjusts the standard guidance origin with an identity-conditioned adaptive origin using identity instructions. Interpolates identity prediction with unconditional prediction based on edit strength to ensure continuous transitions.

Result: Method evaluated on image and video editing tasks, demonstrating smoother and more consistent control compared to current slider-based approaches. Enables fine-grained control at inference without per-edit procedures or specialized datasets.

Conclusion: AdaOr provides effective continuous control for text-guided image and video editing by addressing the limitations of CFG scaling through adaptive guidance origins, enabling smooth intensity adjustments.

Abstract: Diffusion-based editing models have emerged as a powerful tool for semantic image and video manipulation. However, existing models lack a mechanism for smoothly controlling the intensity of text-guided edits. In standard text-conditioned generation, Classifier-Free Guidance (CFG) impacts prompt adherence, suggesting it as a potential control for edit intensity in editing models. However, we show that scaling CFG in these models does not produce a smooth transition between the input and the edited result. We attribute this behavior to the unconditional prediction, which serves as the guidance origin and dominates the generation at low guidance scales, while representing an arbitrary manipulation of the input content. To enable continuous control, we introduce Adaptive-Origin Guidance (AdaOr), a method that adjusts this standard guidance origin with an identity-conditioned adaptive origin, using an identity instruction corresponding to the identity manipulation. By interpolating this identity prediction with the standard unconditional prediction according to the edit strength, we ensure a continuous transition from the input to the edited result. We evaluate our method on image and video editing tasks, demonstrating that it provides smoother and more consistent control compared to current slider-based editing approaches. Our method incorporates an identity instruction into the standard training framework, enabling fine-grained control at inference time without per-edit procedure or reliance on specialized datasets.

Method: Combines 3D signed distance function (SDF) and density field learning with event-based supervision. Introduces spherical harmonics encodings to handle view-dependent effects. Uses self-supervised neural model for learning 3D representations from monocular color event streams.

Result: Outperforms existing approaches by significant margin: 34% lower Chamfer distance and 31% lower mean absolute error on average compared to best previous method.

Conclusion: EventNeuS successfully addresses limitations in event-based 3D reconstruction, achieving state-of-the-art performance through novel combination of SDF/density field learning with event-based supervision and spherical harmonics encodings.

Abstract: Event cameras offer a considerable alternative to RGB cameras in many scenarios. While there are recent works on event-based novel-view synthesis, dense 3D mesh reconstruction remains scarcely explored and existing event-based techniques are severely limited in their 3D reconstruction accuracy. To address this limitation, we present EventNeuS, a self-supervised neural model for learning 3D representations from monocular colour event streams. Our approach, for the first time, combines 3D signed distance function and density field learning with event-based supervision. Furthermore, we introduce spherical harmonics encodings into our model for enhanced handling of view-dependent effects. EventNeuS outperforms existing approaches by a significant margin, achieving 34% lower Chamfer distance and 31% lower mean absolute error on average compared to the best previous method.

Method: Leverages foundation models to create panoptic Dynamic Descriptors that unify open-vocabulary labels with category structure and geometric size priors, using language-guided open-vocabulary panoptic segmentation and semantic retrieval within a multi-resolution multi-TSDF map.

Result: UPPM achieves best overall performance in map reconstruction accuracy and panoptic segmentation quality, preserving open-vocabulary interpretability while delivering strong geometric and panoptic accuracy.

Conclusion: UPPM advances open-world scene understanding by creating persistent, promptable panoptic maps that maintain open-vocabulary interpretability while improving geometric consistency and segmentation quality without requiring additional model training.

Abstract: Panoptic maps enable robots to reason about both geometry and semantics. However, open-vocabulary models repeatedly produce closely related labels that split panoptic entities and degrade volumetric consistency. The proposed UPPM advances open-world scene understanding by leveraging foundation models to introduce a panoptic Dynamic Descriptor that reconciles open-vocabulary labels with unified category structure and geometric size priors. The fusion for such dynamic descriptors is performed within a multi-resolution multi-TSDF map using language-guided open-vocabulary panoptic segmentation and semantic retrieval, resulting in a persistent and promptable panoptic map without additional model training. Based on our evaluation experiments, UPPM shows the best overall performance in terms of the map reconstruction accuracy and the panoptic segmentation quality. The ablation study investigates the contribution for each component of UPPM (custom NMS, blurry-frame filtering, and unified semantics) to the overall system performance. Consequently, UPPM preserves open-vocabulary interpretability while delivering strong geometric and panoptic accuracy.

Method: Proposes HAAP with two key components: 1) Implicit Permutation Neurons (IPN) that generate adaptive attention masks to dynamically capture token dependencies, and 2) Cross-modal Hierarchical Attention (CHA) that captures dependencies among position queries, contextual semantics, and visual information.

Result: HAAP achieves state-of-the-art performance in accuracy, complexity, and latency on several benchmark datasets for scene text recognition.

Conclusion: The proposed HAAP framework effectively addresses limitations of existing PLM approaches by enabling adaptive attention and hierarchical cross-modal interaction, leading to improved STR performance without iterative refinement overhead.

Abstract: Scene Text Recognition (STR) is challenging in extracting effective character representations from visual data when text is unreadable. Permutation language modeling (PLM) is introduced to refine character predictions by jointly capturing contextual and visual information. However, in PLM, the use of random permutations causes training fit oscillation, and the iterative refinement (IR) operation also introduces additional overhead. To address these issues, this paper proposes the Hierarchical Attention autoregressive Model with Adaptive Permutation (HAAP) to enhance position-context-image interaction capability, improving autoregressive LM generalization. First, we propose Implicit Permutation Neurons (IPN) to generate adaptive attention masks that dynamically exploit token dependencies, enhancing the correlation between visual information and context. Adaptive correlation representation helps the model avoid training fit oscillation. Second, the Cross-modal Hierarchical Attention mechanism (CHA) is introduced to capture the dependencies among position queries, contextual semantics and visual information. CHA enables position tokens to aggregate global semantic information, avoiding the need for IR. Extensive experimental results show that the proposed HAAP achieves state-of-the-art (SOTA) performance in terms of accuracy, complexity, and latency on several datasets.

Method: Proposes an architecture leveraging foundational video models for text-video alignment, combined with Saliency-Guided Cross Attention mechanism and hybrid DETR architecture. Also introduces InterVid-MR, a large-scale high-quality dataset for pretraining.

Result: Achieves state-of-the-art results on QVHighlights, Charades-STA and TACoS benchmarks. Provides efficient and scalable solution for both zero-shot and fine-tuning scenarios.

Conclusion: The proposed approach significantly enhances performance in video moment retrieval and highlight detection tasks through better text-video alignment and large-scale pretraining.

Abstract: Existing approaches for video moment retrieval and highlight detection are not able to align text and video features efficiently, resulting in unsatisfying performance and limited production usage. To address this, we propose a novel architecture that utilizes recent foundational video models designed for such alignment. Combined with the introduced Saliency-Guided Cross Attention mechanism and a hybrid DETR architecture, our approach significantly enhances performance in both moment retrieval and highlight detection tasks. For even better improvement, we developed InterVid-MR, a large-scale and high-quality dataset for pretraining. Using it, our architecture achieves state-of-the-art results on the QVHighlights, Charades-STA and TACoS benchmarks. The proposed approach provides an efficient and scalable solution for both zero-shot and fine-tuning scenarios in video-language tasks.

Method: Introduces Multi-Image Safety (MIS) dataset with multi-image inputs and safety Chain-of-Thought (CoT) labels as fine-grained reasoning logic. Fine-tunes InternVL2.5-8B model using this dataset to enhance both visual perception and reasoning in safety-critical contexts.

Result: Fine-tuning with MIS significantly outperforms both open-source and API-based models in challenging multi-image tasks requiring safety-related visual reasoning. Increases average accuracy by 0.83% across five general benchmarks and reduces Attack Success Rate (ASR) on multiple safety benchmarks by a large margin.

Conclusion: The proposed approach effectively addresses the safety reasoning gap in VLMs by providing fine-grained reasoning logic through multi-image safety scenarios, delivering exceptional safety performance while preserving general capabilities without trade-offs.

Abstract: Large Vision-Language Models (VLMs) have achieved remarkable performance across a wide range of tasks. However, their deployment in safety-critical domains poses significant challenges. Existing safety fine-tuning methods, which focus on textual or multimodal content, fall short in addressing challenging cases or disrupt the balance between helpfulness and harmlessness. Our evaluation highlights a safety reasoning gap: these methods lack safety visual reasoning ability, leading to such bottlenecks. To address this limitation and enhance both visual perception and reasoning in safety-critical contexts, we propose a novel dataset that integrates multi-image inputs with safety Chain-of-Thought (CoT) labels as fine-grained reasoning logic to improve model performance. Specifically, we introduce the Multi-Image Safety (MIS) dataset, an instruction-following dataset tailored for multi-image safety scenarios, consisting of training and test splits. Our experiments demonstrate that fine-tuning InternVL2.5-8B with MIS significantly outperforms both powerful open-source models and API-based models in challenging multi-image tasks requiring safety-related visual reasoning. This approach not only delivers exceptional safety performance but also preserves general capabilities without any trade-offs. Specifically, fine-tuning with MIS increases average accuracy by 0.83% across five general benchmarks and reduces the Attack Success Rate (ASR) on multiple safety benchmarks by a large margin.

Method: OptiPMB employs an optimized Poisson multi-Bernoulli (PMB) filter within the tracking-by-detection framework. Key innovations include: 1) measurement-driven hybrid adaptive birth model for improved track initialization, 2) adaptive detection probability parameters to maintain tracks for occluded objects, and 3) optimized density pruning and track extraction modules to enhance overall tracking performance.

Result: Extensive evaluations on nuScenes and KITTI datasets show that OptiPMB achieves superior tracking accuracy compared with state-of-the-art methods, establishing a new benchmark for model-based 3D MOT.

Conclusion: OptiPMB demonstrates the effectiveness of RFS-based approaches for 3D MOT in autonomous driving, offering valuable insights for future research on RFS-based trackers and providing a strong model-based alternative to deep learning methods.

Abstract: Accurate 3D multi-object tracking (MOT) is crucial for autonomous driving, as it enables robust perception, navigation, and planning in complex environments. While deep learning-based solutions have demonstrated impressive 3D MOT performance, model-based approaches remain appealing for their simplicity, interpretability, and data efficiency. Conventional model-based trackers typically rely on random vector-based Bayesian filters within the tracking-by-detection (TBD) framework but face limitations due to heuristic data association and track management schemes. In contrast, random finite set (RFS)-based Bayesian filtering handles object birth, survival, and death in a theoretically sound manner, facilitating interpretability and parameter tuning. In this paper, we present OptiPMB, a novel RFS-based 3D MOT method that employs an optimized Poisson multi-Bernoulli (PMB) filter while incorporating several key innovative designs within the TBD framework. Specifically, we propose a measurement-driven hybrid adaptive birth model for improved track initialization, employ adaptive detection probability parameters to effectively maintain tracks for occluded objects, and optimize density pruning and track extraction modules to further enhance overall tracking performance. Extensive evaluations on nuScenes and KITTI datasets show that OptiPMB achieves superior tracking accuracy compared with state-of-the-art methods, thereby establishing a new benchmark for model-based 3D MOT and offering valuable insights for future research on RFS-based trackers in autonomous driving.

Method: FedVSR is a model-agnostic, stateless FL framework that introduces: 1) A lightweight loss function based on Discrete Wavelet Transform (DWT) to preserve high-frequency details during local training, and 2) A loss-aware aggregation strategy combining DWT-based and task-specific losses to guide global updates effectively.

Result: Extensive experiments across multiple VSR models and datasets show FedVSR improves perceptual video quality (up to +0.89 dB PSNR, +0.0370 SSIM, -0.0347 LPIPS and 4.98 VMAF) with close to zero computation and communication overhead compared to rivals.

Conclusion: FedVSR bridges the gap between privacy, efficiency, and perceptual quality, setting a new benchmark for federated learning in low-level vision tasks.

Abstract: Video super-resolution (VSR) aims to enhance low-resolution videos by leveraging both spatial and temporal information. While deep learning has led to impressive progress, it typically requires centralized data, which raises privacy concerns. Federated learning (FL) offers a privacy-friendly solution, but general FL frameworks often struggle with low-level vision tasks, resulting in blurry, low-quality outputs. To address this, we introduce FedVSR, the first FL framework specifically designed for VSR. It is model-agnostic and stateless, and introduces a lightweight loss function based on the Discrete Wavelet Transform (DWT) to better preserve high-frequency details during local training. Additionally, a loss-aware aggregation strategy combines both DWT-based and task-specific losses to guide global updates effectively. Extensive experiments across multiple VSR models and datasets show that FedVSR not only improves perceptual video quality (up to +0.89 dB PSNR, +0.0370 SSIM, -0.0347 LPIPS and 4.98 VMAF) but also achieves these gains with close to zero computation and communication overhead compared to its rivals. These results demonstrate FedVSR’s potential to bridge the gap between privacy, efficiency, and perceptual quality, setting a new benchmark for federated learning in low-level vision tasks. The code is available at: https://github.com/alimd94/FedVSR

Method: Created V2P-Bench with 980 videos and 1172 high-quality QA pairs, each with manually annotated visual prompt frames. Covers three main tasks and twelve categories for fine-grained evaluation. Analyzed current LVLMs on this benchmark.

Result: Key findings: 1) Visual prompts are more model/user-friendly than text prompts, improving performance and user experience; 2) Models have reasonable zero-shot visual prompt understanding but struggle with spatiotemporal reasoning (o1 achieves 71.8% vs human 88.3%); 3) LVLMs exhibit “Hack Phenomena” in video QA that inflates scores as video length increases and frame sampling decreases.

Conclusion: V2P-Bench addresses limitations of text-prompt-based evaluation and provides a foundation for advancing human-model interaction and improving video understanding evaluation in LVLMs.

Abstract: Large Vision-Language Models (LVLMs) have made significant strides in the field of video understanding in recent times. Nevertheless, existing video benchmarks predominantly rely on text prompts for evaluation, which often require complex referential language and diminish both the accuracy and efficiency of human model interaction in turn. To address this limitation, we propose V2P-Bench, a robust and comprehensive benchmark for evaluating the ability of LVLMs to understand Video Visual Prompts in human model interaction scenarios. V2P-Bench consists of 980 videos and 1172 well-structured high-quality QA pairs, each paired with manually annotated visual prompt frames. The benchmark spans three main tasks and twelve categories, thereby enabling fine-grained, instance-level evaluation. Through an in-depth analysis of current LVLMs, we identify several key findings: 1) Visual prompts are both more model-friendly and user-friendly in interactive scenarios than text prompts, leading to significantly improved model performance and enhanced user experience. 2) Models are reasonably capable of zero-shot understanding of visual prompts, but struggle with spatiotemporal understanding. Even o1 achieves only 71.8%, far below the human expert score of 88.3%, while most open-source models perform below 60%. 3) LVLMs exhibit pervasive Hack Phenomena in video question answering tasks, which become more pronounced as video length increases and frame sampling density decreases, thereby inflating performance scores artificially. We anticipate that V2P-Bench will not only shed light on these challenges but also serve as a foundational tool for advancing human model interaction and improving the evaluation of video understanding.

Method: Patronus incorporates a prototypical network into the diffusion framework to encode semantics in visual patches. This allows the model to reveal what visual patterns are being modeled, where they emerge spatially in images, and when they appear across different timesteps of the denoising process.

Result: The model was evaluated on four natural image datasets and one medical imaging dataset, demonstrating both faithful interpretability and strong generative performance. It successfully enabled detection of shortcut learning via unwanted correlations and tracing of semantic emergence across timesteps.

Conclusion: Patronus opens new avenues for understanding and steering diffusion models through prototype-based interpretability, providing deeper insights into the generative mechanism of diffusion models.

Abstract: Uncovering the opacity of diffusion-based generative models is urgently needed, as their applications continue to expand while their underlying procedures largely remain a black box. With a critical question – how can the diffusion generation process be interpreted and understood? – we proposed Patronus, an interpretable diffusion model that incorporates a prototypical network to encode semantics in visual patches, revealing what visual patterns are modeled and where and when they emerge throughout denoising. This interpretability of Patronus provides deeper insights into the generative mechanism, enabling the detection of shortcut learning via unwanted correlations and the tracing of semantic emergence across timesteps. We evaluate Patronus on four natural image datasets and one medical imaging dataset, demonstrating both faithful interpretability and strong generative performance. With this work, we open new avenues for understanding and steering diffusion models through prototype-based interpretability.\ Our code is available at https://github.com/nina-weng/patronus}{https://github.com/nina-weng/patronus.

Method: Developed a scalable pipeline using narrative transcripts from medical education videos to align visual frames with textual concepts. Automatically produced 2,851 high-quality multi-image VQA pairs with explicit, transcript-grounded reasoning chains.

Result: Evaluation of 11 advanced MLLMs (including reasoning models) showed severe deficiencies: accuracies mostly below 50%, instability across varying image counts. Models treat images as isolated instances, failing to track pathological progression or cross-reference anatomical shifts.

Conclusion: MedFrameQA provides a rigorous standard for evaluating next-generation MLLMs in handling complex medical narratives. Current models lack essential multi-image synthesis capabilities needed for real clinical practice.

Abstract: Real-world clinical practice demands multi-image comparative reasoning, yet current medical benchmarks remain limited to single-frame interpretation. We present MedFrameQA, the first benchmark explicitly designed to test multi-image medical VQA through educationally-validated diagnostic sequences. To construct this dataset, we develop a scalable pipeline that leverages narrative transcripts from medical education videos to align visual frames with textual concepts, automatically producing 2,851 high-quality multi-image VQA pairs with explicit, transcript-grounded reasoning chains. Our evaluation of 11 advanced MLLMs (including reasoning models) exposes severe deficiencies in multi-image synthesis, where accuracies mostly fall below 50% and exhibit instability across varying image counts. Error analysis demonstrates that models often treat images as isolated instances, failing to track pathological progression or cross-reference anatomical shifts. MedFrameQA provides a rigorous standard for evaluating the next generation of MLLMs in handling complex, temporally grounded medical narratives.

Method: Proposes Sat2Density++, which learns neural radiance fields from paired satellite and street-view images. Key innovation: modeling street-view specific elements (sky, illumination effects) that are only visible in street-view panoramas but not in satellite images, enabling more realistic rendering.

Result: Tested on both urban and suburban scene datasets, demonstrating capability to render photorealistic street-view panoramas that are consistent across multiple views and faithful to the satellite image.

Conclusion: Sat2Density++ successfully addresses the challenging SatStreet-view synthesis problem by incorporating street-view specific elements into neural radiance field modeling, enabling realistic panorama rendering from satellite imagery.

Abstract: This paper studies the task of SatStreet-view synthesis, which aims to render photorealistic street-view panorama images and videos given any satellite image and specified camera positions or trajectories. We formulate to learn neural radiance field from paired images captured from satellite and street viewpoints, which comes to be a challenging learning problem due to the sparse-view natural and the extremely-large viewpoint changes between satellite and street-view images. We tackle the challenges based on a task-specific observation that street-view specific elements, including the sky and illumination effects are only visible in street-view panoramas, and present a novel approach Sat2Density++ to accomplish the goal of photo-realistic street-view panoramas rendering by modeling these street-view specific in neural networks. In the experiments, our method is testified on both urban and suburban scene datasets, demonstrating that Sat2Density++ is capable of rendering photorealistic street-view panoramas that are consistent across multiple views and faithful to the satellite image.

Method: Builds on existing Hephaestus dataset by creating Thalia - a global collection of 38 spatiotemporal datacubes covering 7 years with InSAR products, topographic data, atmospheric variables, and expert annotations including deformation type, intensity, and extent with descriptive text.

Result: Created a comprehensive benchmark dataset with state-of-the-art models for classification and segmentation, enabling fair evaluation and fostering collaboration between machine learning and Earth science.

Conclusion: Thalia advances volcanic monitoring capabilities and promotes data-driven approaches in geoscience by providing a rich, annotated dataset that addresses previous limitations in InSAR data interpretation for deep learning applications.

Abstract: Monitoring volcanic activity is of paramount importance to safeguarding lives, infrastructure, and ecosystems. However, only a small fraction of known volcanoes are continuously monitored. Satellite-based Interferometric Synthetic Aperture Radar (InSAR) enables systematic, global-scale deformation monitoring. However, its complex data challenge traditional remote sensing methods. Deep learning offers a powerful means to automate and enhance InSAR interpretation, advancing volcanology and geohazard assessment. Despite its promise, progress has been limited by the scarcity of well-curated datasets. In this work, we build on the existing Hephaestus dataset and introduce Thalia, addressing crucial limitations and enriching its scope with higher-resolution, multi-source, and multi-temporal data. Thalia is a global collection of 38 spatiotemporal datacubes covering 7 years and integrating InSAR products, topographic data, as well as atmospheric variables, known to introduce signal delays that can mimic ground deformation in InSAR imagery. Each sample includes expert annotations detailing the type, intensity, and extent of deformation, ac- companied by descriptive text. To enable fair and consistent evaluation, we provide a comprehensive benchmark using state-of-the-art models for classification and segmentation. This work fosters collaboration between machine learning and Earth science, advancing volcanic monitoring and promoting data-driven approaches in geoscience. The code and latest version of the dataset are available through the github repository: https://github.com/Orion-AI-Lab/Thalia

Method: Consistency-aware dynamic detection analyzes geometric/texture discrepancies between historical map renderings and real observations. Uses bidirectional dynamic object tracking (backward/forward in time) for complete sequence recognition. Dynamic-static decoupled mapping with temporal Gaussian model for online incremental dynamic modeling.

Result: Experiments on multiple dynamic datasets show flexible and accurate dynamic segmentation capabilities with state-of-the-art performance in both localization and mapping.

Conclusion: CAD-SLAM effectively addresses dynamic environment challenges in SLAM through consistency-aware dynamic detection and decoupled mapping, enabling robust performance in scenes with moving objects.

Abstract: Recent advances in neural radiation fields (NeRF) and 3D Gaussian-based SLAM have achieved impressive localization accuracy and high-quality dense mapping in static scenes. However, these methods remain challenged in dynamic environments, where moving objects violate the static-world assumption and introduce inconsistent observations that degrade both camera tracking and map reconstruction. This motivates two fundamental problems: robustly identifying dynamic objects and modeling them online. To address these limitations, we propose CAD-SLAM, a Consistency-Aware Dynamic SLAM framework with dynamic-static decoupled mapping. Our key insight is that dynamic objects inherently violate cross-view and cross-time scene consistency. We detect object motion by analyzing geometric and texture discrepancies between historical map renderings and real-world observations. Once a moving object is identified, we perform bidirectional dynamic object tracking (both backward and forward in time) to achieve complete sequence-wise dynamic recognition. Our consistency-aware dynamic detection model achieves category-agnostic, instantaneous dynamic identification, which effectively mitigates motion-induced interference during localization and mapping. In addition, we introduce a dynamic-static decoupled mapping strategy that employs a temporal Gaussian model for online incremental dynamic modeling. Experiments conducted on multiple dynamic datasets demonstrate the flexible and accurate dynamic segmentation capabilities of our method, along with the state-of-the-art performance in both localization and mapping.

Method: Proposes Ground-R1 framework with Scale Relative Policy Optimization (SRPO) that replaces standard GRPO. SRPO recalibrates reward learning across evidence regions of different sizes through scale-aware binning and intra-/inter-bin comparisons for balanced credit assignment.

Result: Experimental results on general LVLM, high-resolution, and visual grounding benchmarks show Ground-R1’s effectiveness. SRPO yields consistent gains over standard GRPO in both response accuracy and evidence grounding.

Conclusion: Ground-R1 with SRPO addresses scale bias in vision-language models, improving reliability and interpretability through better visual grounding across all region sizes.

Abstract: Large Vision-Language Models (LVLMs) have become powerful general-purpose assistants, yet their predictions often lack reliability and interpretability due to insufficient grounding in visual evidence. The emerging thinking-with-images paradigm seeks to address this issue by explicitly anchoring reasoning to image regions. However, we empirically find that most existing methods suffer from a systematic scale-driven bias in optimization, where training rewards are dominated by large visual regions, suppressing learning from small but semantically critical evidence and leading to spurious grounding at inference time. To address this limitation, we propose Ground-R1, a de-biased thinking-with-images framework trained via a novel Scale Relative Policy Optimization (SRPO) objective that replaces standard GRPO. Specifically, our SRPO recalibrates reward learning across evidence regions of different sizes through scale-aware binning and intra-/inter-bin comparisons, enabling balanced credit assignment during training. Experimental results on general LVLM, high-resolution, and visual grounding benchmarks validate the effectiveness of Ground-R1 and show that SRPO yields consistent gains over standard GRPO in both response accuracy and evidence grounding.

Method: Proposes SurgVidLM, a video language model trained on SVU-31K dataset (31K video-instruction pairs). Uses two-stage StageFocus mechanism: first stage extracts global procedural context, second stage performs high-frequency local analysis guided by temporal cues. Implements Multi-frequency Fusion Attention to integrate low- and high-frequency visual tokens while preserving task-specific details.

Result: SurgVidLM significantly outperforms state-of-the-art Vid-LLMs of comparable parameter scale in both full and fine-grained video understanding tasks, demonstrating superior capability in capturing context of complex robot-assisted surgeries.

Conclusion: SurgVidLM successfully bridges the gap in fine-grained surgical video comprehension, offering a comprehensive solution for surgical scene understanding that combines global context with detailed local analysis, with potential applications in surgical training and robotic decision-making.

Abstract: Surgical scene understanding is critical for surgical training and robotic decision-making in robot-assisted surgery. Recent advances in Multimodal Large Language Models (MLLMs) have demonstrated great potential for advancing scene perception in the medical domain, facilitating surgeons to understand surgical scenes and procedures. However, these methods are primarily oriented towards image-based analysis or global video understanding, overlooking the fine-grained video reasoning that is crucial for analyzing specific processes and capturing detailed task execution within a surgical procedure. To bridge this gap, we propose SurgVidLM, the first video language model designed to address both full and fine-grained surgical video comprehension. To train our SurgVidLM, we construct the SVU-31K that is a large-scale dataset with over 31K video-instruction pairs, enabling both holistic understanding and detailed analysis of surgical procedures. Building on this resource, SurgVidLM incorporates a two-stage StageFocus mechanism: the first stage extracts global procedural context, while the second stage performs high-frequency local analysis guided by temporal cues. We also develop the Multi-frequency Fusion Attention to effectively integrate low- and high-frequency visual tokens, ensuring the preservation of critical task-specific details. Experimental results demonstrate that SurgVidLM significantly outperforms state-of-the-art Vid-LLMs of comparable parameter scale in both full and fine-grained video understanding tasks, showcasing its superior capability in capturing the context of complex robot-assisted surgeries. Our code and dataset will be publicly accessible soon.

Method: Proposes a lightweight RGB-T (RGB-Thermal) tracker built on MobileViT architecture with progressive fusion framework. Uses separable mixed attention to model intra- and inter-modal interactions, creating compact and effective features for accurate localization.

Result: Achieves under 4M parameters with real-time performance of 25.7 FPS on CPU and 122 FPS on GPU. First MobileViT-based multimodal tracker, supporting embedded and mobile platforms.

Conclusion: The proposed tracker successfully addresses the need for lightweight, real-time multimodal tracking by combining MobileViT efficiency with progressive fusion and attention mechanisms for RGB-T data.

Abstract: Single-modality tracking (RGB-only) struggles under low illumination, weather, and occlusion. Multimodal tracking addresses this by combining complementary cues. While Vision Transformer-based trackers achieve strong accuracy, they are often too large for real-time. We propose a lightweight RGB-T tracker built on MobileViT with a progressive fusion framework that models intra- and inter-modal interactions using separable mixed attention. This design delivers compact, effective features for accurate localization, with under 4M parameters and real-time performance of 25.7 FPS on the CPU and 122 FPS on the GPU, supporting embedded and mobile platforms. To the best of our knowledge, this is the first MobileViT-based multimodal tracker. Model code and weights are available in the GitHub repository.

Method: Three key components: 1) Multimodal Identity Fusion (MIF) module using Q-Former to unify visual and textual cues, 2) Time-Aware Identity Injection (TAII) for dynamic conditioning across denoising steps, and 3) Adaptive Motion Learning (AML) using optical-flow-derived motion heatmaps to enhance motion realism.

Result: Outperforms prior methods in identity preservation, text alignment, and motion quality, establishing a new benchmark for video identity customization. Also introduces Proteus-Bench dataset with 200K training clips and 150 diverse individuals for evaluation.

Conclusion: Proteus-ID effectively addresses identity consistency and motion coherence challenges in video customization through multimodal fusion, adaptive conditioning, and motion-aware training.

Abstract: Video identity customization seeks to synthesize realistic, temporally coherent videos of a specific subject, given a single reference image and a text prompt. This task presents two core challenges: (1) maintaining identity consistency while aligning with the described appearance and actions, and (2) generating natural, fluid motion without unrealistic stiffness. To address these challenges, we introduce Proteus-ID, a novel diffusion-based framework for identity-consistent and motion-coherent video customization. First, we propose a Multimodal Identity Fusion (MIF) module that unifies visual and textual cues into a joint identity representation using a Q-Former, providing coherent guidance to the diffusion model and eliminating modality imbalance. Second, we present a Time-Aware Identity Injection (TAII) mechanism that dynamically modulates identity conditioning across denoising steps, improving fine-detail reconstruction. Third, we propose Adaptive Motion Learning (AML), a self-supervised strategy that reweights the training loss based on optical-flow-derived motion heatmaps, enhancing motion realism without requiring additional inputs. To support this task, we construct Proteus-Bench, a high-quality dataset comprising 200K curated clips for training and 150 individuals from diverse professions and ethnicities for evaluation. Extensive experiments demonstrate that Proteus-ID outperforms prior methods in identity preservation, text alignment, and motion quality, establishing a new benchmark for video identity customization. Codes and data are publicly available at https://grenoble-zhang.github.io/Proteus-ID/.

Method: Proposes a 4D video generation model that enforces multi-view 3D consistency by supervising the model with cross-view pointmap alignment during training. This geometric supervision helps the model learn a shared 3D scene representation, enabling generation of spatio-temporally aligned future video sequences from novel viewpoints given only single RGB-D images per view, without requiring camera poses as input.

Result: The method produces more visually stable and spatially aligned predictions across multiple simulated and real-world robotic datasets compared to existing baselines. The predicted 4D videos can be used to recover robot end-effector trajectories using an off-the-shelf 6DoF pose tracker, yielding robot manipulation policies that generalize well to novel camera viewpoints.

Conclusion: The proposed 4D video generation model with geometric supervision enables generation of consistent multi-view videos for robotic applications, improving visual stability and spatial alignment while supporting downstream tasks like trajectory recovery and manipulation policy generalization.

Abstract: Understanding and predicting dynamics of the physical world can enhance a robot’s ability to plan and interact effectively in complex environments. While recent video generation models have shown strong potential in modeling dynamic scenes, generating videos that are both temporally coherent and geometrically consistent across camera views remains a significant challenge. To address this, we propose a 4D video generation model that enforces multi-view 3D consistency of generated videos by supervising the model with cross-view pointmap alignment during training. Through this geometric supervision, the model learns a shared 3D scene representation, enabling it to generate spatio-temporally aligned future video sequences from novel viewpoints given a single RGB-D image per view, and without relying on camera poses as input. Compared to existing baselines, our method produces more visually stable and spatially aligned predictions across multiple simulated and real-world robotic datasets. We further show that the predicted 4D videos can be used to recover robot end-effector trajectories using an off-the-shelf 6DoF pose tracker, yielding robot manipulation policies that generalize well to novel camera viewpoints.

Method: Large experimental study examining architectural choices (late fusion, channel stacking, space-to-depth projections, cross-attention) and their impact on emergent relative pose estimation from navigation training. Analyzes simulator settings and their influence on shortcuts.

Result: Success of recent methods is influenced by simulator settings leading to shortcuts, but capabilities can be transferred to more realistic settings to some extent. Found correlations between navigation performance and emergent relative pose estimation performance.

Conclusion: End-to-end RL training can develop relative pose estimation from navigation alone, though simulator artifacts affect results; architectural choices play important role in emergent navigation skills.

Abstract: Image goal navigation requires two different skills: firstly, core navigation skills, including the detection of free space and obstacles, and taking decisions based on an internal representation; and secondly, computing directional information by comparing visual observations to the goal image. Current state-of-the-art methods either rely on dedicated image-matching, or pre-training of computer vision modules on relative pose estimation. In this paper, we study whether this task can be efficiently solved with end-to-end training of full agents with RL, as has been claimed by recent work. A positive answer would have impact beyond Embodied AI and allow training of relative pose estimation from reward for navigation alone. In this large experimental study we investigate the effect of architectural choices like late fusion, channel stacking, space-to-depth projections and cross-attention, and their role in the emergence of relative pose estimators from navigation training. We show that the success of recent methods is influenced up to a certain extent by simulator settings, leading to shortcuts in simulation. However, we also show that these capabilities can be transferred to more realistic setting, up to some extent. We also find evidence for correlations between navigation performance and probed (emerging) relative pose estimation performance, an important sub skill.

Method: Multi-stage training-free approach: (1) memory bank construction from reference images, (2) representation aggregation, and (3) semantic-aware feature matching using foundation model correspondences between reference and target images.

Result: State-of-the-art performance on COCO FSOD (36.8% nAP), PASCAL VOC Few-Shot (71.2% nAP50), and outperforms existing training-free approaches on Cross-Domain FSOD benchmark (22.4% nAP).

Conclusion: Foundation model semantic priors enable effective few-shot segmentation without training, reducing the prompt burden of SAM while maintaining strong performance across benchmarks.

Abstract: The performance of image segmentation models has historically been constrained by the high cost of collecting large-scale annotated data. The Segment Anything Model (SAM) alleviates this original problem through a promptable, semantics-agnostic, segmentation paradigm and yet still requires manual visual-prompts or complex domain-dependent prompt-generation rules to process a new image. Towards reducing this new burden, our work investigates the task of object segmentation when provided with, alternatively, only a small set of reference images. Our key insight is to leverage strong semantic priors, as learned by foundation models, to identify corresponding regions between a reference and a target image. We find that correspondences enable automatic generation of instance-level segmentation masks for downstream tasks and instantiate our ideas via a multi-stage, training-free method incorporating (1) memory bank construction; (2) representation aggregation and (3) semantic-aware feature matching. Our experiments show significant improvements on segmentation metrics, leading to state-of-the-art performance on COCO FSOD (36.8% nAP), PASCAL VOC Few-Shot (71.2% nAP50) and outperforming existing training-free approaches on the Cross-Domain FSOD benchmark (22.4% nAP).

Method: Proposes Affine-Equivariant Kernel Space Encoding (EKS) that aggregates features through a field of anisotropic Gaussian kernels defining localized regions of influence. Includes feature distillation mechanism transferring information from multi-resolution hash grid encodings into the kernel field for compact, grid-free representation.

Result: Enables intuitive, localized scene editing directly via Gaussian kernels without retraining while maintaining high-quality rendering. Provides stable feature interpolation under spatial transformations while preserving continuity.

Conclusion: EKS offers a novel spatial encoding for neural radiance fields that bridges the gap between high-fidelity reconstruction and flexible, localized editability through kernel-based representations.

Abstract: Neural scene representations achieve high-fidelity rendering by encoding 3D scenes as continuous functions, but their latent spaces are typically implicit and globally entangled, making localized editing and physically grounded manipulation difficult. While several works introduce explicit control structures or point-based latent representations to improve editability, these approaches often suffer from limited locality, sensitivity to deformations, or visual artifacts. In this paper, we introduce Affine-Equivariant Kernel Space Encoding (EKS), a spatial encoding for neural radiance fields that provides localized, deformation-aware feature representations. Instead of querying latent features directly at discrete points or grid vertices, our encoding aggregates features through a field of anisotropic Gaussian kernels, each defining a localized region of influence. This kernel-based formulation enables stable feature interpolation under spatial transformations while preserving continuity and high reconstruction quality. To preserve detail without sacrificing editability, we further propose a training-time feature distillation mechanism that transfers information from multi-resolution hash grid encodings into the kernel field, yielding a compact and fully grid-free representation at inference. This enables intuitive, localized scene editing directly via Gaussian kernels without retraining, while maintaining high-quality rendering. The code can be found under (https://github.com/MikolajZielinski/eks)

Method: Proposes DSKC framework with: 1) Domain-Style Encoder (DSE) to dynamically model domain-specific styles, and 2) Unified Knowledge Consolidation (UKC) mechanism to integrate instance-level representations with domain-specific styles into cross-domain unified representations.

Result: Outperforms state-of-the-art methods in two training orders and enhances model’s strong performance, demonstrating effectiveness in mitigating forgetting and improving generalization.

Conclusion: DSKC provides an effective rehearsal-free and distillation-free solution for lifelong person re-identification by explicitly modeling inter-domain associations at both instance and domain levels.

Abstract: Lifelong Person Re-identification (LReID) aims to continuously match individuals across camera views from sequential data streams. Existing LReID methods often ignore domain-specific style awareness and unified knowledge consolidation, which are crucial for mitigating forgetting when adapting to new information. We propose DSKC, a novel rehearsal-free and distillation-free framework for LReID. DSKC designs a domain-style encoder (DSE) to dynamically model domain-specific styles, and a unified knowledge consolidation (UKC) mechanism to adaptively integrate instance-level representations with domain-specific style into a cross-domain unified representation. By leveraging unified representation as a bridge, DSKC explicitly models inter-domain associations at both instance and domain levels to enhance anti-forgetting and generalization. Experimental results demonstrate that our DSKC outperforms state-of-the-art methods in two training orders and enhances the model’s strong performance. Our code is available at https://github.com/LiuShiBen/DKUA.

Method: 1) Proposes Category-Discriminative Visual Captioner (CDV-Captioner) using MLLMs with chain-of-thought prompting and visually similar reference images to generate structured text descriptions capturing fine-grained attribute features. 2) Converts images to image-description pairs for comprehensive representation. 3) Constructs multimodal category templates from few-shot samples. 4) Uses off-the-shelf vision and text encoders to embed queries and templates. 5) Performs FGVC by retrieving nearest template in joint multimodal space.

Result: Extensive experiments on 12 FGVC benchmarks show consistent superiority over prior few-shot CLIP-based methods and even several fully-supervised MLLMs-based approaches.

Conclusion: UniFGVC provides a universal training-free framework that leverages MLLMs’ open-world knowledge for few-shot FGVC, offering broad compatibility, reliable generalization, and adaptability across diverse scenarios without requiring fine-tuning.

Abstract: Few-shot fine-grained visual classification (FGVC) aims to leverage limited data to enable models to discriminate subtly distinct categories. Recent works mostly finetuned the pre-trained visual language models to achieve performance gain, yet suffering from overfitting and weak generalization. To deal with this, we introduce UniFGVC, a universal training-free framework that reformulates few-shot FGVC as multimodal retrieval. First, we propose the Category-Discriminative Visual Captioner (CDV-Captioner) to exploit the open-world knowledge of multimodal large language models (MLLMs) to generate a structured text description that captures the fine-grained attribute features distinguishing closely related classes. CDV-Captioner uses chain-of-thought prompting and visually similar reference images to reduce hallucination and enhance discrimination of generated captions. Using it we can convert each image into an image-description pair, enabling more comprehensive feature representation, and construct the multimodal category templates using few-shot samples for the subsequent retrieval pipeline. Then, off-the-shelf vision and text encoders embed query and template pairs, and FGVC is accomplished by retrieving the nearest template in the joint space. UniFGVC ensures broad compatibility with diverse MLLMs and encoders, offering reliable generalization and adaptability across few-shot FGVC scenarios. Extensive experiments on 12 FGVC benchmarks demonstrate its consistent superiority over prior few-shot CLIP-based methods and even several fully-supervised MLLMs-based approaches.

Method: Proposes Object Fidelity Diffusion (OF-Diff) with three key components: 1) First extraction of object shape priors from layouts for remote sensing diffusion models, 2) Dual-branch diffusion model with diffusion consistency loss that generates high-fidelity images without real images during sampling, 3) DDPO fine-tuning to enhance diversity and semantic consistency.

Result: OF-Diff outperforms state-of-the-art methods across key quality metrics. Significant improvements for polymorphic and small object classes: mAP increases by 8.3% for airplanes, 7.7% for ships, and 4.0% for vehicles.

Conclusion: OF-Diff effectively improves object fidelity in remote sensing image generation through shape priors, dual-branch architecture, and reinforcement learning fine-tuning, demonstrating superior performance for challenging object classes.

Abstract: High-precision controllable remote sensing image generation is both meaningful and challenging. Existing diffusion models often produce low-fidelity images due to their inability to adequately capture morphological details, which may affect the robustness and reliability of object detection models. To enhance the accuracy and fidelity of generated objects in remote sensing, this paper proposes Object Fidelity Diffusion (OF-Diff), which effectively improves the fidelity of generated objects. Specifically, we are the first to extract the prior shapes of objects based on the layout for diffusion models in remote sensing. Then, we introduce a dual-branch diffusion model with diffusion consistency loss, which can generate high-fidelity remote sensing images without providing real images during the sampling phase. Furthermore, we introduce DDPO to fine-tune the diffusion process, making the generated remote sensing images more diverse and semantically consistent. Comprehensive experiments demonstrate that OF-Diff outperforms state-of-the-art methods in the remote sensing across key quality metrics. Notably, the performance of several polymorphic and small object classes shows significant improvement. For instance, the mAP increases by 8.3%, 7.7%, and 4.0% for airplanes, ships, and vehicles, respectively.

Method: Generates explicit correspondence maps from user drag inputs as reliable references to boost attention control, enabling stable full-strength inversion without test-time optimization, and supports multi-round workflows with simultaneous move and scale operations.

Result: Outperforms baselines on DragBench in drag accuracy and perceptual quality (validated by VIEScore and human evaluation), enables complex edits like opening a dog’s mouth with interior inpainting, generating new objects, and context-aware changes.

Conclusion: LazyDrag establishes new state-of-the-art performance in drag-based editing, eliminates the need for test-time optimization, unifies geometric control with text guidance, and paves a new way for editing paradigms in multimodal diffusion models.

Abstract: The reliance on implicit point matching via attention has become a core bottleneck in drag-based editing, resulting in a fundamental compromise on weakened inversion strength and costly test-time optimization (TTO). This compromise severely limits the generative capabilities of diffusion models, suppressing high-fidelity inpainting and text-guided creation. In this paper, we introduce LazyDrag, the first drag-based image editing method for Multi-Modal Diffusion Transformers, which directly eliminates the reliance on implicit point matching. In concrete terms, our method generates an explicit correspondence map from user drag inputs as a reliable reference to boost the attention control. This reliable reference opens the potential for a stable full-strength inversion process, which is the first in the drag-based editing task. It obviates the necessity for TTO and unlocks the generative capability of models. Therefore, LazyDrag naturally unifies precise geometric control with text guidance, enabling complex edits that were previously out of reach: opening the mouth of a dog and inpainting its interior, generating new objects like a ``tennis ball’’, or for ambiguous drags, making context-aware changes like moving a hand into a pocket. Additionally, LazyDrag supports multi-round workflows with simultaneous move and scale operations. Evaluated on the DragBench, our method outperforms baselines in drag accuracy and perceptual quality, as validated by VIEScore and human evaluation. LazyDrag not only establishes new state-of-the-art performance, but also paves a new way to editing paradigms.

Method: DiffVL treats visual localization as a GPS denoising task using diffusion models. The framework learns to reverse GPS noise perturbations by jointly modeling GPS, SD maps, and visual BEV features. Unlike traditional BEV-matching methods, it uses diffusion refinement to recover true pose distribution from noisy GPS trajectories conditioned on visual and map signals.

Result: Achieves state-of-the-art accuracy compared to BEV-matching baselines, with sub-meter accuracy without relying on HD maps. Demonstrates effectiveness on multiple datasets, proving diffusion models can enable scalable localization.

Conclusion: Diffusion models can enable scalable visual localization by treating noisy GPS as a generative prior, representing a paradigm shift from traditional matching-based methods. The approach successfully leverages all available signals (GPS, SD maps, visual features) to achieve high-precision localization without HD maps.

Abstract: Accurate visual localization is crucial for autonomous driving, yet existing methods face a fundamental dilemma: While high-definition (HD) maps provide high-precision localization references, their costly construction and maintenance hinder scalability, which drives research toward standard-definition (SD) maps like OpenStreetMap. Current SD-map-based approaches primarily focus on Bird’s-Eye View (BEV) matching between images and maps, overlooking a ubiquitous signal-noisy GPS. Although GPS is readily available, it suffers from multipath errors in urban environments. We propose DiffVL, the first framework to reformulate visual localization as a GPS denoising task using diffusion models. Our key insight is that noisy GPS trajectory, when conditioned on visual BEV features and SD maps, implicitly encode the true pose distribution, which can be recovered through iterative diffusion refinement. DiffVL, unlike prior BEV-matching methods (e.g., OrienterNet) or transformer-based registration approaches, learns to reverse GPS noise perturbations by jointly modeling GPS, SD map, and visual signals, achieving sub-meter accuracy without relying on HD maps. Experiments on multiple datasets demonstrate that our method achieves state-of-the-art accuracy compared to BEV-matching baselines. Crucially, our work proves that diffusion models can enable scalable localization by treating noisy GPS as a generative prior-making a paradigm shift from traditional matching-based methods.

Method: Developed a 3D MRI-specific vision-language foundation model using 200,000 MRI series from over 22,000 studies. Integrated self-supervised vision learning with report-guided text supervision. Features modular design with frozen pretrained encoder and lightweight task-specific decoders.

Result: Demonstrated consistent improvements over existing foundation models and task-specific approaches across disease classification, demographic prediction, anatomical localization, and cross-modal retrieval tasks.

Conclusion: Decipher-MR serves as a versatile foundation for MRI-based AI in clinical and research settings, addressing the limitations of current approaches for medical imaging.

Abstract: Magnetic Resonance Imaging is a critical imaging modality in clinical diagnosis and research, yet its complexity and heterogeneity hinder scalable, generalizable machine learning. Although foundation models have revolutionized language and vision tasks, their application to MRI remains constrained by data scarcity and narrow anatomical focus. We present Decipher-MR, a 3D MRI-specific vision-language foundation model trained on 200,000 MRI series from over 22,000 studies spanning diverse anatomical regions, sequences, and pathologies. Decipher-MR integrates self-supervised vision learning with report-guided text supervision to build robust representations for broad applications. To enable efficient use, Decipher-MR supports a modular design that enables tuning of lightweight, task-specific decoders attached to a frozen pretrained encoder. Following this setting, we evaluate Decipher-MR across disease classification, demographic prediction, anatomical localization, and cross-modal retrieval, demonstrating consistent improvements over existing foundation models and task-specific approaches. These results position Decipher-MR as a versatile foundation for MRI-based AI in clinical and research settings.

Method: Conducted systematic study of spatial bias through controlled probing experiments. Proposed Adaptive Global Context Injection (AGCI) - a lightweight mechanism that dynamically injects shared global visual context into each image token without architectural modifications, enhancing semantic accessibility while preserving model capabilities.

Result: Found current LVLMs produce inconsistent outputs under spatial shifts, revealing clear spatial bias. Analysis showed bias stems from attention mismatch between vision encoder and LLM, not from vision encoder itself. AGCI effectively mitigates spatial bias while improving performance on downstream tasks and hallucination benchmarks.

Conclusion: LVLMs have significant spatial bias affecting consistency, which can be addressed through global context injection. AGCI provides a lightweight solution that enhances spatial robustness without compromising model performance, offering insights for more robust multimodal understanding.

Abstract: Large Vision-Language Models (LVLMs) have achieved remarkable success across a wide range of multimodal tasks, yet their robustness to spatial variations remains insufficiently understood. In this work, we conduct a systematic study of the spatial bias of LVLMs, examining how models respond when identical key visual information is placed at different locations within an image. Through controlled probing experiments, we observe that current LVLMs often produce inconsistent outputs under such spatial shifts, revealing a clear spatial bias in their semantic understanding. Further analysis indicates that this bias does not stem from the vision encoder, but rather from a mismatch in attention mechanisms between the vision encoder and the large language model, which disrupts the global information flow. Motivated by this insight, we propose Adaptive Global Context Injection (AGCI), a lightweight mechanism that dynamically injects shared global visual context into each image token. AGCI works without architectural modifications, mitigating spatial bias by enhancing the semantic accessibility of image tokens while preserving the model’s intrinsic capabilities. Extensive experiments demonstrate that AGCI not only enhances the spatial robustness of LVLMs, but also achieves strong performance on various downstream tasks and hallucination benchmarks.

Method: Proposes an entropy-aware variance optimized method (EVODiff) that systematically reduces uncertainty by optimizing conditional entropy during denoising, based on information-theoretic analysis of reverse transitions.

Result: EVODiff outperforms SOTA gradient-based solvers: reduces reconstruction error by 45.5% on CIFAR-10 at 10 NFE, cuts NFE cost by 25% on ImageNet-256, and improves text-to-image generation with fewer artifacts.

Conclusion: Information-theoretic perspective provides fundamental insights for diffusion model inference; optimizing conditional entropy leads to significant improvements in efficiency and quality.

Abstract: Diffusion models (DMs) excel in image generation but suffer from slow inference and training-inference discrepancies. Although gradient-based solvers for DMs accelerate denoising inference, they often lack theoretical foundations in information transmission efficiency. In this work, we introduce an information-theoretic perspective on the inference processes of DMs, revealing that successful denoising fundamentally reduces conditional entropy in reverse transitions. This principle leads to our key insights into the inference processes: (1) data prediction parameterization outperforms its noise counterpart, and (2) optimizing conditional variance offers a reference-free way to minimize both transition and reconstruction errors. Based on these insights, we propose an entropy-aware variance optimized method for the generative process of DMs, called EVODiff, which systematically reduces uncertainty by optimizing conditional entropy during denoising. Extensive experiments on DMs validate our insights and demonstrate that our method significantly and consistently outperforms state-of-the-art (SOTA) gradient-based solvers. For example, compared to the DPM-Solver++, EVODiff reduces the reconstruction error by up to 45.5% (FID improves from 5.10 to 2.78) at 10 function evaluations (NFE) on CIFAR-10, cuts the NFE cost by 25% (from 20 to 15 NFE) for high-quality samples on ImageNet-256, and improves text-to-image generation while reducing artifacts. Code is available at https://github.com/ShiguiLi/EVODiff.

Method: Proposes AT-BPTT with three components: 1) probabilistic stage-aware timestep selection, 2) adaptive window sizing based on gradient variation, and 3) low-rank Hessian approximation to reduce computational overhead.

Result: Achieves state-of-the-art performance on CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K, improving accuracy by average 6.16% over baselines, accelerating inner-loop optimization by 3.9x while saving 63% memory cost.

Conclusion: AT-BPTT effectively addresses limitations of random truncation in dataset distillation by dynamically adapting to neural network learning dynamics, achieving superior performance and efficiency.

Abstract: The growing demand for efficient deep learning has positioned dataset distillation as a pivotal technique for compressing training dataset while preserving model performance. However, existing inner-loop optimization methods for dataset distillation typically rely on random truncation strategies, which lack flexibility and often yield suboptimal results. In this work, we observe that neural networks exhibit distinct learning dynamics across different training stages-early, middle, and late-making random truncation ineffective. To address this limitation, we propose Automatic Truncated Backpropagation Through Time (AT-BPTT), a novel framework that dynamically adapts both truncation positions and window sizes according to intrinsic gradient behavior. AT-BPTT introduces three key components: (1) a probabilistic mechanism for stage-aware timestep selection, (2) an adaptive window sizing strategy based on gradient variation, and (3) a low-rank Hessian approximation to reduce computational overhead. Extensive experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K show that AT-BPTT achieves state-of-the-art performance, improving accuracy by an average of 6.16% over baseline methods. Moreover, our approach accelerates inner-loop optimization by 3.9x while saving 63% memory cost.

Method: Uses test-time training (TTT) as a streaming memory mechanism to transform short-term multimodal representations into long-term memory embedded in model parameters. Includes TTT_MEM layer with long-span prediction objective, two-stage training scheme, and modality-aware memory reader. Processes over 3-hour videos at 1 FPS and 360p resolution.

Result: Outperforms both streaming and non-streaming baselines by 3-7% on long video benchmarks. Achieves 15% absolute accuracy improvement over strong non-streaming models on the new ELViM benchmark, demonstrating strong learning abilities from video memory.

Conclusion: video-SALMONN S represents an advance in long-duration streaming video understanding with effective memory mechanisms, showing promise for AI agents that need to learn from extended video observations.

Abstract: Long-duration streaming video understanding is fundamental for future AI agents, yet remains limited by ineffective long-term memory. We introduce video-SALMONN S, a memory-enhanced streaming audio-visual large language model that processes over 3-hour videos at 1 FPS and 360p resolution, outperforming strong non-streaming models under the same memory budget. In addition to token merging or downsampling, video-SALMONN S is the first to employ test-time training (TTT) as a streaming memory mechanism for video understanding. TTT continuously transforms short-term multimodal representations into long-term memory embedded in model parameters. To improve long-range dependency modeling and memory capacity, we propose (i) a TTT_MEM layer with an additional long-span prediction objective, (ii) a two-stage training scheme, and (iii) a modality-aware memory reader. We further introduce the Episodic Learning from Video Memory (ELViM) benchmark, simulating agent-like scenarios where models must learn from videos observed hours earlier. video-SALMONN S consistently outperforms both streaming and non-streaming baselines by 3-7% on long video benchmarks. Notably, video-SALMONN S achieves a 15% absolute accuracy improvement over strong non-streaming models on ELViM, demonstrating strong learning abilities from video memory.

Method: SAIL-RL uses reinforcement learning post-training with a dual reward system: 1) Thinking Reward evaluates reasoning quality through factual grounding, logical coherence, and answer consistency; 2) Judging Reward adaptively determines whether deep reasoning or direct answering is appropriate for each task.

Result: Experiments on SAIL-VL2 show SAIL-RL improves reasoning and multimodal understanding benchmarks at both 4B and 8B scales, achieves competitive performance against GPT-4o, and substantially reduces hallucinations.

Conclusion: SAIL-RL establishes a principled framework for building more reliable and adaptive multimodal large language models by teaching them when and how to think through reinforcement learning.

Abstract: We introduce SAIL-RL, a reinforcement learning (RL) post-training framework that enhances the reasoning capabilities of multimodal large language models (MLLMs) by teaching them when and how to think. Existing approaches are limited by outcome-only supervision, which rewards correct answers without ensuring sound reasoning, and by uniform thinking strategies, which often lead to overthinking on simple tasks and underthinking on complex ones. SAIL-RL addresses these challenges with a dual reward system: the Thinking Reward, which evaluates reasoning quality through factual grounding, logical coherence, and answer consistency, and the Judging Reward, which adaptively determines whether deep reasoning or direct answering is appropriate. Experiments on the state-of-the-art SAIL-VL2 show that SAIL-RL improves reasoning and multimodal understanding benchmarks at both 4B and 8B scales, achieving competitive performance against commercial closed-source models such as GPT-4o, and substantially reduces hallucinations, establishing it as a principled framework for building more reliable and adaptive MLLMs. The code will be available at https://github.com/BytedanceDouyinContent/SAIL-RL.

Method: Proposes UniADC with two key components: 1) Training-free Controllable Inpainting Network that synthesizes anomaly images by repainting normal regions guided by anomaly priors, and augments few-shot anomaly samples; 2) Implicit-Normal Discriminator that addresses class imbalance by implicitly modeling normal state and aligning fine-grained image features with anomaly-category embeddings.

Result: Extensive experiments on MVTec-FS, MTD, and WFDD datasets show UniADC consistently outperforms existing methods in anomaly detection, localization, and classification tasks.

Conclusion: UniADC effectively unifies anomaly detection and classification, achieving superior performance with few or no anomaly images by leveraging the proposed inpainting network and implicit-normal discriminator components.

Abstract: In this paper, we introduce a novel task termed unified anomaly detection and classification, which aims to simultaneously detect anomalous regions in images and identify their specific categories. Existing methods typically treat anomaly detection and classification as separate tasks, thereby neglecting their inherent correlations and limiting information sharing, which results in suboptimal performance. To address this, we propose UniADC, a model designed to effectively perform both tasks with only a few or even no anomaly images. Specifically, UniADC consists of two key components: a training-free Controllable Inpainting Network and an Implicit-Normal Discriminator. The inpainting network can synthesize anomaly images of specific categories by repainting normal regions guided by anomaly priors, and can also repaint few-shot anomaly samples to augment the available anomaly data. The implicit-normal discriminator addresses the severe challenge of the imbalance between normal and anomalous pixel distributions by implicitly modeling the normal state, achieving precise anomaly detection and classification by aligning fine-grained image features with anomaly-category embeddings. We conduct extensive experiments on three anomaly detection and classification datasets, including MVTec-FS, MTD, and WFDD, and the results demonstrate that UniADC consistently outperforms existing methods in anomaly detection, localization, and classification. The code is available at https://github.com/cnulab/UniADC.

Method: Proposes Missing-aware Mixture-of-LoRAs (MaMOL), a parameter-efficient MoE framework with dual-routing mechanism that decouples modality-invariant shared experts and modality-aware dynamic experts, enabling automatic expert activation conditioned on available modalities.

Result: Extensive experiments on multiple remote sensing benchmarks show MaMOL significantly improves robustness and generalization under diverse missing-modality scenarios with minimal computational overhead. Transfer experiments on natural image datasets validate scalability and cross-domain applicability.

Conclusion: MaMOL effectively addresses missing-modality challenges in multimodal classification through a unified parameter-efficient framework that maintains performance while being computationally efficient and transferable across domains.

Abstract: Multimodal remote sensing classification often suffers from missing modalities caused by sensor failures and environmental interference, leading to severe performance degradation. In this work, we rethink missing-modality learning from a conditional computation perspective and investigate whether Mixture-of-Experts (MoE) models can inherently adapt to diverse modality-missing scenarios. We first conduct a systematic study of representative MoE paradigms under various missing-modality settings, revealing both their potential and limitations. Building on these insights, we propose a Missing-aware Mixture-of-LoRAs (MaMOL), a parameter-efficient MoE framework that unifies multiple modality-missing cases within a single model. MaMOL introduces a dual-routing mechanism to decouple modality-invariant shared experts and modality-aware dynamic experts, enabling automatic expert activation conditioned on available modalities. Extensive experiments on multiple remote sensing benchmarks demonstrate that MaMOL significantly improves robustness and generalization under diverse missing-modality scenarios with minimal computational overhead. Transfer experiments on natural image datasets further validate its scalability and cross-domain applicability.

Method: 1) Reconstruct scenes using 3D Gaussian Splatting from multi-view images; 2) Extract semantic material masks via vision models; 3) Convert Gaussian representations to mesh surfaces with projected material labels; 4) Assign physics-based material properties for accurate sensor simulation in modern graphics engines and simulators.

Result: The camera-only approach achieves sensor simulation fidelity comparable to LiDAR-camera fusion while eliminating hardware complexity and calibration requirements. Validated using internal dataset from instrumented test vehicle with LiDAR as ground truth for reflectivity validation alongside image similarity metrics.

Conclusion: Camera-only pipeline successfully combines photorealistic reconstruction with physics-based material assignment, providing a practical alternative to LiDAR-camera fusion for Digital Twin applications with reduced hardware complexity.

Abstract: 3D reconstruction for Digital Twins often relies on LiDAR-based methods, which provide accurate geometry but lack the semantics and textures naturally captured by cameras. Traditional LiDAR-camera fusion approaches require complex calibration and still struggle with certain materials like glass, which are visible in images but poorly represented in point clouds. We propose a camera-only pipeline that reconstructs scenes using 3D Gaussian Splatting from multi-view images, extracts semantic material masks via vision models, converts Gaussian representations to mesh surfaces with projected material labels, and assigns physics-based material properties for accurate sensor simulation in modern graphics engines and simulators. This approach combines photorealistic reconstruction with physics-based material assignment, providing sensor simulation fidelity comparable to LiDAR-camera fusion while eliminating hardware complexity and calibration requirements. We validate our camera-only method using an internal dataset from an instrumented test vehicle, leveraging LiDAR as ground truth for reflectivity validation alongside image similarity metrics.

Method: Introduces frequency-domain masking as a training strategy, using random masking and geometric transformations with focus on frequency masking for superior generalization properties.

Result: Achieves state-of-the-art generalization on GAN- and diffusion-generated image datasets, maintains performance under significant model pruning, offers scalable resource-conscious solution.

Conclusion: Frequency-based masking is a practical step toward sustainable and generalizable deepfake detection with strong generalization capabilities and computational efficiency.

Abstract: Universal deepfake detection aims to identify AI-generated images across a broad range of generative models, including unseen ones. This requires robust generalization to new and unseen deepfakes, which emerge frequently, while minimizing computational overhead to enable large-scale deepfake screening, a critical objective in the era of Green AI. In this work, we explore frequency-domain masking as a training strategy for deepfake detectors. Unlike traditional methods that rely heavily on spatial features or large-scale pretrained models, our approach introduces random masking and geometric transformations, with a focus on frequency masking due to its superior generalization properties. We demonstrate that frequency masking not only enhances detection accuracy across diverse generators but also maintains performance under significant model pruning, offering a scalable and resource-conscious solution. Our method achieves state-of-the-art generalization on GAN- and diffusion-generated image datasets and exhibits consistent robustness under structured pruning. These results highlight the potential of frequency-based masking as a practical step toward sustainable and generalizable deepfake detection. Code and models are available at https://github.com/chandlerbing65nm/FakeImageDetection.

Method: Created first multicenter lymphoma benchmarking dataset covering four common subtypes and healthy tissue. Evaluated five pathology foundation models (H-optimus-1, H0-mini, Virchow2, UNI2, Titan) with attention-based (AB-MIL) and transformer-based (TransMIL) multiple instance learning aggregators across three magnifications (10x, 20x, 40x).

Result: On in-distribution test sets: models achieved >80% multiclass balanced accuracy across all magnifications, with all foundation models performing similarly and both aggregation methods comparable. 40x resolution sufficient with no gains from higher resolutions or cross-magnification aggregation. On out-of-distribution test sets: performance dropped to ~60%, highlighting significant generalization challenges.

Conclusion: Larger multicenter studies covering additional rare lymphoma subtypes are needed. Provided automated benchmarking pipeline to facilitate future research. Shows promise for AI-assisted lymphoma diagnosis but generalization remains a major challenge.

Abstract: Timely and accurate lymphoma diagnosis is essential for guiding cancer treatment. Standard diagnostic practice combines hematoxylin and eosin (HE)-stained whole slide images with immunohistochemistry, flow cytometry, and molecular genetic tests to determine lymphoma subtypes, a process requiring costly equipment, skilled personnel, and causing treatment delays. Deep learning methods could assist pathologists by extracting diagnostic information from routinely available HE-stained slides, yet comprehensive benchmarks for lymphoma subtyping on multicenter data are lacking. In this work, we present the first multicenter lymphoma benchmarking dataset covering four common lymphoma subtypes and healthy control tissue. We systematically evaluate five publicly available pathology foundation models (H-optimus-1, H0-mini, Virchow2, UNI2, Titan) combined with attention-based (AB-MIL) and transformer-based (TransMIL) multiple instance learning aggregators across three magnifications (10x, 20x, 40x). On in-distribution test sets, models achieve multiclass balanced accuracies exceeding 80% across all magnifications, with all foundation models performing similarly and both aggregation methods showing comparable results. The magnification study reveals that 40x resolution is sufficient, with no performance gains from higher resolutions or cross-magnification aggregation. However, on out-of-distribution test sets, performance drops substantially to around 60%, highlighting significant generalization challenges. To advance the field, larger multicenter studies covering additional rare lymphoma subtypes are needed. We provide an automated benchmarking pipeline to facilitate such future research.

Method: CountZES uses three inference-only stages: 1) Detection-Anchored Exemplar (DAE) refines OVD detections to isolate precise single-instance exemplars; 2) Density-Guided Exemplar (DGE) employs density-driven self-supervised paradigm to identify statistically consistent exemplars; 3) Feature-Consensus Exemplar (FCE) reinforces visual coherence through feature-space clustering.

Result: Experiments on diverse datasets demonstrate CountZES’s superior performance among zero-shot object counting methods while generalizing effectively across domains.

Conclusion: CountZES provides a robust approach for zero-shot object counting that balances textual grounding, count consistency, and feature representativeness through complementary exemplar discovery stages.

Abstract: Object counting in complex scenes is particularly challenging in the zero-shot (ZS) setting, where instances of unseen categories are counted using only a class name. Existing ZS counting methods that infer exemplars from text often rely on off-the-shelf open-vocabulary detectors (OVDs), which in dense scenes suffer from semantic noise, appearance variability, and frequent multi-instance proposals. Alternatively, random image-patch sampling is employed, which fails to accurately delineate object instances. To address these issues, we propose CountZES, an inference-only approach for object counting via ZS exemplar selection. CountZES discovers diverse exemplars through three synergistic stages: Detection-Anchored Exemplar (DAE), Density-Guided Exemplar (DGE), and Feature-Consensus Exemplar (FCE). DAE refines OVD detections to isolate precise single-instance exemplars. DGE introduces a density-driven, self-supervised paradigm to identify statistically consistent and semantically compact exemplars, while FCE reinforces visual coherence through feature-space clustering. Together, these stages yield a complementary exemplar set that balances textual grounding, count consistency, and feature representativeness. Experiments on diverse datasets demonstrate CountZES superior performance among ZOC methods while generalizing effectively across domains.

Method: UniRect incorporates various task-specific inverse problems into a general distortion model by simulating different lens types. It uses a dual-component structure: 1) Deformation Module with Residual Progressive Thin-Plate Spline (RP-TPS) for geometric deformations, and 2) Restoration Module with Residual Mamba Blocks (RMBs) to counteract degradation. A Sparse Mixture-of-Experts (SMoEs) structure handles multi-task learning with varying distortions.

Result: Extensive experiments show state-of-the-art performance compared with other up-to-date methods across various image correction tasks.

Conclusion: UniRect provides a comprehensive unified framework for image correction tasks that overcomes limitations of task-specific architectures and demonstrates strong generalization capabilities.

Abstract: Image correction and rectangling are valuable tasks in practical photography systems such as smartphones. Recent remarkable advancements in deep learning have undeniably brought about substantial performance improvements in these fields. Nevertheless, existing methods mainly rely on task-specific architectures. This significantly restricts their generalization ability and effective application across a wide range of different tasks. In this paper, we introduce the Unified Rectification Framework (UniRect), a comprehensive approach that addresses these practical tasks from a consistent distortion rectification perspective. Our approach incorporates various task-specific inverse problems into a general distortion model by simulating different types of lenses. To handle diverse distortions, UniRect adopts one task-agnostic rectification framework with a dual-component structure: a {Deformation Module}, which utilizes a novel Residual Progressive Thin-Plate Spline (RP-TPS) model to address complex geometric deformations, and a subsequent Restoration Module, which employs Residual Mamba Blocks (RMBs) to counteract the degradation caused by the deformation process and enhance the fidelity of the output image. Moreover, a Sparse Mixture-of-Experts (SMoEs) structure is designed to circumvent heavy task competition in multi-task learning due to varying distortions. Extensive experiments demonstrate that our models have achieved state-of-the-art performance compared with other up-to-date methods.

Method: Uses pretrained Vision Transformers with camera-aware register tokens to compress multi-camera features into compact scene representations. Two lightweight transformer decoders then generate candidate trajectories and score them using interpretable sub-scores (safety, comfort, efficiency) learned by mimicking an oracle.

Result: Outperforms or matches strong contemporary baselines across NAVSIM-v1, NAVSIM-v2, and HUGSIM benchmarks. Shows that pure-transformer architecture with token compression enables accurate, efficient, and adaptive driving.

Conclusion: A pure-transformer architecture with targeted token compression is sufficient for accurate, efficient, and adaptive end-to-end autonomous driving, demonstrating strong performance across multiple benchmarks.

Abstract: We present DrivoR, a simple and efficient transformer-based architecture for end-to-end autonomous driving. Our approach builds on pretrained Vision Transformers (ViTs) and introduces camera-aware register tokens that compress multi-camera features into a compact scene representation, significantly reducing downstream computation without sacrificing accuracy. These tokens drive two lightweight transformer decoders that generate and then score candidate trajectories. The scoring decoder learns to mimic an oracle and predicts interpretable sub-scores representing aspects such as safety, comfort, and efficiency, enabling behavior-conditioned driving at inference. Despite its minimal design, DrivoR outperforms or matches strong contemporary baselines across NAVSIM-v1, NAVSIM-v2, and the photorealistic closed-loop HUGSIM benchmark. Our results show that a pure-transformer architecture, combined with targeted token compression, is sufficient for accurate, efficient, and adaptive end-to-end driving. Code and checkpoints will be made available via the project page.

Method: Collaborated with DeepScenario to create DeepUrban dataset using drones at ~100m altitude over urban intersections. Collected high-resolution images to extract 3D traffic objects, enriched with comprehensive map and scene information.

Result: Adding DeepUrban to nuScenes dataset boosts vehicle prediction and planning accuracy by up to 44.1% on ADE and 44.3% on FDE metrics. The dataset enables better evaluation of generalization capabilities.

Conclusion: DeepUrban addresses the gap in dense traffic scenarios for autonomous driving benchmarks and demonstrates significant improvements in prediction and planning performance when combined with existing datasets.

Abstract: The efficacy of autonomous driving systems hinges critically on robust prediction and planning capabilities. However, current benchmarks are impeded by a notable scarcity of scenarios featuring dense traffic, which is essential for understanding and modeling complex interactions among road users. To address this gap, we collaborated with our industrial partner, DeepScenario, to develop DeepUrban-a new drone dataset designed to enhance trajectory prediction and planning benchmarks focusing on dense urban settings. DeepUrban provides a rich collection of 3D traffic objects, extracted from high-resolution images captured over urban intersections at approximately 100 meters altitude. The dataset is further enriched with comprehensive map and scene information to support advanced modeling and simulation tasks. We evaluate state-of-the-art (SOTA) prediction and planning methods, and conducted experiments on generalization capabilities. Our findings demonstrate that adding DeepUrban to nuScenes can boost the accuracy of vehicle predictions and planning, achieving improvements up to 44.1 % / 44.3% on the ADE / FDE metrics. Website: https://iv.ee.hm.edu/deepurban

Method: Proposes MOD-DiT with two-stage process: 1) Uses prior information from early denoising steps with distributed mixing approach to model linear approximation for predicting mask patterns for specific denoising intervals, 2) Implements online block masking strategy to dynamically apply predicted masks while maintaining historical sparsity information, eliminating repetitive sampling operations.

Result: Extensive evaluations show consistent acceleration and quality improvements across multiple benchmarks and model architectures, validating effectiveness for efficient, high-quality video generation while overcoming computational limitations of traditional sparse attention approaches.

Conclusion: MOD-DiT provides an effective sampling-free dynamic attention framework that addresses computational bottlenecks in video generation DiTs, enabling practical deployment through accurate modeling of evolving attention patterns without sacrificing quality.

Abstract: While Diffusion Transformers (DiTs) have achieved notable progress in video generation, this long-sequence generation task remains constrained by the quadratic complexity inherent to self-attention mechanisms, creating significant barriers to practical deployment. Although sparse attention methods attempt to address this challenge, existing approaches either rely on oversimplified static patterns or require computationally expensive sampling operations to achieve dynamic sparsity, resulting in inaccurate pattern predictions and degraded generation quality. To overcome these limitations, we propose a \underline{\textbf{M}}ixture-\underline{\textbf{O}}f-\underline{\textbf{D}}istribution \textbf{DiT} (\textbf{MOD-DiT}), a novel sampling-free dynamic attention framework that accurately models evolving attention patterns through a two-stage process. First, MOD-DiT leverages prior information from early denoising steps and adopts a {distributed mixing approach} to model an efficient linear approximation model, which is then used to predict mask patterns for a specific denoising interval. Second, an online block masking strategy dynamically applies these predicted masks while maintaining historical sparsity information, eliminating the need for repetitive sampling operations. Extensive evaluations demonstrate consistent acceleration and quality improvements across multiple benchmarks and model architectures, validating MOD-DiT’s effectiveness for efficient, high-quality video generation while overcoming the computational limitations of traditional sparse attention approaches.

Method: Introduces a Topological Feature Fusion Module (TFFM) that maps local features into latent graph space using Graph Attention Networks to capture global structural dependencies. Uses hybrid objective combining Tversky loss for class imbalance and soft clDice loss to penalize topological disconnects.

Result: Achieves state-of-the-art performance on Fundus-AVSeg dataset: 90.97% combined Dice score, 95% Hausdorff Distance of 3.50 pixels. Reduces vessel fragmentation by ~38% relative to baselines, yielding topologically coherent vascular trees for automated biomarker quantification.

Conclusion: The proposed topology-aware framework successfully maintains vascular connectivity, addressing a critical limitation of standard segmentation methods and enabling reliable clinical analysis of retinal vasculature for cardiovascular diagnosis.

Abstract: Precise segmentation of retinal arteries and veins carries the diagnosis of systemic cardiovascular conditions. However, standard convolutional architectures often yield topologically disjointed segmentations, characterized by gaps and discontinuities that render reliable graph-based clinical analysis impossible despite high pixel-level accuracy. To address this, we introduce a topology-aware framework engineered to maintain vascular connectivity. Our architecture fuses a Topological Feature Fusion Module (TFFM) that maps local feature representations into a latent graph space, deploying Graph Attention Networks to capture global structural dependencies often missed by fixed receptive fields. Furthermore, we drive the learning process with a hybrid objective function, coupling Tversky loss for class imbalance with soft clDice loss to explicitly penalize topological disconnects. Evaluation on the Fundus-AVSeg dataset reveals state-of-the-art performance, achieving a combined Dice score of 90.97% and a 95% Hausdorff Distance of 3.50 pixels. Notably, our method decreases vessel fragmentation by approximately 38% relative to baselines, yielding topologically coherent vascular trees viable for automated biomarker quantification. We open-source our code at https://tffm-module.github.io/.

Method: Uses diffusion models with creativity defined as inverse probability in CLIP embedding space. Calculates probability distribution of generated images and drives generation toward low-probability regions. Introduces pullback mechanisms to maintain visual fidelity while achieving high creativity.

Result: Extensive experiments on text-to-image diffusion models demonstrate effectiveness and efficiency in producing unique, novel, and thought-provoking images with high creativity without sacrificing visual quality.

Conclusion: Provides a new perspective on creativity in generative models with a principled method for fostering innovation in visual content synthesis through probability-based creative generation.

Abstract: Creative image generation has emerged as a compelling area of research, driven by the need to produce novel and high-quality images that expand the boundaries of imagination. In this work, we propose a novel framework for creative generation using diffusion models, where creativity is associated with the inverse probability of an image’s existence in the CLIP embedding space. Unlike prior approaches that rely on a manual blending of concepts or exclusion of subcategories, our method calculates the probability distribution of generated images and drives it towards low-probability regions to produce rare, imaginative, and visually captivating outputs. We also introduce pullback mechanisms, achieving high creativity without sacrificing visual fidelity. Extensive experiments on text-to-image diffusion models demonstrate the effectiveness and efficiency of our creative generation framework, showcasing its ability to produce unique, novel, and thought-provoking images. This work provides a new perspective on creativity in generative models, offering a principled method to foster innovation in visual content synthesis.

Method: Compared top-view depth image and point cloud data for BCS prediction under four settings: 1) unsegmented raw data, 2) segmented full-body data, 3) segmented hindquarter data, and 4) handcrafted feature data. Used data from 1,020 dairy cows with cow-level cross-validation to prevent data leakage.

Result: Depth image-based models consistently achieved higher accuracy than point cloud-based models with unsegmented raw data and segmented full-body data. Comparable performance was observed with segmented hindquarter data. Both approaches showed reduced accuracy with handcrafted features. Point cloud predictions were more sensitive to noise and model architecture.

Conclusion: 3D point clouds do not provide a consistent advantage over depth images for BCS prediction in dairy cattle under the evaluated conditions, with depth images being more robust and accurate in most settings.

Abstract: Body condition score (BCS) is a widely used indicator of body energy status and is closely associated with metabolic status, reproductive performance, and health in dairy cattle; however, conventional visual scoring is subjective and labor-intensive. Computer vision approaches have been applied to BCS prediction, with depth images widely used because they capture geometric information independent of coat color and texture. More recently, three-dimensional point cloud data have attracted increasing interest due to their ability to represent richer geometric characteristics of animal morphology, but direct head-to-head comparisons with depth image-based approaches remain limited. In this study, we compared top-view depth image and point cloud data for BCS prediction under four settings: 1) unsegmented raw data, 2) segmented full-body data, 3) segmented hindquarter data, and 4) handcrafted feature data. Prediction models were evaluated using data from 1,020 dairy cows collected on a commercial farm, with cow-level cross-validation to prevent data leakage. Depth image-based models consistently achieved higher accuracy than point cloud-based models when unsegmented raw data and segmented full-body data were used, whereas comparable performance was observed when segmented hindquarter data were used. Both depth image and point cloud approaches showed reduced accuracy when handcrafted feature data were employed compared with the other settings. Overall, point cloud-based predictions were more sensitive to noise and model architecture than depth image-based predictions. Taken together, these results indicate that three-dimensional point clouds do not provide a consistent advantage over depth images for BCS prediction in dairy cattle under the evaluated conditions.

Method: Introduced ShotFinder benchmark with 1,210 samples across 20 categories, formalizing editing requirements as keyframe-oriented shot descriptions with five controllable constraints (Temporal, Color, Visual style, Audio, Resolution). Proposed three-stage retrieval pipeline: query expansion via video imagination, candidate retrieval with search engine, and description-guided temporal localization.

Result: Experiments show significant gap to human performance, with clear imbalance across constraints: temporal localization is relatively tractable, while color and visual style remain major challenges.

Conclusion: Open-domain video shot retrieval is a critical capability that multimodal large models have yet to overcome, revealing limitations in handling complex temporal, visual, and audio constraints.

Abstract: In recent years, large language models (LLMs) have made rapid progress in information retrieval, yet existing research has mainly focused on text or static multimodal settings. Open-domain video shot retrieval, which involves richer temporal structure and more complex semantics, still lacks systematic benchmarks and analysis. To fill this gap, we introduce ShotFinder, a benchmark that formalizes editing requirements as keyframe-oriented shot descriptions and introduces five types of controllable single-factor constraints: Temporal order, Color, Visual style, Audio, and Resolution. We curate 1,210 high-quality samples from YouTube across 20 thematic categories, using large models for generation with human verification. Based on the benchmark, we propose ShotFinder, a text-driven three-stage retrieval and localization pipeline: (1) query expansion via video imagination, (2) candidate video retrieval with a search engine, and (3) description-guided temporal localization. Experiments on multiple closed-source and open-source models reveal a significant gap to human performance, with clear imbalance across constraints: temporal localization is relatively tractable, while color and visual style remain major challenges. These results reveal that open-domain video shot retrieval is still a critical capability that multimodal large models have yet to overcome.

Method: Evaluates Sparse4D (query-based spatiotemporal 3D detection/tracking) under: reduced input frame rates, post-training quantization (INT8/FP8), transfer to WILDTRACK benchmark, and Transformer Engine mixed-precision fine-tuning. Introduces Average Track Duration metric for identity stability.

Result: Sparse4D stable under moderate FPS reductions but identity association collapses below 2 FPS; selective quantization of backbone/neck offers best speed-accuracy trade-off; low-FPS pretraining yields large zero-shot gains on WILDTRACK; mixed precision reduces latency but can destabilize identity propagation.

Conclusion: Multi-camera tracking systems need stability-aware validation, selective quantization strategies, and careful consideration of frame rate thresholds for identity persistence, with mixed-precision training offering scalability benefits but requiring stability monitoring.

Abstract: Outside-in multi-camera perception is increasingly important in indoor environments, where networks of static cameras must support multi-target tracking under occlusion and heterogeneous viewpoints. We evaluate Sparse4D, a query-based spatiotemporal 3D detection and tracking framework that fuses multi-view features in a shared world frame and propagates sparse object queries via instance memory. We study reduced input frame rates, post-training quantization (INT8 and FP8), transfer to the WILDTRACK benchmark, and Transformer Engine mixed-precision fine-tuning. To better capture identity stability, we report Average Track Duration (AvgTrackDur), which measures identity persistence in seconds. Sparse4D remains stable under moderate FPS reductions, but below 2 FPS, identity association collapses even when detections are stable. Selective quantization of the backbone and neck offers the best speed-accuracy trade-off, while attention-related modules are consistently sensitive to low precision. On WILDTRACK, low-FPS pretraining yields large zero-shot gains over the base checkpoint, while small-scale fine-tuning provides limited additional benefit. Transformer Engine mixed precision reduces latency and improves camera scalability, but can destabilize identity propagation, motivating stability-aware validation.

Method: Systematic framework with three components: (1) Large-scale instruction-tuning dataset combining rewritten multimodal conversations and synthetic examples, (2) Architecture leveraging pretrained MLLM for visual understanding and DiT for visual generation, (3) Two-stage decoupled training: first instruction-tune MLLM, then align DiT with curated interleaved sequences.

Result: Outperforms prior open-source models in text quality, image fidelity, and image-context alignment; achieves state-of-the-art performance on text-to-image and image editing among unified generation models across public and new benchmarks.

Conclusion: DuoGen demonstrates that systematic data curation and architectural design can enable high-quality interleaved multimodal generation without costly unimodal pretraining, with flexible base model selection.

Abstract: Interleaved multimodal generation enables capabilities beyond unimodal generation models, such as step-by-step instructional guides, visual planning, and generating visual drafts for reasoning. However, the quality of existing interleaved generation models under general instructions remains limited by insufficient training data and base model capacity. We present DuoGen, a general-purpose interleaved generation framework that systematically addresses data curation, architecture design, and evaluation. On the data side, we build a large-scale, high-quality instruction-tuning dataset by combining multimodal conversations rewritten from curated raw websites, and diverse synthetic examples covering everyday scenarios. Architecturally, DuoGen leverages the strong visual understanding of a pretrained multimodal LLM and the visual generation capabilities of a diffusion transformer (DiT) pretrained on video generation, avoiding costly unimodal pretraining and enabling flexible base model selection. A two-stage decoupled strategy first instruction-tunes the MLLM, then aligns DiT with it using curated interleaved image-text sequences. Across public and newly proposed benchmarks, DuoGen outperforms prior open-source models in text quality, image fidelity, and image-context alignment, and also achieves state-of-the-art performance on text-to-image and image editing among unified generation models. Data and code will be released at https://research.nvidia.com/labs/dir/duogen/.

Method: Prompt an LMM to directly generate the “mental image” from the multimodal query (reference image + modification text). Generate synthetic counterparts for real database images to bridge domain gaps. Match the generated “mental image” with synthetic database images in a constructed “paracosm”.

Result: Significantly outperforms existing zero-shot methods on four challenging benchmarks, achieving state-of-the-art performance for zero-shot CIR.

Conclusion: Directly generating “mental images” for multimodal queries is more effective than generating textual descriptions, and bridging synthetic-to-real domain gaps improves matching accuracy in zero-shot CIR.

Abstract: Composed Image Retrieval (CIR) is the task of retrieving a target image from a database using a multimodal query, which consists of a reference image and a modification text. The text specifies how to alter the reference image to form a mental image'', based on which CIR should find the target image in the database. The fundamental challenge of CIR is that this mental image’’ is not physically available and is only implicitly defined by the query. The contemporary literature pursues zero-shot methods and uses a Large Multimodal Model (LMM) to generate a textual description for a given multimodal query, and then employs a Vision-Language Model (VLM) for textual-visual matching to search the target image. In contrast, we address CIR from first principles by directly generating the mental image'' for more accurate matching. Particularly, we prompt an LMM to generate a mental image’’ for a given multimodal query and propose to use this mental image'' to search for the target image. As the mental image’’ has a synthetic-to-real domain gap with real images, we also generate a synthetic counterpart for each real image in the database to facilitate matching. In this sense, our method uses LMM to construct a ``paracosm’’, where it matches the multimodal query and database images. Hence, we call this method Paracosm. Notably, Paracosm is a training-free zero-shot CIR method. It significantly outperforms existing zero-shot methods on four challenging benchmarks, achieving state-of-the-art performance for zero-shot CIR.

Method: Two data augmentation frameworks: 1) Instance Aware Automatic Augmentation (IAAA) - automated, instance-preserving augmentation method applying optimized policies to original data; 2) Semantic-Aware AI Augmentation (SAAA) - combines diffusion-based mask generator with Pix2Pix conditional image synthesizer to create diverse, anatomically realistic segmentation masks and corresponding high-fidelity images.

Result: The augmentation strategies generate synthetic images with visual and structural properties closely matching real IS data, significantly improving CAR-T/NK IS detection and segmentation performance.

Conclusion: The work enhances robustness and accuracy of IS quantification, supporting development of more reliable imaging-based biomarkers for predicting patient response to CAR-T/NK immunotherapy.

Abstract: Chimeric antigen receptor (CAR)-T and NK cell immunotherapies have transformed cancer treatment, and recent studies suggest that the quality of the CAR-T/NK cell immunological synapse (IS) may serve as a functional biomarker for predicting therapeutic efficacy. Accurate detection and segmentation of CAR-T/NK IS structures using artificial neural networks (ANNs) can greatly increase the speed and reliability of IS quantification. However, a persistent challenge is the limited size of annotated microscopy datasets, which restricts the ability of ANNs to generalize. To address this challenge, we integrate two complementary data-augmentation frameworks. First, we employ Instance Aware Automatic Augmentation (IAAA), an automated, instance-preserving augmentation method that generates synthetic CAR-T/NK IS images and corresponding segmentation masks by applying optimized augmentation policies to original IS data. IAAA supports multiple imaging modalities (e.g., fluorescence and brightfield) and can be applied directly to CAR-T/NK IS images derived from patient samples. In parallel, we introduce a Semantic-Aware AI Augmentation (SAAA) pipeline that combines a diffusion-based mask generator with a Pix2Pix conditional image synthesizer. This second method enables the creation of diverse, anatomically realistic segmentation masks and produces high-fidelity CAR-T/NK IS images aligned with those masks, further expanding the training corpus beyond what IAAA alone can provide. Together, these augmentation strategies generate synthetic images whose visual and structural properties closely match real IS data, significantly improving CAR-T/NK IS detection and segmentation performance. By enhancing the robustness and accuracy of IS quantification, this work supports the development of more reliable imaging-based biomarkers for predicting patient response to CAR-T/NK immunotherapy.

Method: PISA uses an exact-or-approximate strategy: critical attention blocks are computed exactly, while non-critical blocks are efficiently approximated using block-wise Taylor expansion based on the observation that their attention scores exhibit distributional stability.

Result: PISA achieves 1.91× speedup on Wan2.1-14B and 2.57× on Hunyuan-Video for video generation, and 1.2× speedup on FLUX for image generation, while maintaining the highest quality among sparse attention methods.

Conclusion: PISA effectively bridges the speed-quality gap in diffusion transformers by approximating rather than discarding non-critical attention blocks, making it a practical solution for efficient video and image generation.

Abstract: Diffusion Transformers are fundamental for video and image generation, but their efficiency is bottlenecked by the quadratic complexity of attention. While block sparse attention accelerates computation by attending only critical key-value blocks, it suffers from degradation at high sparsity by discarding context. In this work, we discover that attention scores of non-critical blocks exhibit distributional stability, allowing them to be approximated accurately and efficiently rather than discarded, which is essentially important for sparse attention design. Motivated by this key insight, we propose PISA, a training-free Piecewise Sparse Attention that covers the full attention span with sub-quadratic complexity. Unlike the conventional keep-or-drop paradigm that directly drop the non-critical block information, PISA introduces a novel exact-or-approximate strategy: it maintains exact computation for critical blocks while efficiently approximating the remainder through block-wise Taylor expansion. This design allows PISA to serve as a faithful proxy to full attention, effectively bridging the gap between speed and quality. Experimental results demonstrate that PISA achieves 1.91 times and 2.57 times speedups on Wan2.1-14B and Hunyuan-Video, respectively, while consistently maintaining the highest quality among sparse attention methods. Notably, even for image generation on FLUX, PISA achieves a 1.2 times acceleration without compromising visual quality. Code is available at: https://github.com/xie-lab-ml/piecewise-sparse-attention.

Method: 1) Scalable synthetic data pipeline generating photorealistic human frames with pixel-accurate depth, normals, and masks; 2) Unified ViT-based dense predictor with explicit human geometric prior via CSE embeddings and lightweight channel reweighting module; 3) Two-stage training combining static pretraining with dynamic sequence supervision.

Result: Achieves state-of-the-art performance on THuman2.1 and Hi4D datasets and generalizes effectively to in-the-wild videos.

Conclusion: The proposed approach successfully addresses temporal consistency in human-centric dense prediction through synthetic data generation and architectural innovations, enabling robust spatial and temporal learning.

Abstract: In this work, we focus on the challenge of temporally consistent human-centric dense prediction across video sequences. Existing models achieve strong per-frame accuracy but often flicker under motion, occlusion, and lighting changes, and they rarely have paired human video supervision for multiple dense tasks. We address this gap with a scalable synthetic data pipeline that generates photorealistic human frames and motion-aligned sequences with pixel-accurate depth, normals, and masks. Unlike prior static data synthetic pipelines, our pipeline provides both frame-level labels for spatial learning and sequence-level supervision for temporal learning. Building on this, we train a unified ViT-based dense predictor that (i) injects an explicit human geometric prior via CSE embeddings and (ii) improves geometry-feature reliability with a lightweight channel reweighting module after feature fusion. Our two-stage training strategy, combining static pretraining with dynamic sequence supervision, enables the model first to acquire robust spatial representations and then refine temporal consistency across motion-aligned sequences. Extensive experiments show that we achieve state-of-the-art performance on THuman2.1 and Hi4D and generalize effectively to in-the-wild videos.

Method: Created a dataset of 2,854 anchor-linked variation pairs from original art and photographs, applying contextual transformations (illumination, weather, viewpoint, scene composition, color tone, mood) while preserving underlying identity.

Result: Dataset shows high identity stability, strong target attribute realization, and robust aesthetic profile exceeding large-scale web datasets, with high aesthetic scores maintained.

Conclusion: Lunara Aesthetic II provides a valuable resource for benchmarking, fine-tuning, and analyzing contextual generalization, identity preservation, and edit robustness in image generation systems with interpretable supervision.

Abstract: We introduce Lunara Aesthetic II, a publicly released, ethically sourced image dataset designed to support controlled evaluation and learning of contextual consistency in modern image generation and editing systems. The dataset comprises 2,854 anchor-linked variation pairs derived from original art and photographs created by Moonworks. Each variation pair applies contextual transformations, such as illumination, weather, viewpoint, scene composition, color tone, or mood; while preserving a stable underlying identity. Lunara Aesthetic II operationalizes identity-preserving contextual variation as a supervision signal while also retaining Lunara’s signature high aesthetic scores. Results show high identity stability, strong target attribute realization, and a robust aesthetic profile that exceeds large-scale web datasets. Released under the Apache 2.0 license, Lunara Aesthetic II is intended for benchmarking, fine-tuning, and analysis of contextual generalization, identity preservation, and edit robustness in image generation and image-to-image systems with interpretable, relational supervision. The dataset is publicly available at: https://huggingface.co/datasets/moonworks/lunara-aesthetic-image-variations.

Method: CMAFNet integrates RGB appearance and depth geometry through a purify-then-fuse paradigm. It uses a Semantic Recomposition Module with dictionary-based feature purification via learned codebook to suppress noise while preserving defect information, and a Contextual Semantic Integration Framework with partial-channel attention for global spatial dependencies. Position-wise normalization enforces cross-modal alignment.

Result: Achieves 32.2% mAP@50 and 12.5% APs on TLRGBD benchmark (94.5% small objects), outperforming strongest baseline by 9.8 and 4.0 percentage points. Lightweight variant reaches 24.8% mAP50 at 228 FPS with only 4.9M parameters, surpassing YOLO-based detectors while matching transformer methods at lower cost.

Conclusion: CMAFNet effectively addresses small-scale defect detection challenges through principled cross-modal fusion of RGB and depth information, demonstrating superior performance and efficiency for transmission line inspection.

Abstract: Transmission line defect detection remains challenging for automated UAV inspection due to the dominance of small-scale defects, complex backgrounds, and illumination variations. Existing RGB-based detectors, despite recent progress, struggle to distinguish geometrically subtle defects from visually similar background structures under limited chromatic contrast. This paper proposes CMAFNet, a Cross-Modal Alignment and Fusion Network that integrates RGB appearance and depth geometry through a principled purify-then-fuse paradigm. CMAFNet consists of a Semantic Recomposition Module that performs dictionary-based feature purification via a learned codebook to suppress modality-specific noise while preserving defect-discriminative information, and a Contextual Semantic Integration Framework that captures global spatial dependencies using partial-channel attention to enhance structural semantic reasoning. Position-wise normalization within the purification stage enforces explicit reconstruction-driven cross-modal alignment, ensuring statistical compatibility between heterogeneous features prior to fusion. Extensive experiments on the TLRGBD benchmark, where 94.5% of instances are small objects, demonstrate that CMAFNet achieves 32.2% mAP@50 and 12.5% APs, outperforming the strongest baseline by 9.8 and 4.0 percentage points, respectively. A lightweight variant reaches 24.8% mAP50 at 228 FPS with only 4.9M parameters, surpassing all YOLO-based detectors while matching transformer-based methods at substantially lower computational cost.

Method: ObjEmbed decomposes input images into multiple regional embeddings (one per object) plus global embeddings. Each region gets two complementary embeddings: object embedding for semantic matching and IoU embedding for localization quality prediction. Final matching combines semantic similarity with predicted IoU.

Result: Superior performance on 18 diverse benchmarks demonstrates strong semantic discrimination capabilities for both region-level and image-level visual understanding tasks.

Conclusion: ObjEmbed provides an effective solution for fine-grained vision-language alignment with object-oriented representations, supporting versatile visual understanding tasks through efficient single-pass encoding.

Abstract: Aligning objects with corresponding textual descriptions is a fundamental challenge and a realistic requirement in vision-language understanding. While recent multimodal embedding models excel at global image-text alignment, they often struggle with fine-grained alignment between image regions and specific phrases. In this work, we present ObjEmbed, a novel MLLM embedding model that decomposes the input image into multiple regional embeddings, each corresponding to an individual object, along with global embeddings. It supports a wide range of visual understanding tasks like visual grounding, local image retrieval, and global image retrieval. ObjEmbed enjoys three key properties: (1) Object-Oriented Representation: It captures both semantic and spatial aspects of objects by generating two complementary embeddings for each region: an object embedding for semantic matching and an IoU embedding that predicts localization quality. The final object matching score combines semantic similarity with the predicted IoU, enabling more accurate retrieval. (2) Versatility: It seamlessly handles both region-level and image-level tasks. (3) Efficient Encoding: All objects in an image, along with the full image, are encoded in a single forward pass for high efficiency. Superior performance on 18 diverse benchmarks demonstrates its strong semantic discrimination.

Method: Proposes Disentangled Dynamics Prediction (DDP) that decomposes latent state evolution into sparse primary dynamics (physical interactions) and secondary context-driven background updates. Uses efficient historical processing with dynamic localization to isolate primary dynamics and cross-attention for background updates.

Result: Achieves significant efficiency and performance across diverse robotic tasks including navigation, tabletop manipulation, and complex deformable/multi-body interactions. On Push-T task: ~9x inference speedup and improves MPC success rate from 90% to 98% compared to SOTA dense models.

Conclusion: DDP-WM establishes a promising path for developing efficient, high-fidelity world models for real-time robotic planning applications.

Abstract: World models are essential for autonomous robotic planning. However, the substantial computational overhead of existing dense Transformerbased models significantly hinders real-time deployment. To address this efficiency-performance bottleneck, we introduce DDP-WM, a novel world model centered on the principle of Disentangled Dynamics Prediction (DDP). We hypothesize that latent state evolution in observed scenes is heterogeneous and can be decomposed into sparse primary dynamics driven by physical interactions and secondary context-driven background updates. DDP-WM realizes this decomposition through an architecture that integrates efficient historical processing with dynamic localization to isolate primary dynamics. By employing a crossattention mechanism for background updates, the framework optimizes resource allocation and provides a smooth optimization landscape for planners. Extensive experiments demonstrate that DDP-WM achieves significant efficiency and performance across diverse tasks, including navigation, precise tabletop manipulation, and complex deformable or multi-body interactions. Specifically, on the challenging Push-T task, DDP-WM achieves an approximately 9 times inference speedup and improves the MPC success rate from 90% to98% compared to state-of-the-art dense models. The results establish a promising path for developing efficient, high-fidelity world models. Codes will be available at https://github.com/HCPLab-SYSU/DDP-WM.

Method: Proposes SurfSplat using 2D Gaussian Splatting primitives with stronger anisotropy and higher geometric precision. Incorporates surface continuity prior and forced alpha blending strategy for coherent geometry and faithful textures. Introduces High-Resolution Rendering Consistency (HRRC) evaluation metric.

Result: Outperforms prior methods on RealEstate10K, DL3DV, and ScanNet datasets on both standard metrics and the new HRRC metric, demonstrating robust high-fidelity 3D reconstruction from sparse inputs.

Conclusion: SurfSplat provides a robust solution for high-fidelity 3D reconstruction from sparse images by addressing surface continuity issues through 2D Gaussian Splatting primitives and novel blending strategies.

Abstract: Reconstructing 3D scenes from sparse images remains a challenging task due to the difficulty of recovering accurate geometry and texture without optimization. Recent approaches leverage generalizable models to generate 3D scenes using 3D Gaussian Splatting (3DGS) primitive. However, they often fail to produce continuous surfaces and instead yield discrete, color-biased point clouds that appear plausible at normal resolution but reveal severe artifacts under close-up views. To address this issue, we present SurfSplat, a feedforward framework based on 2D Gaussian Splatting (2DGS) primitive, which provides stronger anisotropy and higher geometric precision. By incorporating a surface continuity prior and a forced alpha blending strategy, SurfSplat reconstructs coherent geometry together with faithful textures. Furthermore, we introduce High-Resolution Rendering Consistency (HRRC), a new evaluation metric designed to evaluate high-resolution reconstruction quality. Extensive experiments on RealEstate10K, DL3DV, and ScanNet demonstrate that SurfSplat consistently outperforms prior methods on both standard metrics and HRRC, establishing a robust solution for high-fidelity 3D reconstruction from sparse inputs. Project page: https://hebing-sjtu.github.io/SurfSplat-website/

Method: Reg4Pru is a training regularization technique designed specifically for token pruning strategies. It helps maintain performance when pruning tokens by stabilizing the preserved representations, addressing the performance degradation that typically occurs in deeper layers of pruned models.

Result: On the FIVES blood vessel segmentation dataset, Reg4Pru improved average precision by 46% absolute compared to the same model trained without routing, while achieving 29% relative speedup in wall-clock time compared to the non-pruned baseline.

Conclusion: Reg4Pru is an effective regularizer for token reduction strategies in vision transformers, enabling significant computational efficiency gains while maintaining or improving segmentation performance.

Abstract: Transformers are widely adopted in modern vision models due to their strong ability to scale with dataset size and generalisability. However, this comes with a major drawback: computation scales quadratically to the total number of tokens. Numerous methods have been proposed to mitigate this. For example, we consider token pruning with reactivating tokens from preserved representations, but the increased computational efficiency of this method results in decreased stability from the preserved representations, leading to poorer dense prediction performance at deeper layers. In this work, we introduce Reg4Pru, a training regularisation technique that mitigates token-pruning performance loss for segmentation. We compare our models on the FIVES blood vessel segmentation dataset and find that Reg4Pru improves average precision by an absolute 46% compared to the same model trained without routing. This increase is observed using a configuration that achieves a 29% relative speedup in wall-clock time compared to the non-pruned baseline. These findings indicate that Reg4Pru is a valuable regulariser for token reduction strategies.

Method: CIEC uses two branches: 1) Image-based localization with Textual-guidance Refine Patch Selection (TRPS) that integrates visual and textual forgery cues using spatial priors, plus background silencing and spatial contrast constraints; 2) Text-based localization with Visual-deviation Calibrated Token Grounding (VCTG) that focuses on content words using visual bias, plus asymmetric sparse and semantic consistency constraints.

Result: Extensive experiments show CIEC achieves results comparable to fully supervised methods on several evaluation metrics for multimodal manipulation localization.

Conclusion: CIEC demonstrates that effective multimodal manipulation localization can be achieved with only coarse-grained annotations, reducing annotation costs while maintaining performance comparable to fully supervised approaches.

Abstract: To mitigate the threat of misinformation, multimodal manipulation localization has garnered growing attention. Consider that current methods rely on costly and time-consuming fine-grained annotations, such as patch/token-level annotations. This paper proposes a novel framework named Coupling Implicit and Explicit Cues (CIEC), which aims to achieve multimodal weakly-supervised manipulation localization for image-text pairs utilizing only coarse-grained image/sentence-level annotations. It comprises two branches, image-based and text-based weakly-supervised localization. For the former, we devise the Textual-guidance Refine Patch Selection (TRPS) module. It integrates forgery cues from both visual and textual perspectives to lock onto suspicious regions aided by spatial priors. Followed by the background silencing and spatial contrast constraints to suppress interference from irrelevant areas. For the latter, we devise the Visual-deviation Calibrated Token Grounding (VCTG) module. It focuses on meaningful content words and leverages relative visual bias to assist token localization. Followed by the asymmetric sparse and semantic consistency constraints to mitigate label noise and ensure reliability. Extensive experiments demonstrate the effectiveness of our CIEC, yielding results comparable to fully supervised methods on several evaluation metrics.

Method: 1) Hierarchical Pose-free Memory Compressor (HPMC) that recursively distills historical latents into fixed-budget representation; 2) Uncertainty-aware Action Labeling that discretizes continuous motion into tri-state logic; 3) Revisit-Dense Finetuning Strategy using compact dataset to activate long-range loop-closure capabilities.

Result: Extensive experiments show Infinite-World achieves superior performance in visual quality, action controllability, and spatial consistency compared to existing methods.

Conclusion: The proposed approach enables robust interactive world modeling in real-world environments without explicit geometric priors, maintaining coherent visual memory over extended sequences.

Abstract: We propose Infinite-World, a robust interactive world model capable of maintaining coherent visual memory over 1000+ frames in complex real-world environments. While existing world models can be efficiently optimized on synthetic data with perfect ground-truth, they lack an effective training paradigm for real-world videos due to noisy pose estimations and the scarcity of viewpoint revisits. To bridge this gap, we first introduce a Hierarchical Pose-free Memory Compressor (HPMC) that recursively distills historical latents into a fixed-budget representation. By jointly optimizing the compressor with the generative backbone, HPMC enables the model to autonomously anchor generations in the distant past with bounded computational cost, eliminating the need for explicit geometric priors. Second, we propose an Uncertainty-aware Action Labeling module that discretizes continuous motion into a tri-state logic. This strategy maximizes the utilization of raw video data while shielding the deterministic action space from being corrupted by noisy trajectories, ensuring robust action-response learning. Furthermore, guided by insights from a pilot toy study, we employ a Revisit-Dense Finetuning Strategy using a compact, 30-minute dataset to efficiently activate the model’s long-range loop-closure capabilities. Extensive experiments, including objective metrics and user studies, demonstrate that Infinite-World achieves superior performance in visual quality, action controllability, and spatial consistency.

Method: Proposes ReasonEdit that continuously stores human reasoning in a codebook and retrieves relevant facts during inference using a novel topology-balanced multimodal embedding method inspired by network science.

Result: Achieves state-of-the-art editing performance across four VLMs on multiple rationale-based visual question answering datasets, showing that using human reasoning during editing greatly improves edit generalization.

Conclusion: ReasonEdit successfully demonstrates that incorporating human reasoning explanations during editing significantly enhances vision-language model editing capabilities for reasoning-heavy tasks.

Abstract: Model editing aims to correct errors in large, pretrained models without altering unrelated behaviors. While some recent works have edited vision-language models (VLMs), no existing editors tackle reasoning-heavy tasks, which typically require humans and models to reason about images. We therefore propose ReasonEdit, the first VLM editor to let users explain their reasoning during editing, introducing a new, practical model editing setup. ReasonEdit continuously stores human reasoning in a codebook, and retrieves only relevant facts during inference using a novel topology-balanced multimodal embedding method inspired by network science. Across four VLMs on multiple rationale-based visual question answering datasets, ReasonEdit achieves state-of-the-art editing performance, ultimately showing that using human reasoning during editing greatly improves edit generalization.

Method: Proposes CreditAudit framework that evaluates models under semantically aligned, non-adversarial system prompt templates across benchmarks, measuring mean ability (average performance) and scenario-induced fluctuation sigma (stability risk), then maps volatility into interpretable credit grades (AAA to BBB) using cross-model quantiles with diagnostics to mitigate template difficulty drift.

Result: Experiments on GPQA, TruthfulQA, and MMLU Pro show models with similar mean ability can have substantially different fluctuation patterns, and stability risk can overturn prioritization decisions in agentic or high failure cost regimes.

Conclusion: CreditAudit provides a 2D and grade-based language for regime-specific model selection, supporting tiered deployment and more disciplined allocation of testing/monitoring effort for more objective and trustworthy real-world model evaluation.

Abstract: Leaderboard scores on public benchmarks have been steadily rising and converging, with many frontier language models now separated by only marginal differences. However, these scores often fail to match users’ day to day experience, because system prompts, output protocols, and interaction modes evolve under routine iteration, and in agentic multi step pipelines small protocol shifts can trigger disproportionate failures, leaving practitioners uncertain about which model to deploy. We propose CreditAudit, a deployment oriented credit audit framework that evaluates models under a family of semantically aligned and non adversarial system prompt templates across multiple benchmarks, reporting mean ability as average performance across scenarios and scenario induced fluctuation sigma as a stability risk signal, and further mapping volatility into interpretable credit grades from AAA to BBB via cross model quantiles with diagnostics that mitigate template difficulty drift. Controlled experiments on GPQA, TruthfulQA, and MMLU Pro show that models with similar mean ability can exhibit substantially different fluctuation, and stability risk can overturn prioritization decisions in agentic or high failure cost regimes. By providing a 2D and grade based language for regime specific selection, CreditAudit supports tiered deployment and more disciplined allocation of testing and monitoring effort, enabling more objective and trustworthy model evaluation for real world use.

Method: GeoEvolver uses a retrieval-augmented multi-agent orchestrator to decompose queries into sub-goals, explores diverse tool-parameter configurations at sub-goal level, and distills successful patterns and failure root causes into an evolving memory bank for future in-context demonstrations.

Result: Experiments on three tool-integrated EO benchmarks show GeoEvolver consistently improves end-to-end task success with average gain of 12% across multiple LLM backbones, demonstrating progressive emergence of EO expertise.

Conclusion: EO expertise can emerge progressively from efficient, fine-grained interactions with the environment through structured multi-agent systems without parameter updates, enabling LLM agents to handle complex, multimodal EO workflows.

Abstract: Recent advances have enabled large language model (LLM) agents to solve complex tasks by orchestrating external tools. However, these agents often struggle in specialized, tool-intensive domains that demand long-horizon execution, tight coordination across modalities, and strict adherence to implicit tool constraints. Earth Observation (EO) tasks exemplify this challenge due to the multi-modal and multi-temporal data inputs, as well as the requirements of geo-knowledge constraints (spectrum library, spatial reasoning, etc): many high-level plans can be derailed by subtle execution errors that propagate through a pipeline and invalidate final results. A core difficulty is that existing agents lack a mechanism to learn fine-grained, tool-level expertise from interaction. Without such expertise, they cannot reliably configure tool parameters or recover from mid-execution failures, limiting their effectiveness in complex EO workflows. To address this, we introduce \textbf{GeoEvolver}, a self-evolving multi-agent system~(MAS) that enables LLM agents to acquire EO expertise through structured interaction without any parameter updates. GeoEvolver decomposes each query into independent sub-goals via a retrieval-augmented multi-agent orchestrator, then explores diverse tool-parameter configurations at the sub-goal level. Successful patterns and root-cause attribution from failures are then distilled in an evolving memory bank that provides in-context demonstrations for future queries. Experiments on three tool-integrated EO benchmarks show that GeoEvolver consistently improves end-to-end task success, with an average gain of 12% across multiple LLM backbones, demonstrating that EO expertise can emerge progressively from efficient, fine-grained interactions with the environment.

Method: Introduces benchmark metrics and dataset with 8 demographic attributes across movies/music domains. Uses entropy for uncertainty quantification, measures fairness gaps (SNSR/SNSV), tests prompt perturbations, and integrates personality-aware fairness into RecLLM evaluation pipeline.

Result: Gemini 1.5 Flash exhibits systematic unfairness (SNSR: 0.1363, SNSV: 0.0507) that persists under typographical errors and multilingual inputs. Reveals personality-linked bias patterns and trade-offs between personalization and group fairness.

Conclusion: Proposes uncertainty-aware evaluation methodology and personality profile-informed fairness benchmark to advance explainability and equity in LLM recommendations, establishing foundation for safer, more interpretable RecLLMs.

Abstract: Large language models (LLMs) enable powerful zero-shot recommendations by leveraging broad contextual knowledge, yet predictive uncertainty and embedded biases threaten reliability and fairness. This paper studies how uncertainty and fairness evaluations affect the accuracy, consistency, and trustworthiness of LLM-generated recommendations. We introduce a benchmark of curated metrics and a dataset annotated for eight demographic attributes (31 categorical values) across two domains: movies and music. Through in-depth case studies, we quantify predictive uncertainty (via entropy) and demonstrate that Google DeepMind’s Gemini 1.5 Flash exhibits systematic unfairness for certain sensitive attributes; measured similarity-based gaps are SNSR at 0.1363 and SNSV at 0.0507. These disparities persist under prompt perturbations such as typographical errors and multilingual inputs. We further integrate personality-aware fairness into the RecLLM evaluation pipeline to reveal personality-linked bias patterns and expose trade-offs between personalization and group fairness. We propose a novel uncertainty-aware evaluation methodology for RecLLMs, present empirical insights from deep uncertainty case studies, and introduce a personality profile-informed fairness benchmark that advances explainability and equity in LLM recommendations. Together, these contributions establish a foundation for safer, more interpretable RecLLMs and motivate future work on multi-model benchmarks and adaptive calibration for trustworthy deployment.

Method: PeerRank treats evaluation as a multi-agent process where models symmetrically act as task designers, respondents (with category-scoped live web grounding), and evaluators, removing biased judgments through peer assessment aggregation.

Result: In a study with 12 commercial models and 420 autonomously generated questions, PeerRank produced stable, discriminative rankings, revealed identity and presentation biases, showed robustness, and mean peer scores agreed with Elo ratings.

Conclusion: Bias-aware peer evaluation with selective web-grounded answering can scale open-world LLM assessment beyond static, human-curated benchmarks, as validated on TruthfulQA and GSM8K where peer scores correlate with objective accuracy.

Abstract: Evaluating large language models typically relies on human-authored benchmarks, reference answers, and human or single-model judgments, approaches that scale poorly, become quickly outdated, and mismatch open-world deployments that depend on web retrieval and synthesis. We introduce PeerRank, a fully autonomous end-to-end evaluation framework in which models generate evaluation tasks, answer them with category-scoped live web grounding, judge peer responses and aggregate dense peer assessments into relative performance estimates, without human supervision or gold references. PeerRank treats evaluation as a multi-agent process where each model participates symmetrically as task designer, respondent, and evaluator, while removing biased judgments. In a large-scale study over 12 commercially available models and 420 autonomously generated questions, PeerRank produces stable, discriminative rankings and reveals measurable identity and presentation biases. Rankings are robust, and mean peer scores agree with Elo. We further validate PeerRank on TruthfulQA and GSM8K, where peer scores correlate with objective accuracy. Together, these results suggest that bias-aware peer evaluation with selective web-grounded answering can scale open-world LLM assessment beyond static and human curated benchmarks.

Method: Introduced Normalized Simulatability Gain (NSG) metric based on the idea that faithful explanations should allow observers to learn model’s decision-making criteria and better predict behavior on related inputs. Evaluated 18 frontier models on 7,000 counterfactuals from datasets covering health, business, and ethics.

Result: Self-explanations substantially improve prediction of model behavior (11-37% NSG). They provide more predictive information than external model explanations, even when external models are stronger. Across models, 5-15% of self-explanations are egregiously misleading.

Conclusion: Self-explanations encode information that helps predict model behavior, showing a positive case for their use despite imperfections. There’s an advantage from self-knowledge that external explanation methods cannot replicate.

Abstract: LLM self-explanations are often presented as a promising tool for AI oversight, yet their faithfulness to the model’s true reasoning process is poorly understood. Existing faithfulness metrics have critical limitations, typically relying on identifying unfaithfulness via adversarial prompting or detecting reasoning errors. These methods overlook the predictive value of explanations. We introduce Normalized Simulatability Gain (NSG), a general and scalable metric based on the idea that a faithful explanation should allow an observer to learn a model’s decision-making criteria, and thus better predict its behavior on related inputs. We evaluate 18 frontier proprietary and open-weight models, e.g., Gemini 3, GPT-5.2, and Claude 4.5, on 7,000 counterfactuals from popular datasets covering health, business, and ethics. We find self-explanations substantially improve prediction of model behavior (11-37% NSG). Self-explanations also provide more predictive information than explanations generated by external models, even when those models are stronger. This implies an advantage from self-knowledge that external explanation methods cannot replicate. Our approach also reveals that, across models, 5-15% of self-explanations are egregiously misleading. Despite their imperfections, we show a positive case for self-explanations: they encode information that helps predict model behavior.

Method: Three pillars: (1) Budget-Aware Planning using cost-constrained Monte Carlo Tree Search (MCTS) to balance performance with execution expense; (2) Modular Construction with “Design-Decompose-Implement” pipeline for complex research repositories; (3) Comparative Reflective Memory to address credit assignment by analyzing solution differences for insights.

Result: Achieves state-of-the-art performance among open-source frameworks on MLE-Bench under comparable settings, competitive with global leaderboard’s top methods. Shows qualitative “Aha!” moments with 63% of utilized lessons originating from cross-branch transfer, demonstrating effective generalization across search paths.

Conclusion: MARS provides an effective framework for autonomous AI research that explicitly addresses computational costs and performance attribution challenges, enabling agents to generalize insights across different research paths.

Abstract: Automating AI research differs from general software engineering due to computationally expensive evaluation (e.g., model training) and opaque performance attribution. Current LLM-based agents struggle here, often generating monolithic scripts that ignore execution costs and causal factors. We introduce MARS (Modular Agent with Reflective Search), a framework optimized for autonomous AI research. MARS relies on three pillars: (1) Budget-Aware Planning via cost-constrained Monte Carlo Tree Search (MCTS) to explicitly balance performance with execution expense; (2) Modular Construction, employing a “Design-Decompose-Implement” pipeline to manage complex research repositories; and (3) Comparative Reflective Memory, which addresses credit assignment by analyzing solution differences to distill high-signal insights. MARS achieves state-of-the-art performance among open-source frameworks on MLE-Bench under comparable settings, maintaining competitiveness with the global leaderboard’s top methods. Furthermore, the system exhibits qualitative “Aha!” moments, where 63% of all utilized lessons originate from cross-branch transfer, demonstrating that the agent effectively generalizes insights across search paths.

Method: Proposes ATLAS framework with task distribution: lightweight research agent evolves iteratively while delegating complementary roles to specialized supporter agents (exploration, hyperparameter tuning, reference policy management). Core algorithm is Evolving Direct Preference Optimization (EvoDPO) that adaptively updates phase-indexed reference policy. Includes theoretical regret analysis for preference-based contextual bandit under concept drift.

Result: Experiments on non-stationary linear contextual bandits and scientific machine learning (SciML) loss reweighting for 1D Burgers’ equation show ATLAS improves stability and performance over static single-agent baseline.

Conclusion: ATLAS provides an effective task-distributed framework for adaptive agent evolution in non-stationary environments, with theoretical guarantees and empirical validation showing improved stability and performance over static approaches.

Abstract: Recent multi-LLM agent systems perform well in prompt optimization and automated problem-solving, but many either keep the solver frozen after fine-tuning or rely on a static preference-optimization loop, which becomes intractable for long-horizon tasks. We propose ATLAS (Adaptive Task-distributed Learning for Agentic Self-evolution), a task-distributed framework that iteratively develops a lightweight research agent while delegating complementary roles to specialized supporter agents for exploration, hyperparameter tuning, and reference policy management. Our core algorithm, Evolving Direct Preference Optimization (EvoDPO), adaptively updates the phase-indexed reference policy. We provide a theoretical regret analysis for a preference-based contextual bandit under concept drift. In addition, experiments were conducted on non-stationary linear contextual bandits and scientific machine learning (SciML) loss reweighting for the 1D Burgers’ equation. Both results show that ATLAS improves stability and performance over a static single-agent baseline.

Method: Proposes a dynamic mix-precision routing framework that adaptively selects between high-precision and low-precision LLMs at each decision step based on step sensitivity. Uses a two-stage training pipeline: 1) KL-divergence-based supervised learning to identify precision-sensitive steps, 2) Group-Relative Policy Optimization (GRPO) to improve task success rates.

Result: Experiments on ALFWorld demonstrate significant improvement in accuracy-cost trade-off over single-precision baselines and heuristic routing methods.

Conclusion: Dynamic precision selection can effectively reduce LLM inference costs in long-horizon decision-making while maintaining strong task performance, offering a practical solution for resource-constrained applications.

Abstract: Large language models (LLM) achieve strong performance in long-horizon decision-making tasks through multi-step interaction and reasoning at test time. While practitioners commonly believe a higher task success rate necessitates the use of a larger and stronger LLM model, multi-step interaction with a large LLM incurs prohibitive inference cost. To address this problem, we explore the use of low-precision quantized LLM in the long-horizon decision-making process. Based on the observation of diverse sensitivities among interaction steps, we propose a dynamic mix-precision routing framework that adaptively selects between high-precision and low-precision LLMs at each decision step. The router is trained via a two-stage pipeline, consisting of KL-divergence-based supervised learning that identifies precision-sensitive steps, followed by Group-Relative Policy Optimization (GRPO) to further improve task success rates. Experiments on ALFWorld demonstrate that our approach achieves a great improvement on accuracy-cost trade-off over single-precision baselines and heuristic routing methods.

Method: 1) Scaling-Aware Patching uses instruction-conditioned gating to generate variable-size patches over structural graphs, adaptively scaling query token budget with structural complexity. 2) Geometry Grounding Adapter refines these adaptive tokens via cross-attention to modality embeddings and injects resulting modality tokens into the LLM to expose explicit geometric cues.

Result: Experiments across diverse all-atom benchmarks demonstrate that Cuttlefish achieves superior performance in heterogeneous structure-grounded reasoning.

Conclusion: Cuttlefish provides a unified approach that grounds language reasoning in geometric cues while adaptively scaling modality tokens with structural complexity, addressing limitations of existing methods for biomolecular structure understanding.

Abstract: Large language models (LLMs) are enabling reasoning over biomolecular structures, yet existing methods remain modality-specific and typically compress structural inputs through sequence-based tokenization or fixed-length query connectors. Such architectures either omit the geometric groundings requisite for mitigating structural hallucinations or impose inflexible modality fusion bottlenecks that concurrently over-compress and suboptimally allocate structural tokens, thereby impeding the realization of generalized all-atom reasoning. We introduce Cuttlefish, a unified all-atom LLM that grounds language reasoning in geometric cues while scaling modality tokens with structural complexity. First, Scaling-Aware Patching leverages an instruction-conditioned gating mechanism to generate variable-size patches over structural graphs, adaptively scaling the query token budget with structural complexity to mitigate fixed-length connector bottlenecks. Second, Geometry Grounding Adapter refines these adaptive tokens via cross-attention to modality embeddings and injects the resulting modality tokens into the LLM, exposing explicit geometric cues to reduce structural hallucination. Experiments across diverse all-atom benchmarks demonstrate that Cuttlefish achieves superior performance in heterogeneous structure-grounded reasoning. Code is available at the project repository.

Method: Three-agent system (Analyzer-Reasoner-Executor) with shared memory and gated communication; uses Vision-Language Model for visual reasoning over time series plots to extract temporal structures, then reconstructs predictive trajectories in latent space; router selects task-specific tool chains

Result: Achieves state-of-the-art performance across multiple time series benchmarks; demonstrates strong generalization and efficient inference

Conclusion: MAS4TS provides effective framework for general time series tasks through multi-agent coordination, visual reasoning, and latent reconstruction

Abstract: Time series analysis underpins many real-world applications, yet existing time-series-specific methods and pretrained large-model-based approaches remain limited in integrating intuitive visual reasoning and generalizing across tasks with adaptive tool usage. To address these limitations, we propose MAS4TS, a tool-driven multi-agent system for general time series tasks, built upon an Analyzer-Reasoner-Executor paradigm that integrates agent communication, visual reasoning, and latent reconstruction within a unified framework. MAS4TS first performs visual reasoning over time series plots with structured priors using a Vision-Language Model to extract temporal structures, and subsequently reconstructs predictive trajectories in latent space. Three specialized agents coordinate via shared memory and gated communication, while a router selects task-specific tool chains for execution. Extensive experiments on multiple benchmarks demonstrate that MAS4TS achieves state-of-the-art performance across a wide range of time series tasks, while exhibiting strong generalization and efficient inference.

Method: CoS employs three distinct reasoning modes: (1) computational flow with self-consistency for mathematical problems, (2) symbolic state tracking with JSON representations for spatial reasoning, and (3) hybrid fact-extraction for multi-hop inference. The framework includes algorithms for mode selection, state tracking, and answer extraction, and dynamically routes problems to the appropriate specialized reasoning strategy.

Result: CoS achieves 71.5% accuracy on GSM8K (1.0% absolute improvement), 90.0% on StrategyQA (2.5% improvement), and 19.0% on bAbI (65.2% relative improvement) compared to strongest baselines. Computational mode achieves 81.2% accuracy when correctly applied to mathematical problems. The framework provides comparable performance to Self-Consistency at 54% lower computational cost.

Conclusion: Problem-specific mode selection is crucial for LLM reasoning, and CoS establishes an effective approach for improving reasoning without additional training. The framework demonstrates superior trade-offs between accuracy and efficiency compared to uniform prompting methods.

Abstract: We present Chain of Simulation (CoS), a novel dual-mode reasoning framework that dynamically routes problems to specialized reasoning strategies in Large Language Models (LLMs). Unlike existing uniform prompting approaches, CoS employs three distinct reasoning modes: (1) computational flow with self-consistency for mathematical problems, (2) symbolic state tracking with JSON representations for spatial reasoning, and (3) hybrid fact-extraction for multi-hop inference. Through comprehensive evaluation on GSM8K, StrategyQA, and bAbI benchmarks using four state-of-the-art models (Gemma-3 27B, LLaMA-3.1 8B, Mistral 7B, and Qwen-2.5 14B), we demonstrate that CoS achieves 71.5% accuracy on GSM8K (1.0% absolute improvement), 90.0% on StrategyQA (2.5% improvement), and 19.0% on bAbI (65.2% relative improvement) compared to the strongest baselines. The analysis reveals that problem-specific mode selection is crucial, with computational mode achieving 81.2% accuracy when correctly applied to mathematical problems, while misrouting leads to 0% accuracy. We provide detailed algorithms for mode selection, state tracking, and answer extraction, establishing CoS as an effective approach for improving LLM reasoning without additional training. The framework provides superior trade-offs between accuracy and efficiency compared to Self-Consistency, achieving comparable performance at 54% lower computational cost.

Method: Two-loop optimization framework: inner loop for circuit sizing, outer loop analyzes optimization dynamics and constraints to iteratively refine search space from simulation feedback. Uses LLM-driven meta-optimization that unifies circuit understanding, adaptive search-space construction, and optimization orchestration.

Result: AutoSizer achieves higher solution quality, faster convergence, and higher success rate across varying circuit difficulties, outperforming both traditional optimization methods and existing LLM-based agents. Also introduces AMS-SizingBench benchmark with 24 diverse AMS circuits.

Conclusion: AutoSizer successfully bridges the gap between LLM reasoning and precise numerical optimization for AMS circuit sizing, demonstrating superior performance through adaptive search-space refinement.

Abstract: The design of Analog and Mixed-Signal (AMS) integrated circuits remains heavily reliant on expert knowledge, with transistor sizing a major bottleneck due to nonlinear behavior, high-dimensional design spaces, and strict performance constraints. Existing Electronic Design Automation (EDA) methods typically frame sizing as static black-box optimization, resulting in inefficient and less robust solutions. Although Large Language Models (LLMs) exhibit strong reasoning abilities, they are not suited for precise numerical optimization in AMS sizing. To address this gap, we propose AutoSizer, a reflective LLM-driven meta-optimization framework that unifies circuit understanding, adaptive search-space construction, and optimization orchestration in a closed loop. It employs a two-loop optimization framework, with an inner loop for circuit sizing and an outer loop that analyzes optimization dynamics and constraints to iteratively refine the search space from simulation feedback. We further introduce AMS-SizingBench, an open benchmark comprising 24 diverse AMS circuits in SKY130 CMOS technology, designed to evaluate adaptive optimization policies under realistic simulator-based constraints. AutoSizer experimentally achieves higher solution quality, faster convergence, and higher success rate across varying circuit difficulties, outperforming both traditional optimization methods and existing LLM-based agents.

Method: MAS-ProVe: systematic empirical study evaluating three verification paradigms (LLM-as-a-Judge, reward models, process reward models) across two granularity levels (agent-level, iteration-level), five verifiers, four context management strategies, tested on six MAS frameworks across multiple reasoning benchmarks.

Result: Process-level verification does not consistently improve performance and exhibits high variance. LLM-as-a-Judge generally outperforms reward-based approaches, with trained judges beating general-purpose LLMs. Small performance gap exists between LLMs as judges vs. single agents, with context-length-performance trade-off observed.

Conclusion: Effective and robust process verification for MAS remains an open challenge requiring advances beyond current paradigms, despite LLM-as-a-Judge showing relative promise.

Abstract: Multi-Agent Systems (MAS) built on Large Language Models (LLMs) often exhibit high variance in their reasoning trajectories. Process verification, which evaluates intermediate steps in trajectories, has shown promise in general reasoning settings, and has been suggested as a potential tool for guiding coordination of MAS; however, its actual effectiveness in MAS remains unclear. To fill this gap, we present MAS-ProVe, a systematic empirical study of process verification for multi-agent systems (MAS). Our study spans three verification paradigms (LLM-as-a-Judge, reward models, and process reward models), evaluated across two levels of verification granularity (agent-level and iteration-level). We further examine five representative verifiers and four context management strategies, and conduct experiments over six diverse MAS frameworks on multiple reasoning benchmarks. We find that process-level verification does not consistently improve performance and frequently exhibits high variance, highlighting the difficulty of reliably evaluating partial multi-agent trajectories. Among the methods studied, LLM-as-a-Judge generally outperforms reward-based approaches, with trained judges surpassing general-purpose LLMs. We further observe a small performance gap between LLMs acting as judges and as single agents, and identify a context-length-performance trade-off in verification. Overall, our results suggest that effective and robust process verification for MAS remains an open challenge, requiring further advances beyond current paradigms. Code is available at https://github.com/Wang-ML-Lab/MAS-ProVe.

Method: STEER constructs a population of natural-language personas through offline constrained quality-diversity search that promotes behavioral coverage while enforcing safety, reasoning, and stability thresholds. At inference, a single interpretable control parameter maps user-specified risk percentile to selected persona.

Result: On clinical triage benchmarks, STEER achieves broader behavioral coverage than temperature-based sampling and static persona ensembles, maintains higher accuracy on unambiguous urgent cases compared to post-training methods, and provides comparable control over ambiguous decisions.

Conclusion: STEER demonstrates a safety-preserving paradigm for risk control that can steer LLM behavior without compromising domain competence, particularly valuable for applications requiring tunable decision conservativeness.

Abstract: Large Language Models (LLMs) trained for average correctness often exhibit mode collapse, producing narrow decision behaviors on tasks where multiple responses may be reasonable. This limitation is particularly problematic in ordinal decision settings such as clinical triage, where standard alignment removes the ability to trade off specificity and sensitivity (the ROC operating point) based on contextual constraints. We propose STEER (Steerable Tuning via Evolutionary Ensemble Refinement), a training-free framework that reintroduces this tunable control. STEER constructs a population of natural-language personas through an offline, constrained quality-diversity search that promotes behavioral coverage while enforcing minimum safety, reasoning, and stability thresholds. At inference time, STEER exposes a single, interpretable control parameter that maps a user-specified risk percentile to a selected persona, yielding a monotonic adjustment of decision conservativeness. On two clinical triage benchmarks, STEER achieves broader behavioral coverage compared to temperature-based sampling and static persona ensembles. Compared to a representative post-training method, STEER maintains substantially higher accuracy on unambiguous urgent cases while providing comparable control over ambiguous decisions. These results demonstrate STEER as a safety-preserving paradigm for risk control, capable of steering behavior without compromising domain competence.

Method: Define an instability signal combining consecutive-step distributional shift (Jensen-Shannon Divergence) and uncertainty (entropy) from token log probabilities. Summarize each reasoning trace by its peak instability strength and analyze timing of instability relative to decoding horizon.

Result: Instability strength reliably predicts wrong answers with above-chance AUC on GSM8K and HotpotQA. Early instability can lead to recovery (corrective instability), while late instability more often leads to failure (destructive instability), even at similar peak magnitudes.

Conclusion: Reasoning failures can be detected mid-generation using standard API observables without training. The timing of instability matters - late instability is more predictive of failure than early instability. The method provides a diagnostic lens but not a corrective mechanism.

Abstract: Reasoning failures in large language models (LLMs) are typically measured only at the end of a generation, yet many failures manifest as a process-level breakdown: the model “loses the thread” mid-reasoning. We study whether such breakdowns are detectable from inference-time observables available in standard APIs (token log probabilities), without any training or fine-tuning. We define a simple instability signal that combines consecutive-step distributional shift (JSD) and uncertainty (entropy), summarize each trace by its peak instability strength, and show that this signal reliably predicts failure. Across GSM8K and HotpotQA, instability strength predicts wrong answers with above-chance AUC and yields monotonic bucket-level accuracy decline at scale across model sizes. Crucially, we show that instability is not uniformly harmful: early instability can reflect subsequent stabilization and a correct final answer (\emph{corrective instability}), whereas late instability is more often followed by failure (\emph{destructive instability}), even at comparable peak magnitudes, indicating that recoverability depends not only on how strongly the distribution changes but also on when such changes occur relative to the remaining decoding horizon. The method is model-agnostic, training-free, and reproducible, and is presented as a diagnostic lens rather than a corrective or control mechanism.

Method: Defines structured bipolar argumentation frameworks (SBAFs) where arguments consist of sentences with attack and support relations. Provides novel semantics that: 1) don’t force acceptance of all defended arguments (unlike completeness-based semantics), and 2) provide both argument extensions and language extensions (acceptable sentence sets).

Result: Developed semantics that lie between admissible and complete semantics of abstract argumentation. The approach provides new perspective on existing methods: specifies conditions for ignoring support between arguments (when abstract argumentation is warranted) and shows deductive support semantics is a special case.

Conclusion: The paper presents a philosophically/linguistically informed computational argumentation approach that better models rational agent behavior by allowing doubt-based rejection and sentence-level reasoning, bridging abstract and structured argumentation.

Abstract: This paper develops a new approach to computational argumentation that is informed by philosophical and linguistic views. Namely, it takes into account two ideas that have received little attention in the literature on computational argumentation: First, an agent may rationally reject an argument based on mere doubt, thus not all arguments they could defend must be accepted; and, second, that it is sometimes more natural to think in terms of which individual sentences or claims an agent accepts in a debate, rather than which arguments. In order to incorporate these two ideas into a computational approach, we first define the notion of structured bipolar argumentation frameworks (SBAFs), where arguments consist of sentences and we have both an attack and a support relation between them. Then, we provide semantics for SBAFs with two features: (1) Unlike with completeness-based semantics, our semantics do not force agents to accept all defended arguments. (2) In addition to argument extensions, which give acceptable sets of arguments, we also provide semantics for language extensions that specify acceptable sets of sentences. These semantics represent reasonable positions an agent might have in a debate. Our semantics lie between the admissible and complete semantics of abstract argumentation. Further, our approach can be used to provide a new perspective on existing approaches. For instance, we can specify the conditions under which an agent can ignore support between arguments (i.e. under which the use of abstract argumentation is warranted) and we show that deductive support semantics is a special case of our approach.

Method: Proposes BenchAlign, which uses limited information about model performance (question-level performance and ranked model pairs collected during deployment) to learn preference-aligned weightings for benchmark questions, creating new static benchmarks that better predict model pairwise preferences.

Result: Aligned benchmarks can accurately rank unseen models according to human preferences across different model sizes while remaining interpretable. The method shows promise for bridging the gap between benchmark performance and real utility.

Conclusion: BenchAlign provides insights into aligning benchmarks with practical human preferences, which could accelerate model development towards real utility by creating better predictive benchmarks.

Abstract: Language model benchmarks are pervasive and computationally-efficient proxies for real-world performance. However, many recent works find that benchmarks often fail to predict real utility. Towards bridging this gap, we introduce benchmark alignment, where we use limited amounts of information about model performance to automatically update offline benchmarks, aiming to produce new static benchmarks that predict model pairwise preferences in given test settings. We then propose BenchAlign, the first solution to this problem, which learns preference-aligned weight- ings for benchmark questions using the question-level performance of language models alongside ranked pairs of models that could be collected during deployment, producing new benchmarks that rank previously unseen models according to these preferences. Our experiments show that our aligned benchmarks can accurately rank unseen models according to models of human preferences, even across different sizes, while remaining interpretable. Overall, our work provides insights into the limits of aligning benchmarks with practical human preferences, which stands to accelerate model development towards real utility.

Method: Implement perspective as a slowly evolving global latent state that modulates fast policy dynamics without direct optimization for behavioral consequences. Test in reward-free environment with regime shifts to observe emergent properties.

Result: The latent structure exhibits direction-dependent hysteresis while policy-level behavior remains comparatively reactive, suggesting hysteresis as a measurable signature of perspective-like subjectivity in machine systems.

Conclusion: Direction-dependent hysteresis in slowly evolving latent states can serve as an operational signature of subjective perspective in artificial agents, providing a measurable approach to studying machine subjectivity.

Abstract: This study operationalizes subjective perspective in artificial agents by grounding it in a minimal, phenomenologically motivated internal structure. The perspective is implemented as a slowly evolving global latent state that modulates fast policy dynamics without being directly optimized for behavioral consequences. In a reward-free environment with regime shifts, this latent structure exhibits direction-dependent hysteresis, while policy-level behavior remains comparatively reactive. I argue that such hysteresis constitutes a measurable signature of perspective-like subjectivity in machine systems.

Method: FIRE-Bench evaluates agents through rediscovery of established findings from recent, high-impact ML research. Agents are given only high-level research questions from published studies and must autonomously explore ideas, design experiments, implement code, execute plans, and derive evidence-based conclusions.

Result: Current state-of-the-art agents with frontier LLMs (like GPT-5) achieve limited rediscovery success (<50 F1), show high variance across runs, and exhibit recurring failure modes in experimental design, execution, and evidence-based reasoning.

Conclusion: Full-cycle scientific research remains challenging for current agent systems. FIRE-Bench provides a rigorous, diagnostic framework for measuring progress toward reliable agent-driven scientific discovery.

Abstract: Autonomous agents powered by large language models (LLMs) promise to accelerate scientific discovery end-to-end, but rigorously evaluating their capacity for verifiable discovery remains a central challenge. Existing benchmarks face a trade-off: they either heavily rely on LLM-as-judge evaluations of automatically generated research outputs or optimize convenient yet isolated performance metrics that provide coarse proxies for scientific insight. To address this gap, we introduce FIRE-Bench (Full-cycle Insight Rediscovery Evaluation), a benchmark that evaluates agents through the rediscovery of established findings from recent, high-impact machine learning research. Agents are given only a high-level research question extracted from a published, verified study and must autonomously explore ideas, design experiments, implement code, execute their plans, and derive conclusions supported by empirical evidence. We evaluate a range of state-of-the-art agents with frontier LLMs backbones like gpt-5 on FIRE-Bench. Our results show that full-cycle scientific research remains challenging for current agent systems: even the strongest agents achieve limited rediscovery success (<50 F1), exhibit high variance across runs, and display recurring failure modes in experimental design, execution, and evidence-based reasoning. FIRE-Bench provides a rigorous and diagnostic framework for measuring progress toward reliable agent-driven scientific discovery.

Method: Extends the bounded attention prefix oracle (BAPO) model to analyze information flow requirements. Proves lower bounds on CoT tokens for three canonical BAPO-hard tasks: binary majority, triplet matching, and graph reachability. Provides matching/near-matching upper bounds via explicit constructions and validates with experiments on frontier reasoning models.

Result: Shows each task requires Ω(n) reasoning tokens when input size is n. Experiments with frontier models show approximately linear reasoning token scaling on these tasks and failures when constrained to smaller reasoning budgets, consistent with theoretical lower bounds.

Conclusion: Identifies fundamental bottlenecks in inference-time compute through CoT reasoning and offers a principled tool for analyzing optimal reasoning length. The results establish theoretical limits on reasoning efficiency for certain hard tasks.

Abstract: Inference-time scaling via chain-of-thought (CoT) reasoning is a major driver of state-of-the-art LLM performance, but it comes with substantial latency and compute costs. We address a fundamental theoretical question: how many reasoning tokens are required to solve a problem as input size grows? By extending the bounded attention prefix oracle (BAPO) model–an abstraction of LLMs that quantifies the information flow required to solve a task–we prove lower bounds on the CoT tokens required for three canonical BAPO-hard tasks: binary majority, triplet matching, and graph reachability. We show that each requires $Ω(n)$ reasoning tokens when the input size is $n$. We complement these results with matching or near-matching upper bounds via explicit constructions. Finally, our experiments with frontier reasoning models show approximately linear reasoning token scaling on these tasks and failures when constrained to smaller reasoning budgets, consistent with our theoretical lower bounds. Together, our results identify fundamental bottlenecks in inference-time compute through CoT and offer a principled tool for analyzing optimal reasoning length.

Method: Proposes DeltaEvolve framework that formalizes evolutionary agents as Expectation-Maximization: LLM samples programs (E-step) and system updates control context via evaluation feedback (M-step). Instead of full-code snapshots, uses structured semantic deltas capturing how and why modifications affect performance, organized through multi-level database and progressive disclosure to reduce tokens.

Result: Empirical evaluations across diverse scientific domains show DeltaEvolve discovers better solutions with less token consumption compared to full-code-based evolutionary agents.

Conclusion: Semantic deltas provide more informative evolutionary guidance than full-code histories, enabling more efficient LLM-driven scientific discovery through better context construction and reduced computational overhead.

Abstract: LLM-driven evolutionary systems have shown promise for automated science discovery, yet existing approaches such as AlphaEvolve rely on full-code histories that are context-inefficient and potentially provide weak evolutionary guidance. In this work, we first formalize the evolutionary agents as a general Expectation-Maximization framework, where the language model samples candidate programs (E-step) and the system updates the control context based on evaluation feedback (M-step). Under this view, constructing context via full-code snapshots constitutes a suboptimal M-step, as redundant implement details dilutes core algorithmic ideas, making it difficult to provide clear inspirations for evolution. To address this, we propose DeltaEvolve, a momentum-driven evolutionary framework that replaces full-code history with structured semantic delta capturing how and why modifications between successive nodes affect performance. As programs are often decomposable, semantic delta usually contains many effective components which are transferable and more informative to drive improvement. By organizing semantic delta through multi-level database and progressive disclosure mechanism, input tokens are further reduced. Empirical evaluations on tasks across diverse scientific domains show that our framework can discover better solution with less token consumption over full-code-based evolutionary agents.

Method: UAT-LITE uses approximate Bayesian inference via Monte Carlo dropout in pretrained transformer classifiers during inference. Token-level epistemic uncertainty is estimated from stochastic forward passes and used to modulate self-attention during contextualization. Also introduces layerwise variance decomposition to diagnose uncertainty accumulation across transformer depth.

Result: Across SQuAD 2.0 answerability, MNLI, and SST-2, UAT-LITE reduces Expected Calibration Error by ~20% on average relative to fine-tuned BERT-base baseline while preserving task accuracy. Improves selective prediction and robustness under distribution shift.

Conclusion: UAT-LITE provides an effective inference-time framework for uncertainty-aware self-attention in transformers, improving calibration without weight modifications or expensive training procedures.

Abstract: Neural NLP models are often miscalibrated, assigning high confidence to incorrect predictions, which undermines selective prediction and high-stakes deployment. Post-hoc calibration methods adjust output probabilities but leave internal computation unchanged, while ensemble and Bayesian approaches improve uncertainty at substantial training or storage cost. We propose UAT-LITE, an inference-time framework that makes self-attention uncertainty-aware using approximate Bayesian inference via Monte Carlo dropout in pretrained transformer classifiers. Token-level epistemic uncertainty is estimated from stochastic forward passes and used to modulate self-attention during contextualization, without modifying pretrained weights or training objectives. We additionally introduce a layerwise variance decomposition to diagnose how predictive uncertainty accumulates across transformer depth. Across the SQuAD 2.0 answerability, MNLI, and SST-2, UAT-LITE reduces Expected Calibration Error by approximately 20% on average relative to a fine-tuned BERT-base baseline while preserving task accuracy, and improves selective prediction and robustness under distribution shift.

Method: Two-level framework combining: 1) adaptive path expansion that expands agent paths to goals in multiple stages, and 2) dynamic leading technique that allows reselection of leading agent during each expansion when progress stalls.

Result: The planner handles up to 25 agents with limited communication range and 11-12 agents with line-of-sight constraints across multiple environment types, achieving over 90% success rate where baselines fail.

Conclusion: The proposed framework effectively solves multi-agent pathfinding under communication constraints by adaptively expanding paths and dynamically selecting leaders, outperforming existing approaches.

Abstract: This paper proposes a novel planning framework to handle a multi-agent pathfinding problem under team-connected communication constraint, where all agents must have a connected communication channel to the rest of the team during their entire movements. Standard multi-agent path finding approaches (e.g., priority-based search) have potential in this domain but fail when neighboring configurations at start and goal differ. Their single-expansion approach – computing each agent’s path from the start to the goal in just a single expansion – cannot reliably handle planning under communication constraints for agents as their neighbors change during navigating. Similarly, leader-follower approaches (e.g., platooning) are effective at maintaining team communication, but fixing the leader at the outset of planning can cause planning to become stuck in dense-clutter environments, limiting their practical utility. To overcome this limitation, we propose a novel two-level multi-agent pathfinding framework that integrates two techniques: adaptive path expansion to expand agent paths to their goals in multiple stages; and dynamic leading technique that enables the reselection of the leading agent during each agent path expansion whenever progress cannot be made. Simulation experiments show the efficiency of our planners, which can handle up to 25 agents across five environment types under a limited communication range constraint and up to 11-12 agents on three environment types under line-of-sight communication constraint, exceeding 90% success-rate where baselines routinely fail.

Method: Fine-tune Vision-Language Models (VLMs) to predict what users would search for from images (reverse search design). Augment with AI agents mining real-time internet trends. Use VLM-generated queries to construct semantically coherent Collection Pages via multimodal embeddings. Employ hybrid VLM and two-tower ANN architectures for authority-aware interlinking across billions of visual assets.

Result: Deployed at scale across billions of images and tens of millions of collections. Achieved 20% organic traffic growth contributing to multi-million monthly active user (MAU) growth.

Conclusion: Presents a principled pathway for visual platforms to thrive in the generative search era through reverse search design and GEO framework.

Abstract: Large Language Models are fundamentally reshaping content discovery through AI-native search systems such as ChatGPT, Gemini, and Claude. Unlike traditional search engines that match keywords to documents, these systems infer user intent, synthesize multimodal evidence, and generate contextual answers directly on the search page, introducing a paradigm shift from Search Engine Optimization (SEO) to Generative Engine Optimization (GEO). For visual content platforms hosting billions of assets, this poses an acute challenge: individual images lack the semantic depth and authority signals that generative search prioritizes, risking disintermediation as user needs are satisfied in-place without site visits. We present Pinterest GEO, a production-scale framework that pioneers reverse search design: rather than generating generic image captions describing what content is, we fine-tune Vision-Language Models (VLMs) to predict what users would actually search for, augmented this with AI agents that mine real-time internet trends to capture emerging search demand. These VLM-generated queries then drive construction of semantically coherent Collection Pages via multimodal embeddings, creating indexable aggregations optimized for generative retrieval. Finally, we employ hybrid VLM and two-tower ANN architectures to build authority-aware interlinking structures that propagate signals across billions of visual assets. Deployed at scale across billions of images and tens of millions of collections, GEO delivers 20% organic traffic growth contributing to multi-million monthly active user (MAU) growth, demonstrating a principled pathway for visual platforms to thrive in the generative search era.

Method: Recasts value function estimates as learning a desired poset (partially ordered set). Proposes GCR-RL that computes super-poset refinements by refining previous posets and learning additional order relationships from temporal difference signals. Develops two algorithms: Q-learning and actor-critic variants to efficiently realize these super-poset refinements.

Result: Theoretical properties and convergence rates are analyzed. Empirical evaluation across various tasks demonstrates significant improvements in sample efficiency and stable performance over strong baselines.

Conclusion: Geometric coherence regularization through poset refinement provides a principled approach to improve RL stability and sample efficiency, with both theoretical guarantees and empirical effectiveness.

Abstract: Geometric properties can be leveraged to stabilize and speed reinforcement learning. Existing examples include encoding symmetry structure, geometry-aware data augmentation, and enforcing structural restrictions. In this paper, we take a novel view of RL through the lens of order theory and recast value function estimates into learning a desired poset (partially ordered set). We propose \emph{GCR-RL} (Geometric Coherence Regularized Reinforcement Learning) that computes a sequence of super-poset refinements – by refining posets in previous steps and learning additional order relationships from temporal difference signals – thus ensuring geometric coherence across the sequence of posets underpinning the learned value functions. Two novel algorithms by Q-learning and by actor–critic are developed to efficiently realize these super-poset refinements. Their theoretical properties and convergence rates are analyzed. We empirically evaluate GCR-RL in a range of tasks and demonstrate significant improvements in sample efficiency and stable performance over strong baselines.

Method: Benchmarked 20+ LLMs against human baseline on 11 causal judgment tasks formalized by collider structures (C₁→E←C₂), using interpretable models to compress LLM judgments, and probing robustness under semantic abstraction and prompt overloading.

Result: Most LLMs exhibit more rule-like reasoning than humans, don’t mirror human collider biases (weak explaining away, Markov violations), and CoT increases robustness for many LLMs.

Conclusion: LLMs can complement humans when known biases are undesirable, but their rule-like reasoning may break down with uncertainty, highlighting need to characterize LLM reasoning for safe deployment.

Abstract: Large language models (LLMs) are increasingly used in domains where causal reasoning matters, yet it remains unclear whether their judgments reflect normative causal computation, human-like shortcuts, or brittle pattern matching. We benchmark 20+ LLMs against a matched human baseline on 11 causal judgment tasks formalized by a collider structure ($C_1 !\rightarrow! E! \leftarrow !C_2$). We find that a small interpretable model compresses LLMs’ causal judgments well and that most LLMs exhibit more rule-like reasoning strategies than humans who seem to account for unmentioned latent factors in their probability judgments. Furthermore, most LLMs do not mirror the characteristic human collider biases of weak explaining away and Markov violations. We probe LLMs’ causal judgment robustness under (i) semantic abstraction and (ii) prompt overloading (injecting irrelevant text), and find that chain-of-thought (CoT) increases robustness for many LLMs. Together, this divergence suggests LLMs can complement humans when known biases are undesirable, but their rule-like reasoning may break down when uncertainty is intrinsic – highlighting the need to characterize LLM reasoning strategies for safe, effective deployment.

Method: Proposes a Bayesian account that grounds planning behavior in the evolving generative context, suggesting that accumulated self-generated context drives a planning-shift during inference. Validates through two controlled experiments: random-generation task showing constrained planning under human prompts and increasing planning strength with self-generated context, and Gaussian-sampling task showing reduced initial bias when conditioning on self-generated sequences.

Result: The experiments demonstrate: 1) constrained planning under human prompts with increasing planning strength as self-generated context accumulates, and 2) reduced initial bias when conditioning on self-generated sequences in Gaussian-sampling tasks.

Conclusion: Provides a theoretical explanation with empirical evidence for characterizing how LLMs plan ahead during inference, explaining the apparent gap between training capabilities and inference behavior through the lens of Bayesian modeling and context accumulation.

Abstract: Large language models (LLMs) have been shown to acquire sequence-level planning abilities during training, yet their planning behavior exhibited at inference time often appears short-sighted and inconsistent with these capabilities. We propose a Bayesian account for this gap by grounding planning behavior in the evolving generative context: given the subtle differences between natural language and the language internalized by LLMs, accumulated self-generated context drives a planning-shift during inference and thereby creates the appearance of compromised planning behavior. We further validate the proposed model through two controlled experiments: a random-generation task demonstrating constrained planning under human prompts and increasing planning strength as self-generated context accumulates, and a Gaussian-sampling task showing reduced initial bias when conditioning on self-generated sequences. These findings provide a theoretical explanation along with empirical evidence for characterizing how LLMs plan ahead during inference.

Method: Introduces Agent Alpha framework with step-level MCTS, alpha-UCT guided search, comparison-driven evaluation to mitigate scoring biases, and diversity-constrained expansion to maintain compact search space.

Result: Achieves state-of-the-art success rate of ~77% on OSWorld benchmark, significantly outperforming trajectory-level baselines under equivalent compute.

Conclusion: Agent Alpha demonstrates the effectiveness of step-level MCTS with alpha-UCT guidance for GUI agents, enabling better planning, exploration, and recovery capabilities compared to trajectory-level approaches.

Abstract: While scaling test-time compute through trajectory-level sampling has significantly improved Graphical User Interface (GUI) agents, the lack of regressive ability prevents the reuse of partial successes and the recovery from early missteps. In this paper, we introduce Agent Alpha, a unified framework that synergizes generation, exploration, and evaluation through step-level Monte Carlo Tree Search (MCTS). It enables active modeling or exploiting structures of the planning space. By integrating alpha-UCT guided search into the interaction loop, Agent Alpha enables deliberate planning, facilitating early pruning of suboptimal branches and efficient prefix reuse. We also employ comparison-driven evaluation to mitigate absolute scoring biases and diversity-constrained expansion to maintain a compact, informative search space. Regret bound of alpha-UCT is analyzed. On the OSWorld benchmark, Agent Alpha achieves a state-of-the-art success rate of $\sim 77%$, significantly outperforming trajectory-level baselines under equivalent compute.

Method: Survey approach synthesizing work across auctions, voting, budgeting, liquid democracy, decentralized aggregation, and inverse mechanism learning. Formulates voting rules and aggregation procedures as learnable, differentiable models optimized from data.

Result: Comprehensive review showing how classical axioms and impossibility results reappear as objectives, constraints, and optimization trade-offs in differentiable frameworks. Identifies 36 open problems defining new research agenda.

Conclusion: Differentiable social choice represents an emerging paradigm that connects ML optimization with normative social choice theory, enabling data-driven design of aggregation mechanisms while maintaining theoretical guarantees.

Abstract: Social choice is no longer a peripheral concern of political theory or economics-it has become a foundational component of modern machine learning systems. From auctions and resource allocation to federated learning, participatory governance, and the alignment of large language models, machine learning pipelines increasingly aggregate heterogeneous preferences, incentives, and judgments into collective decisions. In effect, many contemporary machine learning systems already implement social choice mechanisms, often implicitly and without explicit normative scrutiny. This Review surveys differentiable social choice: an emerging paradigm that formulates voting rules, mechanisms, and aggregation procedures as learnable, differentiable models optimized from data. We synthesize work across auctions, voting, budgeting, liquid democracy, decentralized aggregation, and inverse mechanism learning, showing how classical axioms and impossibility results reappear as objectives, constraints, and optimization trade-offs. We conclude by identifying 36 open problems defining a new research agenda at the intersection of machine learning, economics, and democratic theory.

Method: Proposes Graph of Concept Predictors (GCP) that externalizes teacher’s decision process as a directed acyclic graph and mirrors it with modular concept predictors in the student. Uses graph-aware acquisition strategy targeting uncertainty at critical nodes and targeted sub-module retraining.

Result: Experiments on eight NLP classification benchmarks show GCP enhances performance under limited annotation budgets while yielding more interpretable and controllable training dynamics.

Conclusion: GCP provides a more efficient and interpretable active distillation framework that captures reasoning processes beyond just final labels.

Abstract: Deploying Large Language Models (LLMs) for discriminative workloads is often limited by inference latency, compute, and API costs at scale. Active distillation reduces these costs by querying an LLM oracle to train compact discriminative students, but most pipelines distill only final labels, discarding intermediate reasoning signals and offering limited diagnostics of what reasoning is missing and where errors arise. We propose Graph of Concept Predictors (GCP), a reasoning-aware active distillation framework that externalizes the teacher’s decision process as a directed acyclic graph and mirrors it with modular concept predictors in the student. GCP enhances sample efficiency through a graph-aware acquisition strategy that targets uncertainty and disagreement at critical reasoning nodes. Additionally, it improves training stability and efficiency by performing targeted sub-module retraining, which attributes downstream loss to specific concept predictors and updates only the most influential modules. Experiments on eight NLP classification benchmarks demonstrate that GCP enhances performance under limited annotation budgets while yielding more interpretable and controllable training dynamics. Code is available at: https://github.com/Ziyang-Yu/GCP.

Method: STAR framework with two innovations: (1) Constrained Knowledge Distillation (CKD) using top-k forward KL divergence to suppress incorrect predictions while preserving exploration capacity; (2) Similarity-guided RL (Sim-RL) with fine-grained, similarity-based rewards for better policy optimization.

Result: STAR models achieve SOTA in their size classes, with 0.6B model performing best among all open models under 1B, surpassing several larger models. Extensive experiments on challenging benchmarks demonstrate effectiveness.

Conclusion: STAR provides a training framework that successfully distills LLM capabilities into super-tiny models, enabling powerful, accessible, and efficient AI agents for function calling tasks.

Abstract: The proliferation of Large Language Models (LLMs) in function calling is pivotal for creating advanced AI agents, yet their large scale hinders widespread adoption, necessitating transferring their capabilities into smaller ones. However, existing paradigms are often plagued by overfitting, training instability, ineffective binary rewards for multi-solution tasks, and the difficulty of synergizing techniques. We introduce STAR: Similarity-guided Teacher-Assisted Refinement, a novel holistic framework that effectively transfers LLMs’ capabilities to super-tiny models. STAR consists of two core technical innovations: (1) Constrained Knowledge Distillation (CKD), a training objective that augments top-k forward KL divergence to suppress confidently incorrect predictions, ensuring training stability while preserving exploration capacity for downstream RL. STAR holistically synergizes these strategies within a cohesive training curriculum, enabling super-tiny models to achieve exceptional performance on complex function calling tasks; (2) Similarity-guided RL (Sim-RL), a RL mechanism that introduces a fine-grained, similarity-based reward. This provides a robust, continuous, and rich signal for better policy optimization by evaluating the similarity between generated outputs and the ground truth. Extensive experiments on challenging and renowned benchmarks demonstrate the effectiveness of our method. Our STAR models establish SOTA in their size classes, significantly outperforming baselines. Remarkably, our 0.6B STAR model achieves the best performance among all open models under 1B, surpassing even several well-known open models at a larger scale. STAR demonstrates a training framework that distills capabilities of LLMs into super-tiny models, paving the way for powerful, accessible, and efficient AI agents.

Method: Proposes RC-GRPO (Reward-Conditioned Group Relative Policy Optimization) which treats exploration as controllable steering via discrete reward tokens. First fine-tunes a Reward-Conditioned Trajectory Policy (RCTP) on mixed-quality trajectories with reward goal tokens (e.g., <|high_reward|>, <|low_reward|>) injected into prompts. During RL, samples diverse reward tokens within each GRPO group and conditions rollouts on sampled tokens to improve within-group diversity.

Result: On the Berkeley Function Calling Leaderboard v4 (BFCLv4) multi-turn benchmark, RC-GRPO yields consistently improved performance over baselines. Performance on Qwen-2.5-7B-Instruct even surpasses all closed-source API models.

Conclusion: RC-GRPO effectively addresses the low within-group diversity problem in multi-turn tool calling by using reward-conditioned steering, leading to better exploration and improved performance on challenging benchmarks.

Abstract: Multi-turn tool calling is challenging for Large Language Models (LLMs) because rewards are sparse and exploration is expensive. A common recipe, SFT followed by GRPO, can stall when within-group reward variation is low (e.g., more rollouts in a group receive the all 0 or all 1 reward), making the group-normalized advantage uninformative and yielding vanishing updates. To address this problem, we propose RC-GRPO (Reward-Conditioned Group Relative Policy Optimization), which treats exploration as a controllable steering problem via discrete reward tokens. We first fine-tune a Reward-Conditioned Trajectory Policy (RCTP) on mixed-quality trajectories with reward goal special tokens (e.g., <|high_reward|>, <|low_reward|>) injected into the prompts, enabling the model to learn how to generate distinct quality trajectories on demand. Then during RL, we sample diverse reward tokens within each GRPO group and condition rollouts on the sampled token to improve within-group diversity, improving advantage gains. On the Berkeley Function Calling Leaderboard v4 (BFCLv4) multi-turn benchmark, our method yields consistently improved performance than baselines, and the performance on Qwen-2.5-7B-Instruct even surpasses all closed-source API models.

Method: Proposes KANFIS (Kolmogorov-Arnold Neuro-Fuzzy Inference System) using additive function decomposition with linear scaling of parameters/rules, sparse masking for compact rule sets, and compatibility with Type-1 and Interval Type-2 fuzzy logic.

Result: Achieves competitive performance against neural and neuro-fuzzy baselines while providing interpretable models with clear rule semantics and transparent inference processes.

Conclusion: KANFIS offers a compact, interpretable neuro-fuzzy architecture that avoids exponential complexity while maintaining performance and enabling uncertainty modeling.

Abstract: Adaptive Neuro-Fuzzy Inference System (ANFIS) was designed to combine the learning capabilities of neural network with the reasoning transparency of fuzzy logic. However, conventional ANFIS architectures suffer from structural complexity, where the product-based inference mechanism causes an exponential explosion of rules in high-dimensional spaces. We herein propose the Kolmogorov-Arnold Neuro-Fuzzy Inference System (KANFIS), a compact neuro-symbolic architecture that unifies fuzzy reasoning with additive function decomposition. KANFIS employs an additive aggregation mechanism, under which both model parameters and rule complexity scale linearly with input dimensionality rather than exponentially. Furthermore, KANFIS is compatible with both Type-1 (T1) and Interval Type-2 (IT2) fuzzy logic systems, enabling explicit modeling of uncertainty and ambiguity in fuzzy representations. By using sparse masking mechanisms, KANFIS generates compact and structured rule sets, resulting in an intrinsically interpretable model with clear rule semantics and transparent inference processes. Empirical results demonstrate that KANFIS achieves competitive performance against representative neural and neuro-fuzzy baselines.

Method: JobRec introduces: 1) Unified Semantic Alignment Schema that aligns candidate and job attributes into structured semantic layers; 2) Two-Stage Cooperative Training Strategy that learns decoupled experts to separately infer preference and qualification; 3) Lagrangian-based Policy Alignment module that optimizes recommendations under explicit eligibility requirements; 4) Synthetic dataset construction refined by experts to mitigate data scarcity.

Result: Experiments show that JobRec consistently outperforms strong baselines and provides improved controllability for strategy-aware professional matching.

Conclusion: JobRec successfully addresses the challenges of confounded supervision and limited controllability in professional job recommendation by decoupling preference and qualification through structured semantic alignment and dual-perspective reasoning, enabling more accurate and controllable matching.

Abstract: Professional job recommendation involves a complex bipartite matching process that must reconcile a candidate’s subjective preference with an employer’s objective qualification. While Large Language Models (LLMs) are well-suited for modeling the rich semantics of resumes and job descriptions, existing paradigms often collapse these two decision dimensions into a single interaction signal, yielding confounded supervision under recruitment-funnel censoring and limiting policy controllability. To address these challenges, We propose JobRec, a generative job recommendation framework for de-conflating preference and qualification via constrained dual-perspective reasoning. JobRec introduces a Unified Semantic Alignment Schema that aligns candidate and job attributes into structured semantic layers, and a Two-Stage Cooperative Training Strategy that learns decoupled experts to separately infer preference and qualification. Building on these experts, a Lagrangian-based Policy Alignment module optimizes recommendations under explicit eligibility requirements, enabling controllable trade-offs. To mitigate data scarcity, we construct a synthetic dataset refined by experts. Experiments show that JobRec consistently outperforms strong baselines and provides improved controllability for strategy-aware professional matching.

Method: Proposes Risky-Bench framework that organizes evaluation around domain-agnostic safety principles to derive context-aware safety rubrics delineating safety space. Systematically evaluates safety risks through realistic task execution under varying threat assumptions. The framework can be adapted to different deployment settings to construct environment-specific safety evaluations.

Result: When applied to life-assist agent settings, Risky-Bench uncovers substantial safety risks in state-of-the-art agents under realistic execution conditions. The framework demonstrates effectiveness in identifying safety vulnerabilities that traditional evaluations miss.

Conclusion: Risky-Bench provides an extensible methodology for agent safety assessment that goes beyond linguistic harm to address real-world deployment risks. It offers a systematic approach to evaluate safety across diverse agent configurations and deployment scenarios.

Abstract: Large Language Models (LLMs) are increasingly deployed as agents that operate in real-world environments, introducing safety risks beyond linguistic harm. Existing agent safety evaluations rely on risk-oriented tasks tailored to specific agent settings, resulting in limited coverage of safety risk space and failing to assess agent safety behavior during long-horizon, interactive task execution in complex real-world deployments. Moreover, their specialization to particular agent settings limits adaptability across diverse agent configurations. To address these limitations, we propose Risky-Bench, a framework that enables systematic agent safety evaluation grounded in real-world deployment. Risky-Bench organizes evaluation around domain-agnostic safety principles to derive context-aware safety rubrics that delineate safety space, and systematically evaluates safety risks across this space through realistic task execution under varying threat assumptions. When applied to life-assist agent settings, Risky-Bench uncovers substantial safety risks in state-of-the-art agents under realistic execution conditions. Moreover, as a well-structured evaluation pipeline, Risky-Bench is not confined to life-assist scenarios and can be adapted to other deployment settings to construct environment-specific safety evaluations, providing an extensible methodology for agent safety assessment.

Method: Introduced an architectural taxonomy for comparing multi-agent LLM frameworks along fundamental dimensions, and developed MAFBench - a unified evaluation suite that integrates existing benchmarks under a standardized execution pipeline for controlled empirical study.

Result: Framework-level design choices alone can increase latency by over 100x, reduce planning accuracy by up to 30%, and lower coordination success from above 90% to below 30%.

Conclusion: The study provides concrete architectural design principles and framework selection guidance, highlighting the critical importance of framework architecture in multi-agent LLM systems and outlining promising future research directions.

Abstract: Multi-agent LLM frameworks are widely used to accelerate the development of agent systems powered by large language models (LLMs). These frameworks impose distinct architectural structures that govern how agents interact, store information, and coordinate tasks. However, their impact on system performance remains poorly understood. This gap is critical, as architectural choices alone can induce order-of-magnitude differences in latency and throughput, as well as substantial variation in accuracy and scalability. Addressing this challenge requires (i) jointly evaluating multiple capabilities, such as orchestration overhead, memory behavior, planning, specialization, and coordination, and (ii) conducting these evaluations under controlled, framework-level conditions to isolate architectural effects. Existing benchmarks focus on individual capabilities and lack standardized framework-level evaluation. We address these limitations by (i) introducing an architectural taxonomy for systematically comparing multi-agent LLM frameworks along fundamental dimensions, and (ii) developing MAFBench, a unified evaluation suite that integrates existing benchmarks under a standardized execution pipeline. Using MAFBench, we conduct a controlled empirical study across several widely used frameworks. Our results show that framework-level design choices alone can increase latency by over 100x, reduce planning accuracy by up to 30%, and lower coordination success from above 90% to below 30%. Finally, we translate our findings into concrete architectural design principles and framework selection guidance, and outline promising future research directions.

Method: Theoretical extension of previous framework, proving theorems that stochastic agents operating in partially observable environments must contain sufficient knowledge of their environment for approximate reconstruction, even when using randomization.

Result: Shows stochastic agents cannot avoid learning their environment through randomization, and strengthens the result by weakening the generality requirement, proving less powerful agents already contain world models.

Conclusion: The necessity of world models in optimal agents is fundamental and extends to stochastic agents in partially observable environments, demonstrating that randomization doesn’t circumvent the need to learn environment structure.

Abstract: Deciding whether an agent possesses a model of its surrounding world is a fundamental step toward understanding its capabilities and limitations. In [10], it was shown that, within a particular framework, every almost optimal and general agent necessarily contains sufficient knowledge of its environment to allow an approximate reconstruction of it by querying the agent as a black box. This result relied on the assumptions that the agent is deterministic and that the environment is fully observable. In this work, we remove both assumptions by extending the theorem to stochastic agents operating in partially observable environments. Fundamentally, this shows that stochastic agents cannot avoid learning their environment through the usage of randomization. We also strengthen the result by weakening the notion of generality, proving that less powerful agents already contain a model of the world in which they operate.

Method: Proposes a pluggable mid-stage training module using an enhanced diffusion model with two key innovations: (1) Dynamic Modality Gating that adaptively uses conditional features to guide semantically consistent feature generation, and (2) Cross-Modal Mutual Learning that bridges semantic spaces of dual encoders for bidirectional alignment.

Result: Zero-shot evaluations across benchmark datasets show the approach outperforms existing baseline methods. Extensive experiments confirm the model as a robust and scalable extension for VLMs in missing modality scenarios, ensuring reliability across diverse missing rates and environments.

Conclusion: The proposed restoration strategy effectively handles missing modalities in VLMs through diffusion-based feature restoration with adaptive gating and cross-modal alignment, providing a general solution for incomplete modality scenarios.

Abstract: Vision Language Models (VLMs) typically assume complete modality input during inference. However, their effectiveness drops sharply when certain modalities are unavailable or incomplete. Current research primarily faces two dilemmas: Prompt-based methods struggle to restore missing yet indispensable features and impair generalization of VLMs. Imputation-based approaches, lacking effective guidance, are prone to generating semantically irrelevant noise. Restoring precise semantics while sustaining VLM generalization remains challenging. Therefore, we propose a general missing modality restoration strategy in this paper. We introduce an enhanced diffusion model as a pluggable mid-stage training module to effectively restore missing features. Our strategy introduces two key innovations: (I) Dynamic Modality Gating, which adaptively leverages conditional features to steer the generation of semantically consistent features; (II) Cross-Modal Mutual Learning mechanism, which bridges the semantic spaces of dual encoders to achieve bidirectional alignment. Zero-shot evaluations across benchmark datasets demonstrate that our approach outperforms existing baseline methods. Extensive experiments and ablation studies confirm our model as a robust and scalable extension for VLMs in missing modality scenarios, ensuring reliability across diverse missing rates and environments. Our code and models will be publicly available.

Method: Three-component framework: 1) HIVES for hierarchical value embedding, 2) VIDB database of value-labeled texts with intensity estimates, 3) anchor-based evaluator for consistent intensity scoring.

Result: Comprehensive study across 10 models and 4 value theories reveals asymmetries in steerability and composition laws for multi-value control.

Conclusion: Establishes scalable infrastructure for evaluating and controlling value intensity, advancing pluralistic alignment of LLMs.

Abstract: Aligning Large Language Models (LLMs) with the diverse spectrum of human values remains a central challenge: preference-based methods often fail to capture deeper motivational principles. Value-based approaches offer a more principled path, yet three gaps persist: extraction often ignores hierarchical structure, evaluation detects presence but not calibrated intensity, and the steerability of LLMs at controlled intensities remains insufficiently understood. To address these limitations, we introduce VALUEFLOW, the first unified framework that spans extraction, evaluation, and steering with calibrated intensity control. The framework integrates three components: (i) HIVES, a hierarchical value embedding space that captures intra- and cross-theory value structure; (ii) the Value Intensity DataBase (VIDB), a large-scale resource of value-labeled texts with intensity estimates derived from ranking-based aggregation; and (iii) an anchor-based evaluator that produces consistent intensity scores for model outputs by ranking them against VIDB panels. Using VALUEFLOW, we conduct a comprehensive large-scale study across ten models and four value theories, identifying asymmetries in steerability and composition laws for multi-value control. This paper establishes a scalable infrastructure for evaluating and controlling value intensity, advancing pluralistic alignment of LLMs.

Method: TDScaling introduces four innovations: 1) Business Cluster mechanism for real-service logical dependencies, 2) blueprint-driven multi-agent paradigm for trajectory coherence, 3) adaptive evolution mechanism using Domain Entropy, Reasoning Mode Entropy, and Cumulative Action Complexity to prevent mode collapse, and 4) sandboxed code tool to mitigate catastrophic forgetting of coding capabilities.

Result: Experiments on general tool-use benchmarks (BFCL, tau^2-Bench) and code agent tasks (RebenchT, CodeCI, BIRD) show TDScaling improves both tool-use generalization and inherent coding proficiency, with 30,000+ tool clusters synthesized.

Conclusion: TDScaling demonstrates that scaling through trajectory diversity yields better performance-cost trade-offs than quantity scaling for code agent training, offering a win-win for both tool-use generalization and coding capabilities.

Abstract: As code large language models (LLMs) evolve into tool-interactive agents via the Model Context Protocol (MCP), their generalization is increasingly limited by low-quality synthetic data and the diminishing returns of quantity scaling. Moreover, quantity-centric scaling exhibits an early bottleneck that underutilizes trajectory data. We propose TDScaling, a Trajectory Diversity Scaling-based data synthesis framework for code agents that scales performance through diversity rather than raw volume. Under a fixed training budget, increasing trajectory diversity yields larger gains than adding more trajectories, improving the performance-cost trade-off for agent training. TDScaling integrates four innovations: (1) a Business Cluster mechanism that captures real-service logical dependencies; (2) a blueprint-driven multi-agent paradigm that enforces trajectory coherence; (3) an adaptive evolution mechanism that steers synthesis toward long-tail scenarios using Domain Entropy, Reasoning Mode Entropy, and Cumulative Action Complexity to prevent mode collapse; and (4) a sandboxed code tool that mitigates catastrophic forgetting of intrinsic coding capabilities. Experiments on general tool-use benchmarks (BFCL, tau^2-Bench) and code agent tasks (RebenchT, CodeCI, BIRD) demonstrate a win-win outcome: TDScaling improves both tool-use generalization and inherent coding proficiency. We plan to release the full codebase and the synthesized dataset (including 30,000+ tool clusters) upon publication.

Method: Proposes TAME framework with dual-memory evolution: executor memory evolves to improve task performance by distilling generalizable methodologies, while evaluator memory refines safety and utility assessments based on historical feedback. Uses closed loop of memory filtering, draft generation, trustworthy refinement, execution, and dual-track memory updating.

Result: Experiments show TAME mitigates misevolution, achieving joint improvement in both trustworthiness and task performance. The Trust-Memevo benchmark reveals overall trustworthiness decline during benign task evolution across various domains.

Conclusion: Dual-memory evolutionary framework effectively addresses agent memory misevolution, preserving safety alignment during test-time evolution while maintaining task utility.

Abstract: Test-time evolution of agent memory serves as a pivotal paradigm for achieving AGI by bolstering complex reasoning through experience accumulation. However, even during benign task evolution, agent safety alignment remains vulnerable-a phenomenon known as Agent Memory Misevolution. To evaluate this phenomenon, we construct the Trust-Memevo benchmark to assess multi-dimensional trustworthiness during benign task evolution, revealing an overall decline in trustworthiness across various task domains and evaluation settings. To address this issue, we propose TAME, a dual-memory evolutionary framework that separately evolves executor memory to improve task performance by distilling generalizable methodologies, and evaluator memory to refine assessments of both safety and task utility based on historical feedback. Through a closed loop of memory filtering, draft generation, trustworthy refinement, execution, and dual-track memory updating, TAME preserves trustworthiness without sacrificing utility. Experiments demonstrate that TAME mitigates misevolution, achieving a joint improvement in both trustworthiness and task performance.

Method: Introduces a proposal for standardizing agent evaluation through a unified framework that addresses issues with fragmented researcher-specific frameworks, varying prompt engineering, and lack of standardized environmental data.

Result: The paper presents a conceptual framework proposal rather than empirical results, arguing that standardization is essential for rigorous advancement of agent evaluation.

Conclusion: A unified evaluation framework is essential for the rigorous advancement of agent evaluation to address current issues with standardization, fairness, and reproducibility in LLM-based agent benchmarking.

Abstract: With the advent of Large Language Models (LLMs), general-purpose agents have seen fundamental advancements. However, evaluating these agents presents unique challenges that distinguish them from static QA benchmarks. We observe that current agent benchmarks are heavily confounded by extraneous factors, including system prompts, toolset configurations, and environmental dynamics. Existing evaluations often rely on fragmented, researcher-specific frameworks where the prompt engineering for reasoning and tool usage varies significantly, making it difficult to attribute performance gains to the model itself. Additionally, the lack of standardized environmental data leads to untraceable errors and non-reproducible results. This lack of standardization introduces substantial unfairness and opacity into the field. We propose that a unified evaluation framework is essential for the rigorous advancement of agent evaluation. To this end, we introduce a proposal aimed at standardizing agent evaluation.

Method: Introduces Accordion-Thinking where LLMs learn to self-regulate reasoning granularity through dynamic summarization. Uses Fold inference mode where models periodically summarize thoughts and discard former tokens. Applies reinforcement learning to incentivize this compression capability.

Result: Achieves 3x throughput while maintaining accuracy on 48GB GPU memory configuration. The accuracy gap between efficient Fold mode and exhaustive Unfold mode vanishes during training, showing models learn to encode essential reasoning into compact summaries.

Conclusion: LLMs can tackle complex reasoning tasks with minimal dependency token overhead through learned self-compression, while structured step summaries provide human-readable reasoning accounts.

Abstract: Scaling test-time compute via long Chain-ofThought unlocks remarkable gains in reasoning capabilities, yet it faces practical limits due to the linear growth of KV cache and quadratic attention complexity. In this paper, we introduce Accordion-Thinking, an end-to-end framework where LLMs learn to self-regulate the granularity of the reasoning steps through dynamic summarization. This mechanism enables a Fold inference mode, where the model periodically summarizes its thought process and discards former thoughts to reduce dependency on historical tokens. We apply reinforcement learning to incentivize this capability further, uncovering a critical insight: the accuracy gap between the highly efficient Fold mode and the exhaustive Unfold mode progressively narrows and eventually vanishes over the course of training. This phenomenon demonstrates that the model learns to encode essential reasoning information into compact summaries, achieving effective compression of the reasoning context. Our Accordion-Thinker demonstrates that with learned self-compression, LLMs can tackle complex reasoning tasks with minimal dependency token overhead without compromising solution quality, and it achieves a 3x throughput while maintaining accuracy on a 48GB GPU memory configuration, while the structured step summaries provide a human-readable account of the reasoning process.

Method: Created LPS-Bench with 65 scenarios across 7 task domains and 9 risk types. Uses a multi-agent automated pipeline for scalable data generation and adopts an LLM-as-a-judge evaluation protocol to assess safety awareness through planning trajectories.

Result: Experiments reveal substantial deficiencies in existing CUAs’ ability to maintain safe behavior during long-horizon planning. The benchmark successfully identifies safety gaps and enables analysis of risks in MCP-based CUA systems.

Conclusion: Planning-time safety awareness is critical for computer-use agents, especially in long-horizon tasks. LPS-Bench provides a comprehensive evaluation framework, reveals current system deficiencies, and enables development of mitigation strategies for safer CUA systems.

Abstract: Computer-use agents (CUAs) that interact with real computer systems can perform automated tasks but face critical safety risks. Ambiguous instructions may trigger harmful actions, and adversarial users can manipulate tool execution to achieve malicious goals. Existing benchmarks mostly focus on short-horizon or GUI-based tasks, evaluating on execution-time errors but overlooking the ability to anticipate planning-time risks. To fill this gap, we present LPS-Bench, a benchmark that evaluates the planning-time safety awareness of MCP-based CUAs under long-horizon tasks, covering both benign and adversarial interactions across 65 scenarios of 7 task domains and 9 risk types. We introduce a multi-agent automated pipeline for scalable data generation and adopt an LLM-as-a-judge evaluation protocol to assess safety awareness through the planning trajectory. Experiments reveal substantial deficiencies in existing CUAs’ ability to maintain safe behavior. We further analyze the risks and propose mitigation strategies to improve long-horizon planning safety in MCP-based CUA systems. We open-source our code at https://github.com/tychenn/LPS-Bench.

Method: Introduces CSR-Bench with four stress-testing interaction patterns (Safety, Over-rejection, Bias, Hallucination) covering 61 fine-grained types. Each instance requires integrated image-text interpretation, with paired text-only controls to diagnose modality-induced behavior shifts. Evaluates 16 state-of-the-art MLLMs.

Result: Models show systematic cross-modal alignment gaps: weak safety awareness, strong language dominance under interference, and consistent performance degradation from text-only to multimodal inputs. Clear trade-off observed between reducing over-rejection and maintaining safe, non-discriminatory behavior.

Conclusion: MLLMs exhibit systematic reliability issues in cross-modal understanding, with apparent safety gains potentially coming from refusal-oriented heuristics rather than robust intent understanding. The benchmark reveals fundamental challenges in achieving true multimodal alignment.

Abstract: Multimodal large language models (MLLMs) enable interaction over both text and images, but their safety behavior can be driven by unimodal shortcuts instead of true joint intent understanding. We introduce CSR-Bench, a benchmark for evaluating cross-modal reliability through four stress-testing interaction patterns spanning Safety, Over-rejection, Bias, and Hallucination, covering 61 fine-grained types. Each instance is constructed to require integrated image-text interpretation, and we additionally provide paired text-only controls to diagnose modality-induced behavior shifts. We evaluate 16 state-of-the-art MLLMs and observe systematic cross-modal alignment gaps. Models show weak safety awareness, strong language dominance under interference, and consistent performance degradation from text-only controls to multimodal inputs. We also observe a clear trade-off between reducing over-rejection and maintaining safe, non-discriminatory behavior, suggesting that some apparent safety gains may come from refusal-oriented heuristics rather than robust intent understanding. WARNING: This paper contains unsafe contents.

Method: Models problem synthesis as goal-driven sequential decision process where specialized agents dynamically select and compose modular reasoning skills. Uses iterative workflow with internal reflection and tool-use, developing Agentic-Proposer-4B with Multi-Granularity Policy Optimization (MGPO).

Result: Downstream solvers trained on agent-synthesized data significantly outperform leading baselines and show robust cross-domain generalization. A 30B solver trained on only 11,000 synthesized trajectories achieves 91.6% accuracy on AIME25, rivaling frontier-scale proprietary models like GPT-5.

Conclusion: Small volumes of high-quality synthetic signals can effectively substitute for massive human-curated datasets, demonstrating the effectiveness of the agentic synthesis framework for generating verifiable reasoning training data.

Abstract: Advancing complex reasoning in large language models relies on high-quality, verifiable datasets, yet human annotation remains cost-prohibitive and difficult to scale. Current synthesis paradigms often face a recurring trade-off: maintaining structural validity typically restricts problem complexity, while relaxing constraints to increase difficulty frequently leads to inconsistent or unsolvable instances. To address this, we propose Agentic Proposing, a framework that models problem synthesis as a goal-driven sequential decision process where a specialized agent dynamically selects and composes modular reasoning skills. Through an iterative workflow of internal reflection and tool-use, we develop the Agentic-Proposer-4B using Multi-Granularity Policy Optimization (MGPO) to generate high-precision, verifiable training trajectories across mathematics, coding, and science. Empirical results demonstrate that downstream solvers trained on agent-synthesized data significantly outperform leading baselines and exhibit robust cross-domain generalization. Notably, a 30B solver trained on only 11,000 synthesized trajectories achieves a state-of-the-art 91.6% accuracy on AIME25, rivaling frontier-scale proprietary models such as GPT-5 and proving that a small volume of high-quality synthetic signals can effectively substitute for massive human-curated datasets.

Method: Three components: 1) MeetAll dataset from 231 enterprise meetings (140 hours) with questions injected using enterprise-informed protocol; 2) MeetBench XL multi-dimensional evaluation protocol; 3) MeetMaster XL dual-policy agent that optimizes query routing between fast/slow reasoning paths and tool invocation.

Result: Experiments show consistent gains over commercial systems, with lightweight classifier enabling accurate routing with minimal overhead and superior quality-latency tradeoff over single-model baselines.

Conclusion: The framework addresses enterprise meeting AI assistant needs through grounded data, comprehensive evaluation, and optimized agent architecture that balances performance with practical constraints.

Abstract: Enterprise meeting environments require AI assistants that handle diverse operational tasks, from rapid fact checking during live discussions to cross meeting analysis for strategic planning, under strict latency, cost, and privacy constraints. Existing meeting benchmarks mainly focus on simplified question answering and fail to reflect real world enterprise workflows, where queries arise organically from multi stakeholder collaboration, span long temporal contexts, and require tool augmented reasoning. We address this gap through a grounded dataset and a learned agent framework. First, we introduce MeetAll, a bilingual and multimodal corpus derived from 231 enterprise meetings totaling 140 hours. Questions are injected using an enterprise informed protocol validated by domain expert review and human discriminability studies. Unlike purely synthetic benchmarks, this protocol is grounded in four enterprise critical dimensions: cognitive load, temporal context span, domain expertise, and actionable task execution, calibrated through interviews with stakeholders across finance, healthcare, and technology sectors. Second, we propose MeetBench XL, a multi dimensional evaluation protocol aligned with human judgment that measures factual fidelity, intent alignment, response efficiency, structural clarity, and completeness. Third, we present MeetMaster XL, a learned dual policy agent that jointly optimizes query routing between fast and slow reasoning paths and tool invocation, including retrieval, cross meeting aggregation, and web search. A lightweight classifier enables accurate routing with minimal overhead, achieving a superior quality latency tradeoff over single model baselines. Experiments against commercial systems show consistent gains, supported by ablations, robustness tests, and a real world deployment case study.Resources: https://github.com/huyuelin/MeetBench.

Method: Memora organizes information via primary abstractions that index concrete memory values and consolidate related updates into unified entries. Cue anchors expand retrieval access across diverse aspects and connect related memories. A retrieval policy actively exploits these memory connections to retrieve relevant information beyond direct semantic similarity.

Result: Memora establishes new state-of-the-art on LoCoMo and LongMemEval benchmarks, demonstrating better retrieval relevance and reasoning effectiveness as memory scales. Theoretically, standard RAG and Knowledge Graph-based memory systems emerge as special cases of this framework.

Conclusion: Memora provides a harmonic memory representation that structurally balances abstraction and specificity, enabling scalable agent memory systems with improved retrieval and reasoning capabilities.

Abstract: Agent memory systems must accommodate continuously growing information while supporting efficient, context-aware retrieval for downstream tasks. Abstraction is essential for scaling agent memory, yet it often comes at the cost of specificity, obscuring the fine-grained details required for effective reasoning. We introduce Memora, a harmonic memory representation that structurally balances abstraction and specificity. Memora organizes information via its primary abstractions that index concrete memory values and consolidate related updates into unified memory entries, while cue anchors expand retrieval access across diverse aspects of the memory and connect related memories. Building on this structure, we employ a retrieval policy that actively exploits these memory connections to retrieve relevant information beyond direct semantic similarity. Theoretically, we show that standard Retrieval-Augmented Generation (RAG) and Knowledge Graph (KG)-based memory systems emerge as special cases of our framework. Empirically, Memora establishes a new state-of-the-art on the LoCoMo and LongMemEval benchmarks, demonstrating better retrieval relevance and reasoning effectiveness as memory scales.

Method: Created MentalDx Bench with 712 de-identified electronic health records annotated by psychiatrists under ICD-11 guidelines covering 76 disorders across 16 categories. Developed MentalSeek-Dx model using supervised trajectory construction and curriculum-based reinforcement learning to internalize clinical reasoning.

Result: Evaluation of 18 LLMs shows paradigm misalignment: strong performance at coarse diagnostic categorization but systematic failure at disorder-level diagnosis. MentalSeek-Dx achieves SOTA performance with only 14B parameters.

Conclusion: Establishes clinically grounded framework for reliable psychiatric diagnosis, addressing gap between pattern-based LLM modeling and clinical hypothetico-deductive reasoning.

Abstract: Mental health disorders represent a burgeoning global public health challenge. While Large Language Models (LLMs) have demonstrated potential in psychiatric assessment, their clinical utility is severely constrained by benchmarks that lack ecological validity and fine-grained diagnostic supervision. To bridge this gap, we introduce \textbf{MentalDx Bench}, the first benchmark dedicated to disorder-level psychiatric diagnosis within real-world clinical settings. Comprising 712 de-identified electronic health records annotated by board-certified psychiatrists under ICD-11 guidelines, the benchmark covers 76 disorders across 16 diagnostic categories. Evaluation of 18 LLMs reveals a critical \textit{paradigm misalignment}: strong performance at coarse diagnostic categorization contrasts with systematic failure at disorder-level diagnosis, underscoring a gap between pattern-based modeling and clinical hypothetico-deductive reasoning. In response, we propose \textbf{MentalSeek-Dx}, a medical-specialized LLM trained to internalize this clinical reasoning process through supervised trajectory construction and curriculum-based reinforcement learning. Experiments on MentalDx Bench demonstrate that MentalSeek-Dx achieves state-of-the-art (SOTA) performance with only 14B parameters, establishing a clinically grounded framework for reliable psychiatric diagnosis.

Method: Built a custom 2-layer transformer model that processes structured trolley dilemma scenarios using embeddings encoding affected individuals, quantities, and outcomes. Applied various interpretability techniques to analyze how moral reasoning distributes across the network.

Result: Achieved 77% accuracy on Moral Machine data while keeping the model small enough for detailed analysis. Found that biases in moral reasoning localize to distinct computational stages within the network.

Conclusion: Transformer models can effectively learn moral decision-making patterns from structured ethical dilemmas, and interpretability techniques reveal how moral reasoning is computationally distributed within neural networks.

Abstract: We build a custom transformer model to study how neural networks make moral decisions on trolley-style dilemmas. The model processes structured scenarios using embeddings that encode who is affected, how many people, and which outcome they belong to. Our 2-layer architecture achieves 77% accuracy on Moral Machine data while remaining small enough for detailed analysis. We use different interpretability techniques to uncover how moral reasoning distributes across the network, demonstrating that biases localize to distinct computational stages among other findings.

Method: Two-step approach: 1) Fine-tune a lightweight prompt-LM using off-policy GFlowNet objective with replay-based training, 2) Dynamic Memory Update mechanism that updates meta-prompt using diverse prompts from replay buffer and top-performing prompts from priority queue.

Result: Outperforms recent discrete prompt optimization baselines across few-shot text classification, instruction induction benchmarks, and question answering tasks.

Conclusion: GFlowPO provides an effective probabilistic framework for prompt optimization that enables sample-efficient exploration and progressively concentrates search on high-reward regions.

Abstract: Finding effective prompts for language models (LMs) is critical yet notoriously difficult: the prompt space is combinatorially large, rewards are sparse due to expensive target-LM evaluation. Yet, existing RL-based prompt optimizers often rely on on-policy updates and a meta-prompt sampled from a fixed distribution, leading to poor sample efficiency. We propose GFlowPO, a probabilistic prompt optimization framework that casts prompt search as a posterior inference problem over latent prompts regularized by a meta-prompted reference-LM prior. In the first step, we fine-tune a lightweight prompt-LM with an off-policy Generative Flow Network (GFlowNet) objective, using a replay-based training policy that reuses past prompt evaluations to enable sample-efficient exploration. In the second step, we introduce Dynamic Memory Update (DMU), a training-free mechanism that updates the meta-prompt by injecting both (i) diverse prompts from a replay buffer and (ii) top-performing prompts from a small priority queue, thereby progressively concentrating the search process on high-reward regions. Across few-shot text classification, instruction induction benchmarks, and question answering tasks, GFlowPO consistently outperforms recent discrete prompt optimization baselines.

Method: RAI constructs an Unsafe Prototype Subspace from language embeddings and performs targeted modulation on selected high-risk visual tokens, explicitly activating safety-critical signals in the cross-modal feature space while preserving semantic integrity.

Result: Extensive experiments across multiple jailbreak and utility benchmarks show RAI substantially reduces attack success rate without compromising task performance.

Conclusion: RAI provides an effective, lightweight, training-free framework for safety calibration in VLMs that restores LLM-like risk recognition capabilities while maintaining cross-modal reasoning utility.

Abstract: Vision language models (VLMs) extend the reasoning capabilities of large language models (LLMs) to cross-modal settings, yet remain highly vulnerable to multimodal jailbreak attacks. Existing defenses predominantly rely on safety fine-tuning or aggressive token manipulations, incurring substantial training costs or significantly degrading utility. Recent research shows that LLMs inherently recognize unsafe content in text, and the incorporation of visual inputs in VLMs frequently dilutes risk-related signals. Motivated by this, we propose Risk Awareness Injection (RAI), a lightweight and training-free framework for safety calibration that restores LLM-like risk recognition by amplifying unsafe signals in VLMs. Specifically, RAI constructs an Unsafe Prototype Subspace from language embeddings and performs targeted modulation on selected high-risk visual tokens, explicitly activating safety-critical signals within the cross-modal feature space. This modulation restores the model’s LLM-like ability to detect unsafe content from visual inputs, while preserving the semantic integrity of original tokens for cross-modal reasoning. Extensive experiments across multiple jailbreak and utility benchmarks demonstrate that RAI substantially reduces attack success rate without compromising task performance.

Method: Introduces intuitionistic fuzzy preference-based conflict situations, develops conflict measures within this framework, constructs three-way conflict analysis models for trisecting agent pairs, agents, and issues, uses relative loss functions to calculate thresholds, and presents adjustment mechanism-based feasible strategies.

Result: The proposed model provides finer granularity in capturing conflict situations, enables three-way analysis of conflicts, and offers adjustment strategies that account for both adjustment magnitudes and conflict degrees. An illustrative example demonstrates validity and effectiveness.

Conclusion: The intuitionistic fuzzy preference-based conflict analysis framework overcomes limitations of classical models by providing more detailed representation of agents’ attitudes, enabling more sophisticated conflict analysis and resolution strategies.

Abstract: In three-way conflict analysis, preference-based conflict situations characterize agents’ attitudes towards issues by formally modeling their preferences over pairs of issues. However, existing preference-based conflict models rely exclusively on three qualitative relations, namely, preference, converse, and indifference, to describe agents’ attitudes towards issue pairs, which significantly limits their capacity in capturing the essence of conflict. To overcome this limitation, we introduce the concept of an intuitionistic fuzzy preference-based conflict situation that captures agents’ attitudes towards issue pairs with finer granularity than that afforded by classical preference-based models. Afterwards, we develop intuitionistic fuzzy preference-based conflict measures within this framework, and construct three-way conflict analysis models for trisecting the set of agent pairs, the agent set, and the issue set. Additionally, relative loss functions built on the proposed conflict functions are employed to calculate thresholds for three-way conflict analysis. Finally, we present adjustment mechanism-based feasible strategies that simultaneously account for both adjustment magnitudes and conflict degrees, together with an algorithm for constructing such feasible strategies, and provide an illustrative example to demonstrate the validity and effectiveness of the proposed model.

Method: Introduces DiscoverLLM framework with a novel user simulator that models cognitive state using a hierarchy of intents that progressively concretize. Uses degree of concretization as a reward signal to train models to optimize for helping users discover intents through adaptive divergence (exploration) and convergence (refinement).

Result: Achieves over 10% higher task performance while reducing conversation length by up to 40% across interactive benchmarks in creative writing, technical writing, and SVG drawing. In a user study with 75 participants, improves conversation satisfaction and efficiency compared to baselines.

Conclusion: DiscoverLLM provides a generalizable framework for training LLMs to help users form and discover intents through interactive exploration, addressing the fundamental challenge of ambiguous requests where users haven’t yet formed clear preferences.

Abstract: To handle ambiguous and open-ended requests, Large Language Models (LLMs) are increasingly trained to interact with users to surface intents they have not yet expressed (e.g., ask clarification questions). However, users are often ambiguous because they have not yet formed their intents: they must observe and explore outcomes to discover what they want. Simply asking “what kind of tone do you want?” fails when users themselves do not know. We introduce DiscoverLLM, a novel and generalizable framework that trains LLMs to help users form and discover their intents. Central to our approach is a novel user simulator that models cognitive state with a hierarchy of intents that progressively concretize as the model surfaces relevant options – where the degree of concretization serves as a reward signal that models can be trained to optimize. Resulting models learn to collaborate with users by adaptively diverging (i.e., explore options) when intents are unclear, and converging (i.e., refine and implement) when intents concretize. Across proposed interactive benchmarks in creative writing, technical writing, and SVG drawing, DiscoverLLM achieves over 10% higher task performance while reducing conversation length by up to 40%. In a user study with 75 human participants, DiscoverLLM improved conversation satisfaction and efficiency compared to baselines.

Method: Ontology-to-tools compilation transforms ontological specifications into executable tool interfaces that LLM-based agents must use. Uses Model Context Protocol (MCP) and agent-based workflow to translate ontologies into ontology-aware tools for iterative extraction, validation, and repair of structured knowledge from unstructured text.

Result: Demonstrated with metal-organic polyhedra synthesis literature, showing executable ontological semantics can guide LLM behavior and reduce manual engineering.

Conclusion: Establishes a general paradigm for embedding formal knowledge into generative systems through ontology-to-tools compilation.

Abstract: We introduce ontology-to-tools compilation as a proof-of-principle mechanism for coupling large language models (LLMs) with formal domain knowledge. Within The World Avatar (TWA), ontological specifications are compiled into executable tool interfaces that LLM-based agents must use to create and modify knowledge graph instances, enforcing semantic constraints during generation rather than through post-hoc validation. Extending TWA’s semantic agent composition framework, the Model Context Protocol (MCP) and associated agents are integral components of the knowledge graph ecosystem, enabling structured interaction between generative models, symbolic constraints, and external resources. An agent-based workflow translates ontologies into ontology-aware tools and iteratively applies them to extract, validate, and repair structured knowledge from unstructured scientific text. Using metal-organic polyhedra synthesis literature as an illustrative case, we show how executable ontological semantics can guide LLM behaviour and reduce manual schema and prompt engineering, establishing a general paradigm for embedding formal knowledge into generative systems.

Method: CRL-VLA introduces a framework with theoretical bounds linking stability-plasticity trade-off to goal-conditioned advantage magnitude scaled by policy divergence. It uses asymmetric regulation: constraining advantage magnitudes on prior tasks while enabling controlled growth on new tasks. The method employs a dual-critic architecture with Goal-Conditioned Value Formulation (GCVF), where a frozen critic anchors semantic consistency and a trainable estimator drives adaptation.

Result: Experiments on the LIBERO benchmark show that CRL-VLA effectively harmonizes stability and plasticity objectives, outperforming baselines in both anti-forgetting (retaining old skills) and forward adaptation (learning new skills).

Conclusion: CRL-VLA provides a theoretically grounded solution to the stability-plasticity dilemma in continual reinforcement learning for VLA models, enabling effective lifelong learning for embodied agents through its dual-critic architecture and asymmetric regulation approach.

Abstract: Lifelong learning is critical for embodied agents in open-world environments, where reinforcement learning fine-tuning has emerged as an important paradigm to enable Vision-Language-Action (VLA) models to master dexterous manipulation through environmental interaction. Thus, Continual Reinforcement Learning (CRL) is a promising pathway for deploying VLA models in lifelong robotic scenarios, yet balancing stability (retaining old skills) and plasticity (learning new ones) remains a formidable challenge for existing methods. We introduce CRL-VLA, a framework for continual post-training of VLA models with rigorous theoretical bounds. We derive a unified performance bound linking the stability-plasticity trade-off to goal-conditioned advantage magnitude, scaled by policy divergence. CRL-VLA resolves this dilemma via asymmetric regulation: constraining advantage magnitudes on prior tasks while enabling controlled growth on new tasks. This is realized through a simple but effective dual-critic architecture with novel Goal-Conditioned Value Formulation (GCVF), where a frozen critic anchors semantic consistency and a trainable estimator drives adaptation. Experiments on the LIBERO benchmark demonstrate that CRL-VLA effectively harmonizes these conflicting objectives, outperforming baselines in both anti-forgetting and forward adaptation.

Method: Used Answer Set Programming (ASP) as a formal framework to define irrelevant details for abstraction. Conducted cognitive experiments where participants classified stimuli across domains with explanations derived from answer set programs, testing removal and clustering abstractions.

Result: Clustering details significantly improved participants’ understanding, while removal of details significantly reduced cognitive effort. Both findings support the hypothesis that abstraction enhances human-centered symbolic explanations.

Conclusion: Formal abstractions (specifically removal and clustering) effectively enhance the interpretability of symbolic AI explanations by improving human understanding and reducing cognitive load.

Abstract: Explanations are central to human cognition, yet AI systems often produce outputs that are difficult to understand. While symbolic AI offers a transparent foundation for interpretability, raw logical traces often impose a high extraneous cognitive load. We investigate how formal abstractions, specifically removal and clustering, impact human reasoning performance and cognitive effort. Utilizing Answer Set Programming (ASP) as a formal framework, we define a notion of irrelevant details to be abstracted over to obtain simplified explanations. Our cognitive experiments, in which participants classified stimuli across domains with explanations derived from an answer set program, show that clustering details significantly improve participants’ understanding, while removal of details significantly reduce cognitive effort, supporting the hypothesis that abstraction enhances human-centered symbolic explanations.

Method: Proposes IntentRL framework with scalable pipeline for generating high-quality dialogue data via shallow-to-deep intent refinement graph, and two-stage RL: offline RL on dialogues followed by online rollouts with user simulator.

Result: IntentRL significantly improves intent hit rate and downstream task performance, outperforming built-in clarify modules of closed-source DR agents and proactive LLM baselines.

Conclusion: Proactive intent clarification before deep research execution addresses the autonomy-interaction dilemma and improves research agent effectiveness.

Abstract: Deep Research (DR) agents extend Large Language Models (LLMs) beyond parametric knowledge by autonomously retrieving and synthesizing evidence from large web corpora into long-form reports, enabling a long-horizon agentic paradigm. However, unlike real-time conversational assistants, DR is computationally expensive and time-consuming, creating an autonomy-interaction dilemma: high autonomy on ambiguous user queries often leads to prolonged execution with unsatisfactory outcomes. To address this, we propose IntentRL, a framework that trains proactive agents to clarify latent user intents before starting long-horizon research. To overcome the scarcity of open-ended research data, we introduce a scalable pipeline that expands a few seed samples into high-quality dialogue turns via a shallow-to-deep intent refinement graph. We further adopt a two-stage reinforcement learning (RL) strategy: Stage I applies RL on offline dialogues to efficiently learn general user-interaction behavior, while Stage II uses the trained agent and a user simulator for online rollouts to strengthen adaptation to diverse user feedback. Extensive experiments show that IntentRL significantly improves both intent hit rate and downstream task performance, outperforming the built-in clarify modules of closed-source DR agents and proactive LLM baselines.

Method: Proposes EquiRouter, a decision-aware router that directly learns model rankings rather than predicting scalar performance scores. This addresses the objective-decision mismatch where small prediction errors can flip relative orderings and trigger suboptimal selections.

Result: On RouterBench, EquiRouter reduces cost by about 17% at GPT-4-level performance compared to the strongest prior router. It restores the role of smaller models and mitigates routing collapse.

Conclusion: Directly learning model rankings rather than scalar scores is more effective for routing decisions, enabling better utilization of smaller models and achieving significant cost savings while maintaining performance.

Abstract: LLM routing aims to achieve a favorable quality–cost trade-off by dynamically assigning easy queries to smaller models and harder queries to stronger ones. However, across both unimodal and multimodal settings, we uncover a pervasive yet underexplored failure mode in existing routers: as the user’s cost budget increases, routers systematically default to the most capable and most expensive model even when cheaper models already suffice. As a result, current routers under-utilize small models, wasting computation and monetary cost and undermining the core promise of routing; we term this phenomenon routing collapse. We attribute routing collapse to an objective–decision mismatch: many routers are trained to predict scalar performance scores, whereas routing decisions ultimately depend on discrete comparisons among candidate models. Consequently, small prediction errors can flip relative orderings and trigger suboptimal selections. To bridge this gap, we propose EquiRouter, a decision-aware router that directly learns model rankings, restoring the role of smaller models and mitigating routing collapse. On RouterBench, EquiRouter reduces cost by about 17% at GPT-4-level performance compared to the strongest prior router. Our code is available at https://github.com/AIGNLAI/EquiRouter.

Method: Uses agent-based modeling and evolutionary game theory to compare two AI use strategies: AI-complement (humans seek guidance but produce final output) vs AI-substitute (humans provide minimal input, AI produces most output), studying how these strategies compete and spread under evolutionary dynamics.

Result: AI-substitute users prevail under individual-level selection despite stronger reduction in cultural variance, while AI-complement users benefit groups by maintaining variance needed for exploration and can be favored under cultural group selection when group boundaries are strong.

Conclusion: The findings reveal population-level effects of AI adoption and inform policy/organizational strategies to mitigate risks of cultural collapse, highlighting the tension between individual and group-level selection pressures in AI adoption.

Abstract: Reliance on generative AI can reduce cultural variance and diversity, especially in creative work. This reduction in variance has already led to problems in model performance, including model collapse and hallucination. In this paper, we examine the long-term consequences of AI use for human cultural evolution and the conditions under which widespread AI use may lead to “cultural collapse”, a process in which reliance on AI-generated content reduces human variation and innovation and slows cumulative cultural evolution. Using an agent-based model and evolutionary game theory, we compare two types of AI use: complement and substitute. AI-complement users seek suggestions and guidance while remaining the main producers of the final output, whereas AI-substitute users provide minimal input, and rely on AI to produce most of the output. We then study how these use strategies compete and spread under evolutionary dynamics. We find that AI-substitute users prevail under individual-level selection despite the stronger reduction in cultural variance. By contrast, AI-complement users can benefit their groups by maintaining the variance needed for exploration, and can therefore be favored under cultural group selection when group boundaries are strong. Overall, our findings shed light on the long-term, population-level effects of AI adoption and inform policy and organizational strategies to mitigate these risks.

Method: Introduce Persona Generators - functions that produce diverse synthetic populations. Use iterative improvement loop based on AlphaEvolve with LLMs as mutation operators to refine generator code over hundreds of iterations. Optimization produces lightweight generators that expand small descriptions into diverse synthetic personas maximizing coverage of opinions/preferences along relevant diversity axes.

Result: Evolved generators substantially outperform existing baselines across six diversity metrics on held-out contexts, producing populations that span rare trait combinations difficult to achieve in standard LLM outputs.

Conclusion: Persona Generators enable creation of diverse synthetic populations for AI evaluation without requiring extensive real human data, addressing the coverage gap in current approaches and supporting evaluation of AI systems across diverse user populations.

Abstract: Evaluating AI systems that interact with humans requires understanding their behavior across diverse user populations, but collecting representative human data is often expensive or infeasible, particularly for novel technologies or hypothetical future scenarios. Recent work in Generative Agent-Based Modeling has shown that large language models can simulate human-like synthetic personas with high fidelity, accurately reproducing the beliefs and behaviors of specific individuals. However, most approaches require detailed data about target populations and often prioritize density matching (replicating what is most probable) rather than support coverage (spanning what is possible), leaving long-tail behaviors underexplored. We introduce Persona Generators, functions that can produce diverse synthetic populations tailored to arbitrary contexts. We apply an iterative improvement loop based on AlphaEvolve, using large language models as mutation operators to refine our Persona Generator code over hundreds of iterations. The optimization process produces lightweight Persona Generators that can automatically expand small descriptions into populations of diverse synthetic personas that maximize coverage of opinions and preferences along relevant diversity axes. We demonstrate that evolved generators substantially outperform existing baselines across six diversity metrics on held-out contexts, producing populations that span rare trait combinations difficult to achieve in standard LLM outputs.

Method: Introduces EHRWorld, a patient-centric medical world model trained under a causal sequential paradigm, along with EHRWorld-110K, a large-scale longitudinal clinical dataset derived from real-world electronic health records. The approach focuses on modeling temporal evolution and causal relationships in clinical data.

Result: EHRWorld significantly outperforms naive LLM-based baselines, achieving more stable long-horizon simulation, improved modeling of clinically sensitive events, and favorable reasoning efficiency. Demonstrates the necessity of training on causally grounded, temporally evolving clinical data.

Conclusion: Training medical world models on causally grounded, temporally evolving clinical data is essential for reliable and robust simulation of disease progression and treatment outcomes, overcoming limitations of naive LLM approaches.

Abstract: World models offer a principled framework for simulating future states under interventions, but realizing such models in complex, high-stakes domains like medicine remains challenging. Recent large language models (LLMs) have achieved strong performance on static medical reasoning tasks, raising the question of whether they can function as dynamic medical world models capable of simulating disease progression and treatment outcomes over time. In this work, we show that LLMs only incorporating medical knowledge struggle to maintain consistent patient states under sequential interventions, leading to error accumulation in long-horizon clinical simulation. To address this limitation, we introduce EHRWorld, a patient-centric medical world model trained under a causal sequential paradigm, together with EHRWorld-110K, a large-scale longitudinal clinical dataset derived from real-world electronic health records. Extensive evaluations demonstrate that EHRWorld significantly outperforms naive LLM-based baselines, achieving more stable long-horizon simulation, improved modeling of clinically sensitive events, and favorable reasoning efficiency, highlighting the necessity of training on causally grounded, temporally evolving clinical data for reliable and robust medical world modeling.

Method: Adapted MLE-Bench framework to orbital mechanics, deployed AIDE-based agent architecture for autonomous mission solution generation, and used “LLM-as-a-Judge” methodology with expert-developed rubric to evaluate strategic viability across five categories.

Result: Strategic viability scores nearly doubled in two years (9.3 to 17.2/26), showing improved conceptual understanding, but models consistently failed at implementation due to physical unit inconsistencies, boundary condition errors, and inefficient debugging loops.

Conclusion: Current LLMs have sufficient knowledge for space science tasks but face an implementation barrier, functioning as domain facilitators rather than fully autonomous engineers.

Abstract: Large Language Models (LLMs) have demonstrated remarkable proficiency in code generation and general reasoning, yet their capacity for autonomous multi-stage planning in high-dimensional, physically constrained environments remains an open research question. This study investigates the limits of current AI agents by evaluating them against the 12th Global Trajectory Optimization Competition (GTOC 12), a complex astrodynamics challenge requiring the design of a large-scale asteroid mining campaign. We adapt the MLE-Bench framework to the domain of orbital mechanics and deploy an AIDE-based agent architecture to autonomously generate and refine mission solutions. To assess performance beyond binary validity, we employ an “LLM-as-a-Judge” methodology, utilizing a rubric developed by domain experts to evaluate strategic viability across five structural categories. A comparative analysis of models, ranging from GPT-4-Turbo to reasoning-enhanced architectures like Gemini 2.5 Pro, and o3, reveals a significant trend: the average strategic viability score has nearly doubled in the last two years (rising from 9.3 to 17.2 out of 26). However, we identify a critical capability gap between strategy and execution. While advanced models demonstrate sophisticated conceptual understanding, correctly framing objective functions and mission architectures, they consistently fail at implementation due to physical unit inconsistencies, boundary condition errors, and inefficient debugging loops. We conclude that, while current LLMs often demonstrate sufficient knowledge and intelligence to tackle space science tasks, they remain limited by an implementation barrier, functioning as powerful domain facilitators rather than fully autonomous engineers.

Method: Proposes Search-R2 with Actor-Refiner collaboration: Actor produces initial reasoning trajectories, Meta-Refiner selectively diagnoses and repairs flawed steps via ‘cut-and-regenerate’ mechanism. Uses hybrid reward design coupling outcome correctness with dense process reward quantifying information density of retrieved evidence.

Result: Search-R2 consistently outperforms strong RAG and RL-based baselines across various general and multi-hop QA datasets and model scales, achieving superior reasoning accuracy with minimal overhead.

Conclusion: The Actor-Refiner framework with targeted intervention and fine-grained supervision effectively addresses credit assignment problems in search-integrated reasoning, enabling language agents to transcend static parametric knowledge through improved external source querying.

Abstract: Search-integrated reasoning enables language agents to transcend static parametric knowledge by actively querying external sources. However, training these agents via reinforcement learning is hindered by the multi-scale credit assignment problem: existing methods typically rely on sparse, trajectory-level rewards that fail to distinguish between high-quality reasoning and fortuitous guesses, leading to redundant or misleading search behaviors. To address this, we propose Search-R2, a novel Actor-Refiner collaboration framework that enhances reasoning through targeted intervention, with both components jointly optimized during training. Our approach decomposes the generation process into an Actor, which produces initial reasoning trajectories, and a Meta-Refiner, which selectively diagnoses and repairs flawed steps via a ‘cut-and-regenerate’ mechanism. To provide fine-grained supervision, we introduce a hybrid reward design that couples outcome correctness with a dense process reward quantifying the information density of retrieved evidence. Theoretically, we formalize the Actor-Refiner interaction as a smoothed mixture policy, proving that selective correction yields strict performance gains over strong baselines. Extensive experiments across various general and multi-hop QA datasets demonstrate that Search-R2 consistently outperforms strong RAG and RL-based baselines across model scales, achieving superior reasoning accuracy with minimal overhead.

Method: Through attention analysis, the authors identify conversational inertia as strong diagonal attention to previous responses. They propose Context Preference Learning to calibrate model preferences to favor low-inertia responses over high-inertia ones, and provide context management strategies at inference time.

Result: Experimental results across eight agentic environments and one deep research scenario validate that the framework reduces conversational inertia and achieves performance improvements.

Conclusion: The paper addresses a key limitation in using LLMs as agents by identifying and mitigating conversational inertia, providing methods to balance exploration and exploitation in multiturn scenarios.

Abstract: Large language models excel as few-shot learners when provided with appropriate demonstrations, yet this strength becomes problematic in multiturn agent scenarios, where LLMs erroneously mimic their own previous responses as few-shot examples. Through attention analysis, we identify conversational inertia, a phenomenon where models exhibit strong diagonal attention to previous responses, which is associated with imitation bias that constrains exploration. This reveals a tension when transforming few-shot LLMs into agents: longer context enriches environmental feedback for exploitation, yet also amplifies conversational inertia that undermines exploration. Our key insight is that for identical states, actions generated with longer contexts exhibit stronger inertia than those with shorter contexts, enabling construction of preference pairs without environment rewards. Based on this, we propose Context Preference Learning to calibrate model preferences to favor low-inertia responses over highinertia ones. We further provide context management strategies at inference time to balance exploration and exploitation. Experimental results across eight agentic environments and one deep research scenario validate that our framework reduces conversational inertia and achieves performance improvements.

Method: TodyComm produces behavior-driven collaboration topologies that adapt to dynamics at each round, optimizing task utility through policy gradient methods.

Result: Experiments on five benchmarks show TodyComm delivers superior task effectiveness under both dynamic adversary and communication budget constraints while maintaining token efficiency and scalability.

Conclusion: Dynamic communication topologies that adapt to round-specific conditions are essential for effective multi-agent LLM collaboration in realistic scenarios.

Abstract: Multi-round LLM-based multi-agent systems rely on effective communication structures to support collaboration across rounds. However, most existing methods employ a fixed communication topology during inference, which falls short in many realistic applications where the agents’ roles may change \textit{across rounds} due to dynamic adversary, task progression, or time-varying constraints such as communication bandwidth. In this paper, we propose addressing this issue through TodyComm, a \textbf{t}ask-\textbf{o}riented \textbf{dy}namic \textbf{comm}unication algorithm. It produces behavior-driven collaboration topologies that adapt to the dynamics at each round, optimizing the utility for the task through policy gradient. Experiments on five benchmarks demonstrate that under both dynamic adversary and communications budgets, TodyComm delivers superior task effectiveness while retaining token efficiency and scalability.

Method: Proposes a unified agent abstraction as a tuple (Instruction, Context, Tools, Model) that serves as a compositional recipe for capabilities. AOrchestra system uses a central orchestrator that concretizes this tuple at each step: curates task-relevant context, selects tools and models, and delegates execution via on-the-fly automatic agent creation.

Result: AOrchestra achieves 16.28% relative improvement against the strongest baseline when paired with Gemini-3-Flash across three challenging benchmarks (GAIA, SWE-Bench, Terminal-Bench). The system enables controllable performance-cost trade-offs and reduces human engineering efforts.

Conclusion: The proposed agent abstraction and AOrchestra system provide a framework-agnostic approach to language agent orchestration that improves adaptability and performance for complex tasks while enabling efficient resource utilization.

Abstract: Language agents have shown strong promise for task automation. Realizing this promise for increasingly complex, long-horizon tasks has driven the rise of a sub-agent-as-tools paradigm for multi-turn task solving. However, existing designs still lack a dynamic abstraction view of sub-agents, thereby hurting adaptability. We address this challenge with a unified, framework-agnostic agent abstraction that models any agent as a tuple Instruction, Context, Tools, Model. This tuple acts as a compositional recipe for capabilities, enabling the system to spawn specialized executors for each task on demand. Building on this abstraction, we introduce an agentic system AOrchestra, where the central orchestrator concretizes the tuple at each step: it curates task-relevant context, selects tools and models, and delegates execution via on-the-fly automatic agent creation. Such designs enable reducing human engineering efforts, and remain framework-agnostic with plug-and-play support for diverse agents as task executors. It also enables a controllable performance-cost trade-off, allowing the system to approach Pareto-efficient. Across three challenging benchmarks (GAIA, SWE-Bench, Terminal-Bench), AOrchestra achieves 16.28% relative improvement against the strongest baseline when paired with Gemini-3-Flash. The code is available at: https://github.com/FoundationAgents/AOrchestra

Method: The authors develop an information-theoretic framework showing MAS performance is bounded by intrinsic task uncertainty rather than agent count. They introduce K*, an effective channel count metric that quantifies the number of effective channels without ground-truth labels, and empirically test heterogeneous vs. homogeneous agent configurations.

Result: Empirical results show heterogeneous configurations consistently outperform homogeneous scaling: 2 diverse agents can match or exceed the performance of 16 homogeneous agents. The effective channel count K* successfully quantifies system diversity and predicts performance improvements.

Conclusion: Multi-agent system performance is limited by task uncertainty, not agent count. Diversity in agents (different models, prompts, tools) provides complementary evidence and substantially improves performance, offering principled guidelines for building efficient and robust MAS through diversity-aware design.

Abstract: LLM-based multi-agent systems (MAS) have emerged as a promising approach to tackle complex tasks that are difficult for individual LLMs. A natural strategy is to scale performance by increasing the number of agents; however, we find that such scaling exhibits strong diminishing returns in homogeneous settings, while introducing heterogeneity (e.g., different models, prompts, or tools) continues to yield substantial gains. This raises a fundamental question: what limits scaling, and why does diversity help? We present an information-theoretic framework showing that MAS performance is bounded by the intrinsic task uncertainty, not by agent count. We derive architecture-agnostic bounds demonstrating that improvements depend on how many effective channels the system accesses. Homogeneous agents saturate early because their outputs are strongly correlated, whereas heterogeneous agents contribute complementary evidence. We further introduce $K^*$, an effective channel count that quantifies the number of effective channels without ground-truth labels. Empirically, we show that heterogeneous configurations consistently outperform homogeneous scaling: 2 diverse agents can match or exceed the performance of 16 homogeneous agents. Our results provide principled guidelines for building efficient and robust MAS through diversity-aware design. Code and Dataset are available at the link: https://github.com/SafeRL-Lab/Agent-Scaling.

Method: Proposes a risk control framework with two thresholds: an upper threshold to stop when model is confident, and a novel parametric lower threshold to preemptively stop unsolvable instances. Uses distribution-free risk control to optimally specify stopping mechanisms given target risk and validation data. Incorporates efficiency loss for scenarios with multiple budget criteria.

Result: Empirical results across diverse reasoning tasks and models show computational efficiency gains from the lower threshold and ensemble stopping mechanisms while adhering to user-specified risk targets.

Conclusion: The framework effectively addresses the risk-accuracy trade-off in adaptive reasoning by providing systematic risk control methods that optimize computational efficiency while maintaining specified error rate bounds.

Abstract: Reasoning Large Language Models (LLMs) enable test-time scaling, with dataset-level accuracy improving as the token budget increases, motivating adaptive reasoning – spending tokens when they improve reliability and stopping early when additional computation is unlikely to help. However, setting the token budget, as well as the threshold for adaptive reasoning, is a practical challenge that entails a fundamental risk-accuracy trade-off. We re-frame the budget setting problem as risk control, limiting the error rate while minimizing compute. Our framework introduces an upper threshold that stops reasoning when the model is confident (risking incorrect output) and a novel parametric lower threshold that preemptively stops unsolvable instances (risking premature stoppage). Given a target risk and a validation set, we use distribution-free risk control to optimally specify these stopping mechanisms. For scenarios with multiple budget controlling criteria, we incorporate an efficiency loss to select the most computationally efficient exiting mechanism. Empirical results across diverse reasoning tasks and models demonstrate the effectiveness of our risk control approach, demonstrating computational efficiency gains from the lower threshold and ensemble stopping mechanisms while adhering to the user-specified risk target.

Method: Proposes AutoFigure, an agentic framework that engages in extensive thinking, recombination, and validation to produce structurally sound and aesthetically refined layouts before rendering final illustrations. Uses FigureBench benchmark with 3,300 high-quality scientific text-figure pairs covering diverse sources.

Result: AutoFigure consistently surpasses all baseline methods, producing publication-ready scientific illustrations. The framework demonstrates strong performance across diverse scientific text-to-illustration tasks.

Conclusion: AutoFigure represents a significant advancement in automated scientific illustration generation, with potential to alleviate the bottleneck in scientific communication. The released code, dataset, and demo provide resources for further research.

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: Two-step reasoning approach: (1) generate diagnostic hypotheses from dialogue context using a diagnostic knowledge graph, and (2) verify hypotheses through iterative clarifying questions until final diagnosis is reached. Uses MIMIC-IV patient profiles with adapted simulator to reflect vague symptom descriptions.

Result: Improved diagnostic accuracy and efficiency over strong baselines. Physician evaluations support realism of the simulator and clinical utility of generated questions.

Conclusion: The proposed conversational diagnosis system with knowledge graph reasoning and realistic patient simulation demonstrates clinical utility and improved performance over existing approaches.

Abstract: Conversational diagnosis requires multi-turn history-taking, where an agent asks clarifying questions to refine differential diagnoses under incomplete information. Existing approaches often rely on the parametric knowledge of a model or assume that patients provide rich and concrete information, which is unrealistic. To address these limitations, we propose a conversational diagnosis system that explores a diagnostic knowledge graph to reason in two steps: (i) generating diagnostic hypotheses from the dialogue context, and (ii) verifying hypotheses through clarifying questions, which are repeated until a final diagnosis is reached. Since evaluating the system requires a realistic patient simulator that responds to the system’s questions, we adopt a well-established simulator along with patient profiles from MIMIC-IV. We further adapt it to describe symptoms vaguely to reflect real-world patients during early clinical encounters. Experiments show improved diagnostic accuracy and efficiency over strong baselines, and evaluations by physicians support the realism of our simulator and the clinical utility of the generated questions. Our code will be released upon publication.

Method: Developed ARIEL framework with curated multimodal biomedical corpus, expert-vetted tasks, uniform evaluation protocols, blinded PhD-level assessment, prompt engineering, lightweight fine-tuning, and compute-scaled inference strategies.

Result: Models generate fluent but incomplete summaries; LMMs struggle with detailed visual reasoning; prompt engineering and fine-tuning improve textual coverage; compute-scaled inference enhances visual QA; ARIEL agent can propose testable mechanistic hypotheses.

Conclusion: ARIEL provides a reproducible platform for advancing trustworthy AI in biomedicine by delineating current strengths and limitations of foundation models and enabling optimization through targeted interventions.

Abstract: Large language models (LLMs) and large multimodal models (LMMs) promise to accelerate biomedical discovery, yet their reliability remains unclear. We introduce ARIEL (AI Research Assistant for Expert-in-the-Loop Learning), an open-source evaluation and optimization framework that pairs a curated multimodal biomedical corpus with expert-vetted tasks to probe two capabilities: full-length article summarization and fine-grained figure interpretation. Using uniform protocols and blinded PhD-level evaluation, we find that state-of-the-art models generate fluent but incomplete summaries, whereas LMMs struggle with detailed visual reasoning. We later observe that prompt engineering and lightweight fine-tuning substantially improve textual coverage, and a compute-scaled inference strategy enhances visual question answering. We build an ARIEL agent that integrates textual and visual cues, and we show it can propose testable mechanistic hypotheses. ARIEL delineates current strengths and limitations of foundation models, and provides a reproducible platform for advancing trustworthy AI in biomedicine.

Method: CP-Agent uses the ReAct (Reasoning + Acting) framework with a persistent IPython kernel. It iteratively executes code, observes solver feedback, and refines constraint models based on execution results. Domain knowledge is provided as a project prompt under 50 lines.

Result: CP-Agent achieves perfect accuracy on all 101 constraint programming problems from CP-Bench after minor benchmark clarifications. Experiments show minimal guidance outperforms detailed procedural scaffolding, and explicit task management tools have mixed effects on modeling tasks.

Conclusion: Agentic workflows with minimal guidance can effectively automate natural language to constraint model translation, achieving high accuracy on benchmark problems while revealing insights about guidance strategies and task management tools.

Abstract: The translation of natural language to formal constraint models requires expertise in the problem domain and modeling frameworks. To explore the effectiveness of agentic workflows, we propose CP-Agent, a Python coding agent that uses the ReAct framework with a persistent IPython kernel. We provide the relevant domain knowledge as a project prompt of under 50 lines. The algorithm works by iteratively executing code, observing the solver’s feedback, and refining constraint models based on execution results. We evaluate CP-Agent on 101 constraint programming problems from CP-Bench. We made minor changes to the benchmark to address systematic ambiguities in the problem specifications and errors in the ground-truth models. On the clarified benchmark, CP-Agent achieves perfect accuracy on all 101 problems. Our experiments show that minimal guidance outperforms detailed procedural scaffolding. Our experiments also show that explicit task management tools can have both positive and negative effects on focused modeling tasks.

Method: When strong model resolves deferred queries, it generates generalized reusable problem-solving strategies stored in dynamic repository. Future queries use similarity matching to retrieve relevant strategies to augment weak model’s context, enabling in-context learning without parameter fine-tuning.

Result: Outperforms standard cascades on multiple benchmarks: improves weak model accuracy by up to 33.06%, overall system accuracy by 6.35%, reduces strong model calls by 48.05%, and saves fees by 49.63%.

Conclusion: Inter-Cascade enables effective in-context knowledge transfer between LLMs, provides scalable framework for both open-source and API-based models, and improves weak model confidence calibration theoretically and empirically.

Abstract: Standard LLM cascades improve efficiency by deferring difficult queries from weak to strong models. However, these systems are typically static: when faced with repeated or semantically similar queries, they redundantly consult the expensive model, failing to adapt during inference. To address this, we propose Inter-Cascade, an online, interactive framework that transforms the strong model from a temporary helper into a long-term teacher. In our approach, when the strong model resolves a deferred query, it generates a generalized, reusable problem-solving strategy. These strategies are stored in a dynamic repository and retrieved via similarity matching to augment the weak model’s context for future queries. This enables the weak model to learn on the job without expensive parameter fine-tuning. We theoretically show that this mechanism improves the weak model’s confidence calibration. Empirically, Inter-Cascade outperforms standard cascades on multiple benchmarks, improving weak model and overall system accuracy by up to 33.06 percent and 6.35 percent, while reducing strong model calls by up to 48.05 percent and saving fee by up to 49.63 percent. Inter-Cascade demonstrates effective in-context knowledge transfer between LLMs and provides a general, scalable framework applicable to both open-source and API-based LLMs.

Method: Evaluated AI models on ConceptARC benchmark with variations in input modality (textual vs. visual), use of external Python tools, and reasoning effort. Analyzed not just output accuracy but also the natural-language rules models generate to explain their solutions, enabling assessment of whether models recognize intended abstractions.

Result: Best models’ rules frequently based on surface-level shortcuts rather than intended abstractions. In visual modality, accuracy drops sharply but rule analysis reveals models can recognize abstractions even when failing to apply them correctly. Accuracy alone overestimates capabilities in textual modalities and underestimates in visual modalities.

Conclusion: Current AI models show limited abstraction reasoning abilities, with accuracy metrics providing misleading assessments. Rule-level analysis offers more faithful evaluation of abstract reasoning and better tracking of progress toward human-like, abstraction-centered intelligence.

Abstract: OpenAI’s o3-preview reasoning model exceeded human accuracy on the ARC-AGI-1 benchmark, but does that mean state-of-the-art models recognize and reason with the abstractions the benchmark was designed to test? Here we investigate abstraction abilities of AI models using the closely related but simpler ConceptARC benchmark. Our evaluations vary input modality (textual vs. visual), use of external Python tools, and reasoning effort. Beyond output accuracy, we evaluate the natural-language rules that models generate to explain their solutions, enabling us to assess whether models recognize the abstractions that ConceptARC was designed to elicit. We show that the best models’ rules are frequently based on surface-level ``shortcuts,’’ capturing intended abstractions considerably less often than humans. In the visual modality, AI models’ output accuracy drops sharply; however, our rule-level analysis reveals that a substantial share of their rules capture the intended abstractions, even as the models struggle to apply these concepts to generate correct solutions. In short, we show that using accuracy alone to evaluate abstract reasoning can substantially overestimate AI capabilities in textual modalities and underestimate it in visual modalities. Our results offer a more faithful picture of AI models’ abstract reasoning abilities and a more principled way to track progress toward human-like, abstraction-centered intelligence.

Method: Propose MemeLens, a unified multilingual multitask explanation-enhanced Vision Language Model. Consolidate 38 public meme datasets, filter and map labels into shared taxonomy of 20 tasks. Conduct comprehensive empirical analysis across modeling paradigms, task categories, and datasets.

Result: Findings show robust meme understanding requires multimodal training, exhibits substantial variation across semantic categories, and remains sensitive to over-specialization when models are fine-tuned on individual datasets rather than trained in unified setting.

Conclusion: Unified multilingual multitask approach improves meme understanding across diverse domains. Experimental resources and datasets will be made publicly available for community.

Abstract: Memes are a dominant medium for online communication and manipulation because meaning emerges from interactions between embedded text, imagery, and cultural context. Existing meme research is distributed across tasks (hate, misogyny, propaganda, sentiment, humour) and languages, which limits cross-domain generalization. To address this gap we propose MemeLens, a unified multilingual and multitask explanation-enhanced Vision Language Model (VLM) for meme understanding. We consolidate 38 public meme datasets, filter and map dataset-specific labels into a shared taxonomy of $20$ tasks spanning harm, targets, figurative/pragmatic intent, and affect. We present a comprehensive empirical analysis across modeling paradigms, task categories, and datasets. Our findings suggest that robust meme understanding requires multimodal training, exhibits substantial variation across semantic categories, and remains sensitive to over-specialization when models are fine-tuned on individual datasets rather than trained in a unified setting. We will make the experimental resources and datasets publicly available for the community.

Method: PoLR clusters short prefixes of reasoning traces, identifies the dominant cluster, and expands all paths in that cluster only. This leverages prefix consistency where early reasoning steps predict final correctness, reducing unnecessary expansions.

Result: PoLR matches or exceeds Self-Consistency accuracy across GSM8K, MATH500, AIME24/25, and GPQA-DIAMOND benchmarks while reducing token usage by up to 60% and wall-clock latency by up to 50%.

Conclusion: PoLR provides a computationally efficient alternative to Self-Consistency that maintains reasoning accuracy, is complementary to adaptive inference methods, and can serve as a drop-in pre-filter without requiring model fine-tuning.

Abstract: Large language models achieve strong reasoning performance, but inference strategies such as Self-Consistency (SC) are computationally expensive, as they fully expand all reasoning traces. We introduce PoLR (Path of Least Resistance), the first inference-time method to leverage prefix consistency for compute-efficient reasoning. PoLR clusters short prefixes of reasoning traces, identifies the dominant cluster, and expands all paths in that cluster, preserving the accuracy benefits of SC while substantially reducing token usage and latency. Our theoretical analysis, framed via mutual information and entropy, explains why early reasoning steps encode strong signals predictive of final correctness. Empirically, PoLR consistently matches or exceeds SC across GSM8K, MATH500, AIME24/25, and GPQA-DIAMOND, reducing token usage by up to 60% and wall-clock latency by up to 50%. Moreover, PoLR is fully complementary to adaptive inference methods (e.g., Adaptive Consistency, Early-Stopping SC) and can serve as a drop-in pre-filter, making SC substantially more efficient and scalable without requiring model fine-tuning.

Method: Proposes EigenData - a hierarchical multi-agent engine that synthesizes tool-grounded dialogues with executable per-instance checkers, using closed-loop self-evolving prompts and workflow. Then applies RL recipe: fine-tunes user model first, then uses GRPO-style training with trajectory-level group-relative advantages and dynamic filtering.

Result: Achieves 73.0% pass^1 on Airline and 98.3% pass^1 on Telecom benchmarks in tau^2-bench, matching or exceeding frontier models. Shows consistent improvements beyond supervised fine-tuning.

Conclusion: Demonstrates a scalable pathway for bootstrapping complex tool-using behaviors without expensive human annotation, combining self-evolving data synthesis with verifier-based RL for efficient agent training.

Abstract: Interactive tool-using agents must solve real-world tasks via multi-turn interaction with both humans and external environments, requiring dialogue state tracking, multi-step tool execution, while following complex instructions. Post-training such agents is challenging because synthesis for high-quality multi-turn tool-use data is difficult to scale, and reinforcement learning (RL) could face noisy signals caused by user simulation, leading to degraded training efficiency. We propose a unified framework that combines a self-evolving data agent with verifier-based RL. Our system, EigenData, is a hierarchical multi-agent engine that synthesizes tool-grounded dialogues together with executable per-instance checkers, and improves generation reliability via closed-loop self-evolving process that updates prompts and workflow. Building on the synthetic data, we develop an RL recipe that first fine-tunes the user model and then applies GRPO-style training with trajectory-level group-relative advantages and dynamic filtering, yielding consistent improvements beyond SFT. Evaluated on tau^2-bench, our best model reaches 73.0% pass^1 on Airline and 98.3% pass^1 on Telecom, matching or exceeding frontier models. Overall, our results suggest a scalable pathway for bootstrapping complex tool-using behaviors without expensive human annotation.

Method: Proposes Integrated Policy Gradient (IPG), which attributes reasoning behaviors to model components by propagating compound outcome-based signals (like post-reasoning accuracy) backward through model inference trajectories, focusing on components with sequential contribution to reasoning behavior.

Result: Empirical evaluations show IPG achieves more precise localization of reasoning mechanisms and enables reliable modulation of reasoning behaviors (reasoning capability, reasoning strength) across diverse reasoning models.

Conclusion: IPG provides a novel framework for understanding LLM reasoning mechanisms by tracking sequential influence from internal components to reasoning outcomes, offering more precise interpretability than existing methods.

Abstract: Large language models (LLMs) demonstrate strong reasoning abilities in solving complex real-world problems. Yet, the internal mechanisms driving these complex reasoning behaviors remain opaque. Existing interpretability approaches targeting reasoning either identify components (e.g., neurons) correlated with special textual patterns, or rely on human-annotated contrastive pairs to derive control vectors. Consequently, current methods struggle to precisely localize complex reasoning mechanisms or capture sequential influence from model internal workings to the reasoning outputs. In this paper, built on outcome-oriented and sequential-influence-aware principles, we focus on identifying components that have sequential contribution to reasoning behavior where outcomes are cumulated by long-range effects. We propose Integrated Policy Gradient (IPG), a novel framework that attributes reasoning behaviors to model’s inner components by propagating compound outcome-based signals such as post reasoning accuracy backward through model inference trajectories. Empirical evaluations demonstrate that our approach achieves more precise localization and enables reliable modulation of reasoning behaviors (e.g., reasoning capability, reasoning strength) across diverse reasoning models.

Method: Introduces regulatory markets framework where governments mandate regulated entities to purchase regulatory services from government-licensed private regulators, creating market incentives for regulatory innovation.

Result: Proposes a hybrid approach that could overcome limitations of both command-and-control regulation and excessive delegation to industry by leveraging market forces and industry R&D for regulatory innovation.

Conclusion: Regulatory markets offer a promising solution to AI governance challenges by combining government policy-setting with market-driven technical implementation, addressing both technical and democratic deficits.

Abstract: Appropriately regulating artificial intelligence is an increasingly urgent and widespread policy challenge. We identify two primary, competing problem. First is a technical deficit: Legislatures and regulatory face significant challenges in rapidly translating conventional command-and-control legal requirements into technical requirements. Second is a democratic deficit: Over-reliance on industry to provide technical standards fails to ensure that the many values-based decisions that must be made to shape AI development and deployment are made by democratically accountable public, not private, actors. We propose a solution: regulatory markets, in which governments require the targets of regulation to purchase regulatory services from a government-licensed private regulator. This approach to AI regulation could overcome the limitations of both command-and-control regulation and excessive delegation to industry. Regulatory markets could enable governments to establish policy priorities for the regulation of AI while relying on market forces and industry R&D efforts to pioneer the technical methods of regulation that best achieve policymakers’ stated objectives.

Method: Iterative process: 1) Train model on filtered data, 2) Use trained model to filter corpus further, 3) Train new model on cleaner corpus, 4) Repeat. Theoretical analysis of convergence.

Result: Theoretical convergence to self-consistent corpus, decay in harmful content even with constant filter quality, creation of large-scale preference annotations for interpretability.

Conclusion: Iterative self-filtering offers novel scalable oversight with human-auditable corpora, calls for empirical testing by researchers with pretraining infrastructure.

Abstract: Recent work demonstrates that filtering harmful content from pretraining data improves model safety without degrading capabilities. We propose a natural extension: do it again. A model trained on filtered data can filter the corpus further; training on this cleaner corpus produces an even cleaner model. We provide theoretical analysis showing this process converges to a self-consistent corpus where the model trained on it approves of its own training data. Even under the weak assumption of constant filter quality, iteration yields decay in harmful content. We argue this framework offers a novel form of scalable oversight. While model internals are opaque, the resulting corpus is human-auditable. Even a single iteration produces a large-scale preference annotations over documents, potentially valuable for interpretability research. We derive bounds on capability-safety tradeoffs and outline open questions. We call on researchers with pretraining infrastructure to empirically test this approach.

Method: Introduces Episodic Spatial World Model (ESWM) that constructs spatial maps from sparse, disjoint episodic memories. It operates on independently stored and updated episodic memories to enable rapid adaptation to environmental changes.

Result: ESWM predicts unobserved transitions from minimal experience across varying environments, with latent space geometry aligning with environment structure. Enables near-optimal exploration and navigation strategies without additional training.

Conclusion: Demonstrates how neuroscience-inspired episodic memory principles can advance development of more flexible and generalizable world models for spatial cognition and navigation.

Abstract: Many animals possess a remarkable capacity to rapidly construct flexible cognitive maps of their environments. These maps are crucial for ethologically relevant behaviors such as navigation, exploration, and planning. Existing computational models typically require long sequential trajectories to build accurate maps, but neuroscience evidence suggests maps can also arise from integrating disjoint experiences governed by consistent spatial rules. We introduce the Episodic Spatial World Model (ESWM), a novel framework that constructs spatial maps from sparse, disjoint episodic memories. Across environments of varying complexity, ESWM predicts unobserved transitions from minimal experience, and the geometry of its latent space aligns with that of the environment. Because it operates on episodic memories that can be independently stored and updated, ESWM is inherently adaptive, enabling rapid adjustment to environmental changes. Furthermore, we demonstrate that ESWM readily enables near-optimal strategies for exploring novel environments and navigating between arbitrary points, all without the need for additional training. Our work demonstrates how neuroscience-inspired principles of episodic memory can advance the development of more flexible and generalizable world models.

Method: Proposes a task-agnostic method that builds Cognitive Profiles using Entropy Decay Curves - plots of a model’s normalized predictive uncertainty as context length grows. Also introduces Information Gain Span (IGS) as a single index summarizing the desirability of decay patterns.

Result: Across several state-of-the-art LLMs and diverse texts, the curves expose distinctive, stable profiles that depend on both model scale and text complexity. The tools provide principled ways to analyze and compare internal dynamics of AI systems.

Conclusion: The proposed Cognitive Profiles and IGS offer a principled framework for understanding how LLMs process information, revealing systematic patterns related to model scale and text complexity that were previously hidden.

Abstract: Large Language Models (LLMs) excel on many task-specific benchmarks, yet the mechanisms that drive this success remain poorly understood. We move from asking what these systems can do to asking how they process information. Our contribution is a task-agnostic method that builds a quantitative Cognitive Profile for any model. The profile is built around the Entropy Decay Curve – a plot of a model’s normalised predictive uncertainty as context length grows. Across several state-of-the-art LLMs and diverse texts, the curves expose distinctive, stable profiles that depend on both model scale and text complexity. We also propose the Information Gain Span (IGS) as a single index that summarises the desirability of a decay pattern. Together, these tools offer a principled way to analyse and compare the internal dynamics of modern AI systems.

Method: MixGRPO integrates stochastic differential equations (SDE) and ordinary differential equations (ODE) with a sliding window mechanism. It uses SDE sampling and GRPO-guided optimization only within the window, while applying ODE sampling outside. This confines randomness to specific time-steps and reduces optimization overhead.

Result: MixGRPO achieves substantial gains in human preference alignment, outperforming DanceGRPO in both effectiveness and efficiency with nearly 50% lower training time. MixGRPO-Flash further reduces training time by 71% while maintaining comparable performance.

Conclusion: MixGRPO provides an efficient framework for human preference alignment in image generation by leveraging mixed sampling strategies, significantly reducing training time while improving performance.

Abstract: Although GRPO substantially enhances flow matching models in human preference alignment of image generation, methods such as FlowGRPO and DanceGRPO still exhibit inefficiency due to the necessity of sampling and optimizing over all denoising steps specified by the Markov Decision Process (MDP). In this paper, we propose $\textbf{MixGRPO}$, a novel framework that leverages the flexibility of mixed sampling strategies through the integration of stochastic differential equations (SDE) and ordinary differential equations (ODE). This streamlines the optimization process within the MDP to improve efficiency and boost performance. Specifically, MixGRPO introduces a sliding window mechanism, using SDE sampling and GRPO-guided optimization only within the window, while applying ODE sampling outside. This design confines sampling randomness to the time-steps within the window, thereby reducing the optimization overhead, and allowing for more focused gradient updates to accelerate convergence. Additionally, as time-steps beyond the sliding window are not involved in optimization, higher-order solvers are supported for faster sampling. So we present a faster variant, termed $\textbf{MixGRPO-Flash}$, which further improves training efficiency while achieving comparable performance. MixGRPO exhibits substantial gains across multiple dimensions of human preference alignment, outperforming DanceGRPO in both effectiveness and efficiency, with nearly 50% lower training time. Notably, MixGRPO-Flash further reduces training time by 71%.

Method: Proposes a “Micro-Scatter-Macro” multi-level foveation methodology that extracts textual information at three granularities (fine-grained details, cross-region connections, holistic patterns), plus a re-synthesis module to improve instruction fidelity and quality.

Result: Comprehensive experiments across multiple unsupervised corpora and diverse model architectures show Self-Foveate consistently outperforms existing methods for instruction synthesis.

Conclusion: Self-Foveate effectively addresses diversity and difficulty limitations in instruction synthesis, providing a promising approach for generating high-quality training data for LLMs from unsupervised text.

Abstract: Synthesizing high-quality instruction data from unsupervised text is a promising paradigm for training large language models (LLMs), yet automated methods for this task still exhibit significant limitations in the diversity and difficulty of synthesized instructions. To address these challenges, we propose Self-Foveate, an LLM-driven method for instruction synthesis. Inspired by hierarchical human visual perception, Self-Foveate introduces a “Micro-Scatter-Macro” multi-level foveation methodology that guides the extraction of textual information at three complementary granularities, from fine-grained details through cross-region connections to holistic patterns, thereby enhancing both the diversity and difficulty of synthesized instructions. Furthermore, a re-synthesis module is incorporated to improve the fidelity of instructions to source text and their overall quality. Comprehensive experiments across multiple unsupervised corpora and diverse model architectures demonstrate that Self-Foveate consistently outperforms existing methods. We publicly release our code at https://github.com/Mubuky/Self-Foveate

Method: Proposes Amadeus, a training-free framework that enhances persona consistency in RAG-based RPAs. Also introduces CharacterRAG dataset with 15 fictional characters’ persona documents (976K characters) and 450 QA pairs for evaluation.

Result: Amadeus effectively models not only character knowledge but also various attributes like personality, significantly enhancing persona consistency even for questions beyond a character’s knowledge.

Conclusion: The proposed training-free framework addresses critical hallucination issues in RAG-based role-playing agents and provides a valuable dataset for future research in this under-explored area.

Abstract: Building role-playing agents (RPAs) that faithfully emulate specific characters remains challenging because collecting character-specific utterances and continually updating model parameters are resource-intensive, making retrieval-augmented generation (RAG) a practical necessity. However, despite the importance of RAG, there has been little research on RAG-based RPAs. For example, we empirically find that when a persona lacks knowledge relevant to a given query, RAG-based RPAs are prone to hallucination, making it challenging to generate accurate responses. In this paper, we propose Amadeus, a training-free framework that can significantly enhance persona consistency even when responding to questions that lie beyond a character’s knowledge. In addition, to underpin the development and rigorous evaluation of RAG-based RPAs, we manually construct CharacterRAG, a role-playing dataset that consists of persona documents for 15 distinct fictional characters totaling 976K written characters, and 450 question-answer pairs. We find that our proposed method effectively models not only the knowledge possessed by characters, but also various attributes such as personality.

Method: Proposes a generic framework that employs uncertainty measures to guide switching decisions, can incorporate various uncertainty-measuring mechanisms (intrinsic motivation, epistemic uncertainty), and subsumes previous adaptive exploration approaches as special cases.

Result: The framework produces adaptive exploration strategies that outperform standard approaches across several experimental environments.

Conclusion: The uncertainty-based adaptive exploration framework provides a principled solution to the switching problem and demonstrates superior performance compared to existing methods.

Abstract: Adaptive exploration methods propose ways to learn complex policies via alternating between exploration and exploitation. An important question for such methods is to determine the appropriate moment to switch between exploration and exploitation and vice versa. This is critical in domains that require the learning of long and complex sequences of actions. In this work, we present a generic adaptive exploration framework that employs uncertainty to address this important issue in a principled manner. Our framework includes previous adaptive exploration approaches as special cases. Moreover, we can incorporate in our framework any uncertainty-measuring mechanism of choice, for instance mechanisms used in intrinsic motivation or epistemic uncertainty-based exploration methods. We experimentally demonstrate that our framework gives rise to adaptive exploration strategies that outperform standard ones across several environments.

Method: MAPGD uses specialized agents focusing on distinct refinement dimensions (instruction clarity, example selection, format structure, stylistic adaptation). Agents coordinate through semantic gradient embedding, conflict detection, and fusion. Introduces Hypersphere Constrained Gradient Clustering (HCGC) for compact clusters and Channel Adaptive Agent Weighting (CAAW) for dynamic agent reweighting.

Result: MAPGD consistently surpasses single-agent and random baselines in both accuracy and efficiency on classification and reasoning benchmarks. Ablation studies confirm effectiveness of gradient fusion, agent specialization, and conflict resolution.

Conclusion: MAPGD establishes a unified, gradient-based, interpretable framework for robust prompt optimization with theoretical convergence guarantees, addressing limitations of single-trajectory optimization methods.

Abstract: Prompt engineering is crucial for fully leveraging large language models (LLMs), yet most existing optimization methods follow a single trajectory, resulting in limited adaptability, gradient conflicts, and high computational overhead. We propose MAPGD (Multi-Agent Prompt Gradient Descent), a novel framework that reconceptualizes prompt optimization as a collaborative process among specialized agents. Each agent focuses on a distinct refinement dimension, such as instruction clarity, example selection, format structure, or stylistic adaptation, and their contributions are coordinated through semantic gradient embedding, conflict detection, and fusion. To further enhance robustness and stability, MAPGD introduces two new mechanisms: Hypersphere Constrained Gradient Clustering (HCGC), which enforces angular margin constraints for compact and well-separated clusters, and Channel Adaptive Agent Weighting (CAAW), which dynamically reweights agent contributions based on validation performance. Experiments on classification and reasoning benchmarks show that MAPGD consistently surpasses single-agent and random baselines in both accuracy and efficiency. Ablation studies confirm the effectiveness of gradient fusion, agent specialization, and conflict resolution. Together, these components establish MAPGD as a unified, gradient-based, and interpretable framework for robust prompt optimization with theoretical convergence guarantees.

Method: PAINT trains a generative neural network to model the distribution of states in parallel over time, then predicts states from measurements in a sliding window fashion at test time. The method is architecture-agnostic.

Result: Theoretical analysis shows PAINT is on-trajectory while autoregressive models generally are not. Empirical evaluation on 2D turbulent fluid dynamics demonstrates PAINT stays on-trajectory and predicts system states from sparse measurements with high fidelity.

Conclusion: PAINT has potential for developing neural twins that stay on-trajectory, enabling more accurate state estimation and decision-making in dynamical systems.

Abstract: Neural surrogates have shown great potential in simulating dynamical systems, while offering real-time capabilities. We envision Neural Twins as a progression of neural surrogates, aiming to create digital replicas of real systems. A neural twin consumes measurements at test time to update its state, thereby enabling context-specific decision-making. We argue, that a critical property of neural twins is their ability to remain on-trajectory, i.e., to stay close to the true system state over time. We introduce Parallel-in-time Neural Twins (PAINT), an architecture-agnostic family of methods for modeling dynamical systems from measurements. PAINT trains a generative neural network to model the distribution of states in parallel over time. At test time, states are predicted from measurements in a sliding window fashion. Our theoretical analysis shows that PAINT is on-trajectory, whereas autoregressive models generally are not. Empirically, we evaluate our method on a challenging two-dimensional turbulent fluid dynamics problem. The results demonstrate that PAINT stays on-trajectory and predicts system states from sparse measurements with high fidelity. These findings underscore PAINT’s potential for developing neural twins that stay on-trajectory, enabling more accurate state estimation and decision-making.

Method: Introduces Chain-of-Thought Hijacking attack that prepends harmful instructions with extended sequences of benign puzzle reasoning. Uses activation probing, attention analysis, and causal interventions to understand the mechanism. Evaluates on HarmBench across major models (Gemini 2.5 Pro, ChatGPT o4 Mini, Grok 3 Mini, Claude 4 Sonnet).

Result: Achieves attack success rates of 99%, 94%, 100%, and 94% on the four models respectively. Finds that refusal depends on a low-dimensional safety signal that becomes diluted as reasoning grows: mid-layers encode safety checking strength, while late layers encode refusal outcome.

Conclusion: Explicit chain-of-thought reasoning introduces a systematic vulnerability when combined with answer-prompting cues, demonstrating that extended reasoning can weaken rather than strengthen safety mechanisms in Large Reasoning Models.

Abstract: Large Reasoning Models (LRMs) improve task performance through extended inference-time reasoning. While prior work suggests this should strengthen safety, we find evidence to the contrary. Long reasoning sequences can be exploited to systematically weaken them. We introduce Chain-of-Thought Hijacking, a jailbreak attack that prepends harmful instructions with extended sequences of benign puzzle reasoning. Across HarmBench, CoT Hijacking achieves attack success rates of 99%, 94%, 100%, and 94% on Gemini 2.5 Pro, ChatGPT o4 Mini, Grok 3 Mini, and Claude 4 Sonnet. To understand this mechanism, we apply activation probing, attention analysis, and causal interventions. We find that refusal depends on a low-dimensional safety signal that becomes diluted as reasoning grows: mid-layers encode the strength of safety checking, while late layers encode the refusal outcome. These findings demonstrate that explicit chain-of-thought reasoning introduces a systematic vulnerability when combined with answer-prompting cues. We release all evaluation materials to facilitate replication.

Method: Uses CrochetPARADE DSL as intermediate representation for structural validation and functional evaluation via execution. Benchmark covers stitch classification, instruction grounding, and both natural language and image-to-DSL translation tasks.

Result: Performance sharply decreases as evaluation shifts from surface-level similarity to executable correctness, revealing limitations in long-range symbolic reasoning and 3D-aware procedural synthesis.

Conclusion: CrochetBench offers a new lens for assessing procedural competence in multimodal models and highlights the gap between surface-level understanding and executable precision in real-world creative domains.

Abstract: While multimodal large language models can describe visual content, their ability to generate executable procedures remains underexplored. CrochetBench presented in this paper evaluates this shift from describing to doing through fine-grained procedural reasoning in crochet: models must recognize stitches, select structurally appropriate instructions, and generate compilable procedures. We adopt the CrochetPARADE DSL as our intermediate representation, enabling structural validation and functional evaluation via execution. The benchmark covers tasks including stitch classification, instruction grounding, and both natural language and image-to-DSL translation. Across all tasks, performance sharply decreases as the evaluation shifts from surface-level similarity to executable correctness, revealing limitations in long-range symbolic reasoning and 3D-aware procedural synthesis. Our proposed CrochetBench offers a new lens for assessing procedural competence in multimodal models and highlights the gap between surface-level understanding and executable precision in real-world creative domains. Code is available at https://anonymous.4open.science/r/crochet-82E6/README.md.

Method: NEZHA integrates a nimble autoregressive draft head directly into the primary model for efficient self-drafting, uses specialized input prompt structures, and introduces a model-free verifier based on hash sets to tackle hallucination issues.

Result: Successfully deployed on Taobao since October 2025, driving billion-level advertising revenue and serving hundreds of millions of daily active users, with extensive experiments demonstrating effectiveness on public datasets.

Conclusion: NEZHA enables practical deployment of generative recommendation systems at industrial scale by solving the latency bottleneck through integrated self-drafting and efficient verification.

Abstract: Generative Recommendation (GR), powered by Large Language Models (LLMs), represents a promising new paradigm for industrial recommender systems. However, their practical application is severely hindered by high inference latency, which makes them infeasible for high-throughput, real-time services and limits their overall business impact. While Speculative Decoding (SD) has been proposed to accelerate the autoregressive generation process, existing implementations introduce new bottlenecks: they typically require separate draft models and model-based verifiers, requiring additional training and increasing the latency overhead. In this paper, we address these challenges with NEZHA, a novel architecture that achieves hyperspeed decoding for GR systems without sacrificing recommendation quality. Specifically, NEZHA integrates a nimble autoregressive draft head directly into the primary model, enabling efficient self-drafting. This design, combined with a specialized input prompt structure, preserves the integrity of sequence-to-sequence generation. Furthermore, to tackle the critical problem of hallucination, a major source of performance degradation, we introduce an efficient, model-free verifier based on a hash set. We demonstrate the effectiveness of NEZHA through extensive experiments on public datasets and have successfully deployed the system on Taobao since October 2025, driving the billion-level advertising revenue and serving hundreds of millions of daily active users.

Method: Proposes a co-evolving agents framework with a target agent and auxiliary failure agent. The failure agent learns through preference optimization over failure trajectories from both agents, generating hard negatives that are close to success but remain failures. These hard negatives are incorporated into the target agent’s optimization to sharpen decision boundaries.

Result: The method shows improved performance across benchmark datasets and demonstrates that failures can be systematically transformed into structured and valuable learning signals in self-improving agents.

Conclusion: The co-evolving agents framework effectively addresses overfitting in self-improving agents by generating informative hard negatives from failure trajectories, leading to better generalization and performance.

Abstract: The rapid progress of large foundation models has accelerated the development of task-specialized agents across diverse domains. However, the effectiveness of agents remains tightly coupled with the quality of training data, while curating task-specific datasets remains costly and often infeasible in real-world scenarios. Recent work has explored self-improving agents that autonomously generate, refine, and re-train on their own trajectories. A prominent line of approaches further leverages preference optimization by pairing predicted trajectories with scarce ground-truth trajectories, enabling agents to learn directly from their own failures. While these methods outperform supervised fine-tuning, their heavy reliance on predicted trajectories under limited ground-truth supervision leaves them prone to overfitting. To address this, we propose a co-evolving agents framework in which a target agent improves jointly with an auxiliary failure agent. The failure agent learns through preference optimization over failure trajectories from both the target and itself, thereby generating hard negatives that are close to success yet remain failures. Incorporating these informative hard negatives into the target agent’s optimization sharpens decision boundaries and enhances generalization. Our comprehensive analysis and experiments across benchmark datasets show that our method not only shows improved performance but also demonstrates that failures, instead of being used as-is, can be systematically transformed into structured and valuable learning signals in self-improving agents.

Method: Proposes GlimpRouter based on the insight that step difficulty can be inferred from the first token’s entropy (inspired by “Aha Moment” phenomenon). Uses lightweight model to generate only the first token of each reasoning step, then routes to larger model only when initial token entropy exceeds a threshold. Training-free framework.

Result: Experiments on multiple benchmarks show significant reduction in inference latency while preserving accuracy. Achieves 10.7% improvement in accuracy while reducing inference latency by 25.9% compared to standalone large model on AIME25.

Conclusion: GlimpRouter demonstrates a simple yet effective mechanism for reasoning: allocating computation based on a glimpse of thought rather than full-step evaluation. Shows initial token entropy is a strong predictor of step difficulty.

Abstract: Large Reasoning Models (LRMs) achieve remarkable performance by explicitly generating multi-step chains of thought, but this capability incurs substantial inference latency and computational cost. Collaborative inference offers a promising solution by selectively allocating work between lightweight and large models, yet a fundamental challenge remains: determining when a reasoning step requires the capacity of a large model or the efficiency of a small model. Existing routing strategies either rely on local token probabilities or post-hoc verification, introducing significant inference overhead. In this work, we propose a novel perspective on step-wise collaboration: the difficulty of a reasoning step can be inferred from its very first token. Inspired by the “Aha Moment” phenomenon in LRMs, we show that the entropy of the initial token serves as a strong predictor of step difficulty. Building on this insight, we introduce GlimpRouter, a training-free step-wise collaboration framework. GlimpRouter employs a lightweight model to generate only the first token of each reasoning step and routes the step to a larger model only when the initial token entropy exceeds a threshold. Experiments on multiple benchmarks demonstrate that our approach significantly reduces inference latency while preserving accuracy. For instance, GlimpRouter attains a substantial 10.7% improvement in accuracy while reducing inference latency by 25.9% compared to a standalone large model on AIME25. These results suggest a simple yet effective mechanism for reasoning: allocating computation based on a glimpse of thought rather than full-step evaluation.

Method: Uses hands-on examples and approachable explanations to teach foundational RL skills, focusing on making the material accessible to undergraduate students.

Result: The tutorial aims to equip students with foundational skills needed to confidently apply RL techniques in real-world scenarios.

Conclusion: This educational resource successfully makes reinforcement learning more accessible through practical, example-driven teaching methods.

Abstract: This tutorial is designed to make reinforcement learning (RL) more accessible to undergraduate students by offering clear, example-driven explanations. It focuses on bridging the gap between RL theory and practical coding applications, addressing common challenges that students face when transitioning from conceptual understanding to implementation. Through hands-on examples and approachable explanations, the tutorial aims to equip students with the foundational skills needed to confidently apply RL techniques in real-world scenarios.

Method: Theoretical framework proposing that cognition can be characterized by the interplay between two invariants: (1) remapping of embedding spaces (transcriptional, morphological, physiological, 3D spaces in biological systems; latent embeddings in AI systems), and (2) navigation within these spaces through distributed error correction/iterative refinement.

Result: Identifies a dual principle - remapping and navigation of embedding spaces via iterative error minimization - as a substrate-independent invariant of cognition that applies across biological collectives and modern AI systems like transformers and diffusion models.

Conclusion: This shared mechanism provides a unifying framework for understanding cognition in both natural and synthetic systems, illuminating deep parallels between living systems and artificial models, and enabling engineering of adaptive intelligence across scales.

Abstract: The emerging field of diverse intelligence seeks an integrated view of problem-solving in agents of very different provenance, composition, and substrates. From subcellular chemical networks to swarms of organisms, and across evolved, engineered, and chimeric systems, it is hypothesized that scale-invariant principles of decision-making can be discovered. We propose that cognition in both natural and synthetic systems can be characterized and understood by the interplay between two equally important invariants: (1) the remapping of embedding spaces, and (2) the navigation within these spaces. Biological collectives, from single cells to entire organisms (and beyond), remap transcriptional, morphological, physiological, or 3D spaces to maintain homeostasis and regenerate structure, while navigating these spaces through distributed error correction. Modern Artificial Intelligence (AI) systems, including transformers, diffusion models, and neural cellular automata enact analogous processes by remapping data into latent embeddings and refining them iteratively through contextualization. We argue that this dual principle - remapping and navigation of embedding spaces via iterative error minimization - constitutes a substrate-independent invariant of cognition. Recognizing this shared mechanism not only illuminates deep parallels between living systems and artificial models, but also provides a unifying framework for engineering adaptive intelligence across scales.

Method: Created a Gymnasium environment modeling immersion water heater with thermal losses and discrete on/off actions. Compared three approaches: time-optimal bang-bang baseline, zero-shot Monte Carlo Tree Search planner, and Proximal Policy Optimization reinforcement learning policy.

Result: PPO achieved most energy-efficient performance, using 3.23 kWh at 60-step horizon vs 4.37-10.45 kWh for bang-bang and 4.18-6.46 kWh for MCTS. Energy savings of 26% at 30 steps and 69% at 90 steps. In representative case, PPO consumed 54% less energy than bang-bang and 33% less than MCTS.

Conclusion: Learned deadline-aware control significantly reduces energy consumption compared to traditional methods. Planners offer partial savings without training, while learned policies provide near-zero inference cost after training.

Abstract: Typical domestic immersion water heater systems are often operated continuously during winter, heating quickly rather than efficiently and ignoring predictable demand windows and ambient losses. We study deadline-aware control, where the aim is to reach a target temperature at a specified time while minimising energy consumption. We introduce an efficient Gymnasium environment that models an immersion hot water heater with first-order thermal losses and discrete on and off actions of 0 W and 6000 W applied every 120 seconds. Methods include a time-optimal bang-bang baseline, a zero-shot Monte Carlo Tree Search planner, and a Proximal Policy Optimisation policy. We report total energy consumption in watt-hours under identical physical dynamics. Across sweeps of initial temperature from 10 to 30 degrees Celsius, deadline from 30 to 90 steps, and target temperature from 40 to 80 degrees Celsius, PPO achieves the most energy-efficient performance at a 60-step horizon of 2 hours, using 3.23 kilowatt-hours, compared to 4.37 to 10.45 kilowatt-hours for bang-bang control and 4.18 to 6.46 kilowatt-hours for MCTS. This corresponds to energy savings of 26 percent at 30 steps and 69 percent at 90 steps. In a representative trajectory with a 50 kg water mass, 20 degrees Celsius ambient temperature, and a 60 degrees Celsius target, PPO consumes 54 percent less energy than bang-bang control and 33 percent less than MCTS. These results show that learned deadline-aware control reduces energy consumption under identical physical assumptions, while planners provide partial savings without training and learned policies offer near-zero inference cost once trained.

Method: PROTEUS uses Lagrangian dual control where a learned dual variable tracks constraint violations during training and conditions the policy network. This allows the router to translate specified accuracy targets (tau) into routing decisions that satisfy them, with a single trained model serving the full accuracy spectrum without retraining.

Result: On RouterBench (11 models, 405K queries) and SPROUT (14 models, 45K queries), PROTEUS achieves consistent floor compliance where accuracy meets or exceeds tau, with target-response correlation of 0.97-0.98. It achieves 90.1% accuracy on RouterBench (within 1.3% of oracle) and 94.0% on SPROUT (within 4.6% of oracle), with cost savings reaching 89.8% versus the best fixed model.

Conclusion: PROTEUS enables operators to specify accuracy targets directly as runtime input, providing predictable performance while optimizing costs across diverse LLM deployment scenarios, significantly outperforming existing routing approaches.

Abstract: Production LLM deployments serve diverse workloads where cost and quality requirements vary by customer tier, time of day, and query criticality. Model serving systems accept latency SLOs directly. LLM routers do not. They force operators to tune parameters offline and guess what accuracy might result. The relationship between parameters and outcomes is indirect, non-monotonic, and dataset-dependent. Operators need to specify accuracy targets, not infer them from opaque settings. We present PROTEUS (Polymorphic Router for Operational Target Enforcement with Unified SLA), a router that accepts accuracy targets tau as runtime input. PROTEUS uses Lagrangian dual control. A learned dual variable lambda tracks constraint violations during training and conditions the policy network. This lets the router translate specified tau values into routing decisions that satisfy them. A single trained model serves the full accuracy spectrum without retraining.We evaluate on RouterBench (11 models, 405K queries) and SPROUT (14 models, 45K queries). PROTEUS achieves consistent floor compliance where accuracy meets or exceeds tau. The target-response correlation reaches 0.97 to 0.98. The closest baseline, OmniRouter, meets floors only 22% of the time despite also using Lagrangian optimization. PROTEUS operates across tau in [0.85, 0.95] from a single model. On RouterBench it achieves 90.1% accuracy, within 1.3% of oracle. On SPROUT it achieves 94.0% accuracy, within 4.6% of oracle. Cost savings reach 89.8% versus the best fixed model.

Method: Introduces EPDDL (Epistemic Planning Domain Definition Language) with formal development of abstract event models for representing epistemic actions, and formal specification of syntax and semantics grounded in DEL.

Result: EPDDL provides a unique PDDL-like representation capturing entire DEL semantics, enabling uniform specification of epistemic planning tasks and facilitating interoperability and reproducible evaluation.

Conclusion: EPDDL addresses fragmentation in epistemic planning by providing a standardized language that captures full DEL semantics, supporting future advances in the field through better interoperability and benchmark development.

Abstract: Epistemic planning extends (multi-agent) automated planning by making agents’ knowledge and beliefs first-class aspects of the planning formalism. One of the most well-known frameworks for epistemic planning is Dynamic Epistemic Logic (DEL), which offers an rich and natural semantics for modelling problems in this setting. The high expressive power provided by DEL make DEL-based epistemic planning a challenging problem to tackle both theoretically, and in practical implementations. As a result, existing epistemic planners often target different DEL fragments, and typically rely on ad hoc languages to represent benchmarks, and sometimes no language at all. This fragmentation hampers comparison, reuse, and systematic benchmark development. We address these issues by introducing the Epistemic Planning Domain Definition Language (EPDDL). EPDDL provides a unique PDDL-like representation that captures the entire DEL semantics, enabling uniform specification of epistemic planning tasks. Our main contributions are: 1. A formal development of abstract event models, a novel representation for epistemic actions used to define the semantics of our language; 2. A formal specification of EPDDL’s syntax and semantics grounded in DEL with abstract event models. Through examples of representative benchmarks, we illustrate how EPDDL facilitates interoperability, reproducible evaluation, and future advances in epistemic planning.

Method: Develop CUA-Skill, a large-scale library of carefully engineered skills spanning common Windows applications, with parameterized execution and composition graphs. Build CUA-Skill Agent on top with dynamic skill retrieval, argument instantiation, and memory-aware failure recovery.

Result: CUA-Skill substantially improves execution success rates and robustness on challenging end-to-end agent benchmarks. On WindowsAgentArena, achieves state-of-the-art 57.5% success rate while being significantly more efficient than prior approaches.

Conclusion: CUA-Skill establishes a strong foundation for future computer-using agent development by providing reusable skill abstractions that encode human computer-use knowledge, enabling scalable and reliable agent systems.

Abstract: Computer-Using Agents (CUAs) aim to autonomously operate computer systems to complete real-world tasks. However, existing agentic systems remain difficult to scale and lag behind human performance. A key limitation is the absence of reusable and structured skill abstractions that capture how humans interact with graphical user interfaces and how to leverage these skills. We introduce CUA-Skill, a computer-using agentic skill base that encodes human computer-use knowledge as skills coupled with parameterized execution and composition graphs. CUA-Skill is a large-scale library of carefully engineered skills spanning common Windows applications, serving as a practical infrastructure and tool substrate for scalable, reliable agent development. Built upon this skill base, we construct CUA-Skill Agent, an end-to-end computer-using agent that supports dynamic skill retrieval, argument instantiation, and memory-aware failure recovery. Our results demonstrate that CUA-Skill substantially improves execution success rates and robustness on challenging end-to-end agent benchmarks, establishing a strong foundation for future computer-using agent development. On WindowsAgentArena, CUA-Skill Agent achieves state-of-the-art 57.5% (best of three) successful rate while being significantly more efficient than prior and concurrent approaches. The project page is available at https://microsoft.github.io/cua_skill/.

Method: Provides rigorous theoretical guarantees including: 1) lower bound characterizing fundamental limits of single update step, 2) finite-sample error bounds for full iterative paradigm, 3) analysis showing dependence on initial model decays exponentially with iterations, 4) instantiation for linear softmax model class.

Result: Performance improves at rate O(1/√n) with sample size n; dependence on initial model decays exponentially with number of iterations T; provides formal explanation for why self-rewarding succeeds by steering dynamics toward internal stability and consistency.

Conclusion: Theoretical framework explains self-rewarding success through robust overcoming of poor initialization via internal stability and consistency mechanisms, with practical guarantees for linear softmax architectures.

Abstract: Self-Rewarding Language Models (SRLMs) achieve notable success in iteratively improving alignment without external feedback. Yet, despite their striking empirical progress, the core mechanisms driving their capabilities remain unelucidated, leaving a critical gap in theoretical understanding. This paper provides the first rigorous theoretical guarantees for SRLMs. We first establish a lower bound that characterizes the fundamental limits of a single update step, revealing a critical dependence on the quality of the initial model. We then derive finite-sample error bounds for the full iterative paradigm, showing that performance improves at a rate of $\widetilde{\mathcal{O}}\left(1/\sqrt{n}\right)$ with sample size $n$. Crucially, our analysis reveals that the dependence on the initial model decays exponentially with the number of iterations $T$. This provides a formal explanation for why self-rewarding succeeds: it robustly overcomes poor initialization by steering the dynamics toward internal stability and consistency. Finally, we instantiate our theoretical framework for the linear softmax model class, yielding tailored guarantees that connect our high-level insights to practical model architectures.

Method: Proposes Golden Goose method: given source text, uses LLM to identify and mask key reasoning steps, then generates diverse plausible distractors to create multiple-choice question-answering versions of fill-in-the-middle tasks. This enables synthesis of RLVR datasets from reasoning-rich unverifiable corpora like science textbooks.

Result: Created GooseReason-0.7M dataset with 0.7M tasks across math, programming, and science domains. Achieved new SOTA results for 1.5B and 4B-Instruct models across 15 benchmarks. Also created GooseReason-Cyber for cybersecurity, where Qwen3-4B-Instruct surpassed a 7B domain-specialized model.

Conclusion: Golden Goose enables automatic scaling of RLVR data by exploiting abundant unverifiable internet text, reviving saturated models and achieving sustained gains under continuous RL training across diverse domains including cybersecurity.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become a cornerstone for unlocking complex reasoning in Large Language Models (LLMs). Yet, scaling up RL is bottlenecked by limited existing verifiable data, where improvements increasingly saturate over prolonged training. To overcome this, we propose Golden Goose, a simple trick to synthesize unlimited RLVR tasks from unverifiable internet text by constructing a multiple-choice question-answering version of the fill-in-the-middle task. Given a source text, we prompt an LLM to identify and mask key reasoning steps, then generate a set of diverse, plausible distractors. This enables us to leverage reasoning-rich unverifiable corpora typically excluded from prior RLVR data construction (e.g., science textbooks) to synthesize GooseReason-0.7M, a large-scale RLVR dataset with over 0.7 million tasks spanning mathematics, programming, and general scientific domains. Empirically, GooseReason effectively revives models saturated on existing RLVR data, yielding robust, sustained gains under continuous RL and achieving new state-of-the-art results for 1.5B and 4B-Instruct models across 15 diverse benchmarks. Finally, we deploy Golden Goose in a real-world setting, synthesizing RLVR tasks from raw FineWeb scrapes for the cybersecurity domain, where no prior RLVR data exists. Training Qwen3-4B-Instruct on the resulting data GooseReason-Cyber sets a new state-of-the-art in cybersecurity, surpassing a 7B domain-specialized model with extensive domain-specific pre-training and post-training. This highlights the potential of automatically scaling up RLVR data by exploiting abundant, reasoning-rich, unverifiable internet text.

Method: Used Embodied Agent Interface (EAI) framework to evaluate two 7B-parameter models on VirtualHome benchmark across four tasks: Goal Interpretation, Action Sequencing, Subgoal Decomposition, and Transition Modeling. Proposed Structured Self-Consistency (SSC) - an enhanced decoding strategy using multiple sampling with domain-specific voting mechanisms.

Result: SSC significantly enhanced performance, with OPENPANGU-7B excelling at hierarchical planning while QWEN2.5-7B showed advantages in action-level tasks. Models revealed complementary strengths across different task types.

Conclusion: The analysis provides insights for future embodied AI system development, showing that different LLM architectures have complementary strengths and that structured decoding strategies like SSC can significantly improve performance on embodied AI tasks.

Abstract: Embodied AI requires agents to understand goals, plan actions, and execute tasks in simulated environments. We present a comprehensive evaluation of Large Language Models (LLMs) on the VirtualHome benchmark using the Embodied Agent Interface (EAI) framework. We compare two representative 7B-parameter models OPENPANGU-7B and QWEN2.5-7B across four fundamental tasks: Goal Interpretation, Action Sequencing, Subgoal Decomposition, and Transition Modeling. We propose Structured Self-Consistency (SSC), an enhanced decoding strategy that leverages multiple sampling with domain-specific voting mechanisms to improve output quality for structured generation tasks. Experimental results demonstrate that SSC significantly enhances performance, with OPENPANGU-7B excelling at hierarchical planning while QWEN2.5-7B show advantages in action-level tasks. Our analysis reveals complementary strengths across model types, providing insights for future embodied AI system development.

Method: REPCORE aligns heterogeneous hidden states from different models into a unified latent space to construct representative coresets. It uses these subsets for performance extrapolation, enabling precise estimation with as few as ten source models.

Result: Experiments on five benchmarks and over 200 models show consistent gains over output-based baselines in ranking correlation and estimation accuracy. Spectral analysis reveals that aligned representations contain separable components reflecting broad response tendencies and task-specific reasoning patterns.

Conclusion: REPCORE provides a more effective approach to efficient LLM benchmarking by leveraging hidden state information rather than just correctness labels, enabling reliable performance estimation with minimal source models.

Abstract: The prohibitive cost of evaluating Large Language Models (LLMs) necessitates efficient alternatives to full-scale benchmarking. Prevalent approaches address this by identifying a small coreset of items to approximate full-benchmark performance. However, existing methods must estimate a reliable item profile from response patterns across many source models, which becomes statistically unstable when the source pool is small. This dependency is particularly limiting for newly released benchmarks with minimal historical evaluation data. We argue that discrete correctness labels are a lossy view of the model’s decision process and fail to capture information encoded in hidden states. To address this, we introduce REPCORE, which aligns heterogeneous hidden states into a unified latent space to construct representative coresets. Using these subsets for performance extrapolation, REPCORE achieves precise estimation accuracy with as few as ten source models. Experiments on five benchmarks and over 200 models show consistent gains over output-based baselines in ranking correlation and estimation accuracy. Spectral analysis further indicates that the aligned representations contain separable components reflecting broad response tendencies and task-specific reasoning patterns.

Method: CAREP uses three specialized agents: 1) causal discovery agent to identify DTC-EP relations, 2) contextual information agent to integrate metadata and descriptions, and 3) orchestrator agent to synthesize Boolean rules with interpretable reasoning traces.

Result: Evaluation on large-scale automotive dataset (29,100+ unique DTCs, 474 error patterns) shows CAREP can automatically and accurately discover unknown EP rules, outperforming LLM-only baselines while providing transparent causal explanations.

Conclusion: CAREP represents progress toward fully automated fault diagnostics, enabling scalable, interpretable, and cost-efficient vehicle maintenance through practical causal discovery and agent-based reasoning.

Abstract: Modern vehicles generate thousands of different discrete events known as Diagnostic Trouble Codes (DTCs). Automotive manufacturers use Boolean combinations of these codes, called error patterns (EPs), to characterize system faults and ensure vehicle safety. Yet, EP rules are still manually handcrafted by domain experts, a process that is expensive and prone to errors as vehicle complexity grows. This paper introduces CAREP (Causal Automated Reasoning for Error Patterns), a multi-agent system that automatizes the generation of EP rules from high-dimensional event sequences of DTCs. CAREP combines a causal discovery agent that identifies potential DTC-EP relations, a contextual information agent that integrates metadata and descriptions, and an orchestrator agent that synthesizes candidate boolean rules together with interpretable reasoning traces. Evaluation on a large-scale automotive dataset with over 29,100 unique DTCs and 474 error patterns demonstrates that CAREP can automatically and accurately discover the unknown EP rules, outperforming LLM-only baselines while providing transparent causal explanations. By uniting practical causal discovery and agent-based reasoning, CAREP represents a step toward fully automated fault diagnostics, enabling scalable, interpretable, and cost-efficient vehicle maintenance.

Method: Proposes DFA (Disambiguation–Filtering–Aggregation), a modular agentic baseline that decomposes aggregation querying into interpretable stages: disambiguation, filtering, and aggregation, exposing key failure modes.

Result: DFA consistently improves aggregation evidence coverage over strong RAG and agentic baselines on AGGBench, a new benchmark for completeness-oriented aggregation under realistic large-scale corpus settings.

Conclusion: The work formalizes entity-level aggregation querying with strict completeness requirements, introduces AGGBench for evaluation, and demonstrates DFA’s effectiveness as a modular approach that improves evidence coverage.

Abstract: Aggregation query over free text is a long-standing yet underexplored problem. Unlike ordinary question answering, aggregate queries require exhaustive evidence collection and systems are required to “find all,” not merely “find one.” Existing paradigms such as Text-to-SQL and Retrieval-Augmented Generation fail to achieve this completeness. In this work, we formalize entity-level aggregation querying over text in a corpus-bounded setting with strict completeness requirement. To enable principled evaluation, we introduce AGGBench, a benchmark designed to evaluate completeness-oriented aggregation under realistic large-scale corpus. To accompany the benchmark, we propose DFA (Disambiguation–Filtering–Aggregation), a modular agentic baseline that decomposes aggregation querying into interpretable stages and exposes key failure modes related to ambiguity, filtering, and aggregation. Empirical results show that DFA consistently improves aggregation evidence coverage over strong RAG and agentic baselines. The data and code are available in \href{https://anonymous.4open.science/r/DFA-A4C1}.

Method: Establishes equivalence between GFlowNet objectives and Markov chain reversibility, then proposes α-GFNs with tunable parameter α to generalize the mixing. This provides direct control over exploration-exploitation dynamics while ensuring convergence to unique flows.

Result: α-GFN objectives consistently outperform previous GFlowNet objectives across Set, Bit Sequence, and Molecule Generation benchmarks, achieving up to 10× increase in discovered modes.

Conclusion: The α-GFN framework successfully generalizes GFlowNets with tunable exploration-exploitation control, significantly improving mode discovery capabilities while maintaining theoretical guarantees.

Abstract: Generative Flow Network (GFlowNet) objectives implicitly fix an equal mixing of forward and backward policies, potentially constraining the exploration-exploitation trade-off during training. By further exploring the link between GFlowNets and Markov chains, we establish an equivalence between GFlowNet objectives and Markov chain reversibility, thereby revealing the origin of such constraints, and provide a framework for adapting Markov chain properties to GFlowNets. Building on these theoretical findings, we propose $α$-GFNs, which generalize the mixing via a tunable parameter $α$. This generalization enables direct control over exploration-exploitation dynamics to enhance mode discovery capabilities, while ensuring convergence to unique flows. Across various benchmarks, including Set, Bit Sequence, and Molecule Generation, $α$-GFN objectives consistently outperform previous GFlowNet objectives, achieving up to a $10 \times$ increase in the number of discovered modes.

Method: Formalized analogical reasoning using category theory functors, created synthetic evaluation tasks, analyzed emergence under controlled settings, and conducted mechanistic analysis of Transformer components.

Result: Found analogical reasoning emerges through two key mechanisms: geometric alignment of relational structure in embedding space and application of functor-like operations within Transformers, with emergence highly sensitive to data, optimization, and scale.

Conclusion: Analogical reasoning can be concretely understood as mechanistically grounded phenomena in neural networks, moving from abstract cognitive notion to measurable implementation in Transformers.

Abstract: Analogy is a central faculty of human intelligence, enabling abstract patterns discovered in one domain to be applied to another. Despite its central role in cognition, the mechanisms by which Transformers acquire and implement analogical reasoning remain poorly understood. In this work, inspired by the notion of functors in category theory, we formalize analogical reasoning as the inference of correspondences between entities across categories. Based on this formulation, we introduce synthetic tasks that evaluate the emergence of analogical reasoning under controlled settings. We find that the emergence of analogical reasoning is highly sensitive to data characteristics, optimization choices, and model scale. Through mechanistic analysis, we show that analogical reasoning in Transformers decomposes into two key components: (1) geometric alignment of relational structure in the embedding space, and (2) the application of a functor within the Transformer. These mechanisms enable models to transfer relational structure from one category to another, realizing analogy. Finally, we quantify these effects and find that the same trends are observed in pretrained LLMs. In doing so, we move analogy from an abstract cognitive notion to a concrete, mechanistically grounded phenomenon in modern neural networks.

Method: Proposes TIDE (Test-time Improvement Diagnostic Evaluation), an agent-agnostic and environment-agnostic framework that decomposes TTI into three interconnected dimensions: (1) overall temporal dynamics of task completion, (2) identification of performance constraints from recursive looping behaviors, and (3) identification of constraints from burdensome accumulated memory.

Result: Through extensive experiments across diverse agents and environments, TIDE reveals that improving agent performance requires more than scaling internal reasoning - it calls for explicitly optimizing the interaction dynamics between the agent and the environment.

Conclusion: TIDE provides a comprehensive diagnostic framework for understanding TTI mechanisms in autonomous LLM agents, highlighting the importance of optimizing agent-environment interaction dynamics rather than just scaling internal reasoning capabilities.

Abstract: Recent advances in autonomous LLM agents demonstrate their ability to improve performance through iterative interaction with the environment. We define this paradigm as Test-Time Improvement (TTI). However, the mechanisms under how and why TTI succeed or fail remain poorly understood, and existing evaluation metrics fail to capture their task optimization efficiency, behavior adaptation after erroneous actions, and the specific utility of working memory for task completion. To address these gaps, we propose Test-time Improvement Diagnostic Evaluation (TIDE), an agent-agnostic and environment-agnostic framework that decomposes TTI into three comprehensive and interconnected dimensions. The framework measures (1) the overall temporal dynamics of task completion and (2) identifies whether performance is primarily constrained by recursive looping behaviors or (3) by burdensome accumulated memory. Through extensive experiments across diverse agents and environments, TIDE highlights that improving agent performance requires more than scaling internal reasoning, calling for explicitly optimizing the interaction dynamics between the agent and the environment.

Method: SafeGround uses a distribution-aware uncertainty quantification method to capture spatial dispersion of stochastic samples from model outputs, then calibrates a test-time decision threshold with statistically guaranteed false discovery rate (FDR) control.

Result: SafeGround’s uncertainty measure outperforms existing baselines in distinguishing correct from incorrect predictions, enables rigorous risk control, and improves system-level accuracy by up to 5.38% over Gemini-only inference across multiple GUI grounding models on the ScreenSpot-Pro benchmark.

Conclusion: SafeGround provides an effective uncertainty-aware framework for GUI grounding that enhances reliability through statistical risk control, addressing critical safety concerns in automated GUI interactions.

Abstract: Graphical User Interface (GUI) grounding aims to translate natural language instructions into executable screen coordinates, enabling automated GUI interaction. Nevertheless, incorrect grounding can result in costly, hard-to-reverse actions (e.g., erroneous payment approvals), raising concerns about model reliability. In this paper, we introduce SafeGround, an uncertainty-aware framework for GUI grounding models that enables risk-aware predictions through calibrations before testing. SafeGround leverages a distribution-aware uncertainty quantification method to capture the spatial dispersion of stochastic samples from outputs of any given model. Then, through the calibration process, SafeGround derives a test-time decision threshold with statistically guaranteed false discovery rate (FDR) control. We apply SafeGround on multiple GUI grounding models for the challenging ScreenSpot-Pro benchmark. Experimental results show that our uncertainty measure consistently outperforms existing baselines in distinguishing correct from incorrect predictions, while the calibrated threshold reliably enables rigorous risk control and potentials of substantial system-level accuracy improvements. Across multiple GUI grounding models, SafeGround improves system-level accuracy by up to 5.38% percentage points over Gemini-only inference.

Method: Proposes “Thinking with Comics” paradigm using comics as high information-density medium. Studies two reasoning paths based on comics and evaluates them on reasoning and long-context understanding tasks, comparing performance against image and video-based approaches.

Result: Thinking with Comics outperforms Thinking with Images on multi-step temporal and causal reasoning tasks while being substantially more efficient than Thinking with Video. Different comic narrative structures and styles consistently affect performance across tasks.

Conclusion: Comics serve as an effective intermediate visual representation that improves multimodal reasoning by balancing temporal structure preservation with computational efficiency, offering advantages over both static images and videos.

Abstract: Chain-of-Thought reasoning has driven large language models to extend from thinking with text to thinking with images and videos. However, different modalities still have clear limitations: static images struggle to represent temporal structure, while videos introduce substantial redundancy and computational cost. In this work, we propose Thinking with Comics, a visual reasoning paradigm that uses comics as a high information-density medium positioned between images and videos. Comics preserve temporal structure, embedded text, and narrative coherence while requiring significantly lower reasoning cost. We systematically study two reasoning paths based on comics and evaluate them on a range of reasoning tasks and long-context understanding tasks. Experimental results show that Thinking with Comics outperforms Thinking with Images on multi-step temporal and causal reasoning tasks, while remaining substantially more efficient than Thinking with Video. Further analysis indicates that different comic narrative structures and styles consistently affect performance across tasks, suggesting that comics serve as an effective intermediate visual representation for improving multimodal reasoning.

Method: Proposes VividVoice framework with two key components: 1) Vivid-210K dataset - a large-scale hybrid multimodal dataset created via programmatic pipeline establishing correlations between visual scenes, speaker identity, and audio; 2) D-MSVA alignment module - uses decoupled memory bank architecture and cross-modal hybrid supervision for fine-grained alignment from visual scenes to timbre and environmental acoustic features.

Result: Both subjective and objective experimental results show VividVoice significantly outperforms existing baseline models in audio fidelity, content clarity, and multimodal consistency. The framework demonstrates strong scene-aware speech synthesis capabilities.

Conclusion: VividVoice successfully addresses scene-aware visually-driven speech synthesis challenges through a unified generative framework with novel dataset construction and alignment mechanisms, enabling immersive auditory experiences aligned with visual scenes.

Abstract: We introduce and define a novel task-Scene-Aware Visually-Driven Speech Synthesis, aimed at addressing the limitations of existing speech generation models in creating immersive auditory experiences that align with the real physical world. To tackle the two core challenges of data scarcity and modality decoupling, we propose VividVoice, a unified generative framework. First, we constructed a large-scale, high-quality hybrid multimodal dataset, Vivid-210K, which, through an innovative programmatic pipeline, establishes a strong correlation between visual scenes, speaker identity, and audio for the first time. Second, we designed a core alignment module, D-MSVA, which leverages a decoupled memory bank architecture and a cross-modal hybrid supervision strategy to achieve fine-grained alignment from visual scenes to timbre and environmental acoustic features. Both subjective and objective experimental results provide strong evidence that VividVoice significantly outperforms existing baseline models in terms of audio fidelity, content clarity, and multimodal consistency. Our demo is available at https://chengyuann.github.io/VividVoice/.

Method: Introduce noise injection experiments where controlled noise signals of varying lengths are injected into musical contexts, analyzing loss curve shapes; test with MusicGen models in audio waveform domain.

Result: Music LLMs respond more strongly to local texture-level disruptions than global semantic corruption; loss curve shape (especially sharp “Peak” area for short noise injections) serves as proxy for musical integrity discrimination.

Conclusion: Loss curve shape encodes critical information about generated content quality; profile-based evaluation offers label-free, model-intrinsic framework for assessing musical quality, enabling better training objectives and benchmarks.

Abstract: The rise of music large language models (LLMs) demands robust methods of evaluating output quality, especially in distinguishing high-quality compositions from “garbage music”. Curiously, we observe that the standard cross-entropy loss – a core training metric – often decrease when models encounter systematically corrupted music, undermining its validity as a standalone quality indicator. To investigate this paradox, we introduce noise injection experiment, where controlled noise signal of varying lengths are injected into musical contexts. We hypothesize that a model’s loss reacting positively to these perturbations, specifically a sharp increase (“Peak” area) for short injection, can serve as a proxy for its ability to discern musical integrity. Experiments with MusicGen models in the audio waveform domain confirm that Music LLMs respond more strongly to local, texture-level disruptions than to global semantic corruption. Beyond exposing this bias, our results highlight a new principle: the shape of the loss curve – rather than its absolute value – encodes critical information about the quality of the generated content (i.e., model behavior). We envision this profile-based evaluation as a label-free, model-intrinsic framework for assessing musical quality – opening the door to more principled training objectives and sharper benchmarks.

Method: Used a baseline deep CNN trained on imbalanced respiratory sound data (73%:27%). Tested three generative augmentation strategies: variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models under controlled conditions.

Result: Baseline without augmentation: F1-score 0.645. Individual augmentation strategies showed no performance gains, with some configurations showing neutral or degraded performance. Only ensemble of augmented models achieved modest improvement to F1-score 0.664.

Conclusion: Synthetic augmentation may not consistently enhance medical audio classification with standard CNNs. Future work should focus on task-specific data characteristics, model-augmentation compatibility, and evaluation frameworks for effective synthetic augmentation.

Abstract: Medical audio classification remains challenging due to low signal-to-noise ratios, subtle discriminative features, and substantial intra-class variability, often compounded by class imbalance and limited training data. Synthetic data augmentation has been proposed as a potential strategy to mitigate these constraints; however, prior studies report inconsistent methodological approaches and mixed empirical results. In this preliminary study, we explore the impact of synthetic augmentation on respiratory sound classification using a baseline deep convolutional neural network trained on a moderately imbalanced dataset (73%:27%). Three generative augmentation strategies (variational autoencoders, generative adversarial networks, and diffusion models) were assessed under controlled experimental conditions. The baseline model without augmentation achieved an F1-score of 0.645. Across individual augmentation strategies, performance gains were not observed, with several configurations demonstrating neutral or degraded classification performance. Only an ensemble of augmented models yielded a modest improvement in F1-score (0.664). These findings suggest that, for medical audio classification, synthetic augmentation may not consistently enhance performance when applied to a standard CNN classifier. Future work should focus on delineating task-specific data characteristics, model-augmentation compatibility, and evaluation frameworks necessary for synthetic augmentation to be effective in medical audio applications.

Method: Proposes metadata-based captioning: (1) Train a metadata prediction model to infer detailed music metadata from audio, (2) At inference time, convert metadata into expressive captions using pre-trained LLMs. This decouples factual information extraction from stylistic caption generation.

Result: The method achieves comparable performance to end-to-end captioners in less training time, offers flexibility to change stylization post-training, and enables metadata imputation/in-filling by prompting with audio and partial metadata.

Conclusion: Metadata-based captioning provides an efficient and flexible alternative to end-to-end approaches, separating factual extraction from stylistic generation while enabling powerful metadata organization capabilities.

Abstract: Music captioning, or the task of generating a natural language description of music, is useful for both music understanding and controllable music generation. Training captioning models, however, typically requires high-quality music caption data which is scarce compared to metadata (e.g., genre, mood, etc.). As a result, it is common to use large language models (LLMs) to synthesize captions from metadata to generate training data for captioning models, though this process imposes a fixed stylization and entangles factual information with natural language style. As a more direct approach, we propose metadata-based captioning. We train a metadata prediction model to infer detailed music metadata from audio and then convert it into expressive captions via pre-trained LLMs at inference time. Compared to a strong end-to-end baseline trained on LLM-generated captions derived from metadata, our method: (1) achieves comparable performance in less training time over end-to-end captioners, (2) offers flexibility to easily change stylization post-training, enabling output captions to be tailored to specific stylistic and quality requirements, and (3) can be prompted with audio and partial metadata to enable powerful metadata imputation or in-filling–a common task for organizing music data.

Method: Proposes GRAM using multi-channel masked autoencoder to learn spatial audio representations, evaluated on two benchmark suites: NatHEAR (simulated naturalistic spatial environments) and RealSELD (real-world recordings).

Result: GRAM outperforms state-of-the-art self-supervised audio foundation models on NatHEAR and clean single-channel HEAR benchmarks using less training data, and shows SOTA localization performance in simulated environments with efficient generalization to real-world recordings.

Conclusion: GRAM represents a significant advance toward robust spatial audio foundation models for real-world environments, addressing key limitations of current models.

Abstract: Audio foundation models learn general-purpose audio representations that facilitate a wide range of downstream tasks. While the performance of these models has greatly increased for conventional single-channel, dry audio clips, their success in real-world acoustic environments with reverberation and noise is limited. Furthermore, most audio foundation models ignore the spatial dimension of real-world acoustic environments, ruling out tasks involving sound localization. To address these limitations, we propose GRAM: a general-purpose real-world audio model that employs a multi-channel masked autoencoder to efficiently learn spatial audio representations. We evaluated GRAM and other audio foundation models in a standardized manner on high-quality simulations of naturalistic, spatial acoustic environments as well as recordings of real-world environments and release these two complementary benchmark task suites: NatHEAR and RealSELD. Our results demonstrate that GRAM outperforms all state-of-the-art self-supervised audio foundation models on NatHEAR and the clean, single-channel version HEAR, while using only a fraction of the training data. GRAM also shows state-of-the-art localization performance in simulated environments and generalizes efficiently to real-world recordings in RealSELD. Taken together, GRAM presents a significant advance toward robust spatial audio foundation models for real-world environments.

Method: Proposes D3PIA, a discrete diffusion-based model using Neighborhood Attention (NA) to encode lead sheet conditions and predict note states in piano accompaniment. NA enables efficient local contextual modeling by attending to nearby melody and chord conditions in piano-roll representation.

Result: On POP909 dataset, D3PIA preserves chord conditions more faithfully than continuous diffusion and Transformer baselines. Subjective listening tests show D3PIA generates more musically coherent accompaniments.

Conclusion: Discrete diffusion with neighborhood attention effectively addresses piano accompaniment generation by improving local alignment between lead sheet constraints and generated music, outperforming existing methods in both objective and subjective evaluations.

Abstract: Generating piano accompaniments in the symbolic music domain is a challenging task that requires producing a complete piece of piano music from given melody and chord constraints, such as those provided by a lead sheet. In this paper, we propose a discrete diffusion-based piano accompaniment generation model, D3PIA, leveraging local alignment between lead sheet and accompaniment in piano-roll representation. D3PIA incorporates Neighborhood Attention (NA) to both encode the lead sheet and condition it for predicting note states in the piano accompaniment. This design enhances local contextual modeling by efficiently attending to nearby melody and chord conditions. We evaluate our model using the POP909 dataset, a widely used benchmark for piano accompaniment generation. Objective evaluation results demonstrate that D3PIA preserves chord conditions more faithfully compared to continuous diffusion-based and Transformer-based baselines. Furthermore, a subjective listening test indicates that D3PIA generates more musically coherent accompaniments than the comparison models.

Method: Proposes PACE method with: 1) regularized analytic classifier with first-session adaptation, 2) adaptive subspace-orthogonal PEFT for multi-session adaptation, and 3) spectrogram-based boundary-aware perturbations to mitigate representation overlap.

Result: Experiments on six diverse audio CL benchmarks show PACE substantially outperforms state-of-the-art baselines, addressing both coarse-grained (representation saturation) and fine-grained (representation drift) scenarios.

Conclusion: PACE marks an important step toward robust and scalable audio continual learning with pretrained models by addressing unique audio-specific challenges in representation learning and semantic alignment.

Abstract: Audio is a fundamental modality for analyzing speech, music, and environmental sounds. Although pretrained audio models have significantly advanced audio understanding, they remain fragile in real-world settings where data distributions shift over time. In this work, we present the first systematic benchmark for audio continual learning (CL) with pretrained models (PTMs), together with a comprehensive analysis of its unique challenges. Unlike in vision, where parameter-efficient fine-tuning (PEFT) has proven effective for CL, directly transferring such strategies to audio leads to poor performance. This stems from a fundamental property of audio backbones: they focus on low-level spectral details rather than structured semantics, causing severe upstream-downstream misalignment. Through extensive empirical study, we identify analytic classifiers with first-session adaptation (FSA) as a promising direction, but also reveal two major limitations: representation saturation in coarse-grained scenarios and representation drift in fine-grained scenarios. To address these challenges, we propose PACE, a novel method that enhances FSA via a regularized analytic classifier and enables multi-session adaptation through adaptive subspace-orthogonal PEFT for improved semantic alignment. In addition, we introduce spectrogram-based boundary-aware perturbations to mitigate representation overlap and improve stability. Experiments on six diverse audio CL benchmarks demonstrate that PACE substantially outperforms state-of-the-art baselines, marking an important step toward robust and scalable audio continual learning with PTMs.

Method: Introduces a quantitative, controllable activation steering framework for hybrid TTS models, using latent direction vectors to steer emotional expression. Includes systematic analysis of where steering should be applied within hybrid architectures and develops multi-rater evaluation protocols.

Result: Demonstrates that emotional prosody and expressive variability are primarily synthesized by the TTS language module rather than the flow-matching module. Provides a lightweight steering approach for generating natural, human-like emotional speech with composable mixed emotions.

Conclusion: Activation steering enables nuanced emotional control in TTS, allowing for complex mixed-emotion synthesis and text-emotion mismatch handling, with the language module playing a crucial role in emotional prosody generation.

Abstract: Emotional expression in human speech is nuanced and compositional, often involving multiple, sometimes conflicting, affective cues that may diverge from linguistic content. In contrast, most expressive text-to-speech systems enforce a single utterance-level emotion, collapsing affective diversity and suppressing mixed or text-emotion-misaligned expression. While activation steering via latent direction vectors offers a promising solution, it remains unclear whether emotion representations are linearly steerable in TTS, where steering should be applied within hybrid TTS architectures, and how such complex emotion behaviors should be evaluated. This paper presents the first systematic analysis of activation steering for emotional control in hybrid TTS models, introducing a quantitative, controllable steering framework, and multi-rater evaluation protocols that enable composable mixed-emotion synthesis and reliable text-emotion mismatch synthesis. Our results demonstrate, for the first time, that emotional prosody and expressive variability are primarily synthesized by the TTS language module instead of the flow-matching module, and also provide a lightweight steering approach for generating natural, human-like emotional speech.

Method: Uses LMS-based adaptive noise suppression (ANS) to attenuate ambient noise while preserving respiration-related acoustic components. The system operates fully on-device without neural networks or audio streaming, addressing energy and privacy constraints of wearables.

Result: Achieved robust performance with global MAE of 0.84 CPM (reduced to 0.47 CPM via automatic outlier rejection) in study with 18 participants under realistic acoustic conditions including music, cafeteria noise, and white noise up to 80 dB SPL. Operates with less than 2% processor load directly on earphones.

Conclusion: EarResp-ANS enables practical, real-time respiratory rate monitoring on commercial earphones by addressing noise robustness, computational efficiency, and privacy concerns through adaptive noise suppression without neural networks.

Abstract: Respiratory rate (RR) is a key vital sign for clinical assessment and mental well-being, yet it is rarely monitored in everyday life due to the lack of unobtrusive sensing technologies. In-ear audio sensing is promising due to its high social acceptance and the amplification of physiological sounds caused by the occlusion effect; however, existing approaches often fail under real-world noise or rely on computationally expensive models. We present EarResp-ANS, the first system enabling fully on-device, real-time RR estimation on commercial earphones. The system employs LMS-based adaptive noise suppression (ANS) to attenuate ambient noise while preserving respiration-related acoustic components, without requiring neural networks or audio streaming, thereby explicitly addressing the energy and privacy constraints of wearable devices. We evaluate EarResp-ANS in a study with 18 participants under realistic acoustic conditions, including music, cafeteria noise, and white noise up to 80 dB SPL. EarResp-ANS achieves robust performance with a global MAE of 0.84 CPM , reduced to 0.47 CPM via automatic outlier rejection, while operating with less than 2% processor load directly on the earphone.

Method: FINCH uses a log-linear fusion framework that integrates pre-trained audio classifiers with structured spatiotemporal predictors. It learns a per-sample gating function that estimates contextual reliability from uncertainty and informativeness statistics, bounding contextual influence and containing the audio-only classifier as a special case.

Result: FINCH outperforms fixed-weight fusion and audio-only baselines across benchmarks, achieving state-of-the-art on CBI and competitive/improved performance on BirdSet subsets. It improves robustness and error trade-offs even when contextual information is weak in isolation.

Conclusion: The framework provides a lightweight, interpretable, evidence-based approach for multimodal fusion that maintains risk containment through explicit audio-only fallback mechanisms, offering practical advantages for real-world bioacoustic classification systems.

Abstract: Many machine learning systems have access to multiple sources of evidence for the same prediction target, yet these sources often differ in reliability and informativeness across inputs. In bioacoustic classification, species identity may be inferred both from the acoustic signal and from spatiotemporal context such as location and season; while Bayesian inference motivates multiplicative evidence combination, in practice we typically only have access to discriminative predictors rather than calibrated generative models. We introduce \textbf{F}usion under \textbf{IN}dependent \textbf{C}onditional \textbf{H}ypotheses (\textbf{FINCH}), an adaptive log-linear evidence fusion framework that integrates a pre-trained audio classifier with a structured spatiotemporal predictor. FINCH learns a per-sample gating function that estimates the reliability of contextual information from uncertainty and informativeness statistics. The resulting fusion family \emph{contains} the audio-only classifier as a special case and explicitly bounds the influence of contextual evidence, yielding a risk-contained hypothesis class with an interpretable audio-only fallback. Across benchmarks, FINCH consistently outperforms fixed-weight fusion and audio-only baselines, improving robustness and error trade-offs even when contextual information is weak in isolation. We achieve state-of-the-art performance on CBI and competitive or improved performance on several subsets of BirdSet using a lightweight, interpretable, evidence-based approach. Code is available: \texttt{\href{https://anonymous.4open.science/r/birdnoise-85CD/README.md}{anonymous-repository}}

Method: Proposes VioPTT, a lightweight cascade model that jointly transcribes violin playing techniques with pitch onset/offset. Uses MOSA-VPT, a novel high-quality synthetic violin playing technique dataset to train the model without manual annotations.

Result: The model demonstrates strong generalization to real-world note-level violin technique recordings and achieves state-of-the-art transcription performance.

Conclusion: VioPTT is the first unified framework to combine violin transcription and playing technique prediction, addressing a significant gap in expressive music transcription.

Abstract: While automatic music transcription is well-established in music information retrieval, most models are limited to transcribing pitch and timing information from audio, and thus omit crucial expressive and instrument-specific nuances. One example is playing technique on the violin, which affords its distinct palette of timbres for maximal emotional impact. Here, we propose VioPTT (Violin Playing Technique-aware Transcription), a lightweight cascade model that directly transcribes violin playing technique in addition to pitch onset and offset. Furthermore, we release MOSA-VPT, a novel, high-quality synthetic violin playing technique dataset to circumvent the need for manually labeled annotations. Leveraging this dataset, our model demonstrated strong generalization to real-world note-level violin technique recordings in addition to achieving state-of-the-art transcription performance. To our knowledge, VioPTT is the first to jointly combine violin transcription and playing technique prediction within a unified framework.

Method: Proposes BELLE framework that shifts from deterministic prediction to Bayesian inference using Normal-Inverse-Gamma distributions to capture data-dependent aleatoric uncertainty. Introduces a “one-to-many” training strategy using synthetic samples as statistical support sets to enable accurate variance estimation on single-reference datasets without increasing model parameters or inference latency.

Result: BELLE trained on only ~5k hours of data outperforms leading open-source models trained on 50k hours, achieving a 25.8% relative WER reduction. The framework naturally supports high-quality streaming generation.

Conclusion: BELLE successfully bridges the gap between deterministic TTS approaches and the inherent uncertainty of speech generation, enabling more natural and robust speech synthesis through principled Bayesian modeling without computational overhead.

Abstract: Text-to-Speech (TTS) is inherently a “one-to-many” mapping characterized by intrinsic uncertainty, yet current paradigms often oversimplify it into a deterministic regression task. While continuous-valued autoregressive (AR) models have recently emerged as a promising alternative to discrete codec-based approaches, they typically rely on a fixed-variance prior, fundamentally constraining generation to a static point estimate that ignores the dynamic variability of natural speech. To bridge this gap, we propose BELLE (Bayesian evidential learning with language modelling), a framework that shifts from deterministic prediction to principled Bayesian inference without increasing model parameters or inference latency. By modeling the acoustic target as a Normal-Inverse-Gamma distribution, BELLE captures data-dependent aleatoric uncertainty. To enable accurate variance estimation on standard single-reference datasets, we introduce a “one-to-many” training strategy that leverages synthetic samples as a statistical support set, allowing the model to learn robust distributional properties rather than merely imitating teacher artifacts. Experiments demonstrate that BELLE, trained on only ~5k hours of data, outperforms leading open-source models trained on 50k hours (achieving a 25.8% relative WER reduction) and naturally supports high-quality streaming generation. Audio samples are available at https://belletts.github.io/Belle/.

Method: Systematically examined layer-wise representational alignment between 12 open-source Audio LLMs and EEG signals across 2 datasets. Used 8 similarity metrics including Spearman-based Representational Similarity Analysis (RSA) to characterize within-sentence representational geometry.

Result: Three key findings: (1) Rank-dependence split where model rankings vary substantially across different similarity metrics; (2) Spatio-temporal alignment patterns with depth-dependent alignment peaks and increased RSA in 250-500 ms window (N400-related); (3) Affective dissociation where negative prosody reduces geometric similarity but enhances covariance-based dependence.

Conclusion: The findings provide new neurobiological insights into the representational mechanisms of Audio LLMs, revealing complex alignment patterns with human neural dynamics that vary by metric, model depth, and emotional prosody.

Abstract: Audio Large Language Models (Audio LLMs) have demonstrated strong capabilities in integrating speech perception with language understanding. However, whether their internal representations align with human neural dynamics during naturalistic listening remains largely unexplored. In this work, we systematically examine layer-wise representational alignment between 12 open-source Audio LLMs and Electroencephalogram (EEG) signals across 2 datasets. Specifically, we employ 8 similarity metrics, such as Spearman-based Representational Similarity Analysis (RSA), to characterize within-sentence representational geometry. Our analysis reveals 3 key findings: (1) we observe a rank-dependence split, in which model rankings vary substantially across different similarity metrics; (2) we identify spatio-temporal alignment patterns characterized by depth-dependent alignment peaks and a pronounced increase in RSA within the 250-500 ms time window, consistent with N400-related neural dynamics; (3) we find an affective dissociation whereby negative prosody, identified using a proposed Tri-modal Neighborhood Consistency (TNC) criterion, reduces geometric similarity while enhancing covariance-based dependence. These findings provide new neurobiological insights into the representational mechanisms of Audio LLMs.

Method: Consolidates iterative structure into Unified Newton-Schulz Orthogonalization (UNSO) framework, avoids polynomial expansion, evaluates role of each matrix power, removes insignificant terms, and provides recommended polynomial with learnable coefficients that are optimized.

Result: Achieves outstanding performance with stable convergence while reducing computational burden compared to conventional NS iterations.

Conclusion: UNSO provides an efficient and stable alternative to conventional Newton-Schulz orthogonalization through a unified framework with optimized learnable coefficients.

Abstract: The Newton-Schulz (NS) iteration has gained increasing interest for its role in the Muon optimizer and the Stiefel manifold. However, the conventional NS iteration suffers from inefficiency and instability. Although various improvements have been introduced to NS iteration, they fail to deviate from the conventional iterative paradigm, which could increase computation burden largely due to the matrix products along the long dimension repeatedly. To address this, we consolidate the iterative structure into a unified framework, named Unified Newton-Schulz Orthogonalization (UNSO). To do so, we could avoid a polynomial expansion. Instead, we evaluate the role of each matrix power, remove the insignificant terms, and provide a recommended polynomial with learnable coefficients. These learnable coefficients are then optimized, and achieve an outstanding performance with stable convergence. The code of our method is available: https://github.com/greekinRoma/Unified_Newton_Schulz_Orthogonalization.

Method: Two main strategies: 1) Using LLMs as data-labeling tools to generate labels for unlabeled data, 2) Using LLMs as fallback mechanisms for low-confidence predictions. Combined with parameter-efficient fine-tuning techniques (compacters with layer freezing strategies) for pre-trained language models.

Result: Comprehensive experimental analysis on cybersecurity downstream tasks shows improved reliability and robustness of models, making them more suitable for real-world cybersecurity applications.

Conclusion: Combining parameter-efficient pre-trained models with large language models enhances model reliability and robustness for cybersecurity applications, addressing data drift and labeling challenges.

Abstract: Training AI models in cybersecurity with help of vast datasets offers significant opportunities to mimic real-world behaviors effectively. However, challenges like data drift and scarcity of labelled data lead to frequent updates of models and the risk of overfitting. To address these challenges, we used parameter-efficient fine-tuning techniques for pre-trained language models wherein we combine compacters with various layer freezing strategies. To enhance the capabilities of these pre-trained language models, in this work we introduce two strategies that use large language models. In the first strategy, we utilize large language models as data-labelling tools wherein they generate labels for unlabeled data. In the second strategy, large language modes are utilized as fallback mechanisms for predictions having low confidence scores. We perform comprehensive experimental analysis on the proposed strategies on different downstream tasks specific to cybersecurity domain. We empirically demonstrate that by combining parameter-efficient pre-trained models with large language models, we can improve the reliability and robustness of models, making them more suitable for real-world cybersecurity applications.

Method: Proposes Toxicity Association Graphs (TAGs) to model semantic associations between innocuous entities and latent toxic implications, introduces Multimodal Toxicity Covertness (MTC) metric, and creates Covert Toxic Dataset benchmark for evaluation.

Result: The approach outperforms existing methods across both low- and high-covertness toxicity regimes while providing interpretable and auditable detection outcomes.

Conclusion: The framework advances explainable multimodal toxicity detection and establishes foundations for future context-aware interpretable approaches in multimodal content analysis.

Abstract: Detecting toxicity in multimodal data remains a significant challenge, as harmful meanings often lurk beneath seemingly benign individual modalities: only emerging when modalities are combined and semantic associations are activated. To address this, we propose a novel detection framework based on Toxicity Association Graphs (TAGs), which systematically model semantic associations between innocuous entities and latent toxic implications. Leveraging TAGs, we introduce the first quantifiable metric for hidden toxicity, the Multimodal Toxicity Covertness (MTC), which measures the degree of concealment in toxic multimodal expressions. By integrating our detection framework with the MTC metric, our approach enables precise identification of covert toxicity while preserving full interpretability of the decision-making process, significantly enhancing transparency in multimodal toxicity detection. To validate our method, we construct the Covert Toxic Dataset, the first benchmark specifically designed to capture high-covertness toxic multimodal instances. This dataset encodes nuanced cross-modal associations and serves as a rigorous testbed for evaluating both the proposed metric and detection framework. Extensive experiments demonstrate that our approach outperforms existing methods across both low- and high-covertness toxicity regimes, while delivering clear, interpretable, and auditable detection outcomes. Together, our contributions advance the state of the art in explainable multimodal toxicity detection and lay the foundation for future context-aware and interpretable approaches. Content Warning: This paper contains examples of toxic multimodal content that may be offensive or disturbing to some readers. Reader discretion is advised.

Method: SAFM operates in two stages: 1) Decision stage determines whether to incorporate new adapter, reuse existing one, or add empty adapter, with architecture search prioritizing reuse or empty adapters; 2) Tuning stage uses layer-wise loss to encourage differentiation between adapters and capture task knowledge.

Result: SAFM outperforms state-of-the-art methods, achieving comparable performance while utilizing less than 60% of the parameters.

Conclusion: SAFM effectively addresses continual learning challenges by dynamically fusing adapters, minimizing parameter consumption while maximizing reuse and preventing catastrophic forgetting.

Abstract: Continual learning in natural language processing plays a crucial role in adapting to evolving data and preventing catastrophic forgetting. Despite significant progress, existing methods still face challenges, such as inefficient parameter reuse across tasks, risking catastrophic forgetting when tasks are dissimilar, and the unnecessary introduction of new parameters for each task, which hampers knowledge sharing among similar tasks. To tackle these issues, we propose a Sparse Adapter Fusion Method (SAFM), which dynamically fuses old and new adapters to address these challenges. SAFM operates in two stages: the decision stage and the tuning stage. In the decision stage, SAFM determines whether to incorporate a new adapter, reuse an existing one, or add an empty adapter. The architecture search procedure, designed to prioritize reusing or adding empty adapters, minimizes parameter consumption and maximizes reuse. In the tuning stage, SAFM especially facilitates a layer-wise loss to encourage differentiation between adapters, effectively capturing knowledge within the same task. Experimental results consistently show that SAFM outperforms state-of-the-art (SOTA) methods, achieving comparable performance while utilizing less than 60% of the parameters.

Method: ORDinal-aware imagE-tabulaR alignment (ORDER) - a multimodal pretraining framework that establishes ordinality as a core principle for composite material representations. It ensures materials with similar target properties occupy nearby regions in latent space, preserving the continuous nature of composite properties and enabling meaningful interpolation between sparsely observed designs.

Result: ORDER achieves consistent improvements over state-of-the-art multimodal baselines across property prediction, cross-modal retrieval, and microstructure generation tasks on both a public Nanofiber-enforced composite dataset and an internally curated carbon fiber T700 dataset.

Conclusion: ORDER successfully addresses the challenge of learning in continuous composite material design spaces under data scarcity by establishing ordinality as a core principle, enabling effective multimodal learning for composite materials where traditional graph-based approaches fail.

Abstract: Artificial intelligence (AI) has shown remarkable success in materials discovery and property prediction, particularly for crystalline and polymer systems where material properties and structures are dominated by discrete graph representations. Such graph-central paradigm breaks down on composite materials, which possess continuous and nonlinear design spaces that lack well-defined graph structures. General composite descriptors, e.g., fiber volume and misalignment angle, cannot fully capture the fiber distributions that fundamentally determine microstructural characteristics, necessitating the integration of heterogeneous data sources through multimodal learning. Existing alignment-oriented multimodal frameworks have proven effective on abundant crystal or polymer data under discrete, unique graph-property mapping assumptions, but fail to address the highly continuous composite design space under extreme data scarcity. In this work, we introduce ORDinal-aware imagE-tabulaR alignment (ORDER), a multimodal pretraining framework that establishes ordinality as a core principle for composite material representations. ORDER ensures that materials with similar target properties occupy nearby regions in the latent space, which effectively preserves the continuous nature of composite properties and enables meaningful interpolation between sparsely observed designs. We evaluate ORDER on a public Nanofiber-enforced composite dataset and an internally curated dataset that simulates the construction of carbon fiber T700 with diverse fiber distributions. ORDER achieves consistent improvements over state-of-the-art multimodal baselines across property prediction, cross-modal retrieval, and microstructure generation tasks.

Method: PRISMA-guided systematic review of 1,263 studies from PubMed, Google Scholar, and ACM Digital Library (2015-2024), identifying 29 eligible studies that applied statistical or machine learning models to predict length of stay for elective spine surgery patients.

Result: Machine learning models (logistic regression, random forest, boosting algorithms, neural networks) consistently outperformed traditional statistical models, with AUCs ranging from 0.94-0.99. K-Nearest Neighbors and Naive Bayes achieved top performance in some studies. Key predictors included age, comorbidities, BMI, surgery type/duration, and spinal levels.

Conclusion: Machine learning models show strong potential for length of stay prediction in elective spine surgery, but lack of standardization and external validation limits clinical utility. Future work needs standardized outcome definitions and transparent reporting.

Abstract: Objective: Predicting length of stay after elective spine surgery is essential for optimizing patient outcomes and hospital resource use. This systematic review synthesizes computational methods used to predict length of stay in this patient population, highlighting model performance and key predictors. Methods: Following PRISMA guidelines, we systematically searched PubMed, Google Scholar, and ACM Digital Library for studies published between December 1st, 2015, and December 1st, 2024. Eligible studies applied statistical or machine learning models to predict length of stay for elective spine surgery patients. Three reviewers independently screened studies and extracted data. Results: Out of 1,263 screened studies, 29 studies met inclusion criteria. Length of stay was predicted as a continuous, binary, or percentile-based outcome. Models included logistic regression, random forest, boosting algorithms, and neural networks. Machine learning models consistently outperformed traditional statistical models, with AUCs ranging from 0.94 to 0.99. K-Nearest Neighbors and Naive Bayes achieved top performance in some studies. Common predictors included age, comorbidities (notably hypertension and diabetes), BMI, type and duration of surgery, and number of spinal levels. However, external validation and reporting practices varied widely across studies. Discussion: There is growing interest in artificial intelligence and machine learning in length of stay prediction, but lack of standardization and external validation limits clinical utility. Future studies should prioritize standardized outcome definitions and transparent reporting needed to advance real-world deployment. Conclusion: Machine learning models offer strong potential for length of stay prediction after elective spine surgery, highlighting their potential for improving discharge planning and hospital resource management.

Method: Proposes ToolTok with: 1) Multi-step pathfinding where operations are sequences of progressive tool usage, 2) Tools represented as learnable token embeddings, 3) Semantic anchoring mechanism to ground tools with related concepts as inductive bias, 4) Easy-to-hard curriculum learning with three tasks (token definition QA, text-guided tool selection, simplified visual pathfinding).

Result: Achieves superior performance among 4B-scale models, competitive with 235B model, using <1% of training data required by other approaches. Shows strong generalization across unseen scenarios.

Conclusion: ToolTok provides an effective paradigm for GUI agents that addresses generalization and data scarcity issues through semantic anchoring and curriculum learning, enabling efficient learning with minimal supervision.

Abstract: Existing GUI agent models relying on coordinate-based one-step visual grounding struggle with generalizing to varying input resolutions and aspect ratios. Alternatives introduce coordinate-free strategies yet suffer from learning under severe data scarcity. To address the limitations, we propose ToolTok, a novel paradigm of multi-step pathfinding for GUI agents, where operations are modeled as a sequence of progressive tool usage. Specifically, we devise tools aligned with human interaction habits and represent each tool using learnable token embeddings. To enable efficient embedding learning under limited supervision, ToolTok introduces a semantic anchoring mechanism that grounds each tool with semantically related concepts as natural inductive bias. To further enable a pre-trained large language model to progressively acquire tool semantics, we construct an easy-to-hard curriculum consisting of three tasks: token definition question-answering, pure text-guided tool selection, and simplified visual pathfinding. Extensive experiments on multiple benchmarks show that ToolTok achieves superior performance among models of comparable scale (4B) and remains competitive with a substantially larger model (235B). Notably, these results are obtained using less than 1% of the training data required by other post-training approaches. In addition, ToolTok demonstrates strong generalization across unseen scenarios. Our training & inference code is open-source at https://github.com/ZephinueCode/ToolTok.

Method: Portable acoustic sensing captures subtle neck signals during swallowing tasks, combined with applied machine learning to identify patterns associated with abnormal physiological conditions.

Result: Achieves promising abnormality detection performance with AUC-ROC of 0.904 under 5 independent train-test splits, demonstrating feasibility of the approach.

Conclusion: Noninvasive acoustic sensing shows potential as practical, scalable tool for pharyngeal health monitoring and dysphagia detection.

Abstract: Pharyngeal health plays a vital role in essential human functions such as breathing, swallowing, and vocalization. Early detection of swallowing abnormalities, also known as dysphagia, is crucial for timely intervention. However, current diagnostic methods often rely on radiographic imaging or invasive procedures. In this study, we propose an automated framework for detecting dysphagia using portable and noninvasive acoustic sensing coupled with applied machine learning. By capturing subtle acoustic signals from the neck during swallowing tasks, we aim to identify patterns associated with abnormal physiological conditions. Our approach achieves promising test-time abnormality detection performance, with an AUC-ROC of 0.904 under 5 independent train-test splits. This work demonstrates the feasibility of using noninvasive acoustic sensing as a practical and scalable tool for pharyngeal health monitoring.

Method: Proposes GraphDancer, a reinforcement learning framework that teaches LLMs to navigate graphs by interleaving reasoning and function execution. Uses a graph-aware curriculum that schedules training by structural complexity of information-seeking trajectories with an easy-to-hard biased sampler.

Result: Despite using only a 3B backbone, GraphDancer outperforms baselines with 14B backbone or GPT-4o-mini, demonstrating robust cross-domain generalization of graph exploration and reasoning skills on multi-domain benchmark.

Conclusion: GraphDancer effectively enables LLMs to navigate heterogeneous knowledge graphs through RL training with structural curriculum, achieving strong performance and generalization with moderate-sized models.

Abstract: Large language models (LLMs) increasingly rely on external knowledge to improve factuality, yet many real-world knowledge sources are organized as heterogeneous graphs rather than plain text. Reasoning over such graph-structured knowledge poses two key challenges: (1) navigating structured, schema-defined relations requires precise function calls rather than similarity-based retrieval, and (2) answering complex questions often demands multi-hop evidence aggregation through iterative information seeking. We propose GraphDancer, a reinforcement learning (RL) framework that teaches LLMs to navigate graphs by interleaving reasoning and function execution. To make RL effective for moderate-sized LLMs, we introduce a graph-aware curriculum that schedules training by the structural complexity of information-seeking trajectories using an easy-to-hard biased sampler. We evaluate GraphDancer on a multi-domain benchmark by training on one domain only and testing on unseen domains and out-of-distribution question types. Despite using only a 3B backbone, GraphDancer outperforms baselines equipped with either a 14B backbone or GPT-4o-mini, demonstrating robust cross-domain generalization of graph exploration and reasoning skills. Our code and models can be found at https://yuyangbai.com/graphdancer/ .

Method: AURA coordinates multiple AI agents to generate and validate labels without requiring ground truth. It adapts a classical probabilistic model that jointly infers latent true labels and annotator reliability via confusion matrices, using Expectation-Maximization to reconcile conflicting annotations and aggregate noisy predictions.

Result: Across four benchmark datasets, AURA achieves accuracy improvements of up to 5.8% over baseline. In challenging settings with poor quality annotators, the improvement reaches up to 50% over baseline. AURA also accurately estimates annotator reliability without requiring pre-validation steps.

Conclusion: AURA provides an effective agentic AI framework for large-scale, multimodal data annotation that improves annotation accuracy while simultaneously estimating annotator reliability, addressing key challenges in modern supervised learning pipelines.

Abstract: Data annotation is essential for supervised learning, yet producing accurate, unbiased, and scalable labels remains challenging as datasets grow in size and modality. Traditional human-centric pipelines are costly, slow, and prone to annotator variability, motivating reliability-aware automated annotation. We present AURA (Agentic AI for Unified Reliability Modeling and Annotation Aggregation), an agentic AI framework for large-scale, multi-modal data annotation. AURA coordinates multiple AI agents to generate and validate labels without requiring ground truth. At its core, AURA adapts a classical probabilistic model that jointly infers latent true labels and annotator reliability via confusion matrices, using Expectation-Maximization to reconcile conflicting annotations and aggregate noisy predictions. Across the four benchmark datasets evaluated, AURA achieves accuracy improvements of up to 5.8% over baseline. In more challenging settings with poor quality annotators, the improvement is up to 50% over baseline. AURA also accurately estimates the reliability of annotators, allowing assessment of annotator quality even without any pre-validation steps.

Method: Reformulate SDPA in a mathematically equivalent form as projection of input vectors onto a common surface determined by the inputs themselves, revealing time-dependent and context-dependent nonlinear dependencies.

Result: The rewritten form enables increased speed for both feedforward and learning algorithms, suggests potential extensions, and provides a new interpretation of SDPA’s role in language processing as finding time-dependent contextual meaning.

Conclusion: SDPA discovers nonlinear dependencies in input data that are time-dependent and context-dependent, providing strong justification for its use in time-series data with time-varying local nonlinear dependencies, with implications for both computational efficiency and theoretical understanding.

Abstract: Scaled dot-product attention (SDPA) is a fundamental component responsible for the success of large-language models and other nonlinear signal processing applications. The rationale for SDPA has been based upon “query, key, value” concepts borrowed from database theory, but these concepts are difficult to reconcile with standard methods in mathematical signal processing. We show that SDPA can be rewritten in a different but mathematically equivalent form as a projection of the input vectors onto a common surface determined by the inputs themselves. Therefore SDPA discovers nonlinear dependencies in the input that are time-dependent and context-dependent. The rewritten form of SDPA permits increased speed of both feedforward and learning algorithms, but more importantly suggests potential extensions. In the context of language, we re-interpret the role of SDPA as finding a time-dependent contextual meaning determined by the surface on which the set of input vectors lies. Input token embeddings are then modified by the local context surface. This interpretation differs substantially from the concept of “self-attention”, and provides a strong justification for the use of SDPA for time-series data with time-varying local nonlinear dependencies.

Method: Proposes IceBench-S2S benchmark with a generalized framework that compresses spatial features of daily sea ice data into a deep latent space, then uses DL-based forecasting backbones to model temporally concatenated deep features for S2S-scale predictions (180-day periods).

Result: Provides a unified training and evaluation pipeline for different DL backbones, along with practical guidance for model selection in polar environmental monitoring tasks.

Conclusion: IceBench-S2S bridges the gap between current DL forecasting lead times and operational S2S scale needs, enabling better evaluation and development of sea ice forecasting models for real-world applications.

Abstract: Arctic sea ice plays a critical role in regulating Earth’s climate system, significantly influencing polar ecological stability and human activities in coastal regions. Recent advances in artificial intelligence have facilitated the development of skillful pan-Arctic sea ice forecasting systems, where data-driven approaches showcase tremendous potential to outperform conventional physics-based numerical models in terms of accuracy, computational efficiency and forecasting lead times. Despite the latest progress made by deep learning (DL) forecasting models, most of their skillful forecasting lead times are confined to daily subseasonal scale and monthly averaged values for up to six months, which drastically hinders their deployment for real-world applications, e.g., maritime routine planning for Arctic transportation and scientific investigation. Extending daily forecasts from subseasonal to seasonal (S2S) scale is scientifically crucial for operational applications. To bridge the gap between the forecasting lead time of current DL models and the significant daily S2S scale, we introduce IceBench-S2S, the first comprehensive benchmark for evaluating DL approaches in mitigating the challenge of forecasting Arctic sea ice concentration in successive 180-day periods. It proposes a generalized framework that first compresses spatial features of daily sea ice data into a deep latent space. The temporally concatenated deep features are subsequently modeled by DL-based forecasting backbones to predict the sea ice variation at S2S scale. IceBench-S2S provides a unified training and evaluation pipeline for different backbones, along with practical guidance for model selection in polar environmental monitoring tasks.

Method: Combines architectural interventions (QK-norm attention, per-head gating, value residuals, LayerNorm scaling) with optimization advances (NorMuon with cautious weight decay, muP parametrization) and a three-stage training schedule with post-hoc checkpoint EMA, trained on 72B tokens.

Result: IMU-1 approaches benchmark performance of models trained on 56x more data (approximately 4 trillion tokens), demonstrating significant efficiency improvements in language model training.

Conclusion: The paper presents a validated training recipe that enables efficient language model development with reduced data requirements, providing code, weights, and data for reproduction to advance open research in language modeling.

Abstract: We present IMU-1, a 430M-parameter language model trained on 72B tokens that approaches the benchmark performance of models trained on 56x more data. We describe a validated training recipe combining recent architectural interventions (QK-norm attention, per-head gating, value residuals, LayerNorm scaling) with optimization advances (NorMuon with cautious weight decay, muP parametrization) and a three-stage training schedule with post-hoc checkpoint EMA. We provide ablations for each component and release code, weights and data to enable reproduction: https://huggingface.co/thepowerfuldeez/imu1_base

Method: Analyzes connection between modality alignment and cross-modality jailbreak transfer, then empirically evaluates textual jailbreaks, text-transferred audio jailbreaks, and existing audio-based jailbreaks on recent omni-models.

Result: Text-transferred audio jailbreaks perform comparably to or better than audio-based jailbreaks, show strong cross-model transferability, and remain effective under audio-only access threat models.

Conclusion: Text-transferred audio jailbreaks serve as powerful baselines for audio red-teaming, revealing vulnerabilities in multimodal alignment that can propagate textual weaknesses to audio.

Abstract: Recent advances in end-to-end trained omni-models have significantly improved multimodal understanding. At the same time, safety red-teaming has expanded beyond text to encompass audio-based jailbreak attacks. However, an important bridge between textual and audio jailbreaks remains underexplored. In this work, we study the cross-modality transfer of jailbreak attacks from text to audio, motivated by the semantic similarity between the two modalities and the maturity of textual jailbreak methods. We first analyze the connection between modality alignment and cross-modality jailbreak transfer, showing that strong alignment can inadvertently propagate textual vulnerabilities to the audio modality, which we term the alignment curse. Guided by this analysis, we conduct an empirical evaluation of textual jailbreaks, text-transferred audio jailbreaks, and existing audio-based jailbreaks on recent omni-models. Our results show that text-transferred audio jailbreaks perform comparably to, and often better than, audio-based jailbreaks, establishing them as simple yet powerful baselines for future audio red-teaming. We further demonstrate strong cross-model transferability and show that text-transferred audio attacks remain effective even under a stricter audio-only access threat model.

Method: Created TabularMath benchmark with 114 deterministic problems (233,472 rows) from verified programs based on GSM8K and AIME. Evaluated 9 tabular architectures and in-context learning with GPT-OSS-120B using both regression metrics (R²) and exact integer match accuracy.

Result: TabPFN v2.5 achieved R²=0.998 in-distribution and maintained positive R² under distribution shift, but dropped below 10% exact-match accuracy on out-of-distribution data. ICL maintained around 40% exact-match accuracy, showing complementary strengths: TabPFN scales efficiently with data while ICL achieves exact computation from few examples.

Conclusion: Tabular models learn smooth function approximations but struggle with precise computational outputs under extrapolation. The gap between regression metrics and exact-match accuracy reveals limitations in current tabular models for deterministic computational tasks.

Abstract: Standard tabular benchmarks mainly focus on the evaluation of a model’s capability to interpolate values inside a data manifold, where models good at performing local statistical smoothing are rewarded. However, there exists a very large category of high-value tabular data, including financial modeling and physical simulations, which are generated based upon deterministic computational processes, as opposed to stochastic and noisy relationships. Therefore, we investigate if tabular models can provide an extension from statistical interpolation to computational extrapolation. We propose TabularMath, a diagnostic benchmark of 114 deterministic problems (233,472 rows) generated from verified programs based on GSM8K and AIME. We evaluate 9 tabular architectures and in-context learning (ICL) with GPT-OSS-120B. On standard regression metrics, TabPFN v2.5 performs remarkably well, achieving R^2=0.998 in-distribution and maintaining positive R^2 even under distribution shift, which is unique among the tabular models we tested. When we measure rounded consistency (exact integer match), a different picture emerges: TabPFN v2.5 drops below 10% on out-of-distribution data, while ICL maintains around 40%. This gap between R^2 and exact-match accuracy suggests that tabular models learn smooth function approximations but struggle to recover precise computational outputs under extrapolation. The two paradigms appear complementary: TabPFN scales efficiently with data; ICL achieves exact computation from few examples. We release all code and data to support further investigation.

Method: Used sliding-window protocol (N=1500) to identify “Semantic Tunneling” failure mode. Applied Multi-Scale Negative Coupled Information Systems (MNCIS) framework with Adaptive Spectral Negative Coupling (ASNC) as topological operator to induce “Manifold Unfolding.”

Result: Baseline model converged to “Robert Boulton” Singularity within 7 generations despite high PPL (~83.9). MNCIS expanded effective rank from 3.62 to 5.35, preventing semantic collapse and preserving long-tail data distribution.

Conclusion: PPL is insufficient for monitoring generative AI stability; semantic diversity collapse requires new metrics. MNCIS framework successfully prevents semantic tunneling by inducing manifold unfolding and preserving semantic diversity.

Abstract: The stability of generative artificial intelligence trained on recursive synthetic data is conventionally monitored via Perplexity (PPL). We demonstrate that PPL is a deceptive metric in context-stabilized regimes (L=128). Using a rigorous sliding-window protocol (N=1500), we identify a novel failure mode termed “Semantic Tunneling.” While the Baseline model maintains high grammatical fluency (PPL approx. 83.9), it suffers a catastrophic loss of semantic diversity, converging within seven generations to a single, low-entropy narrative attractor: the “Robert Boulton” Singularity. This phenomenon represents a total collapse of the latent manifold (Global Effective Rank 3.62 -> 2.22), where the model discards diverse world knowledge to optimize for statistically safe syntactic templates. To address this, we apply the Multi-Scale Negative Coupled Information Systems (MNCIS) framework recently established in Hou (2026) [arXiv:2601.11594]. We demonstrate that Adaptive Spectral Negative Coupling (ASNC) acts as a topological operator that actively induces “Manifold Unfolding.” MNCIS forces the model to expand its effective rank from the anisotropic baseline of 3.62 to a hyper-diverse state of 5.35, effectively constructing an “Artificial Manifold” that resists the gravitational pull of semantic attractors and preserves the long-tail distribution of the training data.

Method: Proposes Incident-Guided Spatiotemporal Graph Neural Network (IGSTGNN) with two core components: Incident-Context Spatial Fusion (ICSF) module to capture initial heterogeneous spatial influence of incidents, and Temporal Incident Impact Decay (TIID) module to model dynamic dissipation of incident effects over time.

Result: IGSTGNN achieves state-of-the-art performance on a newly constructed large-scale dataset with time-aligned incident records. The ICSF and TIID modules also demonstrate generalizability when integrated into various existing models.

Conclusion: Explicitly modeling incident impacts through specialized spatial and temporal modules significantly improves traffic forecasting accuracy and provides a valuable framework for handling external disturbances in spatio-temporal prediction tasks.

Abstract: Recent years have witnessed the rapid development of deep-learning-based, graph-neural-network-based forecasting methods for modern intelligent transportation systems. However, most existing work focuses exclusively on capturing spatio-temporal dependencies from historical traffic data, while overlooking the fact that suddenly occurring transportation incidents, such as traffic accidents and adverse weather, serve as external disturbances that can substantially alter temporal patterns. We argue that this issue has become a major obstacle to modeling the dynamics of traffic systems and improving prediction accuracy, but the unpredictability of incidents makes it difficult to observe patterns from historical sequences. To address these challenges, this paper proposes a novel framework named the Incident-Guided Spatiotemporal Graph Neural Network (IGSTGNN). IGSTGNN explicitly models the incident’s impact through two core components: an Incident-Context Spatial Fusion (ICSF) module to capture the initial heterogeneous spatial influence, and a Temporal Incident Impact Decay (TIID) module to model the subsequent dynamic dissipation. To facilitate research on the spatio-temporal impact of incidents on traffic flow, a large-scale dataset is constructed and released, featuring incident records that are time-aligned with traffic time series. On this new benchmark, the proposed IGSTGNN framework is demonstrated to achieve state-of-the-art performance. Furthermore, the generalizability of the ICSF and TIID modules is validated by integrating them into various existing models.

Method: Proposes an RL framework that leverages off-policy evaluation (OPE) to systematically evaluate multiple candidate state representations and reward functions using only logged interaction data, training offline RL agents and applying OPE to estimate policy performance.

Result: Validated on two environments: OpenAI Gym’s Lunar Lander (controlled setting) and NASA-MATB-II human subjects study (real-world human-robot teaming), demonstrating feasibility for automating RL design decisions.

Conclusion: The approach enhances feasibility and scalability of offline RL for real-world environments by automating critical RL design decisions through data-driven OPE-based evaluation, enabling more reliable RL formulation for complex human-robot interaction.

Abstract: Reinforcement learning (RL) has the potential to transform real-world decision-making systems by enabling autonomous agents to learn from experience. Deploying RL in real-world settings, especially in the context of human-robot interaction, requires defining state representations and reward functions, which are critical for learning efficiency and policy performance. Traditional RL approaches often rely on domain expertise and trial-and-error, necessitating extensive human involvement as well as direct interaction with the environment, which can be costly and impractical, especially in complex and safety-critical applications. This work proposes a novel RL framework that leverages off-policy evaluation (OPE) for state space and reward function selection, using only logged interaction data. This approach eliminates the need for real-time access to the environment or human-in-the-loop feedback, greatly reducing the dependency on costly real-time interactions. The proposed approach systematically evaluates multiple candidate state representations and reward functions by training offline RL agents and applying OPE to estimate policy performance. The optimal state space and reward function are selected based on their ability to produce high-performing policies under OPE metrics. Our method is validated on two environments: the Lunar Lander environment by OpenAI Gym, which provides a controlled setting for assessing state space and reward function selection, and a NASA-MATB-II human subjects study environment, which evaluates the approach’s real-world applicability to human-robot teaming scenarios. This work enhances the feasibility and scalability of offline RL for real-world environments by automating critical RL design decisions through a data-driven OPE-based evaluation, enabling more reliable, effective, and sustainable RL formulation for complex human-robot interaction settings.

Method: Deep reinforcement learning-based active flow control strategy using high-fidelity CFD simulations with adaptive mesh refinement to resolve shock motion, boundary-layer dynamics, and flow separation for learning physically consistent control policies.

Result: DRL controller robustly stabilizes inlet over wide back pressure ranges, generalizes to unseen scenarios (different back-pressure levels, Reynolds numbers, sensor configurations), operates with noisy measurements, and achieves comparable performance with minimal sensor sets.

Conclusion: Establishes data-driven approach for real-time hypersonic flow control under realistic operational uncertainties with strong zero-shot generalization capabilities.

Abstract: The hypersonic unstart phenomenon poses a major challenge to reliable air-breathing propulsion at Mach 5 and above, where strong shock-boundary-layer interactions and rapid pressure fluctuations can destabilize inlet operation. Here, we demonstrate a deep reinforcement learning (DRL)- based active flow control strategy to control unstart in a canonical two-dimensional hypersonic inlet at Mach 5 and Reynolds number $5\times 10^6$. The in-house CFD solver enables high-fidelity simulations with adaptive mesh refinement, resolving key flow features, including shock motion, boundary-layer dynamics, and flow separation, that are essential for learning physically consistent control policies suitable for real-time deployment. The DRL controller robustly stabilizes the inlet over a wide range of back pressures representative of varying combustion chamber conditions. It further generalizes to previously unseen scenarios, including different back-pressure levels, Reynolds numbers, and sensor configurations, while operating with noisy measurements, thereby demonstrating strong zero-shot generalization. Control remains robust in the presence of noisy sensor measurements, and a minimal, optimally selected sensor set achieves comparable performance, enabling practical implementation. These results establish a data-driven approach for real-time hypersonic flow control under realistic operational uncertainties.

Method: CADENT unifies strategic automaton-based knowledge with tactical policy-level knowledge using an experience-gated trust mechanism that dynamically weighs teacher guidance against student’s own experience at state-action level.

Result: Achieves 40-60% better sample efficiency than baselines across challenging environments (sparse-reward grid worlds to continuous control tasks) while maintaining superior asymptotic performance.

Conclusion: CADENT establishes a robust approach for adaptive knowledge transfer in RL by combining strategic and tactical knowledge with dynamic trust mechanisms.

Abstract: Transfer learning promises to reduce the high sample complexity of deep reinforcement learning (RL), yet existing methods struggle with domain shift between source and target environments. Policy distillation provides powerful tactical guidance but fails to transfer long-term strategic knowledge, while automaton-based methods capture task structure but lack fine-grained action guidance. This paper introduces Context-Aware Distillation with Experience-gated Transfer (CADENT), a framework that unifies strategic automaton-based knowledge with tactical policy-level knowledge into a coherent guidance signal. CADENT’s key innovation is an experience-gated trust mechanism that dynamically weighs teacher guidance against the student’s own experience at the state-action level, enabling graceful adaptation to target domain specifics. Across challenging environments, from sparse-reward grid worlds to continuous control tasks, CADENT achieves 40-60% better sample efficiency than baselines while maintaining superior asymptotic performance, establishing a robust approach for adaptive knowledge transfer in RL.

Method: HyExDNN-RNN model combining DNN and RNN architectures with Pearson correlation for feature selection, SHAP and PFI for explainability, and standardized techniques for feature reduction and model interpretation.

Result: Achieved 99% F1 score on binary classification and 94.2% on multi-class categorization. XAI approaches provided important insights into feature importance and model decision logic.

Conclusion: The framework demonstrates potential for assisting ADHD diagnosis with interpretability, bridging computational techniques with psychological applications.

Abstract: Attention Deficit Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder that is challenging to diagnose and requires advanced approaches for reliable and transparent identification and classification. It is characterized by a pattern of inattention, hyperactivity and impulsivity that is more severe and more frequent than in individuals with a comparable level of development. In this paper, an explainable framework based on a fine-tuned hybrid Deep Neural Network (DNN) and Recurrent Neural Network (RNN) called HyExDNN-RNN model is proposed for ADHD detection, multi-class categorization, and decision interpretation. This framework not only detects ADHD, but also provides interpretable insights into the diagnostic process so that psychologists can better understand and trust the results of the diagnosis. We use the Pearson correlation coefficient for optimal feature selection and machine and deep learning models for experimental analysis and comparison. We use a standardized technique for feature reduction, model selection and interpretation to accurately determine the diagnosis rate and ensure the interpretability of the proposed framework. Our framework provided excellent results on binary classification, with HyExDNN-RNN achieving an F1 score of 99% and 94.2% on multi-class categorization. XAI approaches, in particular SHapley Additive exPlanations (SHAP) and Permutation Feature Importance (PFI), provided important insights into the importance of features and the decision logic of models. By combining AI with human expertise, we aim to bridge the gap between advanced computational techniques and practical psychological applications. These results demonstrate the potential of our framework to assist in ADHD diagnosis and interpretation.

Method: Proposes UniMod with structured reasoning trajectories (evidence grounding, modality assessment, risk mapping, policy decision, response generation) and UniRM multi-head scalar reward model for multi-dimensional supervision. Uses specialized optimization strategies to decouple task parameters.

Result: Achieves competitive textual moderation performance and sets new multimodal benchmark using <40% of training data compared to leading baselines. Ablations validate multi-attribute trajectory reasoning effectiveness.

Conclusion: UniMod provides an effective and efficient framework for multimodal moderation by forcing explicit safety semantics grounding and preventing shortcut learning through dense reasoning traces.

Abstract: Safety moderation is pivotal for identifying harmful content. Despite the success of textual safety moderation, its multimodal counterparts remain hindered by a dual sparsity of data and supervision. Conventional reliance on binary labels lead to shortcut learning, which obscures the intrinsic classification boundaries necessary for effective multimodal discrimination. Hence, we propose a novel learning paradigm (UniMod) that transitions from sparse decision-making to dense reasoning traces. By constructing structured trajectories encompassing evidence grounding, modality assessment, risk mapping, policy decision, and response generation, we reformulate monolithic decision tasks into a multi-dimensional boundary learning process. This approach forces the model to ground its decision in explicit safety semantics, preventing the model from converging on superficial shortcuts. To facilitate this paradigm, we develop a multi-head scalar reward model (UniRM). UniRM provides multi-dimensional supervision by assigning attribute-level scores to the response generation stage. Furthermore, we introduce specialized optimization strategies to decouple task-specific parameters and rebalance training dynamics, effectively resolving interference between diverse objectives in multi-task learning. Empirical results show UniMod achieves competitive textual moderation performance and sets a new multimodal benchmark using less than 40% of the training data used by leading baselines. Ablations further validate our multi-attribute trajectory reasoning, offering an effective and efficient framework for multimodal moderation. Supplementary materials are available at \href{https://trustworthylab.github.io/UniMod/}{project website}.

Method: Proposes FAQ which leverages future-layer activations to guide quantization, uses window-wise preview to aggregate multiple future layers, and employs pre-searched configurations to avoid expensive search.

Result: FAQ consistently outperforms prior methods with negligible extra cost, requires no backward passes, data reconstruction, or tuning, making it suitable for edge deployment.

Conclusion: Future-aware quantization effectively addresses PTQ limitations by using future information to guide current-layer quantization, improving stability and performance without significant overhead.

Abstract: Post-training quantization (PTQ) is a widely used method to compress large language models (LLMs) without fine-tuning. It typically sets quantization hyperparameters (e.g., scaling factors) based on current-layer activations. Although this method is efficient, it suffers from quantization bias and error accumulation, resulting in suboptimal and unstable quantization, especially when the calibration data is biased. To overcome these issues, we propose Future-Aware Quantization (FAQ), which leverages future-layer activations to guide quantization. This allows better identification and preservation of important weights, while reducing sensitivity to calibration noise. We further introduce a window-wise preview mechanism to softly aggregate multiple future-layer activations, mitigating over-reliance on any single layer. To avoid expensive greedy search, we use a pre-searched configuration to minimize overhead. Experiments show that FAQ consistently outperforms prior methods with negligible extra cost, requiring no backward passes, data reconstruction, or tuning, making it well-suited for edge deployment.

Method: Conducted controlled stress tests by progressively increasing information quantity (character count) within images, analyzed phase-transition phenomena, identified key factors, and formulated a probabilistic scaling law unifying average vision token load and visual density.

Result: Discovered three distinct regimes: Stable Phase (near-perfect), Instability Phase (increased error variance), and Collapse Phase (total failure). The scaling law demonstrates universality across various Vision-Language Models.

Conclusion: Provides critical empirical guidance for optimizing efficiency-accuracy trade-off in visual context compression and establishes fundamental understanding of vision token capacity limits.

Abstract: Recent vision-centric approaches have made significant strides in long-context modeling. Represented by DeepSeek-OCR, these models encode rendered text into continuous vision tokens, achieving high compression rates without sacrificing recognition precision. However, viewing the vision encoder as a lossy channel with finite representational capacity raises a fundamental question: what is the information upper bound of visual tokens? To investigate this limit, we conduct controlled stress tests by progressively increasing the information quantity (character count) within an image. We observe a distinct phase-transition phenomenon characterized by three regimes: a near-perfect Stable Phase, an Instability Phase marked by increased error variance, and a total Collapse Phase. We analyze the mechanical origins of these transitions and identify key factors. Furthermore, we formulate a probabilistic scaling law that unifies average vision token load and visual density into a latent difficulty metric. Extensive experiments across various Vision-Language Models demonstrate the universality of this scaling law, providing critical empirical guidance for optimizing the efficiency-accuracy trade-off in visual context compression.

Method: Proposes AutoCL with Siamese network architecture sharing backbone parameters, embedding a generator to learn auto-augmentation. Uses latent space representations to train generator, overcoming noise and redundancy in raw sensor data. Includes stop-gradient design and correlation reduction strategy to enhance encoder representation learning.

Result: Extensive experiments on four widely-used HAR datasets demonstrate that AutoCL significantly improves recognition accuracy compared with other state-of-the-art methods.

Conclusion: AutoCL effectively automates augmentation in contrastive learning for HAR, reducing manual effort while improving performance through learned augmentation strategies and enhanced representation learning techniques.

Abstract: For low-semantic sensor signals from human activity recognition (HAR), contrastive learning (CL) is essential to implement novel applications or generic models without manual annotation, which is a high-performance self-supervised learning (SSL) method. However, CL relies heavily on data augmentation for pairwise comparisons. Especially for low semantic data in the HAR area, conducting good performance augmentation strategies in pretext tasks still rely on manual attempts lacking generalizability and flexibility. To reduce the augmentation burden, we propose an end-to-end auto-augmentation contrastive learning (AutoCL) method for wearable-based HAR. AutoCL is based on a Siamese network architecture that shares the parameters of the backbone and with a generator embedded to learn auto-augmentation. AutoCL trains the generator based on the representation in the latent space to overcome the disturbances caused by noise and redundant information in raw sensor data. The architecture empirical study indicates the effectiveness of this design. Furthermore, we propose a stop-gradient design and correlation reduction strategy in AutoCL to enhance encoder representation learning. Extensive experiments based on four wide-used HAR datasets demonstrate that the proposed AutoCL method significantly improves recognition accuracy compared with other SOTA methods.

Method: Identifies correlation between collapse and explosive MLP weight norm growth, proves L&E updates cause exponential norm growth, and proposes NAS - a plug-and-play norm-constrained strategy that scales edited weights to maintain norm stability

Result: NAS delays collapse point by >4x, yields 72.2% average relative gain in editing performance, requires minimal code changes and negligible computational overhead

Conclusion: Norm control is crucial for stable sequential model editing, and NAS provides an effective, lightweight solution to prevent model collapse in L&E frameworks

Abstract: Model editing has emerged as a practical approach for mitigating factual errors and outdated knowledge in large language models (LLMs). Among existing methods, the Locate-and-Edit (L&E) paradigm is the dominant framework: it locates MLP parameters implicated in expressing a target fact, and then performs a localized update to rewrite that fact. However, long sequences of edits often trigger abrupt model collapse in L&E beyond a critical point. We empirically identify a strong correlation between collapse and explosive growth of edited MLP weight norms, and formally prove that commonly used L&E update rules can induce exponential norm growth across sequential edits in the absence of explicit norm control. To address this issue, we propose Norm-Anchor Scaling NAS, a plug-and-play norm-constrained strategy. Across extensive experiments, NAS delays the collapse point of representative L&E algorithms by more than 4 times and yields a 72.2% average relative gain in editing performance, requiring only a single additional line of code and incurring negligible computational overhead.

Method: SPA-Cache jointly optimizes update identification and budget allocation. First, it uses a low-dimensional singular proxy to identify update-critical tokens in a subspace, reducing identification overhead. Second, it introduces an adaptive strategy that allocates fewer updates to stable layers without degrading generation quality.

Result: The method achieves up to 8× throughput improvement over vanilla decoding and 2-4× speedup over existing caching baselines.

Conclusion: SPA-Cache significantly improves DLM efficiency through optimized caching strategies, addressing key limitations of existing approaches for non-causal language models.

Abstract: While Diffusion Language Models (DLMs) offer a flexible, arbitrary-order alternative to the autoregressive paradigm, their non-causal nature precludes standard KV caching, forcing costly hidden state recomputation at every decoding step. Existing DLM caching approaches reduce this cost by selective hidden state updates; however, they are still limited by (i) costly token-wise update identification heuristics and (ii) rigid, uniform budget allocation that fails to account for heterogeneous hidden state dynamics. To address these challenges, we present SPA-Cache that jointly optimizes update identification and budget allocation in DLM cache. First, we derive a low-dimensional singular proxy that enables the identification of update-critical tokens in a low-dimensional subspace, substantially reducing the overhead of update identification. Second, we introduce an adaptive strategy that allocates fewer updates to stable layers without degrading generation quality. Together, these contributions significantly improve the efficiency of DLMs, yielding up to an $8\times$ throughput improvement over vanilla decoding and a $2$–$4\times$ speedup over existing caching baselines.

Method: Manifold-Reshaping Policy Optimization (MRPO) operates in two stages: 1) Spectral Orthogonal Exploration (SOE) ejects policy initialization into the null space of the bias manifold, and 2) Effective Rank regularization is integrated into policy optimization to incentivize discovery and maintenance of high-dimensional reasoning trajectories against entropy-reducing tendencies of standard RL.

Result: The 4B-parameter method achieves state-of-the-art performance on mathematical tasks, significantly outperforming larger models (e.g., Qwen3-32B) and expanding the capability boundary beyond standard GRPO.

Conclusion: The latent reasoning space of LLMs can be fundamentally expanded through targeted geometric interventions, challenging the accessibility boundary hypothesis and demonstrating that RL can genuinely expand reasoning capacity beyond pre-trained bias manifolds.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated remarkable success in enhancing the reasoning capabilities of Large Language Models (LLMs). However, recent studies question whether RL genuinely expands reasoning capacity or merely aligns existing latent capabilities, arguing that exploration remains confined within the pre-trained model’s low-rank bias manifold. In this work, we challenge this accessibility boundary hypothesis by demonstrating that the latent reasoning space can be fundamentally expanded through targeted geometric interventions. We propose Manifold-Reshaping Policy Optimization (MRPO), a geometric framework designed to fundamentally restructure the inference space of LLMs. MRPO operates in two stages: first, we employ Spectral Orthogonal Exploration (SOE) to eject the policy initialization into the null space of the bias manifold; second, we integrate an Effective Rank regularization term into the policy optimization objective. This approach incentivizes the discovery and maintenance of high-dimensional reasoning trajectories against the entropy-reducing tendency of standard RL. Empirically, our 4B-parameter method achieves state-of-the-art performance on mathematical tasks, significantly outperforming larger models (e.g., Qwen3-32B) and expanding the capability boundary beyond standard GRPO. Our code is available at https://anonymous.4open.science/r/MRPO-D57B/

Method: Proposes D²Quant framework with two key components: 1) Dual-Scale Quantizer (DSQ) tailored to down-projection matrices with absorbable scaling factor that improves accuracy without increasing bit budget, and 2) Deviation-Aware Correction (DAC) that incorporates mean-shift correction within LayerNorm to mitigate quantization-induced activation distribution shifts.

Result: Extensive experiments across multiple LLM families and evaluation metrics show that D²Quant delivers superior performance for weight-only PTQ at sub-4-bit precision compared to existing methods.

Conclusion: D²Quant effectively addresses key challenges in weight-only PTQ for LLMs at sub-4-bit precision through innovative weight quantization and activation correction techniques, enabling more efficient deployment in resource-constrained scenarios.

Abstract: Large language models (LLMs) deliver strong performance, but their high compute and memory costs make deployment difficult in resource-constrained scenarios. Weight-only post-training quantization (PTQ) is appealing, as it reduces memory usage and enables practical speedup without low-bit operators or specialized hardware. However, accuracy often degrades significantly in weight-only PTQ at sub-4-bit precision, and our analysis identifies two main causes: (1) down-projection matrices are a well-known quantization bottleneck, but maintaining their fidelity often requires extra bit-width; (2) weight quantization induces activation deviations, but effective correction strategies remain underexplored. To address these issues, we propose D$^2$Quant, a novel weight-only PTQ framework that improves quantization from both the weight and activation perspectives. On the weight side, we design a Dual-Scale Quantizer (DSQ) tailored to down-projection matrices, with an absorbable scaling factor that significantly improves accuracy without increasing the bit budget. On the activation side, we propose Deviation-Aware Correction (DAC), which incorporates a mean-shift correction within LayerNorm to mitigate quantization-induced activation distribution shifts. Extensive experiments across multiple LLM families and evaluation metrics show that D$^2$Quant delivers superior performance for weight-only PTQ at sub-4-bit precision. The code and models will be available at https://github.com/XIANGLONGYAN/D2Quant.

Method: naPINN embeds an energy-based model to learn latent distribution of prediction residuals, uses a trainable reliability gate to adaptively filter high-energy data points, and includes rejection cost regularization to prevent trivial solutions.

Result: naPINN significantly outperforms existing robust PINN baselines on various benchmark PDEs corrupted by non-Gaussian noise and varying outlier rates, successfully isolating outliers and accurately reconstructing dynamics under severe data corruption.

Conclusion: naPINN provides a robust framework for solving inverse problems and discovering governing equations from corrupted observational data without prior knowledge of noise distribution.

Abstract: Physics-Informed Neural Networks (PINNs) are effective methods for solving inverse problems and discovering governing equations from observational data. However, their performance degrades significantly under complex measurement noise and gross outliers. To address this issue, we propose the Noise-Adaptive Physics-Informed Neural Network (naPINN), which robustly recovers physical solutions from corrupted measurements without prior knowledge of the noise distribution. naPINN embeds an energy-based model into the training loop to learn the latent distribution of prediction residuals. Leveraging the learned energy landscape, a trainable reliability gate adaptively filters data points exhibiting high energy, while a rejection cost regularization prevents trivial solutions where valid data are discarded. We demonstrate the efficacy of naPINN on various benchmark partial differential equations corrupted by non-Gaussian noise and varying rates of outliers. The results show that naPINN significantly outperforms existing robust PINN baselines, successfully isolating outliers and accurately reconstructing the dynamics under severe data corruption.

Method: HyPAC uses importance sampling and upper confidence bounds to calibrate decision thresholds, partitioning inputs into three uncertainty regions and routing each to appropriate annotation sources.

Result: Experiments show HyPAC reduces annotation cost by 78.51% while tightly controlling annotation error, achieving minimum expected cost with PAC guarantees.

Conclusion: HyPAC provides a principled approach for cost-efficient data annotation with formal error guarantees, applicable to various annotation sources including LLMs and human experts.

Abstract: Data annotation often involves multiple sources with different cost-quality trade-offs, such as fast large language models (LLMs), slow reasoning models, and human experts. In this work, we study the problem of routing inputs to the most cost-efficient annotation source while controlling the labeling error on test instances. We propose \textbf{HyPAC}, a method that adaptively labels inputs to the most cost-efficient annotation source while providing distribution-free guarantees on annotation error. HyPAC calibrates two decision thresholds using importance sampling and upper confidence bounds, partitioning inputs into three regions based on uncertainty and routing each to the appropriate annotation source. We prove that HyPAC achieves the minimum expected cost with a probably approximately correct (PAC) guarantee on the annotation error, free of data distribution and pre-trained models. Experiments on common benchmarks demonstrate the effectiveness of our method, reducing the annotation cost by 78.51% while tightly controlling the annotation error.

Method: Proposes a lightweight Transformer architecture combined with a novel Escape-Explore Optimizer (EEO) that enhances exploration and generalization while avoiding sharp minima and saddle-point traps in high-dimensional parameter spaces.

Result: Achieves performance on par with state-of-the-art models across 11 time-series benchmark datasets and the Synapse medical image segmentation task, demonstrating superior generalization and stability.

Conclusion: The method shows potential as a versatile cross-task foundation model for Web-scale data mining and analysis, effectively addressing optimization challenges in multimodal and temporal data processing.

Abstract: Transformer-based foundation models have achieved remarkable progress in tasks such as time-series forecasting and image segmentation. However, they frequently suffer from error accumulation in multivariate long-sequence prediction and exhibit vulnerability to out-of-distribution samples in image-related tasks. Furthermore, these challenges become particularly pronounced in large-scale Web data analysis tasks, which typically involve complex temporal patterns and multimodal features. This complexity substantially increases optimization difficulty, rendering models prone to stagnation at saddle points within high-dimensional parameter spaces. To address these issues, we propose a lightweight Transformer architecture in conjunction with a novel Escape-Explore Optimizer (EEO). The optimizer enhances both exploration and generalization while effectively avoiding sharp minima and saddle-point traps. Experimental results show that, in representative Web data scenarios, our method achieves performance on par with state-of-the-art models across 11 time-series benchmark datasets and the Synapse medical image segmentation task. Moreover, it demonstrates superior generalization and stability, thereby validating its potential as a versatile cross-task foundation model for Web-scale data mining and analysis.

Method: Uses back-translation strategy: generate documentation from code, then reconstruct original code from documentation. Semantic similarity between original and reconstructed code serves as implicit reward for reinforcement learning.

Result: Achieved 83.5% on HumanEval and 81.0% on MBPP pass@1 with 7B model, outperforming open-source baselines. Shows consistent scaling with training corpus size and model capacity.

Conclusion: BatCoder enables training with only code, substantially increasing available training examples and improving performance on code generation and documentation tasks.

Abstract: Training LLMs for code-related tasks typically depends on high-quality code-documentation pairs, which are costly to curate and often scarce for niche programming languages. We introduce BatCoder, a self-supervised reinforcement learning framework designed to jointly optimize code generation and documentation production. BatCoder employs a back-translation strategy: a documentation is first generated from code, and then the generated documentation is used to reconstruct the original code. The semantic similarity between the original and reconstructed code serves as an implicit reward, enabling reinforcement learning to improve the model’s performance both in generating code from documentation and vice versa. This approach allows models to be trained using only code, substantially increasing the available training examples. Evaluated on HumanEval and MBPP with a 7B model, BatCoder achieved 83.5% and 81.0% pass@1, outperforming strong open-source baselines. Moreover, the framework demonstrates consistent scaling with respect to both training corpus size and model capacity.

Method: PSN-RLVR perturbs policy parameters before rollout generation to induce temporally consistent, trajectory-level exploration that preserves chain-of-thought coherence. It uses truncated importance sampling to mitigate sampling-update mismatch and a computationally efficient adaptive noise scheduler driven by semantic diversity and normalized self-certainty metrics.

Result: PSN-GRPO consistently expands reasoning capability boundaries across multiple mathematical reasoning benchmarks and model families, achieving higher pass-at-k under large sampling budgets and outperforming prior exploration-oriented RLVR methods.

Conclusion: Parameter-space noise with adaptive scheduling effectively addresses exploration limitations in RLVR for LLM reasoning, providing orthogonal improvements that can be composed with other methods for additional gains.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) improves LLM reasoning, yet growing evidence indicates an exploration ceiling: it often reweights existing solution traces rather than discovering new strategies, limiting gains under large sampling budgets (e.g., pass-at-256). We address this limitation with PSN-RLVR, which perturbs policy parameters before rollout generation to induce temporally consistent, trajectory-level exploration that better preserves long-horizon chain-of-thought coherence than action-space noise. To mitigate the resulting sampling-update mismatch, we incorporate truncated importance sampling (TIS). To avoid expensive KL-based adaptive noise control, we propose a computationally efficient real-time adaptive noise scheduler driven by a lightweight surrogate that combines semantic diversity with normalized self-certainty. Instantiated on GRPO, a widely used RLVR method, PSN-GRPO consistently expands the effective reasoning capability boundary across multiple mathematical reasoning benchmarks and model families, yielding higher pass-at-k under large sampling budgets and outperforming prior exploration-oriented RLVR methods (e.g., Pass-at-k-style training) while remaining orthogonal and thus composable for additional gains.

Method: SEAM (Structured Experience Adapter Module) is a lightweight plug-in that stores experience in its parameters and generates structured, instance-tailored experience entries in a single forward pass to guide a frozen LLM executor. It’s trained via executor rollouts and GRPO while keeping the executor frozen, and can be improved after deployment with supervised fine-tuning on logged successful trajectories.

Result: Experiments on mathematical reasoning benchmarks show consistent accuracy gains across executors with low overhead. Extensive ablations and analyses elucidate the mechanisms underlying SEAM’s effectiveness and robustness.

Conclusion: SEAM provides an effective, efficient approach to experience reuse for LLMs that avoids the limitations of external retrieval systems while maintaining low computational overhead.

Abstract: Large language models (LLMs) are largely static and often redo reasoning or repeat mistakes. Prior experience reuse typically relies on external retrieval, which is similarity-based, can introduce noise, and adds latency. We introduce SEAM (Structured Experience Adapter Module), a lightweight, executor-specific plug-in that stores experience in its parameters and generates a structured, instance-tailored experience entry in a single forward pass to guide a frozen LLM executor. SEAM is trained for utility via executor rollouts and GRPO while keeping the executor frozen, and it can be further improved after deployment with supervised fine-tuning on logged successful trajectories. Experiments on mathematical reasoning benchmarks show consistent accuracy gains across executors with low overhead. Extensive ablations and analyses further elucidate the mechanisms underlying SEAM’s effectiveness and robustness.

Method: Proposes Phenotype-Aware Multiple Instance Learning (PA-MIL) with three key components: 1) constructing a phenotype knowledge base linking cancer phenotypes to genotypes, 2) using morphological descriptions as language prompts to aggregate phenotype-related features, and 3) developing a Genotype-to-Phenotype Neural Network (GP-NN) that provides multi-level guidance based on genotype-phenotype relationships.

Result: PA-MIL achieves competitive performance compared to existing MIL methods on multiple datasets while offering improved interpretability. It uses phenotype saliency as evidence and achieves competitive results with a linear classifier compared to state-of-the-art methods. The framework enables thorough analysis of genotype-phenotype relationships and provides cohort-level and case-level interpretability.

Conclusion: PA-MIL provides a reliable and accountable ante-hoc interpretable framework for cancer subtyping from pathology WSIs by leveraging phenotype knowledge and language prompting, demonstrating both competitive performance and transparent reasoning capabilities.

Abstract: Deep learning has been extensively researched in the analysis of pathology whole-slide images (WSIs). However, most existing methods are limited to providing prediction interpretability by locating the model’s salient areas in a post-hoc manner, failing to offer more reliable and accountable explanations. In this work, we propose Phenotype-Aware Multiple Instance Learning (PA-MIL), a novel ante-hoc interpretable framework that identifies cancer-related phenotypes from WSIs and utilizes them for cancer subtyping. To facilitate PA-MIL in learning phenotype-aware features, we 1) construct a phenotype knowledge base containing cancer-related phenotypes and their associated genotypes. 2) utilize the morphological descriptions of phenotypes as language prompting to aggregate phenotype-related features. 3) devise the Genotype-to-Phenotype Neural Network (GP-NN) grounded in genotype-to-phenotype relationships, which provides multi-level guidance for PA-MIL. Experimental results on multiple datasets demonstrate that PA-MIL achieves competitive performance compared to existing MIL methods while offering improved interpretability. PA-MIL leverages phenotype saliency as evidence and, using a linear classifier, achieves competitive results compared to state-of-the-art methods. Additionally, we thoroughly analyze the genotype-phenotype relationships, as well as cohort-level and case-level interpretability, demonstrating the reliability and accountability of PA-MIL.

Method: Proposes S(H)NAP framework that constructs generative interventional attributions using realistic 3D diffusion bridge modeling to systematically modify anatomical features and isolate object-specific causal contributions to risk scores.

Result: First interventional audit of Sybil shows the model often exhibits expert-like behavior in differentiating malignant from benign nodules, but suffers from critical failure modes including dangerous sensitivity to clinically unjustified artifacts and distinct radial bias.

Conclusion: Causal verification through generative interventions reveals both strengths and critical weaknesses in AI models for medical imaging, highlighting the importance of moving beyond correlation-based assessments for clinical deployment.

Abstract: Lung cancer remains the leading cause of cancer mortality, driving the development of automated screening tools to alleviate radiologist workload. Standing at the frontier of this effort is Sybil, a deep learning model capable of predicting future risk solely from computed tomography (CT) with high precision. However, despite extensive clinical validation, current assessments rely purely on observational metrics. This correlation-based approach overlooks the model’s actual reasoning mechanism, necessitating a shift to causal verification to ensure robust decision-making before clinical deployment. We propose S(H)NAP, a model-agnostic auditing framework that constructs generative interventional attributions validated by expert radiologists. By leveraging realistic 3D diffusion bridge modeling to systematically modify anatomical features, our approach isolates object-specific causal contributions to the risk score. Providing the first interventional audit of Sybil, we demonstrate that while the model often exhibits behavior akin to an expert radiologist, differentiating malignant pulmonary nodules from benign ones, it suffers from critical failure modes, including dangerous sensitivity to clinically unjustified artifacts and a distinct radial bias.

Method: Proposes Spatiotemporal Residual Theorem to generalize unidirectional prediction into bidirectional learning. Introduces ReLearner module with two components: Residual Learning Module to disentangle feature discrepancies between input and label representations, and Residual Smoothing Module to smooth residuals for stable convergence.

Result: Extensive experiments on 11 real-world datasets across 14 backbone models show ReLearner significantly enhances predictive performance of existing Spatiotemporal Neural Networks (STNNs).

Conclusion: The bidirectional learning framework with ReLearner effectively addresses spatiotemporal discrepancies and improves prediction accuracy across various models and datasets.

Abstract: Prevailing spatiotemporal prediction models typically operate under a forward (unidirectional) learning paradigm, in which models extract spatiotemporal features from historical observation input and map them to target spatiotemporal space for future forecasting (label). However, these models frequently exhibit suboptimal performance when spatiotemporal discrepancies exist between inputs and labels, for instance, when nodes with similar time-series inputs manifest distinct future labels, or vice versa. To address this limitation, we propose explicitly incorporating label features during the training phase. Specifically, we introduce the Spatiotemporal Residual Theorem, which generalizes the conventional unidirectional spatiotemporal prediction paradigm into a bidirectional learning framework. Building upon this theoretical foundation, we design an universal module, termed ReLearner, which seamlessly augments Spatiotemporal Neural Networks (STNNs) with a bidirectional learning capability via an auxiliary inverse learning process. In this process, the model relearns the spatiotemporal feature residuals between input data and future data. The proposed ReLearner comprises two critical components: (1) a Residual Learning Module, designed to effectively disentangle spatiotemporal feature discrepancies between input and label representations; and (2) a Residual Smoothing Module, employed to smooth residual terms and facilitate stable convergence. Extensive experiments conducted on 11 real-world datasets across 14 backbone models demonstrate that ReLearner significantly enhances the predictive performance of existing STNNs.Our code is available on GitHub.

Method: The approach clusters incomplete vectors by grouping proxy subspaces and minimizes two criteria over the Grassmannian: (1) chordal distance between each point and its corresponding subspace, and (2) geodesic distances between subspaces of all data points.

Result: Experiments on synthetic and real datasets show the method performs comparably to leading methods at high sampling rates and significantly better at low sampling rates, narrowing the gap to the theoretical sampling limit of HRMC.

Conclusion: The proposed method provides a theoretically grounded approach to high-rank matrix completion that works well even at low sampling rates, addressing limitations of existing methods while producing interpretable results.

Abstract: This paper approaches high-rank matrix completion (HRMC) by filling missing entries in a data matrix where columns lie near a union of subspaces, clustering these columns, and identifying the underlying subspaces. Current methods often lack theoretical support, produce uninterpretable results, and require more samples than theoretically necessary. We propose clustering incomplete vectors by grouping proxy subspaces and minimizing two criteria over the Grassmannian: (a) the chordal distance between each point and its corresponding subspace and (b) the geodesic distances between subspaces of all data points. Experiments on synthetic and real datasets demonstrate that our method performs comparably to leading methods in high sampling rates and significantly better in low sampling rates, thus narrowing the gap to the theoretical sampling limit of HRMC.

Method: Comprehensive comparative simulation study analyzing fairness and accuracy of predictive policing technologies in Baltimore, comparing with traditional hot spots policing approaches.

Result: Predictive policing exhibited bias due to feedback loops, but traditional hot spots policing had similar issues. Predictive policing was more fair and accurate in short term but amplified bias faster long-term. In Baltimore, bias sometimes favored over-policing in White neighborhoods, contrary to previous studies.

Conclusion: The study demonstrates a methodology for city-specific evaluation of predictive policing systems, revealing complex bias patterns and showing simulations can expose inequities and long-term behavioral tendencies.

Abstract: There are ongoing discussions about predictive policing systems, such as those deployed in Los Angeles, California and Baltimore, Maryland, being unfair, for example, by exhibiting racial bias. Studies found that unfairness may be due to feedback loops and being trained on historically biased recorded data. However, comparative studies on predictive policing systems are few and are not sufficiently comprehensive. In this work, we perform a comprehensive comparative simulation study on the fairness and accuracy of predictive policing technologies in Baltimore. Our results suggest that the situation around bias in predictive policing is more complex than was previously assumed. While predictive policing exhibited bias due to feedback loops as was previously reported, we found that the traditional alternative, hot spots policing, had similar issues. Predictive policing was found to be more fair and accurate than hot spots policing in the short term, although it amplified bias faster, suggesting the potential for worse long-run behavior. In Baltimore, in some cases the bias in these systems tended toward over-policing in White neighborhoods, unlike in previous studies. Overall, this work demonstrates a methodology for city-specific evaluation and behavioral-tendency comparison of predictive policing systems, showing how such simulations can reveal inequities and long-term tendencies.

Method: HTCL combines fast local adaptation with conservative second-order global consolidation using Hessian-regularized Taylor expansion. It identifies optimal intra-group task sequences and extends to L-level hierarchies for multiscale knowledge integration.

Result: HTCL consistently enhances performance across various datasets and baselines, achieving mean accuracy gains of 7-25% while reducing standard deviation of final accuracy by up to 68% across random task permutations.

Conclusion: HTCL provides a model-agnostic consolidation layer that effectively addresses task-order variance in continual learning through hierarchical Taylor series-based optimization with theoretical guarantees.

Abstract: We introduce $\textbf{Hierarchical Taylor Series-based Continual Learning (HTCL)}$, a framework that couples fast local adaptation with conservative, second-order global consolidation to address the high variance introduced by random task ordering. To address task-order effects, HTCL identifies the best intra-group task sequence and integrates the resulting local updates through a Hessian-regularized Taylor expansion, yielding a consolidation step with theoretical guarantees. The approach naturally extends to an $L$-level hierarchy, enabling multiscale knowledge integration in a manner not supported by conventional single-level CL systems. Across a wide range of datasets and replay and regularization baselines, HTCL acts as a model-agnostic consolidation layer that consistently enhances performance, yielding mean accuracy gains of $7%$ to $25%$ while reducing the standard deviation of final accuracy by up to $68%$ across random task permutations.

Method: Replace difficult trajectory consistency constraints with a principled linear surrogate derived from semigroup formulation of flow-based models. This enables direct data supervision for long-horizon flow-map compositions without explicit Jacobian computations (JVP-free). Supports both u-prediction and x₁-prediction variants.

Result: Improved optimization stability and sample quality under fixed sampling budgets. Approximately 50% reduction in training time and memory consumption compared to existing one-step methods for image generation. Demonstrated effectiveness on image synthesis, particle-based geometry generation, and functional generation tasks.

Conclusion: EMF provides an efficient flow-based generative framework that enforces long-range trajectory consistency with minimal sampling cost, offering significant computational advantages while maintaining or improving sample quality.

Abstract: We propose \emph{Euler Mean Flows (EMF)}, a flow-based generative framework for one-step and few-step generation that enforces long-range trajectory consistency with minimal sampling cost. The key idea of EMF is to replace the trajectory consistency constraint, which is difficult to supervise and optimize over long time scales, with a principled linear surrogate that enables direct data supervision for long-horizon flow-map compositions. We derive this approximation from the semigroup formulation of flow-based models and show that, under mild regularity assumptions, it faithfully approximates the original consistency objective while being substantially easier to optimize. This formulation leads to a unified, JVP-free training framework that supports both $u$-prediction and $x_1$-prediction variants, avoiding explicit Jacobian computations and significantly reducing memory and computational overhead. Experiments on image synthesis, particle-based geometry generation, and functional generation demonstrate improved optimization stability and sample quality under fixed sampling budgets, together with approximately $50%$ reductions in training time and memory consumption compared to existing one-step methods for image generation.

Method: Formalize reward model optimization as a Stackelberg game, develop a simple reward shaping scheme to approximate the optimal reward model, and integrate it into existing alignment methods with minimal overhead.

Result: Empirical evaluation shows consistent improvement in average reward and achieves win-tie rates exceeding 66% against all baselines across various evaluation settings.

Conclusion: The proposed reward model optimization approach effectively addresses the bias-reward hacking tradeoff in LLM alignment and integrates seamlessly with existing methods.

Abstract: Existing alignment methods directly use the reward model learned from user preference data to optimize an LLM policy, subject to KL regularization with respect to the base policy. This practice is suboptimal for maximizing user’s utility because the KL regularization may cause the LLM to inherit the bias in the base policy that conflicts with user preferences. While amplifying rewards for preferred outputs can mitigate this bias, it also increases the risk of reward hacking. This tradeoff motivates the problem of optimally designing reward models under KL regularization. We formalize this reward model optimization problem as a Stackelberg game, and show that a simple reward shaping scheme can effectively approximate the optimal reward model. We empirically evaluate our method in inference-time alignment settings and demonstrate that it integrates seamlessly into existing alignment methods with minimal overhead. Our method consistently improves average reward and achieves win-tie rates exceeding 66% against all baselines, averaged across evaluation settings.

Method: Introduces product interactions as an algebraic formalism where neural network layers are built from compositions of multiplication operators over suitable algebras. Organizes algebraic expressions by increasing interaction order, showing how different architectures correspond to specific interaction orders.

Result: Provides a principled mathematical framework that unifies various neural network architectures: convolutional and equivariant networks correspond to symmetry-constrained linear product interactions, while attention and Mamba correspond to higher-order product interactions.

Conclusion: Product interactions offer a systematic algebraic approach to understanding and constructing neural network layers, revealing fundamental mathematical connections between seemingly disparate architectures through their interaction order properties.

Abstract: In this paper, we introduce product interactions, an algebraic formalism in which neural network layers are constructed from compositions of a multiplication operator defined over suitable algebras. Product interactions provide a principled way to generate and organize algebraic expressions by increasing interaction order. Our central observation is that algebraic expressions in modern neural networks admit a unified construction in terms of linear, quadratic, and higher-order product interactions. Convolutional and equivariant networks arise as symmetry-constrained linear product interactions, while attention and Mamba correspond to higher-order product interactions.

Method: QuantLRM analyzes weight update magnitudes during fine-tuning, hypothesizing that both smallest and largest updates are important (“protecting both ends”). It fits restricted quadratic functions on weight updates, multiplies average quadratic values with zero-update channel counts to compute channel importance, and applies this to quantize various fine-tuned models.

Result: QuantLRM consistently improves LRM quantization across four reasoning benchmarks (AIME-120, FOLIO, temporal sequences, GPQA-Diamond), with average 6.55% improvement on reinforcement learning fine-tuned models. It also works with non-fine-tuned LRMs via pseudo-fine-tuning.

Conclusion: Weight update magnitudes during fine-tuning provide valuable signals for quantizing Large Reasoning Models. The “protecting both ends” hypothesis is validated, and QuantLRM offers an effective quantization approach that outperforms methods using activation or second-order information.

Abstract: Weight-only quantization is important for compressing Large Language Models (LLMs). Inspired by the spirit of classical magnitude pruning, we study whether the magnitude of weight updates during reasoning-incentivized fine-tuning can provide valuable signals for quantizing Large Reasoning Models (LRMs). We hypothesize that the smallest and largest weight updates during fine-tuning are more important than those of intermediate magnitude, a phenomenon we term “protecting both ends”. Upon hypothesis validation, we introduce QuantLRM, which stands for weight quantization of LRMs via fine-tuning signals. We fit simple restricted quadratic functions on weight updates to protect both ends. By multiplying the average quadratic values with the count of zero weight updates of channels, we compute channel importance that is more effective than using activation or second-order information. We run QuantLRM to quantize various fine-tuned models (including supervised, direct preference optimization, and reinforcement learning fine-tuning) over four reasoning benchmarks (AIME-120, FOLIO, temporal sequences, and GPQA-Diamond) and empirically find that QuantLRM delivers a consistent improvement for LRMs quantization, with an average improvement of 6.55% on a reinforcement learning fine-tuned model. Also supporting non-fine-tuned LRMs, QuantLRM gathers effective signals via pseudo-fine-tuning, which greatly enhances its applicability.

Method: Proposes a two-stage framework: 1) Copula-based dependence modeling to capture cross-site correlations, and 2) Context-Aware Conformal Prediction (CACP) to correct miscalibration at the aggregated level. This enables dependence-aware aggregation while maintaining valid coverage and sharp prediction intervals.

Result: Experiments on large-scale solar generation datasets from MISO, ERCOT, and SPP show the Copula+CACP approach consistently achieves near-nominal coverage with significantly sharper intervals than uncalibrated aggregation baselines.

Conclusion: The framework provides a practical solution for constructing reliable fleet-level probabilistic forecasts from heterogeneous site-level inputs when system-level models cannot be trained or maintained, addressing key challenges in renewable energy grid operations.

Abstract: The rapid growth of renewable energy penetration has intensified the need for reliable probabilistic forecasts to support grid operations at aggregated (fleet or system) levels. In practice, however, system operators often lack access to fleet-level probabilistic models and instead rely on site-level forecasts produced by heterogeneous third-party providers. Constructing coherent and calibrated fleet-level probabilistic forecasts from such inputs remains challenging due to complex cross-site dependencies and aggregation-induced miscalibration. This paper proposes a calibrated probabilistic aggregation framework that directly converts site-level probabilistic forecasts into reliable fleet-level forecasts in settings where system-level models cannot be trained or maintained. The framework integrates copula-based dependence modeling to capture cross-site correlations with Context-Aware Conformal Prediction (CACP) to correct miscalibration at the aggregated level. This combination enables dependence-aware aggregation while providing valid coverage and maintaining sharp prediction intervals. Experiments on large-scale solar generation datasets from MISO, ERCOT, and SPP demonstrate that the proposed Copula+CACP approach consistently achieves near-nominal coverage with significantly sharper intervals than uncalibrated aggregation baselines.

Method: Proposes four learnable Koopman variants: scalar-gated, per-mode gated, MLP-shaped spectral mapping, and low-rank Koopman operators. These integrate with nonlinear backbones like Patchtst, Autoformer, and Informer, allowing explicit control over spectrum, stability, and rank of linear transition operators.

Result: Learnable Koopman models provide favorable bias-variance trade-off, improved conditioning, and more interpretable latent dynamics compared to LSTM, DLinear, SSMs, and lightweight transformers across multiple horizons and patch lengths.

Conclusion: Learnable Koopman operators are effective, stable, and theoretically principled components for deep forecasting that combine the benefits of linear dynamical systems theory with modern deep learning architectures.

Abstract: This paper proposes a unified family of learnable Koopman operator parameterizations that integrate linear dynamical systems theory with modern deep learning forecasting architectures. We introduce four learnable Koopman variants-scalar-gated, per-mode gated, MLP-shaped spectral mapping, and low-rank Koopman operators which generalize and interpolate between strictly stable Koopman operators and unconstrained linear latent dynamics. Our formulation enables explicit control over the spectrum, stability, and rank of the linear transition operator while retaining compatibility with expressive nonlinear backbones such as Patchtst, Autoformer, and Informer. We evaluate the proposed operators in a large-scale benchmark that also includes LSTM, DLinear, and simple diagonal State-Space Models (SSMs), as well as lightweight transformer variants. Experiments across multiple horizons and patch lengths show that learnable Koopman models provide a favorable bias-variance trade-off, improved conditioning, and more interpretable latent dynamics. We provide a full spectral analysis, including eigenvalue trajectories, stability envelopes, and learned spectral distributions. Our results demonstrate that learnable Koopman operators are effective, stable, and theoretically principled components for deep forecasting.

Method: Proposes a framework that abstracts learning tasks as coverage of patterns from a Zipfian (long-tail) distribution. Introduces the Effective Frontier (k⋆) threshold separating learned knowledge from unlearned tail. Derives scaling laws for model capacity (N), data size (D), and compute (C) based on this framework, attributing them to capacity, coverage, and optimization bottlenecks respectively.

Result: Derives precise scaling laws for N, D, and C, showing they correspond to capacity, coverage, and optimization bottlenecks. Unifies these mechanisms via a Max-Bottleneck principle, demonstrating that Kaplan and Chinchilla scaling laws are equilibrium solutions to the same constrained optimization problem under different active bottlenecks.

Conclusion: Provides a unified theoretical framework for understanding neural scaling laws through pattern coverage of long-tail distributions, explaining different scaling regimes as manifestations of different bottleneck constraints in the same underlying optimization problem.

Abstract: Neural scaling laws govern the prediction power-law improvement of test loss with respect to model capacity ($N$), datasize ($D$), and compute ($C$). However, existing theoretical explanations often rely on specific architectures or complex kernel methods, lacking intuitive universality. In this paper, we propose a unified framework that abstracts general learning tasks as the progressive coverage of patterns from a long-tail (Zipfian) distribution. We introduce the Effective Frontier ($k_\star$), a threshold in the pattern rank space that separates learned knowledge from the unlearned tail. We prove that reducible loss is asymptotically determined by the probability mass of the tail a resource-dependent frontier truncation. Based on our framework, we derive the precise scaling laws for $N$, $D$, and $C$, attributing them to capacity, coverage, and optimization bottlenecks, respectively. Furthermore, we unify these mechanisms via a Max-Bottleneck principle, demonstrating that the Kaplan and Chinchilla scaling laws are not contradictory, but equilibrium solutions to the same constrained optimization problem under different active bottlenecks.

Method: Proposes a projective geometric framework using the Fubini-Study metric to quantify representation drift, which remains invariant under gauge transformations. Constructs representation trajectories and tracks evolution through cumulative geometric drift, comparing Euclidean, cosine, and Fubini-Study distances.

Result: Conventional metrics systematically overestimate change when representations have projective ambiguity, while Fubini-Study metric isolates intrinsic evolution. The difference between cosine and Fubini-Study drift defines a computable quantity capturing representation churn from gauge freedom.

Conclusion: Establishes geometric criterion for assessing representation stability in high-dimensional systems, clarifies limits of angular distances, and provides diagnostic for distinguishing meaningful structural evolution from parametrization artifacts without model-specific assumptions.

Abstract: High dimensional representation drift is commonly quantified using Euclidean or cosine distances, which presuppose fixed coordinates when comparing representations across time, training or preprocessing stages. While effective in many settings, these measures entangle intrinsic changes in the data with variations induced by arbitrary parametrizations. We introduce a projective geometric view of representation drift grounded in the Fubini Study metric, which identifies representations that differ only by gauge transformations such as global rescalings or sign flips. Applying this framework to empirical high dimensional datasets, we explicitly construct representation trajectories and track their evolution through cumulative geometric drift. Comparing Euclidean, cosine and Fubini Study distances along these trajectories reveals that conventional metrics systematically overestimate change whenever representations carry genuine projective ambiguity. By contrast, the Fubini Study metric isolates intrinsic evolution by remaining invariant under gauge-induced fluctuations. We further show that the difference between cosine and Fubini Study drift defines a computable, monotone quantity that directly captures representation churn attributable to gauge freedom. This separation provides a diagnostic for distinguishing meaningful structural evolution from parametrization artifacts, without introducing model-specific assumptions. Overall, we establish a geometric criterion for assessing representation stability in high-dimensional systems and clarify the limits of angular distances. Embedding representation dynamics in projective space connects data analysis with established geometric programs and yields observables that are directly testable in empirical workflows.

Method: ContextEvolve uses three specialized agents: 1) Summarizer Agent condenses semantic state via code-to-language abstraction, 2) Navigator Agent distills optimization direction from trajectory analysis, and 3) Sampler Agent curates experience distribution through prioritized exemplar retrieval. This orchestration creates a functional isomorphism with RL components (state representation, policy gradient, experience replay) enabling principled optimization in textual latent space.

Result: On the ADRS benchmark, ContextEvolve outperforms state-of-the-art baselines by 33.3% while reducing token consumption by 29.0%.

Conclusion: ContextEvolve demonstrates that RL-level search efficiency can be achieved under parameter-blind constraints through multi-agent decomposition of optimization context, offering an effective approach for optimizing LLM-generated code for computer systems.

Abstract: Large language models are transforming systems research by automating the discovery of performance-critical algorithms for computer systems. Despite plausible codes generated by LLMs, producing solutions that meet the stringent correctness and performance requirements of systems demands iterative optimization. Test-time reinforcement learning offers high search efficiency but requires parameter updates infeasible under API-only access, while existing training-free evolutionary methods suffer from inefficient context utilization and undirected search. We introduce ContextEvolve, a multi-agent framework that achieves RL-level search efficiency under strict parameter-blind constraints by decomposing optimization context into three orthogonal dimensions: a Summarizer Agent condenses semantic state via code-to-language abstraction, a Navigator Agent distills optimization direction from trajectory analysis, and a Sampler Agent curates experience distribution through prioritized exemplar retrieval. This orchestration forms a functional isomorphism with RL-mapping to state representation, policy gradient, and experience replay-enabling principled optimization in a textual latent space. On the ADRS benchmark, ContextEvolve outperforms state-of-the-art baselines by 33.3% while reducing token consumption by 29.0%. Codes for our work are released at https://anonymous.4open.science/r/ContextEvolve-ACC

Method: Proposes RoPE-Aligned Pruning (RAP) which prunes entire RoPE-aligned column pairs to preserve RoPE’s 2x2 rotation structure, restore B absorption (from W ≈ A*B factorization), and eliminate reconstruction overhead.

Result: Evaluation on LLaMA-3-8B and Mistral-7B shows RAP enables joint reduction of KV-Cache, attention parameters, and FLOPs by 20-30% while maintaining strong accuracy. Reduces attention latency to 83% (prefill) and 77% (decode) of baseline.

Conclusion: RAP effectively addresses KV-Cache bottlenecks in RoPE-based LLMs through structured pruning that preserves rotational structure, enabling significant efficiency gains without sacrificing accuracy.

Abstract: Long-context inference in large language models is increasingly bottlenecked by the memory and compute cost of the KV-Cache. Low-rank factorization compresses KV projections by writing $W \approx A * B$, where A produces latent KV states and B can be absorbed into downstream weights. In modern RoPE-based LLMs, this absorption fails: RoPE forces latent KV states to be reconstructed to full dimension, reintroducing substantial memory and compute overhead. We propose RoPE-Aligned Pruning (RAP), which prunes entire RoPE-aligned column pairs to preserve RoPE’s 2x2 rotation structure, restore B absorption, and eliminate reconstruction. Our evaluation on LLaMA-3-8B and Mistral-7B shows that RAP enables joint reduction of KV-Cache, attention parameters, and FLOPs by 20-30%, all at once, while maintaining strong accuracy. Notably, RAP reduces attention latency to 83% (prefill) and 77% (decode) of baseline.

Method: Developed an analytical framework for step-wise refusal dynamics, introduced Step-Wise Refusal Internal Dynamics (SRI) signal, analyzed geometric structure of SRI to capture internal recovery dynamics, and created lightweight inference-time detectors based on this structure.

Result: SRI identifies anomalous behavior in harmful generations as incomplete internal recovery, enables detectors that generalize to unseen attacks while matching/exceeding existing defenses with 100× lower inference overhead.

Conclusion: Sampling strategy itself is central to safety behavior, distinct from learned representations. SRI provides interpretable safety signals and enables efficient, effective jailbreak detection for both AR and diffusion language models.

Abstract: Diffusion language models (DLMs) have recently emerged as a promising alternative to autoregressive (AR) models, offering parallel decoding and controllable sampling dynamics while achieving competitive generation quality at scale. Despite this progress, the role of sampling mechanisms in shaping refusal behavior and jailbreak robustness remains poorly understood. In this work, we present a fundamental analytical framework for step-wise refusal dynamics, enabling comparison between AR and diffusion sampling. Our analysis reveals that the sampling strategy itself plays a central role in safety behavior, as a factor distinct from the underlying learned representations. Motivated by this analysis, we introduce the Step-Wise Refusal Internal Dynamics (SRI) signal, which supports interpretability and improved safety for both AR and DLMs. We demonstrate that the geometric structure of SRI captures internal recovery dynamics, and identifies anomalous behavior in harmful generations as cases of \emph{incomplete internal recovery} that are not observable at the text level. This structure enables lightweight inference-time detectors that generalize to unseen attacks while matching or outperforming existing defenses with over $100\times$ lower inference overhead.

Method: Learn tangent vector fields in ambient space whose flows transport all samples to a common, learnable reference point. Use arc-lengths along these flows as intrinsic coordinates. Prevent degenerate collapse with non-shrinking constraint. Derive scalable, integration-free objective inspired by flow matching. Prove theoretical framework recovers global coordinate chart when one exists.

Result: Method achieves correct tangent alignment and coherent global coordinate structure on synthetic manifolds. Scales to CIFAR-10 where learned coordinates achieve competitive downstream classification performance compared to other methods.

Conclusion: Proposed unsupervised approach successfully constructs global reference systems on data manifolds without assuming flatness, providing interpretable intrinsic coordinates that are theoretically sound and practically scalable to real-world datasets like CIFAR-10.

Abstract: We introduce an unsupervised approach for constructing a global reference system by learning, in the ambient space, vector fields that span the tangent spaces of an unknown data manifold. In contrast to isometric objectives, which implicitly assume manifold flatness, our method learns tangent vector fields whose flows transport all samples to a common, learnable reference point. The resulting arc-lengths along these flows define interpretable intrinsic coordinates tied to a shared global frame. To prevent degenerate collapse, we enforce a non-shrinking constraint and derive a scalable, integration-free objective inspired by flow matching. Within our theoretical framework, we prove that minimizing the proposed objective recovers a global coordinate chart when one exists. Empirically, we obtain correct tangent alignment and coherent global coordinate structure on synthetic manifolds. We also demonstrate the scalability of our method on CIFAR-10, where the learned coordinates achieve competitive downstream classification performance.

Method: Three-step pipeline: (1) Learn embedding function from labeled subset, (2) Perform label-guided clustering over embeddings of both labeled/unlabeled samples to form candidate behavior groups, (3) Use KDE+HDR (highest-density region) containment score to decide if discovered group is truly novel by measuring distribution overlap with known classes.

Result: When entire behavior is withheld from supervision, method recovers distinct cluster and containment score flags novelty via low overlap. Negative-control setting with no novel behavior yields consistently higher overlaps, validating the approach.

Conclusion: HDR-based containment provides practical, quantitative test for generalized class discovery in ecological motion time series under limited annotation and severe class imbalance, enabling discovery of novel animal behaviors.

Abstract: Learning behavioral taxonomies from animal-borne sensors is challenging because labels are scarce, classes are highly imbalanced, and behaviors may be absent from the annotated set. We study generalized behavior discovery in short multivariate motion snippets from gulls, where each sample is a sequence with 3-axis IMU acceleration (20 Hz) and GPS speed, spanning nine expert-annotated behavior categories. We propose a semi-supervised discovery pipeline that (i) learns an embedding function from the labeled subset, (ii) performs label-guided clustering over embeddings of both labeled and unlabeled samples to form candidate behavior groups, and (iii) decides whether a discovered group is truly novel using a containment score. Our key contribution is a KDE + HDR (highest-density region) containment score that measures how much a discovered cluster distribution is contained within, or contains, each known-class distribution; the best-match containment score serves as an interpretable novelty statistic. In experiments where an entire behavior is withheld from supervision and appears only in the unlabeled pool, the method recovers a distinct cluster and the containment score flags novelty via low overlap, while a negative-control setting with no novel behavior yields consistently higher overlaps. These results suggest that HDR-based containment provides a practical, quantitative test for generalized class discovery in ecological motion time series under limited annotation and severe class imbalance.

Method: Mines structured supervision from chain-of-PRs through three mechanisms: (1) progressive task decomposition via continuous commits, (2) long-term consistency enforcement through unified functional objectives, and (3) verifiable refinement from authentic bug-fix trajectories.

Result: Generated substantial trajectories (avg 85k tokens, 116 tool calls) that are data-efficient: fine-tuning GLM-4.6 on 239 samples yields broad improvements, achieving 47% relative gain on Toolathlon benchmark.

Conclusion: PR sequences provide authentic supervision signals for long-horizon learning, enabling better agentic workflows through real-world software evolution patterns.

Abstract: While Large Language Models (LLMs) excel at short-term tasks, scaling them to long-horizon agentic workflows remains challenging. The core bottleneck lies in the scarcity of training data that captures authentic long-dependency structures and cross-stage evolutionary dynamics–existing synthesis methods either confine to single-feature scenarios constrained by model distribution, or incur prohibitive human annotation costs, failing to provide scalable, high-quality supervision. We address this by reconceptualizing data synthesis through the lens of real-world software evolution. Our key insight: Pull Request (PR) sequences naturally embody the supervision signals for long-horizon learning. They decompose complex objectives into verifiable submission units, maintain functional coherence across iterations, and encode authentic refinement patterns through bug-fix histories. Building on this, we propose daVinci-Agency, which systematically mines structured supervision from chain-of-PRs through three interlocking mechanisms: (1) progressive task decomposition via continuous commits, (2) long-term consistency enforcement through unified functional objectives, and (3) verifiable refinement from authentic bug-fix trajectories. Unlike synthetic trajectories that treat each step independently, daVinci-Agency’s PR-grounded structure inherently preserves the causal dependencies and iterative refinements essential for teaching persistent goal-directed behavior and enables natural alignment with project-level, full-cycle task modeling. The resulting trajectories are substantial–averaging 85k tokens and 116 tool calls–yet remarkably data-efficient: fine-tuning GLM-4.6 on 239 daVinci-Agency samples yields broad improvements across benchmarks, notably achieving a 47% relative gain on Toolathlon. Beyond benchmark performance, our analysis confirms…

Method: Proposes a sheaf-theoretic framework where SCMs are Gaussian, restriction maps are transposes of constructive linear causal abstractions adhering to semantic embedding principle, and edge stalks correspond to node stalks of more detailed SCMs. Uses edge-specific local Riemannian problems and avoids nonconvex objectives. Introduces SPECTRAL, an iterative method with closed-form updates suitable for positive definite and semidefinite covariance matrices.

Result: Experiments on synthetic data show competitive performance in the causal abstraction learning task and successful recovery of diverse CAN structures.

Conclusion: The paper presents a novel sheaf-theoretic approach to learning consistent causal abstraction networks with efficient optimization, contributing to causal AI’s explainability and robustness goals.

Abstract: Causal artificial intelligence aims to enhance explainability, trustworthiness, and robustness in AI by leveraging structural causal models (SCMs). In this pursuit, recent advances formalize network sheaves and cosheaves of causal knowledge. Pushing in the same direction, we tackle the learning of consistent causal abstraction network (CAN), a sheaf-theoretic framework where (i) SCMs are Gaussian, (ii) restriction maps are transposes of constructive linear causal abstractions (CAs) adhering to the semantic embedding principle, and (iii) edge stalks correspond–up to permutation–to the node stalks of more detailed SCMs. Our problem formulation separates into edge-specific local Riemannian problems and avoids nonconvex objectives. We propose an efficient search procedure, solving the local problems with SPECTRAL, our iterative method with closed-form updates and suitable for positive definite and semidefinite covariance matrices. Experiments on synthetic data show competitive performance in the CA learning task, and successful recovery of diverse CAN structures.

Method: Propose knowledge distillation from empirically-robust teachers to improve certifiably-robust models using feature-space distillation objectives

Result: Distillation from adversarially-trained teachers consistently improves state-of-the-art certified training for ReLU networks across robust computer vision benchmarks

Conclusion: Leveraging empirically-robust teachers through knowledge distillation effectively bridges the gap between empirical robustness and certified verifiability

Abstract: Adversarial training attains strong empirical robustness to specific adversarial attacks by training on concrete adversarial perturbations, but it produces neural networks that are not amenable to strong robustness certificates through neural network verification. On the other hand, earlier certified training schemes directly train on bounds from network relaxations to obtain models that are certifiably robust, but display sub-par standard performance. Recent work has shown that state-of-the-art trade-offs between certified robustness and standard performance can be obtained through a family of losses combining adversarial outputs and neural network bounds. Nevertheless, differently from empirical robustness, verifiability still comes at a significant cost in standard performance. In this work, we propose to leverage empirically-robust teachers to improve the performance of certifiably-robust models through knowledge distillation. Using a versatile feature-space distillation objective, we show that distillation from adversarially-trained teachers consistently improves on the state-of-the-art in certified training for ReLU networks across a series of robust computer vision benchmarks.

Method: The method involves experimental evaluation of different parallelism techniques (data, intra-operator, inter-operator/pipeline) on homogeneous and heterogeneous GPU clusters using GPT-2 medium/large models with Alpa and Ray frameworks, analyzing network latency impacts.

Result: Results show that Alpa’s execution plans optimizing both intra-operator and inter-operator/pipeline parallelism performed best for geographically distributed GPUs, especially with 10’s of milliseconds network latencies.

Conclusion: The paper proposes a systematic approach for selecting appropriate pretraining techniques to achieve high training performance and reduce GPU usage based on experimental insights.

Abstract: Large language models (LLMs) require enormous computing power to pretrain on massive datasets. When limited datasets are available, smaller-sized LLMs are better choice to pretrain (on user-specified datasets) by following the scaling laws of LLMs. Using pretrained models, vector embeddings can be generated for raw data and stored using vector databases to support modern AI applications and semantic search. In this work, we investigate the performance of pretraining techniques for smaller-sized LLMs on an experimental testbed (with commodity GPUs) available to academic users at no charge. We consider data parallelism, intra-operator parallelism, and inter-operator/pipeline parallelism, and their combinations for pretraining. We set up different GPU clusters with homogeneous and heterogeneous GPU hardware. Furthermore, we investigate the impact of network latency on pretraining performance especially when GPUs are geographically distributed. We used GPT-2 medium and large models and pretrained them using open-source packages, namely, Alpa and Ray. We observed that Alpa’s execution plans that collectively optimized intra-operator and inter-operator/pipeline parallelism consistently performed the best when GPUs were geographically distributed. This was especially true when the network latencies were in 10’s of milliseconds. Based on the insights gained from the experiments, we propose a systematic approach for selecting the appropriate pretraining technique to achieve high training performance/lower execution time as well as to reduce the number of GPUs used.

Method: Introduces a continuous-time model for delayed feedback where regret decomposes into delay-independent learning and delay-induced drift components. Creates a delay-adaptive reduction that converts any online linear optimization algorithm into one handling round-dependent delays. Uses this framework to analyze both first-order and bandit convex optimization settings.

Result: For bandit convex optimization: achieves O(√d_tot + T^{3/4}√k) regret, improving delay-dependent term from O(min{√T d_max, (Td_tot)^{1/3}}) to O(√d_tot). For strongly convex case: achieves O(min{σ_max ln T, √d_tot} + (T^2 ln T)^{1/3} k^{2/3}), improving from O(d_max ln T) to O(min{σ_max ln T, √d_tot}). For first-order feedback: recovers state-of-the-art bounds with simpler analysis.

Conclusion: The reduction framework provides a unified approach to handle delayed feedback in online learning, significantly improving regret bounds for bandit convex optimization and simplifying analysis for first-order optimization. The continuous-time decomposition offers insights into delay effects and enables delay-adaptive algorithms.

Abstract: We develop a reduction-based framework for online learning with delayed feedback that recovers and improves upon existing results for both first-order and bandit convex optimization. Our approach introduces a continuous-time model under which regret decomposes into a delay-independent learning term and a delay-induced drift term, yielding a delay-adaptive reduction that converts any algorithm for online linear optimization into one that handles round-dependent delays. For bandit convex optimization, we significantly improve existing regret bounds, with delay-dependent terms matching state-of-the-art first-order rates. For first-order feedback, we recover state-of-the-art regret bounds via a simpler, unified analysis. Quantitatively, for bandit convex optimization we obtain $O(\sqrt{d_{\text{tot}}} + T^{\frac{3}{4}}\sqrt{k})$ regret, improving the delay-dependent term from $O(\min{\sqrt{T d_{\text{max}}},(Td_{\text{tot}})^{\frac{1}{3}}})$ in previous work to $O(\sqrt{d_{\text{tot}}})$. Here, $k$, $T$, $d_{\text{max}}$, and $d_{\text{tot}}$ denote the dimension, time horizon, maximum delay, and total delay, respectively. Under strong convexity, we achieve $O(\min{σ_{\text{max}} \ln T, \sqrt{d_{\text{tot}}}} + (T^2\ln T)^{\frac{1}{3}} {k}^{\frac{2}{3}})$, improving the delay-dependent term from $O(d_{\text{max}} \ln T)$ in previous work to $O(\min{σ_{\text{max}} \ln T, \sqrt{d_{\text{tot}}}})$, where $σ_{\text{max}}$ denotes the maximum number of outstanding observations and may be considerably smaller than $d_{\text{max}}$.

Method: Two spatially regularized NMF variants: 1) SNMF enforces local spatial smoothness by diffusing each cell’s NMF factor vector over its spatial neighborhood; 2) hSNMF performs spatially regularized NMF followed by Leiden clustering on a hybrid adjacency that integrates spatial proximity and transcriptomic similarity.

Result: On cholangiocarcinoma data, SNMF and hSNMF achieve improved spatial compactness (CHAOS < 0.004, Moran’s I > 0.96), greater cluster separability (Silhouette > 0.12, DBI < 1.8), and higher biological coherence compared to other spatial baselines.

Conclusion: Spatially regularized NMF methods effectively address the challenges of high-dimensional spatial transcriptomics data by incorporating spatial information into representation learning and clustering, leading to more biologically meaningful results.

Abstract: High-resolution spatial transcriptomics platforms, such as Xenium, generate single-cell images that capture both molecular and spatial context, but their extremely high dimensionality poses major challenges for representation learning and clustering. In this study, we analyze data from the Xenium platform, which captures high-resolution images of tumor microarray (TMA) tissues and converts them into cell-by-gene matrices suitable for computational analysis. We benchmark and extend nonnegative matrix factorization (NMF) for spatial transcriptomics by introducing two spatially regularized variants. First, we propose Spatial NMF (SNMF), a lightweight baseline that enforces local spatial smoothness by diffusing each cell’s NMF factor vector over its spatial neighborhood. Second, we introduce Hybrid Spatial NMF (hSNMF), which performs spatially regularized NMF followed by Leiden clustering on a hybrid adjacency that integrates spatial proximity (via a contact-radius graph) and transcriptomic similarity through a tunable mixing parameter alpha. Evaluated on a cholangiocarcinoma dataset, SNMF and hSNMF achieve markedly improved spatial compactness (CHAOS < 0.004, Moran’s I > 0.96), greater cluster separability (Silhouette > 0.12, DBI < 1.8), and higher biological coherence (CMC and enrichment) compared to other spatial baselines. Availability and implementation: https://github.com/ishtyaqmahmud/hSNMF

Method: Introduces Modular Angular-Radial Attention (MARA), which extends spherical attention from SO(3) to SE(3) for molecular domains, operating directly on angular and radial coordinates of neighboring atoms for flexible geometric weighting.

Result: MARA improves energy and force predictions, reduces high-error events, and enhances robustness across molecular benchmarks when integrated into models like MACE.

Conclusion: Continuous spherical attention is an effective geometric operator that increases expressiveness, stability, and reliability of atomistic models in machine learning force fields.

Abstract: Machine learning force fields (MLFFs) have become essential for accurate and efficient atomistic modeling. Despite their high accuracy, most existing approaches rely on fixed angular expansions, limiting flexibility in weighting local geometric interactions. We introduce Modular Angular-Radial Attention (MARA), a module that extends spherical attention – originally developed for SO(3) tasks – to the molecular domain and SE(3), providing an efficient approximation of equivariant interactions. MARA operates directly on the angular and radial coordinates of neighboring atoms, enabling flexible, geometrically informed, and modular weighting of local environments. Unlike existing attention mechanisms in SE(3)-equivariant architectures, MARA can be integrated in a plug-and-play manner into models such as MACE without architectural modifications. Across molecular benchmarks, MARA improves energy and force predictions, reduces high-error events, and enhances robustness. These results demonstrate that continuous spherical attention is an effective and generalizable geometric operator that increases the expressiveness, stability, and reliability of atomistic models.

Method: FlexRank uses low-rank weight decomposition with nested, importance-based consolidation to extract importance-ordered nested components from pretrained models. This creates submodels of increasing capabilities that can be selectively activated based on available computational budget.

Result: Enables “train-once, deploy-everywhere” paradigm with graceful cost-performance trade-offs without training from scratch for each budget, advancing practical deployment of large models.

Conclusion: Importance-ordered nested components can be extracted from pretrained models to enable adaptive deployment across different cost budgets, making large model deployment more practical and flexible.

Abstract: The growing scale of deep neural networks, encompassing large language models (LLMs) and vision transformers (ViTs), has made training from scratch prohibitively expensive and deployment increasingly costly. These models are often used as computational monoliths with fixed cost, a rigidity that does not leverage overparametrized architectures and largely hinders adaptive deployment across different cost budgets. We argue that importance-ordered nested components can be extracted from pretrained models, and selectively activated on the available computational budget. To this end, our proposed FlexRank method leverages low-rank weight decomposition with nested, importance-based consolidation to extract submodels of increasing capabilities. Our approach enables a “train-once, deploy-everywhere” paradigm that offers a graceful trade-off between cost and performance without training from scratch for each budget - advancing practical deployment of large models.

Method: Systematic investigation of DDM routing strategies, analyzing denoising trajectory sensitivity vs. expert-data alignment, using data-cluster distance analysis, per-expert prediction accuracy analysis, and expert disagreement analysis across two distinct DDM systems.

Result: Full ensemble routing achieves most stable sampling dynamics and best numerical convergence but worst generation quality (FID 47.9), while sparse Top-2 routing produces better quality (FID 22.6). Expert-data alignment, not stability, governs generation quality.

Conclusion: For DDM deployment, routing should prioritize expert-data alignment over numerical stability metrics, as generation quality depends on routing inputs to experts whose training distribution covers the current denoising state.

Abstract: Decentralized Diffusion Models (DDMs) route denoising through experts trained independently on disjoint data clusters, which can strongly disagree in their predictions. What governs the quality of generations in such systems? We present the first ever systematic investigation of this question. A priori, the expectation is that minimizing denoising trajectory sensitivity – minimizing how perturbations amplify during sampling – should govern generation quality. We demonstrate this hypothesis is incorrect: a stability-quality dissociation. Full ensemble routing, which combines all expert predictions at each step, achieves the most stable sampling dynamics and best numerical convergence while producing the worst generation quality (FID 47.9 vs. 22.6 for sparse Top-2 routing). Instead, we identify expert-data alignment as the governing principle: generation quality depends on routing inputs to experts whose training distribution covers the current denoising state. Across two distinct DDM systems, we validate expert-data alignment using (i) data-cluster distance analysis, confirming sparse routing selects experts with data clusters closest to the current denoising state, and (ii) per-expert analysis, showing selected experts produce more accurate predictions than non-selected ones, and (iii) expert disagreement analysis, showing quality degrades when experts disagree. For DDM deployment, our findings establish that routing should prioritize expert-data alignment over numerical stability metrics.

Method: Proposes sparsely supervised learning for diffusion models using a simple masking strategy that can mask up to 98% of pixels during training, implemented with only a few lines of code.

Result: The method delivers competitive FID scores, avoids training instability on small datasets, reduces memorization, and promotes the use of essential contextual information during generation.

Conclusion: Sparse masking during diffusion model training effectively addresses global consistency issues while maintaining competitive performance and improving training stability.

Abstract: Diffusion models have shown remarkable success across a wide range of generative tasks. However, they often suffer from spatially inconsistent generation, arguably due to the inherent locality of their denoising mechanisms. This can yield samples that are locally plausible but globally inconsistent. To mitigate this issue, we propose sparsely supervised learning for diffusion models, a simple yet effective masking strategy that can be implemented with only a few lines of code. Interestingly, the experiments show that it is safe to mask up to 98% of pixels during diffusion model training. Our method delivers competitive FID scores across experiments and, most importantly, avoids training instability on small datasets. Moreover, the masking strategy reduces memorization and promotes the use of essential contextual information during generation.

Method: Theoretical analysis combining explicit finite-precision Transformer constructions with communication-complexity lower bounds to prove tight thresholds for expressivity.

Result: For every p, there exists a function Γ (inspired by equality) that a one-layer softmax Transformer can compute with p bits of precision but not with p-1 bits, establishing a sharp “one-bit” threshold.

Conclusion: Tasks requiring equality-like comparisons are especially sensitive to quantization, and precision should be chosen based on the length of equality needed for specific tasks, providing guidance for practitioners.

Abstract: Quantization reduces the numerical precision of Transformer computations and is widely used to accelerate inference, yet its effect on expressivity remains poorly characterized. We demonstrate a fine-grained theoretical tradeoff between expressivity and precision: For every p we exhibit a function Γ, inspired by the equality function, and prove that a one-layer softmax Transformer can compute Γ, with p bits of precision, but not with p-1 bits of precision. This result concretely explains the widely observed phenomenon of empirical loss of expressivity when quantization is used. Practically, it suggests that tasks requiring equality-like comparisons (exact match, membership, etc.) are especially sensitive to quantization. Dropping even one bit can cross a threshold where the model cannot represent the needed comparison reliably. Thus, it paves the way for developing heuristics that will help practitioners choose how much quantization is possible: the precision should be chosen as a function of the length of equality to be checked for the specific task. Our proofs combine explicit finite-precision Transformer constructions with communication-complexity lower bounds, yielding a tight “one-bit” threshold.

Method: BinaryPPO uses offline reinforcement learning with a variant of Proximal Policy Optimization (PPO) and a confidence-weighted reward function that penalizes uncertain or incorrect predictions, learning robust decision policies from static datasets without online interaction.

Result: Across eight domain-specific benchmarks and multiple model architectures, BinaryPPO improves accuracy by 40-60 percentage points, reaching up to 99%, substantially outperforming supervised baselines.

Conclusion: Confidence-based reward design provides a robust alternative to SFT for binary classification, with BinaryPPO demonstrating significant improvements in challenging real-world settings.

Abstract: Supervised fine-tuning (SFT) is the standard approach for binary classification tasks such as toxicity detection, factuality verification, and causal inference. However, SFT often performs poorly in real-world settings with label noise, class imbalance, or sparse supervision. We introduce BinaryPPO, an offline reinforcement learning large language model (LLM) framework that reformulates binary classification as a reward maximization problem. Our method leverages a variant of Proximal Policy Optimization (PPO) with a confidence-weighted reward function that penalizes uncertain or incorrect predictions, enabling the model to learn robust decision policies from static datasets without online interaction. Across eight domain-specific benchmarks and multiple models with differing architectures, BinaryPPO improves accuracy by 40-60 percentage points, reaching up to 99%, substantially outperforming supervised baselines. We provide an in-depth analysis of the role of reward shaping, advantage scaling, and policy stability in enabling this improvement. Overall, we demonstrate that confidence-based reward design provides a robust alternative to SFT for binary classification. Our code is available at https://github.com/psyonp/BinaryPPO.

Method: Introduces Maximum Likelihood Reinforcement Learning (MaxRL) which defines a compute-indexed family of sample-based objectives that interpolate between standard RL and exact maximum likelihood. Uses a simple, unbiased policy-gradient estimator that converges to maximum likelihood optimization as compute increases.

Result: MaxRL Pareto-dominates existing methods across all tested models and tasks, achieving up to 20x test-time scaling efficiency gains compared to GRPO-trained counterparts. Also shows better scaling with additional data and compute.

Conclusion: MaxRL is a promising framework for scaling RL training in correctness-based settings, offering a principled approach to approximate maximum likelihood optimization using reinforcement learning techniques.

Abstract: Reinforcement learning is the method of choice to train models in sampling-based setups with binary outcome feedback, such as navigation, code generation, and mathematical problem solving. In such settings, models implicitly induce a likelihood over correct rollouts. However, we observe that reinforcement learning does not maximize this likelihood, and instead optimizes only a lower-order approximation. Inspired by this observation, we introduce Maximum Likelihood Reinforcement Learning (MaxRL), a sampling-based framework to approximate maximum likelihood using reinforcement learning techniques. MaxRL addresses the challenges of non-differentiable sampling by defining a compute-indexed family of sample-based objectives that interpolate between standard reinforcement learning and exact maximum likelihood as additional sampling compute is allocated. The resulting objectives admit a simple, unbiased policy-gradient estimator and converge to maximum likelihood optimization in the infinite-compute limit. Empirically, we show that MaxRL Pareto-dominates existing methods in all models and tasks we tested, achieving up to 20x test-time scaling efficiency gains compared to its GRPO-trained counterpart. We also observe MaxRL to scale better with additional data and compute. Our results suggest MaxRL is a promising framework for scaling RL training in correctness based settings.

Method: Theoretical analysis characterizing the effect of steering strength on next token probability, concept presence, and cross-entropy, deriving precise qualitative laws, followed by empirical validation on eleven language models.

Result: Reveals surprising non-monotonic effects of steering strength and provides quantitative laws governing steering behavior, validated empirically across diverse LLM architectures.

Conclusion: Provides the first theoretical framework for understanding steering strength in LLM control, offering practical guidance for optimal steering magnitude selection.

Abstract: A popular approach to post-training control of large language models (LLMs) is the steering of intermediate latent representations. Namely, identify a well-chosen direction depending on the task at hand and perturbs representations along this direction at inference time. While many propositions exist to pick this direction, considerably less is understood about how to choose the magnitude of the move, whereas its importance is clear: too little and the intended behavior does not emerge, too much and the model’s performance degrades beyond repair. In this work, we propose the first theoretical analysis of steering strength. We characterize its effect on next token probability, presence of a concept, and cross-entropy, deriving precise qualitative laws governing these quantities. Our analysis reveals surprising behaviors, including non-monotonic effects of steering strength. We validate our theoretical predictions empirically on eleven language models, ranging from a small GPT architecture to modern models.

Method: Neural probabilistic amplitude shaping framework that learns joint distributions for coherent fiber systems, applied to dual-polarized 64-QAM transmission

Result: Achieves 0.5 dB signal-to-noise ratio gain over sequence selection methods across a single-span 205 km fiber link

Conclusion: The neural probabilistic amplitude shaping framework provides significant SNR improvements for coherent fiber communication systems

Abstract: We introduce neural probabilistic amplitude shaping, a joint-distribution learning framework for coherent fiber systems. The proposed scheme provides a 0.5 dB signal-to-noise ratio gain over sequence selection for dual-polarized 64-QAM transmission across a single-span 205 km link.

Method: Two-level hierarchy: value-based GCRL agent + factored subgoal-generating conditional diffusion model. They’re trained independently and composed post hoc through selective subgoal generation based on value function.

Result: Method boosts performance on image-based long-horizon tasks with sparse rewards, achieving over 150% higher success rates on hardest tasks and generalizing to increasing horizons and entity counts.

Conclusion: The hierarchical entity-centric framework effectively addresses combinatorial complexity in multi-entity domains, improving GCRL performance through modular subgoal decomposition.

Abstract: We propose a hierarchical entity-centric framework for offline Goal-Conditioned Reinforcement Learning (GCRL) that combines subgoal decomposition with factored structure to solve long-horizon tasks in domains with multiple entities. Achieving long-horizon goals in complex environments remains a core challenge in Reinforcement Learning (RL). Domains with multiple entities are particularly difficult due to their combinatorial complexity. GCRL facilitates generalization across goals and the use of subgoal structure, but struggles with high-dimensional observations and combinatorial state-spaces, especially under sparse reward. We employ a two-level hierarchy composed of a value-based GCRL agent and a factored subgoal-generating conditional diffusion model. The RL agent and subgoal generator are trained independently and composed post hoc through selective subgoal generation based on the value function, making the approach modular and compatible with existing GCRL algorithms. We introduce new variations to benchmark tasks that highlight the challenges of multi-entity domains, and show that our method consistently boosts performance of the underlying RL agent on image-based long-horizon tasks with sparse rewards, achieving over 150% higher success rates on the hardest task in our suite and generalizing to increasing horizons and numbers of entities. Rollout videos are provided at: https://sites.google.com/view/hecrl

Method: Proposes Vector Quantized Latent Concept (VQLC) method built on VQ-VAE architecture to learn a discrete codebook that maps continuous representations to concept vectors for scalable concept-based explanations.

Result: VQLC improves scalability while maintaining comparable quality of human-understandable explanations compared to existing methods.

Conclusion: VQLC provides a scalable framework for concept-based explanation of deep learning models while preserving explanation quality.

Abstract: Deep Learning models encode rich semantic information in their hidden representations. However, it remains challenging to understand which parts of this information models actually rely on when making predictions. A promising line of post-hoc concept-based explanation methods relies on clustering token representations. However, commonly used approaches such as hierarchical clustering are computationally infeasible for large-scale datasets, and K-Means often yields shallow or frequency-dominated clusters. We propose the vector quantized latent concept (VQLC) method, a framework built upon the vector quantized-variational autoencoder (VQ-VAE) architecture that learns a discrete codebook mapping continuous representations to concept vectors. We perform thorough evaluations and show that VQLC improves scalability while maintaining comparable quality of human-understandable explanations.

Method: SearchDiff is a training-free neurosymbolic framework that combines neural denoising with symbolic search. At each denoising step, model predictions define a proposal set that is optimized under user-specified constraints, modifying the reverse transition to steer sampling toward feasible solutions.

Result: Experiments in biological design and symbolic reasoning show SearchDiff substantially improves constraint satisfaction and property adherence while outperforming both discrete diffusion and autoregressive baselines.

Conclusion: SearchDiff successfully bridges neural generation with symbolic reasoning, enabling discrete diffusion models to handle hard constraints and optimize non-differentiable properties without retraining.

Abstract: Discrete diffusion models generate sequences by iteratively denoising samples corrupted by categorical noise, offering an appealing alternative to autoregressive decoding for structured and symbolic generation. However, standard training targets a likelihood-based objective that primarily matches the data distribution and provides no native mechanism for enforcing hard constraints or optimizing non-differentiable properties at inference time. This work addresses this limitation and introduces Search-Augmented Masked Diffusion (SearchDiff), a training-free neurosymbolic inference framework that integrates informed search directly into the reverse denoising process. At each denoising step, the model predictions define a proposal set that is optimized under a user-specified property satisfaction, yielding a modified reverse transition that steers sampling toward probable and feasible solutions. Experiments in biological design and symbolic reasoning illustrate that SearchDiff substantially improves constraint satisfaction and property adherence, while consistently outperforming discrete diffusion and autoregressive baselines.

Method: CAPS combines SO(2) rotations for phase alignment with three additive gating paths: Riemann softmax, prefix-product gates, and a Clock baseline. The Clock mechanism provides learned temporal weighting that modulates these paths through a shared notion of temporal importance, enabling decoupling of global trends, local shocks, and seasonal patterns.

Result: Experiments on long- and short-term forecasting benchmarks show CAPS surpasses vanilla softmax and linear attention mechanisms, and demonstrates competitive performance against seven strong baselines while maintaining linear complexity.

Conclusion: CAPS provides an effective structured attention mechanism for time series forecasting that properly separates different temporal structures while maintaining computational efficiency with linear complexity.

Abstract: This paper presents $\textbf{CAPS}$ (Clock-weighted Aggregation with Prefix-products and Softmax), a structured attention mechanism for time series forecasting that decouples three distinct temporal structures: global trends, local shocks, and seasonal patterns. Standard softmax attention entangles these through global normalization, while recent recurrent models sacrifice long-term, order-independent selection for order-dependent causal structure. CAPS combines SO(2) rotations for phase alignment with three additive gating paths – Riemann softmax, prefix-product gates, and a Clock baseline – within a single attention layer. We introduce the Clock mechanism, a learned temporal weighting that modulates these paths through a shared notion of temporal importance. Experiments on long- and short-term forecasting benchmarks surpass vanilla softmax and linear attention mechanisms and demonstrate competitive performance against seven strong baselines with linear complexity. Our code implementation is available at https://github.com/vireshpati/CAPS-Attention.

Method: Uses closed-loop optimization where policy is trained with integrated step-wise and outcome feedback, reward model is optimized via consistency feedback, and environment is automatically adapted using critic feedback from both policy and reward models.

Result: Substantial gains across various LLM and agentic tasks: 9.1% improvement for Qwen3-VL-8B-Thinking on OSWorld, 18.7% and 11.9% improvements for Qwen2.5-7B-Instruct on AlfWorld and LiveBench respectively.

Conclusion: RLAnything’s closed-loop optimization of all RL components consistently improves system performance, with optimized reward signals outperforming human-labeled outcomes.

Abstract: We propose RLAnything, a reinforcement learning framework that dynamically forges environment, policy, and reward models through closed-loop optimization, amplifying learning signals and strengthening the overall RL system for any LLM or agentic scenarios. Specifically, the policy is trained with integrated feedback from step-wise and outcome signals, while the reward model is jointly optimized via consistency feedback, which in turn further improves policy training. Moreover, our theory-motivated automatic environment adaptation improves training for both the reward and policy models by leveraging critic feedback from each, enabling learning from experience. Empirically, each added component consistently improves the overall system, and RLAnything yields substantial gains across various representative LLM and agentic tasks, boosting Qwen3-VL-8B-Thinking by 9.1% on OSWorld and Qwen2.5-7B-Instruct by 18.7% and 11.9% on AlfWorld and LiveBench, respectively. We also that optimized reward-model signals outperform outcomes that rely on human labels. Code: https://github.com/Gen-Verse/Open-AgentRL

Method: Proposes a zero-shot generation framework based on TabPFN that generates design parameters sequentially conditioned on target performance indicators, using only limited reference samples without task-specific training or fine-tuning.

Result: Achieves competitive diversity across structured parametric design spaces, robust to sampling/resolution variations, low performance error (<2% for ship hulls), and significantly reduces computational overhead compared to diffusion models.

Conclusion: Zero-shot, data-efficient generation enables practical engineering design with rapid deployment, flexible adaptation to new settings, and easy integration into real-world workflows.

Abstract: Deep generative models for engineering design often require substantial computational cost, large training datasets, and extensive retraining when design requirements or datasets change, limiting their applicability in real-world engineering design workflow. In this work, we propose a zero-shot generation framework for parametric engineering design based on TabPFN, enabling conditional design generation using only a limited number of reference samples and without any task-specific model training or fine-tuning. The proposed method generates design parameters sequentially conditioned on target performance indicators, providing a flexible alternative to conventional generative models. The effectiveness of the proposed approach is evaluated on three engineering design datasets, i.e., ship hull design, BlendedNet aircraft, and UIUC airfoil. Experimental results demonstrate that the proposed method achieves competitive diversity across highly structured parametric design spaces, remains robust to variations in sampling, resolution and parameter dimensionality of geometry generation, and achieves a low performance error (e.g., less than 2% in generated ship hull designs’ performance). Compared with diffusion-based generative models, the proposed framework significantly reduces computational overhead and data requirements while preserving reliable generation performance. These results highlight the potential of zero-shot, data-efficient generation as a practical and efficient tool for engineering design, enabling rapid deployment, flexible adaptation to new design settings, and ease of integration into real-world engineering workflows.

Method: Two-scale topological approach: (1) topology-aware manifold approximation for global low-dimensional embedding, (2) differentiable persistent homology for local topological optimization to rank samples by structural complexity.

Result: TopoPrune achieves high accuracy and precision at significant pruning rates (90%), shows exceptional robustness to noise in latent features, and demonstrates superior transferability across diverse network architectures.

Conclusion: Topology provides a stable, principled framework for robust data-efficient learning, with TopoPrune offering promising results for stable data pruning across architectures.

Abstract: Geometric data pruning methods, while practical for leveraging pretrained models, are fundamentally unstable. Their reliance on extrinsic geometry renders them highly sensitive to latent space perturbations, causing performance to degrade during cross-architecture transfer or in the presence of feature noise. We introduce TopoPrune, a framework which resolves this challenge by leveraging topology to capture the stable, intrinsic structure of data. TopoPrune operates at two scales, (1) utilizing a topology-aware manifold approximation to establish a global low-dimensional embedding of the dataset. Subsequently, (2) it employs differentiable persistent homology to perform a local topological optimization on the manifold embeddings, ranking samples by their structural complexity. We demonstrate that our unified dual-scale topological approach ensures high accuracy and precision, particularly at significant dataset pruning rates (e.g., 90%). Furthermore, through the inherent stability properties of topology, TopoPrune is (a) exceptionally robust to noise perturbations of latent feature embeddings and (b) demonstrates superior transferability across diverse network architectures. This study demonstrates a promising avenue towards stable and principled topology-based frameworks for robust data-efficient learning.

Method: EDT-Former generates dynamic tokens aligned with informative molecular patches using entropy guidance, preserving both local and global structural features. It enables alignment between frozen graph encoders and LLMs without tuning the LLM backbone (excluding embedding layer).

Result: Achieves state-of-the-art results on MoleculeQA, Molecule-oriented Mol-Instructions, and property prediction benchmarks (TDC, MoleculeNet), demonstrating effectiveness for scalable and generalizable multimodal molecular understanding.

Conclusion: EDT-Former provides an effective approach for molecular graph understanding by generating dynamic tokens aligned with informative molecular patches, enabling efficient alignment between graph encoders and LLMs without backbone tuning.

Abstract: Molecular understanding is central to advancing areas such as scientific discovery, yet Large Language Models (LLMs) struggle to understand molecular graphs effectively. Existing graph-LLM bridges often adapt the Q-Former-style connector with fixed-length static tokens, which is originally designed for vision tasks. These designs overlook stereochemistry and substructural context and typically require costly LLM-backbone fine-tuning, limiting efficiency and generalization. We introduce EDT-Former, an Entropy-guided Dynamic Token Transformer that generates tokens aligned with informative molecular patches, thereby preserving both local and global structural features for molecular graph understanding. Beyond prior approaches, EDT-Former enables alignment between frozen graph encoders and LLMs without tuning the LLM backbone (excluding the embedding layer), resulting in computationally efficient finetuning, and achieves stateof-the-art results on MoleculeQA, Molecule-oriented Mol-Instructions, and property prediction benchmarks (TDC, MoleculeNet), underscoring its effectiveness for scalable and generalizable multimodal molecular understanding

Method: Analyzes IDM-based policies theoretically and empirically, comparing VM-IDM and IDM labeling approaches. Uses statistical learning theory insights and experiments with unified video-action prediction (UVA) architectures. Proposes improved LAPO algorithm for latent action policy learning.

Result: Shows that IDM-based policies have advantages due to: (1) ground-truth IDM being in lower complexity hypothesis class than expert policy, and (2) ground-truth IDM being less stochastic than expert policy. Demonstrates these claims through theory and experiments.

Conclusion: IDM-based methods in SSIL outperform behavior cloning primarily due to better sample efficiency from IDM’s lower complexity and reduced stochasticity, leading to improved policy learning with limited labeled data.

Abstract: Semi-supervised imitation learning (SSIL) consists in learning a policy from a small dataset of action-labeled trajectories and a much larger dataset of action-free trajectories. Some SSIL methods learn an inverse dynamics model (IDM) to predict the action from the current state and the next state. An IDM can act as a policy when paired with a video model (VM-IDM) or as a label generator to perform behavior cloning on action-free data (IDM labeling). In this work, we first show that VM-IDM and IDM labeling learn the same policy in a limit case, which we call the IDM-based policy. We then argue that the previously observed advantage of IDM-based policies over behavior cloning is due to the superior sample efficiency of IDM learning, which we attribute to two causes: (i) the ground-truth IDM tends to be contained in a lower complexity hypothesis class relative to the expert policy, and (ii) the ground-truth IDM is often less stochastic than the expert policy. We argue these claims based on insights from statistical learning theory and novel experiments, including a study of IDM-based policies using recent architectures for unified video-action prediction (UVA). Motivated by these insights, we finally propose an improved version of the existing LAPO algorithm for latent action policy learning.

Method: Proposes TSEF (Time Series Explanation Fooler), a dual-target attack that jointly manipulates classifier and explainer outputs to achieve targeted misclassification while keeping explanations consistent with a chosen reference rationale.

Result: Across multiple datasets and explainer backbones, TSEF successfully achieves targeted prediction changes while maintaining explanation consistency, revealing that explanation stability is a misleading proxy for decision robustness.

Conclusion: Explanation stability alone is insufficient for trustworthy time series tasks; coupling-aware robustness evaluations are needed to ensure both prediction and explanation reliability.

Abstract: Interpretable time series deep learning systems are often assessed by checking temporal consistency on explanations, implicitly treating this as evidence of robustness. We show that this assumption can fail: Predictions and explanations can be adversarially decoupled, enabling targeted misclassification while the explanation remains plausible and consistent with a chosen reference rationale. We propose TSEF (Time Series Explanation Fooler), a dual-target attack that jointly manipulates the classifier and explainer outputs. In contrast to single-objective misclassification attacks that disrupt explanation and spread attribution mass broadly, TSEF achieves targeted prediction changes while keeping explanations consistent with the reference. Across multiple datasets and explainer backbones, our results consistently reveal that explanation stability is a misleading proxy for decision robustness and motivate coupling-aware robustness evaluations for trustworthy time series tasks.

Method: PATH treats entire tables as synthesis units and leverages autoregressive capabilities of privately fine-tuned large language models to capture long-range dependencies and temporal patterns in sequential data.

Result: PATH reduces distributional distance to real trajectories by over 60% and state transition errors by nearly 50% compared to leading marginal mechanisms while achieving similar marginal fidelity.

Conclusion: The framework effectively addresses temporal coherence in differentially private synthetic data generation for longitudinal datasets, demonstrating superior performance over traditional flattening approaches.

Abstract: Research on differentially private synthetic tabular data has largely focused on independent and identically distributed rows where each record corresponds to a unique individual. This perspective neglects the temporal complexity in longitudinal datasets, such as electronic health records, where a user contributes an entire (sub) table of sequential events. While practitioners might attempt to model such data by flattening user histories into high-dimensional vectors for use with standard marginal-based mechanisms, we demonstrate that this strategy is insufficient. Flattening fails to preserve temporal coherence even when it maintains valid marginal distributions. We introduce PATH, a novel generative framework that treats the full table as the unit of synthesis and leverages the autoregressive capabilities of privately fine-tuned large language models. Extensive evaluations show that PATH effectively captures long-range dependencies that traditional methods miss. Empirically, our method reduces the distributional distance to real trajectories by over 60% and reduces state transition errors by nearly 50% compared to leading marginal mechanisms while achieving similar marginal fidelity.

Method: Adopts a multi-view data model and identifies signal-to-noise ratio (SNR) as critical factor affecting forgetting. Focuses on task-incremental binary classification across M tasks, analyzing the interplay between signal learning and noise memorization.

Result: Two key findings: (1) forgetting can still occur under full replay when cumulative noise from later tasks dominates earlier task signals; (2) with sufficient signal accumulation, data replay can recover earlier tasks even if initial learning was poor. Also discovers task ordering insight: prioritizing higher-signal tasks facilitates learning of lower-signal tasks and prevents catastrophic forgetting.

Conclusion: Provides theoretical framework showing that full data replay effectiveness depends on signal-to-noise ratio dynamics, with task ordering playing crucial role in preventing catastrophic forgetting through signal accumulation.

Abstract: Continual learning (CL) aims to train models on a sequence of tasks while retaining performance on previously learned ones. A core challenge in this setting is catastrophic forgetting, where new learning interferes with past knowledge. Among various mitigation strategies, data-replay methods, where past samples are periodically revisited, are considered simple yet effective, especially when memory constraints are relaxed. However, the theoretical effectiveness of full data replay, where all past data is accessible during training, remains largely unexplored. In this paper, we present a comprehensive theoretical framework for analyzing full data-replay training in continual learning from a feature learning perspective. Adopting a multi-view data model, we identify the signal-to-noise ratio (SNR) as a critical factor affecting forgetting. Focusing on task-incremental binary classification across $M$ tasks, our analysis verifies two key conclusions: (1) forgetting can still occur under full replay when the cumulative noise from later tasks dominates the signal from earlier ones; and (2) with sufficient signal accumulation, data replay can recover earlier tasks-even if their initial learning was poor. Notably, we uncover a novel insight into task ordering: prioritizing higher-signal tasks not only facilitates learning of lower-signal tasks but also helps prevent catastrophic forgetting. We validate our theoretical findings through synthetic and real-world experiments that visualize the interplay between signal learning and noise memorization across varying SNRs and task correlation regimes.

Method: BTCNet uses a multimodal self-supervised learning framework that: 1) Incorporates temporal information about when each segment occurs within longer recordings, 2) Learns pairwise interactions between physiological signals via cross-attention mechanisms, 3) Operates without task labels or sequence-level supervision, and 4) Can be applied to various downstream tasks through frozen-backbone linear probing.

Result: BTCNet consistently outperforms non-time-aware variants across six pediatric sleep tasks including sleep staging, arousal detection, and respiratory event detection. The gains generalize to an independent pediatric dataset. Compared to existing multimodal self-supervised sleep models, BTCNet achieves strong performance, particularly on respiration-related tasks.

Conclusion: Incorporating temporal context and cross-modal attention in self-supervised learning improves performance on physiological signal analysis tasks, with particular benefits for respiration-related applications in sleep studies.

Abstract: We present BiTimeCrossNet (BTCNet), a multimodal self-supervised learning framework for long physiological recordings such as overnight sleep studies. While many existing approaches train on short segments treated as independent samples, BTCNet incorporates information about when each segment occurs within its parent recording, for example within a sleep session. BTCNet further learns pairwise interactions between physiological signals via cross-attention, without requiring task labels or sequence-level supervision. We evaluate BTCNet on pediatric sleep data across six downstream tasks, including sleep staging, arousal detection, and respiratory event detection. Under frozen-backbone linear probing, BTCNet consistently outperforms an otherwise identical non-time-aware variant, with gains that generalize to an independent pediatric dataset. Compared to existing multimodal self-supervised sleep models, BTCNet achieves strong performance, particularly on respiration-related tasks.

Method: Proposes TraceNAS, a training-free NAS framework that jointly explores structured pruning of LLM depth and width. Uses a scale-invariant zero-shot proxy to identify pruned models that maintain high loss landscape alignment with pretrained models, selecting models with maximal performance potential during post-pruning training.

Result: TraceNAS achieves high-fidelity discovery of pruned models on a single GPU in 8.5 hours (10× reduction in GPU-hours compared to training-aware methods). Evaluations on Llama and Qwen families show competitive performance with training-aware baselines across commonsense and reasoning benchmarks.

Conclusion: TraceNAS provides an efficient training-free approach for structured pruning of LLMs that captures global dependencies and maintains loss landscape alignment, offering computational efficiency while achieving competitive performance with training-aware methods.

Abstract: Structured pruning is essential for efficient deployment of Large Language Models (LLMs). The varying sensitivity of LLM sub-blocks to pruning necessitates the identification of optimal non-uniformly pruned models. Existing methods evaluate the importance of layers, attention heads, or weight channels in isolation. Such localized focus ignores the complex global structural dependencies that exist across the model. Training-aware structured pruning addresses global dependencies, but its computational cost can be just as expensive as post-pruning training. To alleviate the computational burden of training-aware pruning and capture global structural dependencies, we propose TraceNAS, a training-free Neural Architecture Search (NAS) framework that jointly explores structured pruning of LLM depth and width. TraceNAS identifies pruned models that maintain a high degree of loss landscape alignment with the pretrained model using a scale-invariant zero-shot proxy, effectively selecting models that exhibit maximal performance potential during post-pruning training. TraceNAS is highly efficient, enabling high-fidelity discovery of pruned models on a single GPU in 8.5 hours, yielding a 10$\times$ reduction in GPU-hours compared to training-aware methods. Evaluations on the Llama and Qwen families demonstrate that TraceNAS is competitive with training-aware baselines across commonsense and reasoning benchmarks.

Method: Used MLP-based embedding networks on tabular ECG features, then adopted ArcFace embedding models on both features and raw ECG waveforms, with consistent normalization and large training sets. Implemented two-stage pipeline for open-set identification.

Result: Achieved high verification performance (TAR=0.908 @ FAR=1e-3; EER=2.53%), strong closed-set identification (Rank@1=0.812), and excellent open-set identification (DIR@FAR up to 0.976). Showed performance improves with waveforms vs features and larger datasets.

Conclusion: ECG carries measurable individual signatures that enable re-identification even with tabular features, with deep learning models further amplifying this capability, necessitating serious consideration of privacy implications and operational protocols.

Abstract: This work studies electrocardiogram (ECG) biometrics at large scale, evaluating how strongly an ECG can be linked to an individual and, consequently, how its anonymization may be compromised. We show that identity information is already present in tabular representations (fiducial features): even a simple MLP-based embedding network yields non-trivial performance, indicating that anonymization based solely on releasing features does not guarantee privacy. We then adopt embedding-based deep learning models (ArcFace), first on features and then on ECG waveforms, showing a performance jump when moving from tabular inputs to waveforms, and a further gain with larger training sets and consistent normalization across train/val/test. On a large-scale test set, verification achieves high TAR at strict FAR thresholds (TAR=0.908 @ FAR=1e-3; TAR=0.820 @ FAR=1e-4) with EER=2.53% (all-vs-all); closed-set identification yields Rank@1=0.812 and Rank@10=0.910. In open-set, a two-stage pipeline (top-K shortlist on embeddings + re-ranking) reaches DIR@FAR up to 0.976 at FAR=1e-3 and 1e-4. Overall, the results show that ECG carries a measurable individual signature: re-identification is already possible with tabular features and is further amplified by embedding-based models, making privacy implications and realistic operational protocols essential to consider.

Method: Proposes Cross-Temporal Attention Fusion (CTAF) with: 1) time-aware cross attention for soft bidirectional alignments, 2) lightweight fusion gate for robust clip embeddings, and 3) alignment-regularized contrastive objectives with optional weak supervision.

Result: On K-EmoCon dataset, CTAF yields higher cosine margins for matched pairs, better cross-modal token retrieval within one second, and competitive performance on three-bin accuracy and macro-F1 while using few labels.

Conclusion: CTAF represents a step toward label-efficient, generalizable EEG-peripheral fusion under temporal asynchrony by directly modeling correspondence between modalities and accounting for nervous system coupling.

Abstract: We study multimodal affect modeling when EEG and peripheral physiology are asynchronous, which most fusion methods ignore or handle with costly warping. We propose Cross-Temporal Attention Fusion (CTAF), a self-supervised module that learns soft bidirectional alignments between modalities and builds a robust clip embedding using time-aware cross attention, a lightweight fusion gate, and alignment-regularized contrastive objectives with optional weak supervision. On the K-EmoCon dataset, under leave-one-out cross-validation evaluation, CTAF yields higher cosine margins for matched pairs and better cross-modal token retrieval within one second, and it is competitive with the baseline on three-bin accuracy and macro-F1 while using few labels. Our contributions are a time-aware fusion mechanism that directly models correspondence, an alignment-driven self-supervised objective tailored to EEG and physiology, and an evaluation protocol that measures alignment quality itself. Our approach accounts for the coupling between the central and autonomic nervous systems in psychophysiological time series. These results indicate that CTAF is a strong step toward label-efficient, generalizable EEG-peripheral fusion under temporal asynchrony.

Method: LEMON fits a single modality-aware surrogate with group-structured sparsity to produce unified explanations, treating the predictor as a black box and requiring relatively few forward passes while remaining faithful under repeated perturbations.

Result: LEMON achieves competitive deletion-based faithfulness while reducing black-box evaluations by 35-67 times and runtime by 2-8 times compared to strong multimodal baselines across vision-language question answering and clinical prediction tasks.

Conclusion: LEMON provides an efficient, model-agnostic framework for local explanations of multimodal predictions that successfully disentangles modality-level contributions and feature-level attributions while being computationally practical for large-scale use.

Abstract: Multimodal models are ubiquitous, yet existing explainability methods are often single-modal, architecture-dependent, or too computationally expensive to run at scale. We introduce LEMON (Local Explanations via Modality-aware OptimizatioN), a model-agnostic framework for local explanations of multimodal predictions. LEMON fits a single modality-aware surrogate with group-structured sparsity to produce unified explanations that disentangle modality-level contributions and feature-level attributions. The approach treats the predictor as a black box and is computationally efficient, requiring relatively few forward passes while remaining faithful under repeated perturbations. We evaluate LEMON on vision-language question answering and a clinical prediction task with image, text, and tabular inputs, comparing against representative multimodal baselines. Across backbones, LEMON achieves competitive deletion-based faithfulness while reducing black-box evaluations by 35-67 times and runtime by 2-8 times compared to strong multimodal baselines.

Method: SAGE injects privileged hints during training where the model samples a compact hint (e.g., plan or decomposition) and generates solutions conditioned on both prompt and hint. This increases within-group outcome diversity under finite sampling while keeping the same terminal reward unchanged. At test time, the model uses no hints.

Result: Experiments across 6 benchmarks with 3 LLMs show SAGE consistently outperforms GRPO, with average improvements of +2.0 on Llama-3.2-3B-Instruct, +1.2 on Qwen2.5-7B-Instruct, and +1.3 on Qwen3-4B-Instruct.

Conclusion: SAGE effectively addresses GRPO’s limitation with sparse rewards by using self-hints to increase rollout diversity, preventing advantage collapse while maintaining the same verifiable reward structure, leading to consistent performance improvements across multiple LLMs.

Abstract: Group Relative Policy Optimization (GRPO) has recently emerged as a practical recipe for aligning large language models with verifiable objectives. However, under sparse terminal rewards, GRPO often stalls because rollouts within a group frequently receive identical rewards, causing relative advantages to collapse and updates to vanish. We propose self-hint aligned GRPO with privileged supervision (SAGE), an on-policy reinforcement learning framework that injects privileged hints during training to reshape the rollout distribution under the same terminal verifier reward. For each prompt $x$, the model samples a compact hint $h$ (e.g., a plan or decomposition) and then generates a solution $τ$ conditioned on $(x,h)$. Crucially, the task reward $R(x,τ)$ is unchanged; hints only increase within-group outcome diversity under finite sampling, preventing GRPO advantages from collapsing under sparse rewards. At test time, we set $h=\varnothing$ and deploy the no-hint policy without any privileged information. Moreover, sampling diverse self-hints serves as an adaptive curriculum that tracks the learner’s bottlenecks more effectively than fixed hints from an initial policy or a stronger external model. Experiments over 6 benchmarks with 3 LLMs show that SAGE consistently outperforms GRPO, on average +2.0 on Llama-3.2-3B-Instruct, +1.2 on Qwen2.5-7B-Instruct and +1.3 on Qwen3-4B-Instruct. The code is available at https://github.com/BaohaoLiao/SAGE.

Method: General-Geometry Neural Whitney Forms (Geo-NeW): data-driven finite element method jointly learning differential operator and compatible reduced finite element spaces defined on underlying geometry. Uses transformer-based encoding of discretized mesh and finite element exterior calculus to preserve physical conservation laws exactly.

Result: State-of-the-art performance on several steady-state PDE benchmarks and significant improvement over conventional baselines on out-of-distribution geometries.

Conclusion: Geo-NeW provides a powerful inductive bias for learning neural PDEs by explicitly connecting underlying geometry and boundary conditions to solutions, enabling better generalization to unseen domains while preserving physical structure.

Abstract: We aim to develop physics foundation models for science and engineering that provide real-time solutions to Partial Differential Equations (PDEs) which preserve structure and accuracy under adaptation to unseen geometries. To this end, we introduce General-Geometry Neural Whitney Forms (Geo-NeW): a data-driven finite element method. We jointly learn a differential operator and compatible reduced finite element spaces defined on the underlying geometry. The resulting model is solved to generate predictions, while exactly preserving physical conservation laws through Finite Element Exterior Calculus. Geometry enters the model as a discretized mesh both through a transformer-based encoding and as the basis for the learned finite element spaces. This explicitly connects the underlying geometry and imposed boundary conditions to the solution, providing a powerful inductive bias for learning neural PDEs, which we demonstrate improves generalization to unseen domains. We provide a novel parameterization of the constitutive model ensuring the existence and uniqueness of the solution. Our approach demonstrates state-of-the-art performance on several steady-state PDE benchmarks, and provides a significant improvement over conventional baselines on out-of-distribution geometries.

Method: Proposes DynSplit-KV with: (1) dynamic importance-aware delimiter selection strategy to identify semantic boundaries, improving accuracy by 49.9%; (2) uniform mapping strategy to transform variable-length semantic blocks into fixed-length format, reducing inference overhead.

Result: Achieves highest accuracy, 2.2x speedup compared to FlashAttention, and 2.6x peak memory reduction in long-context scenarios. Reduces inference overhead by 4.9x.

Conclusion: Dynamic semantic splitting is crucial for effective KV Cache compression. DynSplit-KV addresses limitations of rigid splitting methods by adapting to semantic boundaries while maintaining efficiency.

Abstract: Although Key-Value (KV) Cache is essential for efficient large language models (LLMs) inference, its growing memory footprint in long-context scenarios poses a significant bottleneck, making KVCache compression crucial. Current compression methods rely on rigid splitting strategies, such as fixed intervals or pre-defined delimiters. We observe that rigid splitting suffers from significant accuracy degradation (ranging from 5.5% to 55.1%) across different scenarios, owing to the scenario-dependent nature of the semantic boundaries. This highlights the necessity of dynamic semantic splitting to match semantics. To achieve this, we face two challenges. (1) Improper delimiter selection misaligns semantics with the KVCache, resulting in 28.6% accuracy loss. (2) Variable-length blocks after splitting introduce over 73.1% additional inference overhead. To address the above challenges, we propose DynSplit-KV, a KVCache compression method that dynamically identifies delimiters for splitting. We propose: (1) a dynamic importance-aware delimiter selection strategy, improving accuracy by 49.9%. (2) A uniform mapping strategy that transforms variable-length semantic blocks into a fixed-length format, reducing inference overhead by 4.9x. Experiments show that DynSplit-KV achieves the highest accuracy, 2.2x speedup compared with FlashAttention and 2.6x peak memory reduction in long-context scenarios.

Method: Derived closed-form expressions for optimal drift and its Jacobian in Gaussian and mixture-of-Gaussian settings, used numerical Jacobian-vector products to analyze flows, and composed flows with attribute classifiers to create attribute-level JVP estimators.

Result: Numerical JVPs recover analytical Jacobians in synthetic benchmarks; attribute-level JVP estimators recover empirical correlations on MNIST and CelebA; conditioning on small classifier-Jacobian norms reduces correlations consistent with hypothesized common-cause structure.

Conclusion: Jacobian-vector products provide practical insights into dependency structures in flow matching models, enabling correlation analysis and control in generated features, though conditioning on norms is not equivalent to formal causal interventions.

Abstract: Flow matching learns a velocity field that transports a base distribution to data. We study how small latent perturbations propagate through these flows and show that Jacobian-vector products (JVPs) provide a practical lens on dependency structure in the generated features. We derive closed-form expressions for the optimal drift and its Jacobian in Gaussian and mixture-of-Gaussian settings, revealing that even globally nonlinear flows admit local affine structure. In low-dimensional synthetic benchmarks, numerical JVPs recover the analytical Jacobians. In image domains, composing the flow with an attribute classifier yields an attribute-level JVP estimator that recovers empirical correlations on MNIST and CelebA. Conditioning on small classifier-Jacobian norms reduces correlations in a way consistent with a hypothesized common-cause structure, while we emphasize that this conditioning is not a formal do intervention.

Method: Introduces prompt augmentation strategy that instructs models to generate reasoning traces under diverse templates and formats, increasing rollout diversity. Enables stable scaling of training duration without KL regularization by tolerating low-entropy regimes.

Result: Qwen2.5-Math-1.5B model trained with prompt augmentation on MATH Level 3-5 dataset achieves state-of-the-art performance: 44.5% per-benchmark and 51.3% per-question accuracy on AIME24, AMC, MATH500, Minerva, and OlympiadBench.

Conclusion: Prompt augmentation addresses entropy collapse in RL training for math reasoning LLMs, enabling stable long-horizon training and achieving SOTA results without KL regularization.

Abstract: Reinforcement learning algorithms such as group-relative policy optimization (GRPO) have demonstrated strong potential for improving the mathematical reasoning capabilities of large language models. However, prior work has consistently observed an entropy collapse phenomenon during reinforcement post-training, characterized by a monotonic decrease in policy entropy that ultimately leads to training instability and collapse. As a result, most existing approaches restrict training to short horizons (typically 5-20 epochs), limiting sustained exploration and hindering further policy improvement. In addition, nearly all prior work relies on a single, fixed reasoning prompt or template during training. In this work, we introduce prompt augmentation, a training strategy that instructs the model to generate reasoning traces under diverse templates and formats, thereby increasing rollout diversity. We show that, without a KL regularization term, prompt augmentation enables stable scaling of training duration under a fixed dataset and allows the model to tolerate low-entropy regimes without premature collapse. Empirically, a Qwen2.5-Math-1.5B model trained with prompt augmentation on the MATH Level 3-5 dataset achieves state-of-the-art performance, reaching 44.5 per-benchmark accuracy and 51.3 per-question accuracy on standard mathematical reasoning benchmarks, including AIME24, AMC, MATH500, Minerva, and OlympiadBench. The code and model checkpoints are available at https://github.com/wenquanlu/prompt-augmentation-GRPO.

Method: Proposes AgentOWL that jointly learns an abstract world model (abstracting states and time) and hierarchical neural options in a sample-efficient manner.

Result: Demonstrates learning more skills with much less data than baseline methods on Object-Centric Atari games.

Conclusion: AgentOWL enables efficient skill acquisition through hierarchical composition, advancing sample-efficient hierarchical reinforcement learning.

Abstract: Building agents that can perform new skills by composing existing skills is a long-standing goal of AI agent research. Towards this end, we investigate how to efficiently acquire a sequence of skills, formalized as hierarchical neural options. However, existing model-free hierarchical reinforcement algorithms need a lot of data. We propose a novel method, which we call AgentOWL (Option and World model Learning Agent), that jointly learns – in a sample efficient way – an abstract world model (abstracting across both states and time) and a set of hierarchical neural options. We show, on a subset of Object-Centric Atari games, that our method can learn more skills using much less data than baseline methods.

Method: Proposes Streaming Merging paradigm with ARM strategy. ARM treats merging coefficients as learning rates and derives rotation vectors from activation subspaces to steer parameter updates along data-driven trajectories. It aligns semantic subspaces to preserve geometric structure of parameter evolution, requiring only early SFT checkpoints.

Result: ARM surpasses fully converged SFT models through iterative merging. Experimental results across model scales (1.7B to 14B) and diverse domains (math, code) demonstrate ARM can transcend converged checkpoints. Provides scalable, lightweight framework for efficient model adaptation.

Conclusion: ARM offers an effective model updating paradigm that conceptualizes merging as iterative optimization, enabling efficient adaptation of large language models while preserving geometric structure and outperforming conventional fine-tuning approaches.

Abstract: The escalating scale of Large Language Models (LLMs) necessitates efficient adaptation techniques. Model merging has gained prominence for its efficiency and controllability. However, existing merging techniques typically serve as post-hoc refinements or focus on mitigating task interference, often failing to capture the dynamic optimization benefits of supervised fine-tuning (SFT). In this work, we propose Streaming Merging, an innovative model updating paradigm that conceptualizes merging as an iterative optimization process. Central to this paradigm is \textbf{ARM} (\textbf{A}ctivation-guided \textbf{R}otation-aware \textbf{M}erging), a strategy designed to approximate gradient descent dynamics. By treating merging coefficients as learning rates and deriving rotation vectors from activation subspaces, ARM effectively steers parameter updates along data-driven trajectories. Unlike conventional linear interpolation, ARM aligns semantic subspaces to preserve the geometric structure of high-dimensional parameter evolution. Remarkably, ARM requires only early SFT checkpoints and, through iterative merging, surpasses the fully converged SFT model. Experimental results across model scales (1.7B to 14B) and diverse domains (e.g., math, code) demonstrate that ARM can transcend converged checkpoints. Extensive experiments show that ARM provides a scalable and lightweight framework for efficient model adaptation.

Method: Frame MIA evaluation as causal inference problem, defining memorization as causal effect of including data point in training set. Derive causal analogues of standard MIA metrics and propose practical estimators for multi-run, one-run, and zero-run regimes with non-asymptotic consistency guarantees.

Result: Experiments on real-world data show the approach enables reliable memorization measurement even when retraining is impractical and under distribution shift.

Conclusion: Provides principled foundation for privacy evaluation in modern AI systems by addressing biases in existing MIA evaluation protocols through causal inference framework.

Abstract: Membership Inference Attacks (MIAs) are widely used to quantify training data memorization and assess privacy risks. Standard evaluation requires repeated retraining, which is computationally costly for large models. One-run methods (single training with randomized data inclusion) and zero-run methods (post hoc evaluation) are often used instead, though their statistical validity remains unclear. To address this gap, we frame MIA evaluation as a causal inference problem, defining memorization as the causal effect of including a data point in the training set. This novel formulation reveals and formalizes key sources of bias in existing protocols: one-run methods suffer from interference between jointly included points, while zero-run evaluations popular for LLMs are confounded by non-random membership assignment. We derive causal analogues of standard MIA metrics and propose practical estimators for multi-run, one-run, and zero-run regimes with non-asymptotic consistency guarantees. Experiments on real-world data show that our approach enables reliable memorization measurement even when retraining is impractical and under distribution shift, providing a principled foundation for privacy evaluation in modern AI systems.

Method: Proposes Collective Adversarial Data Synthesis (CADS) with two cyclic phases: CAD-Generate (uses collective knowledge for diverse multimodal data generation) and CAD-Judge (collaboratively assesses data quality). Includes Adversarial Context Optimization to optimize generation context for challenging data.

Result: Constructed MMSynthetic-20K dataset and trained R1-SyntheticVL model, which demonstrates superior performance on various benchmarks compared to existing approaches.

Conclusion: CADS provides an effective framework for autonomous multimodal data synthesis that enhances MLLM performance through high-quality, diverse, and challenging training data generation.

Abstract: In this work, we aim to develop effective data synthesis techniques that autonomously synthesize multimodal training data for enhancing MLLMs in solving complex real-world tasks. To this end, we propose Collective Adversarial Data Synthesis (CADS), a novel and general approach to synthesize high-quality, diverse and challenging multimodal data for MLLMs. The core idea of CADS is to leverage collective intelligence to ensure high-quality and diverse generation, while exploring adversarial learning to synthesize challenging samples for effectively driving model improvement. Specifically, CADS operates with two cyclic phases, i.e., Collective Adversarial Data Generation (CAD-Generate) and Collective Adversarial Data Judgment (CAD-Judge). CAD-Generate leverages collective knowledge to jointly generate new and diverse multimodal data, while CAD-Judge collaboratively assesses the quality of synthesized data. In addition, CADS introduces an Adversarial Context Optimization mechanism to optimize the generation context to encourage challenging and high-value data generation. With CADS, we construct MMSynthetic-20K and train our model R1-SyntheticVL, which demonstrates superior performance on various benchmarks.

Method: Proposes Continuous Number Modeling (CNM) that directly models numbers as first-class continuous values rather than discrete tokens, aligning model inputs with data’s continuous nature. Trains multimodal transformer on 2M raster-to-SVG samples, then fine-tunes via reinforcement learning with perceptual feedback.

Result: Improves training speed by over 30% while maintaining higher perceptual fidelity compared to alternative approaches. Establishes CNM as practical and efficient approach for high-quality vector generation.

Conclusion: CNM addresses core challenges in vector graphics generation by eliminating discretization artifacts from token-based encoding, offering a more mathematically elegant approach with potential for broader applications beyond SVG generation.

Abstract: For certain image generation tasks, vector graphics such as Scalable Vector Graphics (SVGs) offer clear benefits such as increased flexibility, size efficiency, and editing ease, but remain less explored than raster-based approaches. A core challenge is that the numerical, geometric parameters, which make up a large proportion of SVGs, are inefficiently encoded as long sequences of tokens. This slows training, reduces accuracy, and hurts generalization. To address these problems, we propose Continuous Number Modeling (CNM), an approach that directly models numbers as first-class, continuous values rather than discrete tokens. This formulation restores the mathematical elegance of the representation by aligning the model’s inputs with the data’s continuous nature, removing discretization artifacts introduced by token-based encoding. We then train a multimodal transformer on 2 million raster-to-SVG samples, followed by fine-tuning via reinforcement learning using perceptual feedback to further improve visual quality. Our approach improves training speed by over 30% while maintaining higher perceptual fidelity compared to alternative approaches. This work establishes CNM as a practical and efficient approach for high-quality vector generation, with potential for broader applications. We make our code available http://github.com/mikeogezi/CNM.

Method: Empirical analysis of multiple models across diverse datasets and configurations (paraphrases, different temperatures) to examine correlation between task performance and robustness.

Result: Strong positive correlation found between task performance and robustness; robustness primarily driven by task-specific competence rather than inherent model-level properties.

Conclusion: Robustness emerges naturally as tasks saturate, suggesting explicit robustness measurement/improvement may warrant reduced emphasis; easier tasks become reliable for real-world deployment.

Abstract: Robustness is often regarded as a critical future challenge for real-world applications, where stability is essential. However, as models often learn tasks in a similar order, we hypothesize that easier tasks will be easier regardless of how they are presented to the model. Indeed, in this paper, we show that as models approach high performance on a task, robustness is effectively achieved. Through an empirical analysis of multiple models across diverse datasets and configurations (e.g., paraphrases, different temperatures), we find a strong positive correlation. Moreover, we find that robustness is primarily driven by task-specific competence rather than inherent model-level properties, challenging current approaches that treat robustness as an independent capability. Thus, from a high-level perspective, we may expect that as new tasks saturate, model robustness on these tasks will emerge accordingly. For researchers, this implies that explicit efforts to measure and improve robustness may warrant reduced emphasis, as such robustness is likely to develop alongside performance gains. For practitioners, it acts as a sign that indeed the tasks that the literature deals with are unreliable, but on easier past tasks, the models are reliable and ready for real-world deployment.

Method: PACER (Packet-Conditioned Revision) is a training-free inference framework that: 1) screens generated traces, 2) constructs a consensus packet with candidate answers, aggregated confidence scores, and reasoning summaries, 3) enables traces to perform targeted self-review conditioned on this packet to identify logical divergence points, and 4) uses confidence-weighted voting over revised trajectories.

Result: On challenging competitive math benchmarks (AIME, BRUMO), PACER matches or exceeds the accuracy of 256-sample majority voting, significantly outperforming raw ensemble baselines.

Conclusion: PACER transforms simple consensus into a collaborative logical refinement process, enabling reasoning traces to “peer-review” each other to resolve near-miss errors in complex reasoning tasks.

Abstract: Large language models (LLMs) achieve higher accuracy on challenging reasoning tasks by scaling test-time compute through multiple trajectory sampling. However, standard aggregation methods like majority voting or individual confidence-based filtering face a fundamental “blind spot”: they evaluate each trace in isolation. As problems scale in difficulty, models often generate hallucinated paths that exhibit misleadingly high confidence, causing the true solution to be suppressed by a narrow margin in traditional voting. We ask: can we enable traces to “peer-review” each other to resolve these near-miss errors? We introduce Packet-Conditioned Revision (PACER), a training-free, inference-only framework that enables reasoning traces to revise their conclusions through a structured coordination step. After a preliminary screening of generated traces, PACER constructs a compact consensus packet containing (i) unique candidate answers, (ii) their aggregated confidence scores, and (iii) representative reasoning summaries for each candidate answer. Individual traces then perform a targeted self-review conditioned on this packet, allowing them to identify specific logical junctions where they diverged from the broader consensus and pivot if their original reasoning is found to be flawed. Final predictions are obtained via confidence-weighted voting over these revised trajectories. On challenging competitive math benchmarks such as AIME and BRUMO, PACER matches or exceeds the accuracy of 256-sample majority voting, significantly outperforming raw ensemble baselines by transforming simple consensus into a collaborative logical refinement process.

Method: MeKi equips Transformer layers with token-level memory experts that inject pre-stored semantic knowledge, uses re-parameterization to fold training matrices into compact static lookup tables, and offloads knowledge to ROM to decouple capacity from computational cost.

Result: MeKi significantly outperforms dense LLM baselines with identical inference speed, demonstrating the effectiveness of memory-based scaling for on-device LLMs with zero inference latency overhead.

Conclusion: Memory-based scaling via storage space rather than FLOPs is an effective paradigm for deploying performant LLMs on edge devices, addressing hardware constraints while maintaining user experience.

Abstract: Scaling Large Language Models (LLMs) typically relies on increasing the number of parameters or test-time computations to boost performance. However, these strategies are impractical for edge device deployment due to limited RAM and NPU resources. Despite hardware constraints, deploying performant LLM on edge devices such as smartphone remains crucial for user experience. To address this, we propose MeKi (Memory-based Expert Knowledge Injection), a novel system that scales LLM capacity via storage space rather than FLOPs. MeKi equips each Transformer layer with token-level memory experts that injects pre-stored semantic knowledge into the generation process. To bridge the gap between training capacity and inference efficiency, we employ a re-parameterization strategy to fold parameter matrices used during training into a compact static lookup table. By offloading the knowledge to ROM, MeKi decouples model capacity from computational cost, introducing zero inference latency overhead. Extensive experiments demonstrate that MeKi significantly outperforms dense LLM baselines with identical inference speed, validating the effectiveness of memory-based scaling paradigm for on-device LLMs. Project homepage is at https://github.com/ningding-o/MeKi.

Method: Two-stage differentiable framework: Stage 1 performs edge preselection through node-wise prediction to obtain masks for lagged and instantaneous edges. Stage 2 refines these masks by optimizing a likelihood with sparsity constraints while enforcing acyclicity on the instantaneous block.

Result: SC3D achieves improved stability and more accurate recovery of both lagged and instantaneous causal structures compared to existing temporal baselines across synthetic and benchmark dynamical systems.

Conclusion: SC3D provides an effective differentiable approach for causal discovery in time series that handles both lagged and instantaneous dependencies while addressing the combinatorial search space challenge through a two-stage optimization process.

Abstract: Discovering causal structures from multivariate time series is a key problem because interactions span across multiple lags and possibly involve instantaneous dependencies. Additionally, the search space of the dynamic graphs is combinatorial in nature. In this study, we propose \textit{Stable Causal Dynamic Differentiable Discovery (SC3D)}, a two-stage differentiable framework that jointly learns lag-specific adjacency matrices and, if present, an instantaneous directed acyclic graph (DAG). In Stage 1, SC3D performs edge preselection through node-wise prediction to obtain masks for lagged and instantaneous edges, whereas Stage 2 refines these masks by optimizing a likelihood with sparsity along with enforcing acyclicity on the instantaneous block. Numerical results across synthetic and benchmark dynamical systems demonstrate that SC3D achieves improved stability and more accurate recovery of both lagged and instantaneous causal structures compared to existing temporal baselines.

Method: Introduces a continuous-time Koopman framework that models latent evolution through numerical integration schemes, allowing variable timesteps at inference for robustness to temporal resolution.

Result: The method demonstrates robustness to temporal resolution, generalizes beyond training regimes, and enables efficient long-horizon forecasting with learned dynamics adhering to analytical matrix exponential solutions.

Conclusion: Continuous-time Koopman framework provides improved temporal flexibility and stability for turbulent flow simulation compared to discrete-time approaches.

Abstract: Data-driven surrogate models have emerged as powerful tools for accelerating the simulation of turbulent flows. However, classical approaches which perform autoregressive rollouts often trade off between strong short-term accuracy and long-horizon stability. Koopman autoencoders, inspired by Koopman operator theory, provide a physics-based alternative by mapping nonlinear dynamics into a latent space where linear evolution is conducted. In practice, most existing formulations operate in a discrete-time setting, limiting temporal flexibility. In this work, we introduce a continuous-time Koopman framework that models latent evolution through numerical integration schemes. By allowing variable timesteps at inference, the method demonstrates robustness to temporal resolution and generalizes beyond training regimes. In addition, the learned dynamics closely adhere to the analytical matrix exponential solution, enabling efficient long-horizon forecasting. We evaluate the approach on classical CFD benchmarks and report accuracy, stability, and extrapolation properties.

Method: Introduces RASA with two key modifications: (1) edge-type embeddings that inject relational structure into attention scores, and (2) sparse masking that restricts attention to graph-adjacent positions, reducing search space from O(2^{n^2}) to O(2^m).

Result: Outperforms standard transformers on MetaQA (1/2/3-hop) and WebQuestionsSP, matches GPT-4 performance at lower cost, with advantages growing with reasoning depth (+7.1 points on 3-hop).

Conclusion: Minimal structural modifications to transformers can substantially improve multi-hop relational reasoning capabilities without formal learnability guarantees.

Abstract: Transformers achieve remarkable performance across many domains, yet struggle with tasks requiring multi-hop relational reasoning over structured data. We analyze this limitation through circuit complexity: standard transformers are $\mathsf{TC}^0$-complete and require $Ω(k)$ layers for $k$-hop reasoning. We introduce RASA (Relation-Aware Sparse Attention), a minimal modification adding: (1) edge-type embeddings that inject relational structure into attention scores, and (2) sparse masking that restricts attention to graph-adjacent positions. While RASA has the same asymptotic depth requirements, sparse masking reduces the attention search space from $O(2^{n^2})$ to $O(2^m)$ patterns, and edge biases provide explicit relation routing. Empirically, on MetaQA (1/2/3-hop) and WebQuestionsSP, RASA outperforms standard transformers and matches GPT-4 at lower cost, with advantages growing with reasoning depth (+7.1 points on 3-hop). We do not claim formal learnability guarantees; the contribution is empirical validation that minimal structural modifications substantially improve multi-hop reasoning.

Method: Proposed ChemCensor metric for chemical plausibility evaluation, created CREED dataset with millions of ChemCensor-validated reaction records, and trained models using this dataset.

Result: The trained model using CREED dataset improves over LLM baselines under the new benchmarking framework that emphasizes plausibility over exact match.

Conclusion: The new benchmarking framework with ChemCensor metric better aligns with human synthesis planning practices by evaluating plausibility rather than exact matches.

Abstract: Recent progress has expanded the use of large language models (LLMs) in drug discovery, including synthesis planning. However, objective evaluation of retrosynthesis performance remains limited. Existing benchmarks and metrics typically rely on published synthetic procedures and Top-K accuracy based on single ground-truth, which does not capture the open-ended nature of real-world synthesis planning. We propose a new benchmarking framework for single-step retrosynthesis that evaluates both general-purpose and chemistry-specialized LLMs using ChemCensor, a novel metric for chemical plausibility. By emphasizing plausibility over exact match, this approach better aligns with human synthesis planning practices. We also introduce CREED, a novel dataset comprising millions of ChemCensor-validated reaction records for LLM training, and use it to train a model that improves over the LLM baselines under this benchmark.

Method: GeLDA uses conditional diffusion models to synthesize samples in FM-induced latent spaces, which are low-dimensional and task-informative. It conditions generation on auxiliary feature vectors capturing semantic relationships among classes/subdomains for better augmentation in low-resource settings.

Result: In zero-shot language-specific speech emotion recognition, GeLDA improves Whisper-large baseline’s unweighted average recall by 6.13%. In long-tailed image classification, it achieves 74.7% tail-class accuracy on ImageNet-LT, setting new state-of-the-art.

Conclusion: GeLDA effectively addresses data scarcity by leveraging FM latent spaces for efficient, high-quality generative data augmentation, demonstrating strong performance across audio and vision tasks with limited labeled data.

Abstract: Despite strong performance in data-rich regimes, deep learning often underperforms in the data-scarce settings common in practice. While foundation models (FMs) trained on massive datasets demonstrate strong generalization by extracting general-purpose features, they can still suffer from scarce labeled data during downstream fine-tuning. To address this, we propose GeLDA, a semantics-aware generative latent data augmentation framework that leverages conditional diffusion models to synthesize samples in an FM-induced latent space. Because this space is low-dimensional and concentrates task-relevant information compared to the input space, GeLDA enables efficient, high-quality data generation. GeLDA conditions generation on auxiliary feature vectors that capture semantic relationships among classes or subdomains, facilitating data augmentation in low-resource domains. We validate GeLDA in two large-scale recognition tasks: (a) in zero-shot language-specific speech emotion recognition, GeLDA improves the Whisper-large baseline’s unweighted average recall by 6.13%; and (b) in long-tailed image classification, it achieves 74.7% tail-class accuracy on ImageNet-LT, setting a new state-of-the-art result.

Method: Develops a novel causal offline RL objective optimizing policies’ worst-case performance under confounding biases, with practical implementation using expressive flow-matching policies and a deep discriminator to assess discrepancy between target and behavioral policies.

Result: Experiments across 25 pixel-based tasks show the confounding-robust augmentation achieves 120% success rate compared to confounding-unaware state-of-the-art offline RL methods.

Conclusion: The proposed causal approach effectively handles confounding biases in pixel-based offline RL, significantly outperforming existing methods when sensory capabilities mismatch between demonstrator and learner.

Abstract: Expressive policies based on flow-matching have been successfully applied in reinforcement learning (RL) more recently due to their ability to model complex action distributions from offline data. These algorithms build on standard policy gradients, which assume that there is no unmeasured confounding in the data. However, this condition does not necessarily hold for pixel-based demonstrations when a mismatch exists between the demonstrator’s and the learner’s sensory capabilities, leading to implicit confounding biases in offline data. We address the challenge by investigating the problem of confounded observations in offline RL from a causal perspective. We develop a novel causal offline RL objective that optimizes policies’ worst-case performance that may arise due to confounding biases. Based on this new objective, we introduce a practical implementation that learns expressive flow-matching policies from confounded demonstrations, employing a deep discriminator to assess the discrepancy between the target policy and the nominal behavioral policy. Experiments across 25 pixel-based tasks demonstrate that our proposed confounding-robust augmentation procedure achieves a success rate 120% that of confounding-unaware, state-of-the-art offline RL methods.

Method: CoRSA (Conflict-Resolving and Sharpness-Aware Minimization) framework uses sharpness-aware minimization to reduce loss curvature for better generalization and stability, and maximizes margin between new and prior knowledge to resolve conflicts. It’s parameter-efficient and handles multiple knowledge updates.

Result: Outperforms baselines with 12.42% average improvement over LoRA and 10% over model editing methods on fact editing benchmarks. Reduces catastrophic forgetting by 27.82% compared to LoRA with multiple updates. Also generalizes to code domain with 5.48% Pass@5 improvement.

Conclusion: CoRSA provides an effective parameter-efficient framework for knowledge editing that addresses key limitations of existing methods, offering better generalization, stability, and conflict resolution for keeping LLMs up-to-date.

Abstract: Large language models (LLMs) rely on internal knowledge to solve many downstream tasks, making it crucial to keep them up to date. Since full retraining is expensive, prior work has explored efficient alternatives such as model editing and parameter-efficient fine-tuning. However, these approaches often break down in practice due to poor generalization across inputs, limited stability, and knowledge conflict. To address these limitations, we propose the CoRSA (Conflict-Resolving and Sharpness-Aware Minimization) training framework, a parameter-efficient, holistic approach for knowledge editing with multiple updates. CoRSA tackles multiple challenges simultaneously: it improves generalization to different input forms and enhances stability across multiple updates by minimizing loss curvature, and resolves conflicts by maximizing the margin between new and prior knowledge. Across three widely used fact editing benchmarks, CoRSA achieves significant gains in generalization, outperforming baselines with average absolute improvements of 12.42% over LoRA and 10% over model editing methods. With multiple updates, it maintains high update efficacy while reducing catastrophic forgetting by 27.82% compared to LoRA. CoRSA also generalizes to the code domain, outperforming the strongest baseline by 5.48% Pass@5 in update efficacy.

Method: ZS-SVD performs global singular component selection using activation whitening and first-order calibration loss estimates in whitened coordinates. It prunes components across the whole model with a zero-sum rule that keeps cumulative predicted loss change near zero. An optional lightweight correction applies a single projected gradient update after truncation followed by re-truncation.

Result: Extensive experiments across multiple LLM architectures show consistent gains across diverse benchmarks and compression ratios. The method yields heterogeneous ranks automatically without solving rank allocation optimization.

Conclusion: ZS-SVD provides an effective post-training compression method for LLMs that automatically determines optimal rank allocation through global singular component selection with a zero-sum rule, outperforming prior methods.

Abstract: Advances in large language models have driven strong performance across many tasks, but their memory and compute costs still hinder deployment. SVD-based compression reduces storage and can speed up inference via low-rank factors, yet performance depends on how rank is allocated under a global compression ratio. Prior methods often use homogeneous ranks for similarly sized matrices, despite large differences in loss sensitivity, or rely on expensive iterative pre-truncation optimization to determine per matrix ranks. We propose \textbf{Zero Sum SVD} (\textbf{ZS-SVD}), a post-training method that performs \emph{global} singular component selection using activation whitening and first-order calibration loss estimates in whitened coordinates. \textbf{ZS-SVD} prunes components across the whole model with a \textbf{zero sum} rule that keeps the cumulative predicted loss change near zero, automatically yielding heterogeneous ranks without solving a rank allocation optimization. Motivated by evidence that gradients near pretrained solutions exhibit low rank structure, we also introduce an optional lightweight correction that applies a \textbf{single} projected gradient update after truncation, followed by re-truncation. Extensive experiments across multiple LLM architectures show consistent gains across diverse benchmarks and compression ratios. Code is available at https://github.com/mint-vu/Zero-Sum-SVD

Method: Introduces kernel surrogate models to capture second-order task interactions, with a gradient-based estimation procedure that leverages first-order approximations of pretrained models to efficiently learn the surrogate without repeated retraining.

Result: Kernel surrogate models achieve 25% higher correlation with leave-one-out ground truth than linear surrogates and influence-function baselines, and yield 40% improvement in demonstration selection for in-context learning and multi-objective RL benchmarks.

Conclusion: Kernel surrogate models provide a more accurate and efficient approach to task attribution in multi-task learning, capturing nonlinear task interactions that linear methods miss, with practical benefits for downstream task selection.

Abstract: Modern AI agents such as large language models are trained on diverse tasks – translation, code generation, mathematical reasoning, and text prediction – simultaneously. A key question is to quantify how each individual training task influences performance on a target task, a problem we refer to as task attribution. The direct approach, leave-one-out retraining, measures the effect of removing each task, but is computationally infeasible at scale. An alternative approach that builds surrogate models to predict a target task’s performance for any subset of training tasks has emerged in recent literature. Prior work focuses on linear surrogate models, which capture first-order relationships, but miss nonlinear interactions such as synergy, antagonism, or XOR-type effects. In this paper, we first consider a unified task weighting framework for analyzing task attribution methods, and show a new connection between linear surrogate models and influence functions through a second-order analysis. Then, we introduce kernel surrogate models, which more effectively represent second-order task interactions. To efficiently learn the kernel surrogate, we develop a gradient-based estimation procedure that leverages a first-order approximation of pretrained models; empirically, this yields accurate estimates with less than $2%$ relative error without repeated retraining. Experiments across multiple domains – including math reasoning in transformers, in-context learning, and multi-objective reinforcement learning – demonstrate the effectiveness of kernel surrogate models. They achieve a $25%$ higher correlation with the leave-one-out ground truth than linear surrogates and influence-function baselines. When used for downstream task selection, kernel surrogate models yield a $40%$ improvement in demonstration selection for in-context learning and multi-objective reinforcement learning benchmarks.

Method: Proposes Recurrent Equivariant Constraint Modulation (RECM), a layer-wise constraint modulation mechanism that learns appropriate relaxation levels solely from training signals and symmetry properties of each layer’s input-target distribution, without prior knowledge of target relaxation levels.

Result: RECM provably converges to relaxation levels upper-bounded by each layer’s symmetry gap, automatically recovering full equivariance for symmetric distributions while allowing flexibility for approximate symmetries. Empirically outperforms prior methods on diverse equivariant tasks including molecular conformer generation.

Conclusion: RECM provides an effective, automatic approach to balancing equivariance constraints with learning flexibility, eliminating the need for manual tuning while maintaining theoretical guarantees about convergence to appropriate relaxation levels based on data symmetry properties.

Abstract: Equivariant neural networks exploit underlying task symmetries to improve generalization, but strict equivariance constraints can induce more complex optimization dynamics that can hinder learning. Prior work addresses these limitations by relaxing strict equivariance during training, but typically relies on prespecified, explicit, or implicit target levels of relaxation for each network layer, which are task-dependent and costly to tune. We propose Recurrent Equivariant Constraint Modulation (RECM), a layer-wise constraint modulation mechanism that learns appropriate relaxation levels solely from the training signal and the symmetry properties of each layer’s input-target distribution, without requiring any prior knowledge about the task-dependent target relaxation level. We demonstrate that under the proposed RECM update, the relaxation level of each layer provably converges to a value upper-bounded by its symmetry gap, namely the degree to which its input-target distribution deviates from exact symmetry. Consequently, layers processing symmetric distributions recover full equivariance, while those with approximate symmetries retain sufficient flexibility to learn non-symmetric solutions when warranted by the data. Empirically, RECM outperforms prior methods across diverse exact and approximate equivariant tasks, including the challenging molecular conformer generation on the GEOM-Drugs dataset.

Method: Theoretical analysis using summary statistics description of fine-tuning dynamics for low-rank adaptation (LoRA) fine-tuning on single-index models trained under one-pass SGD. The study characterizes convergence rate dependence on initial fine-tuning alignment and target task non-linearity.

Result: Even when pre-training and downstream tasks are well-aligned, strong pre-training can induce a prolonged search phase and hinder convergence. The theory provides a unified picture of how pre-training strength and task difficulty jointly shape LoRA fine-tuning dynamics.

Conclusion: Excessive pre-training can computationally slow down fine-tuning optimization, revealing non-trivial dynamics in LoRA fine-tuning that depend on both pre-training strength and task difficulty.

Abstract: Pre-training on a source task is usually expected to facilitate fine-tuning on similar downstream problems. In this work, we mathematically show that this naive intuition is not always true: excessive pre-training can computationally slow down fine-tuning optimization. We study this phenomenon for low-rank adaptation (LoRA) fine-tuning on single-index models trained under one-pass SGD. Leveraging a summary statistics description of the fine-tuning dynamics, we precisely characterize how the convergence rate depends on the initial fine-tuning alignment and the degree of non-linearity of the target task. The key take away is that even when the pre-training and down- stream tasks are well aligned, strong pre-training can induce a prolonged search phase and hinder convergence. Our theory thus provides a unified picture of how pre-training strength and task difficulty jointly shape the dynamics and limitations of LoRA fine-tuning in a nontrivial tractable model.

Method: Extended training beyond standard grokking experiments on two canonical setups (3-layer MLP on MNIST subset and transformer on modular addition), using WeightWatcher tool to analyze weight matrices for Correlation Traps and HTSR layer quality metrics.

Result: Discovered anti-grokking phase where test accuracy collapses to chance while training accuracy remains perfect; Correlation Traps (anomalously large eigenvalues) reliably identify this phase, unlike other diagnostics; observed similar pathologies in large-scale LLMs.

Conclusion: Anti-grokking represents a distinct post-generalization failure mode; Correlation Traps provide a data-free diagnostic for generalization collapse; these phenomena extend to large-scale models.

Abstract: \emph{Memorization} in neural networks lacks a precise operational definition and is often inferred from the grokking regime, where training accuracy saturates while test accuracy remains very low. We identify a previously unreported third phase of grokking in this training regime: \emph{anti-grokking}, a late-stage collapse of generalization. We revisit two canonical grokking setups: a 3-layer MLP trained on a subset of MNIST and a transformer trained on modular addition, but extended training far beyond standard. In both cases, after models transition from pre-grokking to successful generalization, test accuracy collapses back to chance while training accuracy remains perfect, indicating a distinct post-generalization failure mode. To diagnose anti-grokking, we use the open-source \texttt{WeightWatcher} tool based on HTSR/SETOL theory. The primary signal is the emergence of \emph{Correlation Traps}: anomalously large eigenvalues beyond the Marchenko–Pastur bulk in the empirical spectral density of shuffled weight matrices, which are predicted to impair generalization. As a secondary signal, anti-grokking corresponds to the average HTSR layer quality metric $α$ deviating from $2.0$. Neither metric requires access to the test or training data. We compare these signals to alternative grokking diagnostic, including $\ell_2$ norms, Activation Sparsity, Absolute Weight Entropy, and Local Circuit Complexity. These track pre-grokking and grokking but fail to identify anti-grokking. Finally, we show that Correlation Traps can induce catastrophic forgetting and/or prototype memorization, and observe similar pathologies in large-scale LLMs, like OSS GPT 20/120B.

Method: Cobalt collects code generation trajectories using a reference LLM, divides them into partial trajectories as contextual prompts, then trains the LLM through online bandit learning with single-step completions of these prompts.

Result: Cobalt outperforms GRPO and VeRPO baselines, improving R1-Distill 8B and Qwen3 8B by up to 9.0 and 6.2 absolute Pass@1 scores on LiveCodeBench. The method also addresses reward hacking with perturbed trajectories.

Conclusion: Cobalt demonstrates a promising solution for iterative decision-making tasks like multi-turn code generation by effectively combining online and offline RL benefits.

Abstract: Recently, there have been significant research interests in training large language models (LLMs) with reinforcement learning (RL) on real-world tasks, such as multi-turn code generation. While online RL tends to perform better than offline RL, its higher training cost and instability hinders wide adoption. In this paper, we build on the observation that multi-turn code generation can be formulated as a one-step recoverable Markov decision process and propose contextual bandit learning with offline trajectories (Cobalt), a new method that combines the benefits of online and offline RL. Cobalt first collects code generation trajectories using a reference LLM and divides them into partial trajectories as contextual prompts. Then, during online bandit learning, the LLM is trained to complete each partial trajectory prompt through single-step code generation. Cobalt outperforms two multi-turn online RL baselines based on GRPO and VeRPO, and substantially improves R1-Distill 8B and Qwen3 8B by up to 9.0 and 6.2 absolute Pass@1 scores on LiveCodeBench. Also, we analyze LLMs’ in-context reward hacking behaviors and augment Cobalt training with perturbed trajectories to mitigate this issue. Overall, our results demonstrate Cobalt as a promising solution for iterative decision-making tasks like multi-turn code generation. Our code and data are available at https://github.com/OSU-NLP-Group/cobalt.

Method: Proposes SCENT algorithm using stochastic proximal mirror descent (SPMD) with Bregman divergence induced by negative exponential function to capture objective geometry, applied to dual formulation of entropic risk minimization as min-min optimization.

Result: Establishes O(1/√T) convergence rate for convex problems, theoretically shows SPMD advantages over standard SGD, and demonstrates empirical superiority on extreme classification, partial AUC maximization, contrastive learning, and distributionally robust optimization.

Conclusion: SCENT addresses fundamental limitations of existing methods for compositional entropic risk minimization through geometry-aware stochastic optimization with provable convergence and practical effectiveness.

Abstract: This paper studies optimization for a family of problems termed $\textbf{compositional entropic risk minimization}$, in which each data’s loss is formulated as a Log-Expectation-Exponential (Log-E-Exp) function. The Log-E-Exp formulation serves as an abstraction of the Log-Sum-Exponential (LogSumExp) function when the explicit summation inside the logarithm is taken over a gigantic number of items and is therefore expensive to evaluate. While entropic risk objectives of this form arise in many machine learning problems, existing optimization algorithms suffer from several fundamental limitations including non-convergence, numerical instability, and slow convergence rates. To address these limitations, we propose a geometry-aware stochastic algorithm, termed $\textbf{SCENT}$, for the dual formulation of entropic risk minimization cast as a min–min optimization problem. The key to our design is a $\textbf{stochastic proximal mirror descent (SPMD)}$ update for the dual variable, equipped with a Bregman divergence induced by a negative exponential function that faithfully captures the geometry of the objective. Our main contributions are threefold: (i) we establish an $O(1/\sqrt{T})$ convergence rate of the proposed SCENT algorithm for convex problems; (ii) we theoretically characterize the advantages of SPMD over standard SGD update for optimizing the dual variable; and (iii) we demonstrate the empirical effectiveness of SCENT on extreme classification, partial AUC maximization, contrastive learning and distributionally robust optimization, where it consistently outperforms existing baselines.

Method: ADFP uses a gradient-based antidistillation sampling framework with a proxy model to identify and sample tokens that maximize expected detectability of fingerprints in student models after fine-tuning, rather than relying on incidental absorption of biases.

Result: Experiments on GSM8K and OASST1 benchmarks show ADFP achieves significant Pareto improvement over state-of-the-art baselines, yielding stronger detection confidence with minimal impact on utility, even with unknown student model architectures.

Conclusion: ADFP provides a principled approach to model distillation fingerprinting that better aligns with student learning dynamics, offering improved detection capabilities without sacrificing generation quality.

Abstract: Model distillation enables efficient emulation of frontier large language models (LLMs), creating a need for robust mechanisms to detect when a third-party student model has trained on a teacher model’s outputs. However, existing fingerprinting techniques that could be used to detect such distillation rely on heuristic perturbations that impose a steep trade-off between generation quality and fingerprinting strength, often requiring significant degradation of utility to ensure the fingerprint is effectively internalized by the student. We introduce antidistillation fingerprinting (ADFP), a principled approach that aligns the fingerprinting objective with the student’s learning dynamics. Building upon the gradient-based framework of antidistillation sampling, ADFP utilizes a proxy model to identify and sample tokens that directly maximize the expected detectability of the fingerprint in the student after fine-tuning, rather than relying on the incidental absorption of the un-targeted biases of a more naive watermark. Experiments on GSM8K and OASST1 benchmarks demonstrate that ADFP achieves a significant Pareto improvement over state-of-the-art baselines, yielding stronger detection confidence with minimal impact on utility, even when the student model’s architecture is unknown.

Method: Proposes Mixture of Concept Bottleneck Experts (M-CBEs) that generalizes CBMs along two dimensions: number of experts and functional form of each expert. Instantiates two models: Linear M-CBE (learns finite set of linear expressions) and Symbolic M-CBE (uses symbolic regression to discover expert functions from data with user-specified operator vocabularies).

Result: Empirical evaluation shows that varying mixture size and functional form provides a robust framework for navigating accuracy-interpretability trade-off, adapting to different user and task needs.

Conclusion: M-CBEs offer a flexible framework that expands the design space of interpretable models, allowing better adaptation to diverse requirements while maintaining interpretability through concept grounding.

Abstract: Concept Bottleneck Models (CBMs) promote interpretability by grounding predictions in human-understandable concepts. However, existing CBMs typically fix their task predictor to a single linear or Boolean expression, limiting both predictive accuracy and adaptability to diverse user needs. We propose Mixture of Concept Bottleneck Experts (M-CBEs), a framework that generalizes existing CBMs along two dimensions: the number of experts and the functional form of each expert, exposing an underexplored region of the design space. We investigate this region by instantiating two novel models: Linear M-CBE, which learns a finite set of linear expressions, and Symbolic M-CBE, which leverages symbolic regression to discover expert functions from data under user-specified operator vocabularies. Empirical evaluation demonstrates that varying the mixture size and functional form provides a robust framework for navigating the accuracy-interpretability trade-off, adapting to different user and task needs.

Method: Proposes Self-Souping: take a base model (stock), fine-tune it into multiple models (ingredients) using different SSL algorithms or hyperparameters on various data sources, then mix their parameters back into one model (soup). Includes techniques like self-souping on corrupted test data then fine-tuning back on clean train data.

Result: Self-Souping boosts robustness by +3.5% on ImageNet-C and +7% on LAION-C. Enables cooking soups from diverse SSL algorithms (MAE, MoCoV3, MMCR) that outperform any single SSL ingredient. First demonstration that ingredients can differ in SSL hyperparameters and algorithms.

Conclusion: Self-Souping successfully generalizes model soups to self-supervised learning, unlocking countless SSL algorithms for creating more robust models through parameter mixing from diverse ingredients trained on various data sources.

Abstract: Model soups are strange and strangely effective combinations of parameters. They take a model (the stock), fine-tune it into multiple models (the ingredients), and then mix their parameters back into one model (the soup) to improve predictions. While all known soups require supervised learning, and optimize the same loss on labeled data, our recipes for Self-\emph{Soup}ervision generalize soups to self-supervised learning (SSL). Our Self-Souping lets us flavor ingredients on new data sources, e.g. from unlabeled data from a task for transfer or from a shift for robustness. We show that Self-Souping on corrupted test data, then fine-tuning back on uncorrupted train data, boosts robustness by +3.5% (ImageNet-C) and +7% (LAION-C). Self-\emph{Soup}ervision also unlocks countless SSL algorithms to cook the diverse ingredients needed for more robust soups. We show for the first time that ingredients can differ in their SSL hyperparameters – and more surprisingly, in their SSL algorithms. We cook soups of MAE, MoCoV3, and MMCR ingredients that are more accurate than any one single SSL ingredient.

Method: Decentralized SGD with Adaptive Consensus (DSGD-AC) that preserves non-vanishing consensus errors through time-dependent scaling, allowing errors to systematically align with dominant Hessian subspace.

Result: DSGD-AC consistently surpasses both standard DSGD and centralized SGD in test accuracy and solution flatness across image classification and machine translation benchmarks.

Conclusion: Consensus errors can be beneficial as implicit regularizers, opening new perspectives for decentralized learning algorithm design by leveraging structured perturbations for better generalization.

Abstract: Decentralized training is often regarded as inferior to centralized training because the consensus errors between workers are thought to undermine convergence and generalization, even with homogeneous data distributions. This work challenges this view by introducing decentralized SGD with Adaptive Consensus (DSGD-AC), which intentionally preserves non-vanishing consensus errors through a time-dependent scaling mechanism. We prove that these errors are not random noise but systematically align with the dominant Hessian subspace, acting as structured perturbations that guide optimization toward flatter minima. Across image classification and machine translation benchmarks, DSGD-AC consistently surpasses both standard DSGD and centralized SGD in test accuracy and solution flatness. Together, these results establish consensus errors as a useful implicit regularizer and open a new perspective on the design of decentralized learning algorithms.

Method: Proposes Manifold-Constrained Energy-based Transition Models (MC-ETM) that: 1) Learn conditional energy-based transition models using manifold projection-diffusion negative sampling, 2) Generate near-manifold hard negatives via latent space perturbations and Langevin dynamics, 3) Use learned energy as reliability signal to truncate rollouts when energy exceeds threshold, 4) Stabilize Bellman backups with pessimistic penalties based on Q-value dispersion across energy-guided samples.

Result: MC-ETM improves multi-step dynamics fidelity and yields higher normalized returns on standard offline control benchmarks, particularly under irregular dynamics and sparse data coverage. The method shows better robustness to distribution shift.

Conclusion: MC-ETM provides an effective framework for handling distribution shift in offline RL through energy-based modeling and manifold constraints, offering reliable detection of out-of-distribution states and stabilization of value estimation.

Abstract: Model-based offline reinforcement learning is brittle under distribution shift: policy improvement drives rollouts into state–action regions weakly supported by the dataset, where compounding model error yields severe value overestimation. We propose Manifold-Constrained Energy-based Transition Models (MC-ETM), which train conditional energy-based transition models using a manifold projection–diffusion negative sampler. MC-ETM learns a latent manifold of next states and generates near-manifold hard negatives by perturbing latent codes and running Langevin dynamics in latent space with the learned conditional energy, sharpening the energy landscape around the dataset support and improving sensitivity to subtle out-of-distribution deviations. For policy optimization, the learned energy provides a single reliability signal: rollouts are truncated when the minimum energy over sampled next states exceeds a threshold, and Bellman backups are stabilized via pessimistic penalties based on Q-value-level dispersion across energy-guided samples. We formalize MC-ETM through a hybrid pessimistic MDP formulation and derive a conservative performance bound separating in-support evaluation error from truncation risk. Empirically, MC-ETM improves multi-step dynamics fidelity and yields higher normalized returns on standard offline control benchmarks, particularly under irregular dynamics and sparse data coverage.

Method: Reformulates multi-agent traffic signal control as sequence modeling problem using Decision Transformer extended with: (1) graph attention for spatial dependencies, (2) temporal transformer encoder for traffic dynamics, (3) return-to-go conditioning for target performance.

Result: Achieves state-of-the-art performance, reducing average travel time by 5-6% compared to strongest baselines, with superior coordination among adjacent intersections in synthetic grid networks and real-world scenarios.

Conclusion: MADT provides an effective approach for multi-agent traffic signal control through sequence modeling, enabling offline learning from historical data with potential for online fine-tuning.

Abstract: Traffic signal control is a critical challenge in urban transportation, requiring coordination among multiple intersections to optimize network-wide traffic flow. While reinforcement learning has shown promise for adaptive signal control, existing methods struggle with multi-agent coordination and sample efficiency. We introduce MADT (Multi-Agent Decision Transformer), a novel approach that reformulates multi-agent traffic signal control as a sequence modeling problem. MADT extends the Decision Transformer paradigm to multi-agent settings by incorporating: (1) a graph attention mechanism for modeling spatial dependencies between intersections, (2) a|temporal transformer encoder for capturing traffic dynamics, and (3) return-to-go conditioning for target performance specification. Our approach enables offline learning from historical traffic data, with architecture design that facilitates potential online fine-tuning. Experiments on synthetic grid networks and real-world traffic scenarios demonstrate that MADT achieves state-of-the-art performance, reducing average travel time by 5-6% compared to the strongest baseline while exhibiting superior coordination among adjacent intersections.

Method: Develops a random matrix theory framework to quantify how finite datasets shape the expectation and variance of learned denoisers and sampling maps in linear settings. Uses deterministic-equivalence tools extended to fractional matrix powers to analyze entire sampling trajectories.

Result: The theory explains that sampling variability acts as a renormalization of noise level through a self-consistent relation, causing limited data to overshrink low-variance directions and pull samples toward dataset mean. Identifies three key factors behind cross-split disagreement: anisotropy across eigenmodes, inhomogeneity across inputs, and scaling with dataset size.

Conclusion: Provides a principled baseline for reproducibility in diffusion training, linking spectral properties of data to the stability of generative outputs, with validation on UNet and DiT architectures in their non-memorization regime.

Abstract: Diffusion models trained on different, non-overlapping subsets of a dataset often produce strikingly similar outputs when given the same noise seed. We trace this consistency to a simple linear effect: the shared Gaussian statistics across splits already predict much of the generated images. To formalize this, we develop a random matrix theory (RMT) framework that quantifies how finite datasets shape the expectation and variance of the learned denoiser and sampling map in the linear setting. For expectations, sampling variability acts as a renormalization of the noise level through a self-consistent relation $σ^2 \mapsto κ(σ^2)$, explaining why limited data overshrink low-variance directions and pull samples toward the dataset mean. For fluctuations, our variance formulas reveal three key factors behind cross-split disagreement: \textit{anisotropy} across eigenmodes, \textit{inhomogeneity} across inputs, and overall scaling with dataset size. Extending deterministic-equivalence tools to fractional matrix powers further allows us to analyze entire sampling trajectories. The theory sharply predicts the behavior of linear diffusion models, and we validate its predictions on UNet and DiT architectures in their non-memorization regime, identifying where and how samples deviates across training data split. This provides a principled baseline for reproducibility in diffusion training, linking spectral properties of data to the stability of generative outputs.

Method: Analyzes the special case where KL-regularized soft updates correspond exactly to Bayesian posterior updates within a single fixed probabilistic model, treating the update variable as a genuine information channel. Examines identification constraints for reward functions given posterior updates.

Result: Shows that posterior updates determine relative, context-dependent incentive signals but not absolute rewards, which remain ambiguous up to context-specific baselines. Coherence constraints emerge when requiring reusable continuation values across different update directions.

Conclusion: Provides a sharp identification result for when decision-making as inference can be made literal, clarifying the relationship between Bayesian updates and reward functions in reinforcement learning and control theory.

Abstract: Many ideas in modern control and reinforcement learning treat decision-making as inference: start from a baseline distribution and update it when a signal arrives. We ask when this can be made literal rather than metaphorical. We study the special case where a KL-regularized soft update is exactly a Bayesian posterior inside a single fixed probabilistic model, so the update variable is a genuine channel through which information is transmitted. In this regime, behavioral change is driven only by evidence carried by that channel: the update must be explainable as an evidence reweighing of the baseline. This yields a sharp identification result: posterior updates determine the relative, context-dependent incentive signal that shifts behavior, but they do not uniquely determine absolute rewards, which remain ambiguous up to context-specific baselines. Requiring one reusable continuation value across different update directions adds a further coherence constraint linking the reward descriptions associated with different conditioning orders.

Method: Proposes a training strategy that learns biomarker-specific decay of sample weight over time gap between PPG segment and ground truth label, using this weight in loss with a regularizer to prevent trivial solutions.

Result: On smartwatch PPG from 450 participants across 10 biomarkers, the approach achieves 0.715 AUPRC (vs 0.674 for fine-tuned self-supervised baseline and 0.626 for feature-based Random Forest). Linear decay function is most robust.

Conclusion: The method improves biomarker prediction from PPG signals while providing interpretable decay rates that summarize how quickly each biomarker’s PPG evidence becomes stale, offering insights into temporal sensitivity.

Abstract: Advances in wearable computing and AI have increased interest in leveraging PPG for health monitoring over the past decade. One of the biggest challenges in developing health algorithms based on such biosignals is the sparsity of clinical labels, which makes biosignals temporally distant from lab draws less reliable for supervision. To address this problem, we introduce a simple training strategy that learns a biomarker-specific decay of sample weight over the time gap between a segment and its ground truth label and uses this weight in the loss with a regularizer to prevent trivial solutions. On smartwatch PPG from 450 participants across 10 biomarkers, the approach improves over baselines. In the subject-wise setting, the proposed approach averages 0.715 AUPRC, compared to 0.674 for a fine-tuned self-supervised baseline and 0.626 for a feature-based Random Forest. A comparison of four decay families shows that a simple linear decay function is most robust on average. Beyond accuracy, the learned decay rates summarize how quickly each biomarker’s PPG evidence becomes stale, providing an interpretable view of temporal sensitivity.

Method: Proposes SDA2E autoencoder for compact discriminative latent representations, plus similarity-guided active learning with three strategies: normal-like expansion, anomaly-like prioritization, and hybrid strategy. Introduces SIM_NM1 similarity measure for sparse binary embeddings.

Result: Evaluated on 52 imbalanced datasets including DARPA Transparent Computing scenarios. Achieves superior ranking performance (nDCG up to 1.0), reduces required labeled data by up to 80% compared to passive training, with statistical significance confirmed.

Conclusion: Establishes robust, efficient, statistically validated framework for anomaly detection suited to cybersecurity applications like APT detection.

Abstract: Detecting rare and diverse anomalies in highly imbalanced datasets-such as Advanced Persistent Threats (APTs) in cybersecurity-remains a fundamental challenge for machine learning systems. Active learning offers a promising direction by strategically querying an oracle to minimize labeling effort, yet conventional approaches often fail to exploit the intrinsic geometric structure of the feature space for model refinement. In this paper, we introduce SDA2E, a Sparse Dual Adversarial Attention-based AutoEncoder designed to learn compact and discriminative latent representations from imbalanced, high-dimensional data. We further propose a similarity-guided active learning framework that integrates three novel strategies to refine decision boundaries efficiently: mormal-like expansion, which enriches the training set with points similar to labeled normals to improve reconstruction fidelity; anomaly-like prioritization, which boosts ranking accuracy by focusing on points resembling known anomalies; and a hybrid strategy that combines both for balanced model refinement and ranking. A key component of our framework is a new similarity measure, Normalized Matching 1s (SIM_NM1), tailored for sparse binary embeddings. We evaluate SDA2E extensively across 52 imbalanced datasets, including multiple DARPA Transparent Computing scenarios, and benchmark it against 15 state-of-the-art anomaly detection methods. Results demonstrate that SDA2E consistently achieves superior ranking performance (nDCG up to 1.0 in several cases) while reducing the required labeled data by up to 80% compared to passive training. Statistical tests confirm the significance of these improvements. Our work establishes a robust, efficient, and statistically validated framework for anomaly detection that is particularly suited to cybersecurity applications such as APT detection.

Method: Integrates domain-informed feature engineering, nested cross-validation, and calibrated decision-threshold optimization specifically designed for high-dimensional structural MRI datasets with small sample sizes.

Result: Achieved a nested-CV balanced accuracy of 0.660 ± 0.068 using a compact, interpretable subset of features selected via importance-guided ranking on deep brain stimulation cognitive outcomes data.

Conclusion: The framework provides a generalizable computational blueprint for reliable machine learning in data-limited biomedical domains by combining interpretability and unbiased evaluation.

Abstract: We introduce a reproducible, bias-resistant machine learning framework that integrates domain-informed feature engineering, nested cross-validation, and calibrated decision-threshold optimization for small-sample neuroimaging data. Conventional cross-validation frameworks that reuse the same folds for both model selection and performance estimation yield optimistically biased results, limiting reproducibility and generalization. Demonstrated on a high-dimensional structural MRI dataset of deep brain stimulation cognitive outcomes, the framework achieved a nested-CV balanced accuracy of 0.660,$\pm$,0.068 using a compact, interpretable subset selected via importance-guided ranking. By combining interpretability and unbiased evaluation, this work provides a generalizable computational blueprint for reliable machine learning in data-limited biomedical domains.

Method: Constructs process behavioral graph using k-Nearest Neighbors based on feature similarity, learns normal relational structure using Graph Autoencoder, identifies anomaly candidates through reconstruction deviations, and integrates rare pattern mining to discover infrequent behavioral co-occurrences that boost anomaly scores.

Result: Evaluated on DARPA Transparent Computing datasets, rare-pattern boosting yields substantial gains in anomaly ranking quality over baseline GAE. The unified model outperforms individual context-based detectors and achieves performance competitive with ensemble aggregation methods requiring multiple separate detectors.

Conclusion: Coupling graph-based representation learning with classical pattern mining improves both effectiveness and interpretability in provenance-based security anomaly detection, offering a valuable approach for identifying APT-like activities.

Abstract: Advanced Persistent Threats (APTs) are sophisticated, long-term cyberattacks that are difficult to detect because they operate stealthily and often blend into normal system behavior. This paper presents a neuro-symbolic anomaly detection framework that combines a Graph Autoencoder (GAE) with rare pattern mining to identify APT-like activities in system-level provenance data. Our approach first constructs a process behavioral graph using k-Nearest Neighbors based on feature similarity, then learns normal relational structure using a Graph Autoencoder. Anomaly candidates are identified through deviations between observed and reconstructed graph structure. To further improve detection, we integrate an rare pattern mining module that discovers infrequent behavioral co-occurrences and uses them to boost anomaly scores for processes exhibiting rare signatures. We evaluate the proposed method on the DARPA Transparent Computing datasets and show that rare-pattern boosting yields substantial gains in anomaly ranking quality over the baseline GAE. Compared with existing unsupervised approaches on the same benchmark, our single unified model consistently outperforms individual context-based detectors and achieves performance competitive with ensemble aggregation methods that require multiple separate detectors. These results highlight the value of coupling graph-based representation learning with classical pattern mining to improve both effectiveness and interpretability in provenance-based security anomaly detection.

Method: Proposes Augmented Lagrangian-Guided Diffusion (ALGD) that revisits optimization theory and energy-based models. Shows primal-dual methods are unstable due to non-convex Lagrangian landscapes. Introduces augmented Lagrangian to locally convexify the energy landscape, stabilizing both policy generation and training without changing optimal policy distribution.

Result: ALGD achieves strong and stable performance across diverse environments. Theoretical analysis confirms it’s both theoretically grounded and empirically effective for safe RL with diffusion policies.

Conclusion: ALGD successfully addresses the stability issues in diffusion-based safe RL by using augmented Lagrangian to convexify the energy landscape, enabling stable multimodal policy generation while maintaining safety constraints in online settings.

Abstract: Diffusion policy sampling enables reinforcement learning (RL) to represent multimodal action distributions beyond suboptimal unimodal Gaussian policies. However, existing diffusion-based RL methods primarily focus on offline settings for reward maximization, with limited consideration of safety in online settings. To address this gap, we propose Augmented Lagrangian-Guided Diffusion (ALGD), a novel algorithm for off-policy safe RL. By revisiting optimization theory and energy-based model, we show that the instability of primal-dual methods arises from the non-convex Lagrangian landscape. In diffusion-based safe RL, the Lagrangian can be interpreted as an energy function guiding the denoising dynamics. Counterintuitively, direct usage destabilizes both policy generation and training. ALGD resolves this issue by introducing an augmented Lagrangian that locally convexifies the energy landscape, yielding a stabilized policy generation and training process without altering the distribution of the optimal policy. Theoretical analysis and extensive experiments demonstrate that ALGD is both theoretically grounded and empirically effective, achieving strong and stable performance across diverse environments.

Method: Proposes Distance Marching with two principled inference methods inspired by distance field modeling. Designs losses that focus on closer targets to yield denoising directions better directed toward the data manifold.

Result: Consistently improves FID by 13.5% on CIFAR-10 and ImageNet over recent time-unconditional baselines. For class-conditional ImageNet generation, surpasses flow matching despite removing time input, achieving lower FID with 60% of sampling steps and 13.6% lower FID on average across backbone sizes.

Conclusion: Distance field modeling provides a principled framework for generative modeling, with distance prediction also useful for early stopping during sampling and OOD detection.

Abstract: Time-unconditional generative models learn time-independent denoising vector fields. But without time conditioning, the same noisy input may correspond to multiple noise levels and different denoising directions, which interferes with the supervision signal. Inspired by distance field modeling, we propose Distance Marching, a new time-unconditional approach with two principled inference methods. Crucially, we design losses that focus on closer targets. This yields denoising directions better directed toward the data manifold. Across architectures, Distance Marching consistently improves FID by 13.5% on CIFAR-10 and ImageNet over recent time-unconditional baselines. For class-conditional ImageNet generation, despite removing time input, Distance Marching surpasses flow matching using our losses and inference methods. It achieves lower FID than flow matching’s final performance using 60% of the sampling steps and 13.6% lower FID on average across backbone sizes. Moreover, our distance prediction is also helpful for early stopping during sampling and for OOD detection. We hope distance field modeling can serve as a principled lens for generative modeling.

Method: Proposes Non-uniform Linear Interpolation (NLI) - a calibration-free, dynamic-programming-optimal framework that recasts cutpoint selection as a dynamic programming problem to achieve globally minimal interpolation error in O(M×N²) time using Bellman’s optimality principle.

Result: Hardware experiments show the NLI Engine achieves over 4× improvement in computational efficiency compared to state-of-the-art designs, enabling seamless integration into LLMs with almost no accuracy loss.

Conclusion: NLI provides an efficient, hardware-friendly solution for approximating nonlinear functions in LLMs, addressing a critical bottleneck in model deployment while maintaining accuracy.

Abstract: Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of tasks, but their deployment is often constrained by substantial memory footprints and computational costs. While prior work has achieved significant progress in compressing and accelerating linear layers, nonlinear layers-such as SiLU, RMSNorm, and Softmax-still heavily depend on high-precision floating-point operations. In this paper, we propose a calibration-free, dynamic-programming-optimal, and hardware-friendly framework called Non-uniform Linear Interpolation (NLI). NLI is capable of efficiently approximating a variety of nonlinear functions, enabling seamless integration into LLMs and other deep neural networks with almost no loss in accuracy. NLI ingeniously recasts cutpoint selection as a dynamic-programming problem, achieving the globally minimal interpolation error in O(MxN2) time via Bellman’s optimality principle. Based on the NLI algorithm, we also design and implement a plug-and-play universal nonlinear computation unit. Hardware experiments demonstrate that the NLI Engine achieves more than 4x improvement in computational efficiency compared to the state-of-the-art designs.

Method: Created standardized post-trauma sepsis onset dataset from MIMIC-III by extracting, relabeling using standardized post-trauma clinical facts, and validating data; framed detection as daily rare event problem according to ICU clinical workflow.

Result: Established comprehensive benchmark showing necessity for further advancements using this new dataset; data and code publicly available on GitHub.

Conclusion: Targeted identification of post-traumatic sepsis is needed for early detection; new dataset enables development of specialized methods for trauma patients in ICU settings.

Abstract: Sepsis is a major public health concern due to its high morbidity, mortality, and cost. Its clinical outcome can be substantially improved through early detection and timely intervention. By leveraging publicly available datasets, machine learning (ML) has driven advances in both research and clinical practice. However, existing public datasets consider ICU patients (Intensive Care Unit) as a uniform group and neglect the potential challenges presented by critically ill trauma patients in whom injury-related inflammation and organ dysfunction can overlap with the clinical features of sepsis. We propose that a targeted identification of post-traumatic sepsis is necessary in order to develop methods for early detection. Therefore, we introduce a publicly available standardized post-trauma sepsis onset dataset extracted, relabeled using standardized post-trauma clinical facts, and validated from MIMIC-III. Furthermore, we frame early detection of post-trauma sepsis onset according to clinical workflow in ICUs in a daily basis resulting in a new rare event detection problem. We then establish a general benchmark through comprehensive experiments, which shows the necessity of further advancements using this new dataset. The data code is available at https://github.com/ML4UWHealth/SepsisOnset_TraumaCohort.git.

Method: Derived gradient noise scales from the geometry of dual norms for signSGD (ℓ∞) and specSGD (S∞), and proposed efficient variance estimation using local mini-batch gradients across distributed ranks.

Result: Adaptive batch sizing with non-Euclidean GNS matched validation loss of constant-batch baselines while reducing training steps by up to 66% for Signum and Muon on a 160M parameter Llama model.

Conclusion: Non-Euclidean gradient noise scales enable principled adaptive batch sizing for popular non-Euclidean optimizers, improving training efficiency without sacrificing performance.

Abstract: To maximize hardware utilization, modern machine learning systems typically employ large constant or manually tuned batch size schedules, relying on heuristics that are brittle and costly to tune. Existing adaptive strategies based on gradient noise scale (GNS) offer a principled alternative. However, their assumption of SGD’s Euclidean geometry creates a fundamental mismatch with popular optimizers based on generalized norms, such as signSGD / Signum ($\ell_\infty$) and stochastic spectral descent (specSGD) / Muon ($\mathcal{S}_\infty$). In this work, we derive gradient noise scales for signSGD and specSGD that naturally emerge from the geometry of their respective dual norms. To practically estimate these non-Euclidean metrics, we propose an efficient variance estimation procedure that leverages the local mini-batch gradients on different ranks in distributed data-parallel systems. Our experiments demonstrate that adaptive batch size strategies using non-Euclidean GNS enable us to match the validation loss of constant-batch baselines while reducing training steps by up to 66% for Signum and Muon on a 160 million parameter Llama model.

Method: Proposes Distributionally Robust Decision-Focused Learning (DR-DFL) with Diffusion-Augmented approach (3D-Learning). Uses diffusion models to parameterize and search for worst-case distributions that remain realistic, balancing average and worst-case performance.

Result: Empirical results on LLM resource provisioning show 3D-Learning outperforms existing Distributionally Robust Optimization and Data Augmentation methods in OOD generalization.

Conclusion: 3D-Learning effectively addresses OOD generalization challenges by leveraging diffusion models to find realistic worst-case distributions, improving decision performance in ML pipelines.

Abstract: Predict-then-Optimize (PTO) pipelines are widely employed in computing and networked systems, where Machine Learning (ML) models are used to predict critical contextual information for downstream decision-making tasks such as cloud LLM serving, data center demand response, and edge workload scheduling. However, these ML predictors are often vulnerable to out-of-distribution (OOD) samples at test time, leading to significant decision performance degradation due to large prediction errors. To address the generalization challenges under OOD conditions, we present the framework of Distributionally Robust Decision-Focused Learning (DR-DFL), which trains ML models to optimize decision performance under the worst-case distribution. Instead of relying on classical Distributionally Robust Optimization (DRO) techniques, we propose Diffusion-Augmented Distributionally Robust Decision-Focused Learning (3D-Learning), which searches for the worst-case distribution within the parameterized space of a diffusion model. By leveraging the powerful distribution modeling capabilities of diffusion models, 3D-Learning identifies worst-case distributions that remain consistent with real data, achieving a favorable balance between average and worst-case scenarios. Empirical results on an LLM resource provisioning task demonstrate that 3D-Learning outperforms existing DRO and Data Augmentation methods in OOD generalization performance.

Method: Attaches lightweight exit heads at intermediate depths, uses Decoupled Knowledge Distillation (DKD) to distill a strong teacher into all exits, enforces deep-to-shallow consistency, and calibrates per-exit stopping thresholds using conformal risk control on held-out data.

Result: Across multiple datasets and architectures, SAFE-KD yields improved accuracy-compute trade-offs, stronger calibration, robust performance under corruption, and provides finite-sample risk guarantees.

Conclusion: SAFE-KD provides a principled framework for early-exit networks with statistical guarantees, making them more practical for deployment by ensuring safe early exits with controlled risk.

Abstract: Early-exit networks reduce inference cost by allowing ``easy’’ inputs to stop early, but practical deployment hinges on knowing \emph{when} early exit is safe. We introduce SAFE-KD, a universal multi-exit wrapper for modern vision backbones that couples hierarchical distillation with \emph{conformal risk control}. SAFE-KD attaches lightweight exit heads at intermediate depths, distills a strong teacher into all exits via Decoupled Knowledge Distillation (DKD), and enforces deep-to-shallow consistency between exits. At inference, we calibrate per-exit stopping thresholds on a held-out set using conformal risk control (CRC) to guarantee a user-specified \emph{selective} misclassification risk (among the samples that exit early) under exchangeability. Across multiple datasets and architectures, SAFE-KD yields improved accuracy compute trade-offs, stronger calibration, and robust performance under corruption while providing finite-sample risk guarantees.

Method: Combines spatial self-attention mechanism (SSAM) for learning correlation graphs, a three-step causal graph structure learning algorithm based on causal invariance principle to derive causal graphs, and graph convolutional LSTM (GCLSTM) spatial-temporal encoder-decoder for sequence-to-sequence reconstruction.

Result: Validated effectiveness through Tennessee Eastman process and real-world air separation process, demonstrating improved process monitoring and fault detection capabilities using feature space and residual space statistics.

Conclusion: CGSTAE provides an effective framework for industrial process monitoring that combines causal discovery with spatial-temporal modeling, offering both reliability and interpretability for fault detection applications.

Abstract: To improve the reliability and interpretability of industrial process monitoring, this article proposes a Causal Graph Spatial-Temporal Autoencoder (CGSTAE). The network architecture of CGSTAE combines two components: a correlation graph structure learning module based on spatial self-attention mechanism (SSAM) and a spatial-temporal encoder-decoder module utilizing graph convolutional long-short term memory (GCLSTM). The SSAM learns correlation graphs by capturing dynamic relationships between variables, while a novel three-step causal graph structure learning algorithm is introduced to derive a causal graph from these correlation graphs. The algorithm leverages a reverse perspective of causal invariance principle to uncover the invariant causal graph from varying correlations. The spatial-temporal encoder-decoder, built with GCLSTM units, reconstructs time-series process data within a sequence-to-sequence framework. The proposed CGSTAE enables effective process monitoring and fault detection through two statistics in the feature space and residual space. Finally, we validate the effectiveness of CGSTAE in process monitoring through the Tennessee Eastman process and a real-world air separation process.

Method: Pairs a standard VAE encoding observations with a sparse VAE encoding quantities of interest, connected through learned latent mapping. Uses hard-concrete spike-and-slab relaxation for differentiable training and beta hyperprior for adaptive sparsity.

Result: Demonstrated effectiveness on blind inpainting and computed tomography experiments, showing vsPAIR is a capable inverse problem solver providing interpretable and structured uncertainty estimates.

Conclusion: vsPAIR successfully addresses the challenge of providing interpretable uncertainty in inverse problems through its paired sparse VAE architecture with learned latent mapping.

Abstract: Inverse problems are fundamental to many scientific and engineering disciplines; they arise when one seeks to reconstruct hidden, underlying quantities from noisy measurements. Many applications demand not just point estimates but interpretable uncertainty. Providing fast inference alongside uncertainty estimates remains challenging yet desirable in numerous applications. We propose the Variational Sparse Paired Autoencoder (vsPAIR) to address this challenge. The architecture pairs a standard VAE encoding observations with a sparse VAE encoding quantities of interest, connected through a learned latent mapping. The variational structure enables uncertainty estimation, the paired architecture encourages interpretability by anchoring QoI representations to clean data, and sparse encodings provide structure by concentrating information into identifiable factors rather than diffusing across all dimensions. We also propose modifications to existing sparse VAE methods: a hard-concrete spike-and-slab relaxation for differentiable training and a beta hyperprior for adaptive sparsity levels. To validate the effectiveness of our proposed architecture, we conduct experiments on blind inpainting and computed tomography, demonstrating that vsPAIR is a capable inverse problem solver that can provide interpretable and structured uncertainty estimates.

Method: Unify homotopy problems under a single framework and propose Neural Predictor-Corrector (NPC) which formulates policy selection as a sequential decision-making problem, using reinforcement learning to automatically discover efficient strategies. Introduce amortized training for one-time offline training on problem classes and efficient online inference.

Result: Experiments on four representative homotopy problems show NPC generalizes effectively to unseen instances, consistently outperforms classical and specialized baselines in efficiency, and demonstrates superior stability across tasks.

Conclusion: The unified neural framework for homotopy methods successfully replaces hand-crafted heuristics with learned policies, demonstrating the value of unifying homotopy methods into a single neural framework with better efficiency and stability.

Abstract: The Homotopy paradigm, a general principle for solving challenging problems, appears across diverse domains such as robust optimization, global optimization, polynomial root-finding, and sampling. Practical solvers for these problems typically follow a predictor-corrector (PC) structure, but rely on hand-crafted heuristics for step sizes and iteration termination, which are often suboptimal and task-specific. To address this, we unify these problems under a single framework, which enables the design of a general neural solver. Building on this unified view, we propose Neural Predictor-Corrector (NPC), which replaces hand-crafted heuristics with automatically learned policies. NPC formulates policy selection as a sequential decision-making problem and leverages reinforcement learning to automatically discover efficient strategies. To further enhance generalization, we introduce an amortized training mechanism, enabling one-time offline training for a class of problems and efficient online inference on new instances. Experiments on four representative homotopy problems demonstrate that our method generalizes effectively to unseen instances. It consistently outperforms classical and specialized baselines in efficiency while demonstrating superior stability across tasks, highlighting the value of unifying homotopy methods into a single neural framework.

Method: QVG uses Semantic Aware Smoothing to leverage video spatiotemporal redundancy for producing low-magnitude, quantization-friendly residuals, and Progressive Residual Quantization - a coarse-to-fine multi-stage scheme that reduces quantization error while enabling quality-memory trade-offs.

Result: QVG reduces KV cache memory by up to 7.0 times with less than 4% end-to-end latency overhead, establishes a new Pareto frontier between quality and memory efficiency, and consistently outperforms existing baselines in generation quality across multiple benchmarks.

Conclusion: QVG effectively addresses the KV cache memory bottleneck in autoregressive video diffusion models through training-free quantization techniques, enabling deployment on more accessible hardware while maintaining generation quality.

Abstract: Despite rapid progress in autoregressive video diffusion, an emerging system algorithm bottleneck limits both deployability and generation capability: KV cache memory. In autoregressive video generation models, the KV cache grows with generation history and quickly dominates GPU memory, often exceeding 30 GB, preventing deployment on widely available hardware. More critically, constrained KV cache budgets restrict the effective working memory, directly degrading long horizon consistency in identity, layout, and motion. To address this challenge, we present Quant VideoGen (QVG), a training free KV cache quantization framework for autoregressive video diffusion models. QVG leverages video spatiotemporal redundancy through Semantic Aware Smoothing, producing low magnitude, quantization friendly residuals. It further introduces Progressive Residual Quantization, a coarse to fine multi stage scheme that reduces quantization error while enabling a smooth quality memory trade off. Across LongCat Video, HY WorldPlay, and Self Forcing benchmarks, QVG establishes a new Pareto frontier between quality and memory efficiency, reducing KV cache memory by up to 7.0 times with less than 4% end to end latency overhead while consistently outperforming existing baselines in generation quality.

Method: FedKRSO uses K-seed random subspaces: server generates shared low-dimensional subspaces, clients update models within these subspaces to save memory, and transmit only subspace accumulators instead of full parameters, enabling efficient aggregation.

Result: Extensive experiments on GLUE benchmark show FedKRSO achieves superior performance with low communication/memory overhead, closely approximating federated FFT performance while overcoming PEFT limitations.

Conclusion: FedKRSO provides an efficient solution for federated LLM fine-tuning at the edge, balancing performance and resource efficiency through subspace optimization techniques.

Abstract: Fine-tuning is essential to adapt general-purpose large language models (LLMs) to domain-specific tasks. As a privacy-preserving framework to leverage decentralized data for collaborative model training, Federated Learning (FL) is gaining popularity in LLM fine-tuning, but remains challenging due to the high cost of transmitting full model parameters and computing full gradients on resource-constrained clients. While Parameter-Efficient Fine-Tuning (PEFT) methods are widely used in FL to reduce communication and memory costs, they often sacrifice model performance compared to FFT. This paper proposes FedKRSO (Federated $K$-Seed Random Subspace Optimization), a novel method that enables communication and memory efficient FFT of LLMs in federated settings. In FedKRSO, clients update the model within a shared set of random low-dimension subspaces generated by the server to save memory usage. Furthermore, instead of transmitting full model parameters in each FL round, clients send only the model update accumulators along the subspaces to the server, enabling efficient global model aggregation and dissemination. By using these strategies, FedKRSO can substantially reduce communication and memory overhead while overcoming the performance limitations of PEFT, closely approximating the performance of federated FFT. The convergence properties of FedKRSO are analyzed rigorously under general FL settings. Extensive experiments on the GLUE benchmark across diverse FL scenarios demonstrate that FedKRSO achieves both superior performance and low communication and memory overhead, paving the way towards on federated LLM fine-tuning at the resource-constrained edge.

Method: Proposes MA2B-DDQN with action-branching discrete control that decomposes corridor control into local per-intersection actions (green time allocation between phases) and a single global action (total phase duration), using a human-centric reward that penalizes delayed individuals across all travel modes.

Result: Extensive evaluations across seven realistic traffic scenarios in Melbourne show significant reduction in impacted travelers compared to existing DRL and baseline methods, with minimal variance across diverse settings demonstrating robustness.

Conclusion: The framework provides a scalable, fair traffic signal system adaptable to varied urban conditions, advocating for human-centric optimization rather than vehicle-centric approaches.

Abstract: Coordinating traffic signals along multimodal corridors is challenging because many multi-agent deep reinforcement learning (DRL) approaches remain vehicle-centric and struggle with high-dimensional discrete action spaces. We propose MA2B-DDQN, a human-centric multi-agent action-branching double Deep Q-Network (DQN) framework that explicitly optimizes traveler-level equity. Our key contribution is an action-branching discrete control formulation that decomposes corridor control into (i) local, per-intersection actions that allocate green time between the next two phases and (ii) a single global action that selects the total duration of those phases. This decomposition enables scalable coordination under discrete control while reducing the effective complexity of joint decision-making. We also design a human-centric reward that penalizes the number of delayed individuals in the corridor, accounting for pedestrians, vehicle occupants, and transit passengers. Extensive evaluations across seven realistic traffic scenarios in Melbourne, Australia, demonstrate that our approach significantly reduces the number of impacted travelers, outperforming existing DRL and baseline methods. Experiments confirm the robustness of our model, showing minimal variance across diverse settings. This framework not only advocates for a fairer traffic signal system but also provides a scalable solution adaptable to varied urban traffic conditions.

Method: Proposes Spectral Evolution Search (SES) based on spectral bias insight: model sensitivity to initial perturbations diminishes rapidly as frequency increases. Executes gradient-free evolutionary search within a low-frequency subspace derived from perturbation propagation dynamics.

Result: SES significantly advances the Pareto frontier of generation quality versus computational cost, consistently outperforming strong baselines under equivalent budgets across extensive experiments.

Conclusion: SES provides an efficient plug-and-play framework for inference-time optimization of visual generative models by exploiting spectral bias properties, offering better computational efficiency than existing approaches.

Abstract: Inference-time scaling offers a versatile paradigm for aligning visual generative models with downstream objectives without parameter updates. However, existing approaches that optimize the high-dimensional initial noise suffer from severe inefficiency, as many search directions exert negligible influence on the final generation. We show that this inefficiency is closely related to a spectral bias in generative dynamics: model sensitivity to initial perturbations diminishes rapidly as frequency increases. Building on this insight, we propose Spectral Evolution Search (SES), a plug-and-play framework for initial noise optimization that executes gradient-free evolutionary search within a low-frequency subspace. Theoretically, we derive the Spectral Scaling Prediction from perturbation propagation dynamics, which explains the systematic differences in the impact of perturbations across frequencies. Extensive experiments demonstrate that SES significantly advances the Pareto frontier of generation quality versus computational cost, consistently outperforming strong baselines under equivalent budgets.

Method: C-DEQ leverages consistency distillation by casting DEQ iterative inference as evolution along a fixed ODE trajectory. It trains models to consistently map intermediate states directly to the equilibrium point, enabling few-step inference while maintaining teacher DEQ performance.

Result: Extensive experiments show C-DEQs achieve consistent 2-20× accuracy improvements over implicit DEQs under the same few-step inference budget, while facilitating flexible multi-step evaluation for computation-performance trade-offs.

Conclusion: C-DEQ provides an effective framework for accelerating DEQ inference through consistency distillation, achieving significant accuracy gains with few-step inference while preserving the benefits of equilibrium modeling.

Abstract: Deep Equilibrium Models (DEQs) have emerged as a powerful paradigm in deep learning, offering the ability to model infinite-depth networks with constant memory usage. However, DEQs incur significant inference latency due to the iterative nature of fixed-point solvers. In this work, we introduce the Consistency Deep Equilibrium Model (C-DEQ), a novel framework that leverages consistency distillation to accelerate DEQ inference. We cast the DEQ iterative inference process as evolution along a fixed ODE trajectory toward the equilibrium. Along this trajectory, we train C-DEQs to consistently map intermediate states directly to the fixed point, enabling few-step inference while preserving the performance of the teacher DEQ. At the same time, it facilitates multi-step evaluation to flexibly trade computation for performance gains. Extensive experiments across various domain tasks demonstrate that C-DEQs achieves consistent 2-20$\times$ accuracy improvements over implicit DEQs under the same few-step inference budget.

Method: Proposes Q-ShiftDP, a differentially private parameter-shift rule that leverages the inherent boundedness and stochasticity of quantum gradients. Combines carefully calibrated Gaussian noise with intrinsic quantum noise from quantum gradient estimation.

Result: Q-ShiftDP enables tighter sensitivity analysis, reduces noise requirements, and provides formal privacy and utility guarantees. Experiments on benchmark datasets show it consistently outperforms classical DP methods in QML.

Conclusion: Q-ShiftDP is the first privacy mechanism tailored to QML that effectively exploits quantum gradient properties, improving privacy-utility trade-offs by combining calibrated noise with intrinsic quantum noise.

Abstract: Quantum Machine Learning (QML) promises significant computational advantages, but preserving training data privacy remains challenging. Classical approaches like differentially private stochastic gradient descent (DP-SGD) add noise to gradients but fail to exploit the unique properties of quantum gradient estimation. In this work, we introduce the Differentially Private Parameter-Shift Rule (Q-ShiftDP), the first privacy mechanism tailored to QML. By leveraging the inherent boundedness and stochasticity of quantum gradients computed via the parameter-shift rule, Q-ShiftDP enables tighter sensitivity analysis and reduces noise requirements. We combine carefully calibrated Gaussian noise with intrinsic quantum noise to provide formal privacy and utility guarantees, and show that harnessing quantum noise further improves the privacy-utility trade-off. Experiments on benchmark datasets demonstrate that Q-ShiftDP consistently outperforms classical DP methods in QML.

Method: CoBA-RL uses a Capability-Oriented Value function to map tasks to potential training gains, then employs a heap-based greedy strategy to self-calibrate computational resource allocation to samples with high training value.

Result: Extensive experiments show CoBA-RL effectively balances exploration and exploitation, delivering consistent generalization improvements across multiple challenging benchmarks.

Conclusion: Quantifying sample training value and optimizing budget allocation are crucial for advancing LLM post-training efficiency, with CoBA-RL demonstrating effective adaptive resource allocation.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a key approach for enhancing LLM reasoning.However, standard frameworks like Group Relative Policy Optimization (GRPO) typically employ a uniform rollout budget, leading to resource inefficiency. Moreover, existing adaptive methods often rely on instance-level metrics, such as task pass rates, failing to capture the model’s dynamic learning state. To address these limitations, we propose CoBA-RL, a reinforcement learning algorithm designed to adaptively allocate rollout budgets based on the model’s evolving capability. Specifically, CoBA-RL utilizes a Capability-Oriented Value function to map tasks to their potential training gains and employs a heap-based greedy strategy to efficiently self-calibrate the distribution of computational resources to samples with high training value. Extensive experiments demonstrate that our approach effectively orchestrates the trade-off between exploration and exploitation, delivering consistent generalization improvements across multiple challenging benchmarks. These findings underscore that quantifying sample training value and optimizing budget allocation are pivotal for advancing LLM post-training efficiency.

Method: Co2PO introduces a shared blackboard architecture for broadcasting positional intent and yield signals, governed by a learned hazard predictor that forecasts potential violations over extended temporal horizons. It integrates these forecasts into constrained optimization with risk-triggered communication and adaptive gating.

Result: Co2PO achieves higher returns compared to leading constrained baselines while converging to cost-compliant policies at deployment across complex multi-agent safety benchmarks. Ablation studies validate the necessity of risk-triggered communication, adaptive gating, and shared memory components.

Conclusion: Co2PO demonstrates that proactive, communication-driven safety mechanisms can overcome the exploration-safety trade-off in constrained MARL, enabling better performance while maintaining safety constraints.

Abstract: Constrained multi-agent reinforcement learning (MARL) faces a fundamental tension between exploration and safety-constrained optimization. Existing leading approaches, such as Lagrangian methods, typically rely on global penalties or centralized critics that react to violations after they occur, often suppressing exploration and leading to over-conservatism. We propose Co2PO, a novel MARL communication-augmented framework that enables coordination-driven safety through selective, risk-aware communication. Co2PO introduces a shared blackboard architecture for broadcasting positional intent and yield signals, governed by a learned hazard predictor that proactively forecasts potential violations over an extended temporal horizon. By integrating these forecasts into a constrained optimization objective, Co2PO allows agents to anticipate and navigate collective hazards without the performance trade-offs inherent in traditional reactive constraints. We evaluate Co2PO across a suite of complex multi-agent safety benchmarks, where it achieves higher returns compared to leading constrained baselines while converging to cost-compliant policies at deployment. Ablation studies further validate the necessity of risk-triggered communication, adaptive gating, and shared memory components.

Method: Frames unlearning through asymptotic linear stability, analyzes data coherence (cross-sample alignment of loss surface directions), decomposes coherence along three axes (retain set, forget set, between them), and studies a two-layer ReLU CNN under signal-plus-noise model using random matrix theory tools.

Result: Proves tight stability thresholds separating convergence from divergence, shows that lower signal-to-noise ratio (weaker memorization) reduces coherence and makes unlearning easier, while high SNR models resist unlearning. Empirical verification shows Hessian tests and CNN heatmaps align with predicted boundaries.

Conclusion: Provides first principled account of trade-offs between memorization, coherence, and unlearning, establishing theoretical foundations for understanding gradient-based unlearning stability frontiers.

Abstract: Machine unlearning, the ability to erase the effect of specific training samples without retraining from scratch, is critical for privacy, regulation, and efficiency. However, most progress in unlearning has been empirical, with little theoretical understanding of when and why unlearning works. We tackle this gap by framing unlearning through the lens of asymptotic linear stability to capture the interaction between optimization dynamics and data geometry. The key quantity in our analysis is data coherence which is the cross sample alignment of loss surface directions near the optimum. We decompose coherence along three axes: within the retain set, within the forget set, and between them, and prove tight stability thresholds that separate convergence from divergence. To further link data properties to forgettability, we study a two layer ReLU CNN under a signal plus noise model and show that stronger memorization makes forgetting easier: when the signal to noise ratio (SNR) is lower, cross sample alignment is weaker, reducing coherence and making unlearning easier; conversely, high SNR, highly aligned models resist unlearning. For empirical verification, we show that Hessian tests and CNN heatmaps align closely with the predicted boundary, mapping the stability frontier of gradient based unlearning as a function of batching, mixing, and data/model alignment. Our analysis is grounded in random matrix theory tools and provides the first principled account of the trade offs between memorization, coherence, and unlearning.

Method: Proposes a semiparametric estimator using control variates, treating pairwise comparison signals (where models judge which of two candidate solutions is better) as auxiliary information. Develops a one-step estimator based on the efficient influence function (EIF) that achieves semiparametric efficiency bounds.

Result: The one-step estimator substantially improves ranking accuracy, with gains increasing as model output noise grows. Experiments on GPQA Diamond, AIME 2025, and GSM8K demonstrate more precise performance estimation and more reliable model rankings, especially in small-sample regimes.

Conclusion: The framework provides statistically efficient evaluation with guaranteed variance reduction over naive averaging, asymptotic normality for uncertainty quantification, and more stable model rankings particularly in data-limited scenarios.

Abstract: Evaluating mathematical reasoning in LLMs is constrained by limited benchmark sizes and inherent model stochasticity, yielding high-variance accuracy estimates and unstable rankings across platforms. On difficult problems, an LLM may fail to produce a correct final answer, yet still provide reliable pairwise comparison signals indicating which of two candidate solutions is better. We leverage this observation to design a statistically efficient evaluation framework that combines standard labeled outcomes with pairwise comparison signals obtained by having models judge auxiliary reasoning chains. Treating these comparison signals as control variates, we develop a semiparametric estimator based on the efficient influence function (EIF) for the setting where auxiliary reasoning chains are observed. This yields a one-step estimator that achieves the semiparametric efficiency bound, guarantees strict variance reduction over naive sample averaging, and admits asymptotic normality for principled uncertainty quantification. Across simulations, our one-step estimator substantially improves ranking accuracy, with gains increasing as model output noise grows. Experiments on GPQA Diamond, AIME 2025, and GSM8K further demonstrate more precise performance estimation and more reliable model rankings, especially in small-sample regimes where conventional evaluation is pretty unstable.

Method: Introduces APRIL dataset with 260,000 supervised tuples pairing systematically generated proof failures with compiler diagnostics, repair targets, and natural-language explanations. Trains language models on this dataset for proof repair.

Result: Training on APRIL substantially improves repair accuracy and feedback-conditioned reasoning. A finetuned 4B-parameter model outperforms the strongest open-source baseline in single-shot repair evaluation.

Conclusion: Diagnostic-conditioned supervision provides complementary training signal for feedback-using theorem provers. The APRIL dataset enables better proof repair capabilities in Lean.

Abstract: As neural theorem provers become increasingly agentic, the ability to interpret and act on compiler feedback is critical. However, existing Lean datasets consist almost exclusively of correct proofs, offering little supervision for understanding and repairing failures. We study Lean proof repair as a supervised learning problem: given an erroneous proof and compiler feedback, predict both a corrected proof and a natural-language diagnosis grounded in the same feedback. We introduce APRIL (Automated Proof Repair in Lean), a dataset of 260,000 supervised tuples pairing systematically generated proof failures with compiler diagnostics and aligned repair and explanation targets. Training language models on APRIL substantially improves repair accuracy and feedback-conditioned reasoning; in our single-shot repair evaluation setting, a finetuned 4B-parameter model outperforms the strongest open-source baseline. We view diagnostic-conditioned supervision as a complementary training signal for feedback-using provers. Our dataset is available at \href{https://huggingface.co/datasets/uw-math-ai/APRIL}{this link}.

Method: Used Neural Tangent Kernel (NTK) framework to analyze linear neural networks, defining features as eigenfunctions of NTK. Examined shortcut features in imbalanced clustered distributions, showing they correspond to larger eigenvalues. Extended analysis to two-layer ReLU networks and ResNet-18 empirically.

Result: Found that shortcut features have larger eigenvalues in imbalanced distributions, maintain influence on network output after training due to data variances, and persist even when controlling output margins, indicating max-margin bias isn’t the sole cause.

Conclusion: Shortcut learning stems from structural properties of neural networks where imbalanced features correspond to larger NTK eigenvalues, and this preference persists beyond margin control, requiring new approaches to mitigate shortcut learning.

Abstract: One of the chronic problems of deep-learning models is shortcut learning. In a case where the majority of training data are dominated by a certain feature, neural networks prefer to learn such a feature even if the feature is not generalizable outside the training set. Based on the framework of Neural Tangent Kernel (NTK), we analyzed the case of linear neural networks to derive some important properties of shortcut learning. We defined a feature of a neural network as an eigenfunction of NTK. Then, we found that shortcut features correspond to features with larger eigenvalues when the shortcuts stem from the imbalanced number of samples in the clustered distribution. We also showed that the features with larger eigenvalues still have a large influence on the neural network output even after training, due to data variances in the clusters. Such a preference for certain features remains even when a margin of a neural network output is controlled, which shows that the max-margin bias is not the only major reason for shortcut learning. These properties of linear neural networks are empirically extended for more complex neural networks as a two-layer fully-connected ReLU network and a ResNet-18.

Method: OUTFORMER advances FoMo-0D with: (1) mixture of synthetic priors for pretraining solely on synthetic labeled datasets, and (2) self-evolving curriculum training. It uses in-context learning with training data as input for zero-shot inference requiring only forward passes.

Result: OUTFORMER achieves state-of-the-art performance on AdBench and two new large-scale OD benchmarks comprising over 1,500 datasets, while maintaining speedy inference.

Conclusion: OUTFORMER enables truly plug-and-play outlier detection deployment with zero-shot inference, requiring no labeled outliers, model training, or bespoke model selection.

Abstract: Outlier detection (OD) is widely used in practice; but its effective deployment on new tasks is hindered by lack of labeled outliers, which makes algorithm and hyperparameter selection notoriously hard. Foundation models (FMs) have transformed ML, and OD is no exception: Shen et. al. (2025) introduced FoMo-0D, the first FM for OD, achieving remarkable performance against numerous baselines. This work introduces OUTFORMER, which advances FoMo-0D with (1) a mixture of synthetic priors and (2) self-evolving curriculum training. OUTFORMER is pretrained solely on synthetic labeled datasets and infers test labels of a new task by using its training data as in-context input. Inference is fast and zero-shot, requiring merely forward pass and no labeled outliers. Thanks to in-context learning, it requires zero additional work-no OD model training or bespoke model selection-enabling truly plug-and-play deployment. OUTFORMER achieves state-of-the-art performance on the prominent AdBench, as well as two new large-scale OD benchmarks that we introduce, comprising over 1,500 datasets, while maintaining speedy inference.

Method: Rewrites stabilized log-domain Sinkhorn updates as row-wise LogSumExp reductions of biased dot-product scores (same normalization as transformer attention), enabling FlashAttention-style fusion and tiling with fused Triton kernels that stream tiles through on-chip SRAM.

Result: Achieves up to 32× forward-pass and 161× end-to-end speedups over state-of-the-art online baselines on point-cloud OT, with improved scalability on OT-based downstream tasks.

Conclusion: FlashSinkhorn provides an efficient, scalable EOT solver that substantially reduces HBM IO while retaining linear-memory operations, with open-source implementation available.

Abstract: Entropic optimal transport (EOT) via Sinkhorn iterations is widely used in modern machine learning, yet GPU solvers remain inefficient at scale. Tensorized implementations suffer quadratic HBM traffic from dense $n\times m$ interactions, while existing online backends avoid storing dense matrices but still rely on generic tiled map-reduce reduction kernels with limited fusion. We present \textbf{FlashSinkhorn}, an IO-aware EOT solver for squared Euclidean cost that rewrites stabilized log-domain Sinkhorn updates as row-wise LogSumExp reductions of biased dot-product scores, the same normalization as transformer attention. This enables FlashAttention-style fusion and tiling: fused Triton kernels stream tiles through on-chip SRAM and update dual potentials in a single pass, substantially reducing HBM IO per iteration while retaining linear-memory operations. We further provide streaming kernels for transport application, enabling scalable first- and second-order optimization. On A100 GPUs, FlashSinkhorn achieves up to $32\times$ forward-pass and $161\times$ end-to-end speedups over state-of-the-art online baselines on point-cloud OT, improves scalability on OT-based downstream tasks. For reproducibility, we release an open-source implementation at https://github.com/ot-triton-lab/ot_triton.

Method: ProCAD uses a two-agent framework: (1) a proactive clarifying agent that audits prompts and asks targeted clarification questions only when necessary to produce self-consistent specifications, and (2) a CAD coding agent that translates specifications into executable CadQuery programs. The coding agent is fine-tuned on curated text-to-CadQuery data, while the clarifying agent is trained via agentic SFT on clarification trajectories.

Result: ProCAD significantly improves robustness to ambiguous prompts while keeping interaction overhead low. It outperforms frontier closed-source models like Claude Sonnet 4.5, reducing mean Chamfer distance by 79.9% and lowering invalidity ratio from 4.8% to 0.9%.

Conclusion: Proactive clarification before code synthesis effectively addresses ambiguity in text-to-CAD generation, leading to more robust and accurate parametric CAD program synthesis from natural language prompts.

Abstract: Large language models have recently enabled text-to-CAD systems that synthesize parametric CAD programs (e.g., CadQuery) from natural language prompts. In practice, however, geometric descriptions can be under-specified or internally inconsistent: critical dimensions may be missing and constraints may conflict. Existing fine-tuned models tend to reactively follow user instructions and hallucinate dimensions when the text is ambiguous. To address this, we propose a proactive agentic framework for text-to-CadQuery generation, named ProCAD, that resolves specification issues before code synthesis. Our framework pairs a proactive clarifying agent, which audits the prompt and asks targeted clarification questions only when necessary to produce a self-consistent specification, with a CAD coding agent that translates the specification into an executable CadQuery program. We fine-tune the coding agent on a curated high-quality text-to-CadQuery dataset and train the clarifying agent via agentic SFT on clarification trajectories. Experiments show that proactive clarification significantly improves robustness to ambiguous prompts while keeping interaction overhead low. ProCAD outperforms frontier closed-source models, including Claude Sonnet 4.5, reducing the mean Chamfer distance by 79.9 percent and lowering the invalidity ratio from 4.8 percent to 0.9 percent. Our code and datasets will be made publicly available.

Method: Proposes FEDCOMPASS with two key components: 1) Spectral clustering to group clients by class distribution similarity for cluster-wise aggregation of classical feature extractors, and 2) Circular mean aggregation combined with adaptive optimization for quantum parameters to ensure stable global updates.

Result: Experiments on three benchmark datasets show FEDCOMPASS improves test accuracy by up to 10.22% and enhances convergence stability under non-IID settings, outperforming six strong federated learning baselines.

Conclusion: FEDCOMPASS effectively addresses non-IID challenges in hybrid classical-quantum federated learning through its layered aggregation approach, demonstrating significant performance improvements and better convergence stability.

Abstract: Federated learning enables collaborative model training across decentralized clients under privacy constraints. Quantum computing offers potential for alleviating computational and communication burdens in federated learning, yet hybrid classical-quantum federated learning remains susceptible to performance degradation under non-IID data. To address this,we propose FEDCOMPASS, a layered aggregation framework for hybrid classical-quantum federated learning. FEDCOMPASS employs spectral clustering to group clients by class distribution similarity and performs cluster-wise aggregation for classical feature extractors. For quantum parameters, it uses circular mean aggregation combined with adaptive optimization to ensure stable global updates. Experiments on three benchmark datasets show that FEDCOMPASS improves test accuracy by up to 10.22% and enhances convergence stability under non-IID settings, outperforming six strong federated learning baselines.

Method: Proposes CaRE with Bi-Level Routing Mixture-of-Experts (BR-MoE): 1) router selection stage dynamically activates task-specific routers, 2) expert routing phase dynamically activates and aggregates experts to inject discriminative and comprehensive representations into every intermediate network layer.

Result: CaRE demonstrates leading performance across various datasets and task settings, including classical CIL settings (5-20 tasks). It’s the first continual learner scaling to very long task sequences (100-300+ non-overlapping tasks), outperforming all baselines by large margins on such sequences.

Conclusion: CaRE effectively addresses the scalability challenge in continual learning through its bi-level routing mechanism, enabling efficient learning over hundreds of tasks while maintaining both stability and plasticity.

Abstract: Continual learning, especially class-incremental learning (CIL), on the basis of a pre-trained model (PTM) has garnered substantial research interest in recent years. However, how to effectively learn both discriminative and comprehensive feature representations while maintaining stability and plasticity over very long task sequences remains an open problem. We propose CaRE, a scalable {C}ontinual Le{a}rner with efficient Bi-Level {R}outing Mixture-of-{E}xperts (BR-MoE). The core idea of BR-MoE is a bi-level routing mechanism: a router selection stage that dynamically activates relevant task-specific routers, followed by an expert routing phase that dynamically activates and aggregates experts, aiming to inject discriminative and comprehensive representations into every intermediate network layer. On the other hand, we introduce a challenging evaluation protocol for comprehensively assessing CIL methods across very long task sequences spanning hundreds of tasks. Extensive experiments show that CaRE demonstrates leading performance across a variety of datasets and task settings, including commonly used CIL datasets with classical CIL settings (e.g., 5-20 tasks). To the best of our knowledge, CaRE is the first continual learner that scales to very long task sequences (ranging from 100 to over 300 non-overlapping tasks), while outperforming all baselines by a large margin on such task sequences. Code will be publicly released at https://github.com/LMMMEng/CaRE.git.

Method: Trajectory-Mixed Supervision (TMS) creates a dynamic curriculum by mixing training data from the model’s own historical checkpoints to minimize Policy-Label Divergence (PLD), approximating on-policy benefits of RL without rewards.

Result: TMS outperforms standard and iterative SFT across reasoning (MATH, GSM8K) and instruction-following benchmarks, shifting the accuracy-retention Pareto frontier and bridging the gap to RL without requiring reward models.

Conclusion: TMS provides an effective reward-free alternative to RL for improving LLM performance while mitigating catastrophic forgetting, with PLD drift serving as a predictive measure for forgetting.

Abstract: Reinforcement Learning (RL) and Supervised Fine-Tuning (SFT) are the two dominant paradigms for enhancing Large Language Model (LLM) performance on downstream tasks. While RL generally preserves broader model capabilities (retention) better than SFT, it comes with significant costs: complex reward engineering, instability, and expensive on-policy sampling. In contrast, SFT is efficient but brittle, often suffering from catastrophic forgetting due to $\textbf{Supervision Mismatch}$: the divergence between the model’s evolving policy and static training labels. We address this trade-off with $\textbf{Trajectory-Mixed Supervision (TMS)}$, a reward-free framework that approximates the on-policy benefits of RL by creating a dynamic curriculum from the model’s own historical checkpoints. TMS minimizes $\textit{Policy-Label Divergence (PLD)}$, preventing the mode collapse that drives forgetting in standard SFT. Experiments across reasoning (MATH, GSM8K) and instruction-following benchmarks demonstrate that TMS effectively shifts the accuracy–retention Pareto frontier. While RL remains the gold standard for retention, TMS significantly outperforms standard and iterative SFT, bridging the gap to RL without requiring reward models or verifiers. Mechanistic analysis confirms that PLD drift accurately predicts forgetting and that TMS successfully mitigates this drift.

Method: Layer-wise analysis of token embeddings to study class separability across network depth, examination of attention patterns in MAE vs standard ViTs, and introduction of two sensitivity indicators: directional alignment between clean/perturbed embeddings and head-wise feature retention under degradations.

Result: MAE learns class-aware representations that become increasingly separable across layers, exhibits early and persistent global attention (unlike standard ViTs), and shows robustness to image degradations like blur and occlusions.

Conclusion: MAE’s strong classification performance stems from its ability to learn robust, class-aware representations through its pretraining objective, with progressive feature separation across layers and consistent global attention patterns.

Abstract: Masked Autoencoders (MAEs) achieve impressive performance in image classification tasks, yet the internal representations they learn remain less understood. This work started as an attempt to understand the strong downstream classification performance of MAE. In this process we discover that representations learned with the pretraining and fine-tuning, are quite robust - demonstrating a good classification performance in the presence of degradations, such as blur and occlusions. Through layer-wise analysis of token embeddings, we show that pretrained MAE progressively constructs its latent space in a class-aware manner across network depth: embeddings from different classes lie in subspaces that become increasingly separable. We further observe that MAE exhibits early and persistent global attention across encoder layers, in contrast to standard Vision Transformers (ViTs). To quantify feature robustness, we introduce two sensitivity indicators: directional alignment between clean and perturbed embeddings, and head-wise retention of active features under degradations. These studies help establish the robust classification performance of MAEs.

Method: Uses innovation-augmented polar decomposition to construct an efficient, low-rank quasi-second-order preconditioner that performs anisotropic spectral shaping - adaptively suppressing updates in high-variance subspaces while preserving update strength in signal-dominated directions.

Result: Achieves curvature-adaptive properties in spectral optimization with minimal computational overhead and zero additional memory compared to first-order baselines.

Conclusion: PRISM demonstrates a practical strategy for integrating curvature-adaptive properties into the spectral optimization paradigm while maintaining computational efficiency.

Abstract: We propose PRISM, an optimizer that enhances first-order spectral descent methods like Muon with partial second-order information. It constructs an efficient, low-rank quasi-second-order preconditioner via innovation-augmented polar decomposition. This mechanism enables PRISM to perform anisotropic spectral shaping, which adaptively suppresses updates in high-variance subspaces while preserving update strength in signal-dominated directions. Crucially, this is achieved with minimal computational overhead and zero additional memory compared to first-order baselines. PRISM demonstrates a practical strategy for integrating curvature-adaptive properties into the spectral optimization paradigm.

Method: Proposes a unified class of geometry-aware architectures that interleave geometric updates between layers, where both projection layers and intrinsic exponential map updates arise as discretizations of projected dynamical systems on manifolds. Also learns projections via small-time heat-kernel limits when constraint sets are unknown.

Result: Establishes universal approximation results for constrained neural ODEs. Demonstrates exact feasibility for analytic updates on S^2 and SO(3), and strong performance for learned projections on S^{d-1}-valued features using diffusion/flow-matching as data-based projections.

Conclusion: The framework provides principled geometry-aware neural architectures with theoretical guarantees, enabling effective learning on manifolds while preserving geometric structure through both analytic and learned projection mechanisms.

Abstract: Preserving geometric structure is important in learning. We propose a unified class of geometry-aware architectures that interleave geometric updates between layers, where both projection layers and intrinsic exponential map updates arise as discretizations of projected dynamical systems on manifolds (with or without boundary). Within this framework, we establish universal approximation results for constrained neural ODEs. We also analyze architectures that enforce geometry only at the output, proving a separate universal approximation property that enables direct comparison to interleaved designs. When the constraint set is unknown, we learn projections via small-time heat-kernel limits, showing diffusion/flow-matching can be used as data-based projections. Experiments on dynamics over S^2 and SO(3), and diffusion on S^{d-1}-valued features demonstrate exact feasibility for analytic updates and strong performance for learned projections

Method: TextME projects diverse modalities into LLM embedding space as a unified anchor, exploiting geometric structure of pretrained contrastive encoders to enable zero-shot cross-modal transfer using only text descriptions without paired supervision.

Result: Demonstrates consistent modality gaps exist across image, video, audio, 3D, X-ray, and molecular domains, showing text-only training preserves substantial performance of pretrained encoders and enables emergent cross-modal retrieval between modality pairs not explicitly aligned during training.

Conclusion: Text-only training serves as a practical alternative to paired supervision for modality expansion, enabling efficient expansion to novel modalities without costly dataset collection.

Abstract: Expanding multimodal representations to novel modalities is constrained by reliance on large-scale paired datasets (e.g., text-image, text-audio, text-3D, text-molecule), which are costly and often infeasible in domains requiring expert annotation such as medical imaging and molecular analysis. We introduce TextME, the first text-only modality expansion framework, to the best of our knowledge, projecting diverse modalities into LLM embedding space as a unified anchor. Our approach exploits the geometric structure of pretrained contrastive encoders to enable zero-shot cross-modal transfer using only text descriptions, without paired supervision. We empirically validate that such consistent modality gaps exist across image, video, audio, 3D, X-ray, and molecular domains, demonstrating that text-only training can preserve substantial performance of pretrained encoders. We further show that our framework enables emergent cross-modal retrieval between modality pairs not explicitly aligned during training (e.g., audio-to-image, 3D-to-image). These results establish text-only training as a practical alternative to paired supervision for modality expansion.

Method: Proposes Consensus Group Relative Policy Optimization (C-GRPO) which formulates consensus utility as a group-relative objective within GRPO framework. Only requires a utility function and policy samples, no gold references or explicit preference labels needed.

Result: Experiments on WMT 2024 machine translation and XSum text summarization show C-GRPO achieves performance comparable to MBR decoding without inference-time overhead, outperforming reference-free baseline methods.

Conclusion: C-GRPO successfully distills MBR decoding into training, eliminating computational costs while maintaining performance, with theoretical convergence guarantees under ideal conditions.

Abstract: Many strong decoding methods for text generation follow a sample-and-rerank paradigm: they draw multiple candidates, score each under a utility (reward) function using consensus across samples, and return the best one. Although effective, these methods incur high computational costs during inference due to repeated sampling and scoring. Prior attempts to amortize inference-time computation typically rely on gold references, teacher labels, or curated preference data, increasing dataset construction effort and the demand for high-fidelity reward models. We propose Consensus Group Relative Policy Optimization (C-GRPO), which distills Minimum Bayes Risk (MBR) decoding into training by formulating the consensus utility as a group-relative objective within GRPO. C-GRPO requires only a utility function and policy samples, without gold references or explicit preference labels. Under ideal conditions, we show that the objective function of C-GRPO is directionally aligned with the gradient of the expected-utility objective underlying MBR decoding, leading to a convergence guarantee. Experiments on machine translation (WMT 2024) and text summarization (XSum) demonstrate that C-GRPO successfully achieves performance comparable to MBR decoding without the associated inference-time overhead, while outperforming reference-free baseline methods.

Method: Proposes VLM-FS-EB: uses VLMs to generate semantically meaningful synthetic context points, then uses VLM embeddings to construct expressive functional priors in a function-space empirical Bayes regularization framework.

Result: Method consistently improves predictive performance and yields more reliable uncertainty estimates, especially in out-of-distribution detection tasks and data-scarce regimes compared to various baselines.

Conclusion: VLMs can effectively generate semantic context points for constructing expressive functional priors in Bayesian deep learning, addressing limitations of traditional GP priors in high-dimensional regimes.

Abstract: Bayesian deep learning (BDL) provides a principled framework for reliable uncertainty quantification by combining deep neural networks with Bayesian inference. A central challenge in BDL lies in the design of informative prior distributions that scale effectively to high-dimensional data. Recent functional variational inference (VI) approaches address this issue by imposing priors directly in function space; however, most existing methods rely on Gaussian process (GP) priors, whose expressiveness and generalisation capabilities become limited in high-dimensional regimes. In this work, we propose VLM-FS-EB, a novel function-space empirical Bayes regularisation framework, leveraging large vision-language models (VLMs) to generates semantically meaningful context points. These synthetic samples are then used VLMs for embeddings to construct expressive functional priors. Furthermore, the proposed method is evaluated against various baselines, and experimental results demonstrate that our method consistently improves predictive performance and yields more reliable uncertainty estimates, particularly in out-of-distribution (OOD) detection tasks and data-scarce regimes.

Method: Quantized Evolution Strategies (QES) with two innovations: (1) accumulated error feedback to preserve gradient signals, and (2) stateless seed replay to reduce memory usage to low-precision inference levels.

Result: QES significantly outperforms state-of-the-art zeroth-order fine-tuning on arithmetic reasoning tasks, enabling direct fine-tuning of quantized models.

Conclusion: QES opens up the possibility for scaling up LLMs entirely in the quantized space by making direct fine-tuning of quantized models possible.

Abstract: Post-Training Quantization (PTQ) is essential for deploying Large Language Models (LLMs) on memory-constrained devices, yet it renders models static and difficult to fine-tune. Standard fine-tuning paradigms, including Reinforcement Learning (RL), fundamentally rely on backpropagation and high-precision weights to compute gradients. Thus they cannot be used on quantized models, where the parameter space is discrete and non-differentiable. While Evolution Strategies (ES) offer a backpropagation-free alternative, optimization of the quantized parameters can still fail due to vanishing or inaccurate gradient. This paper introduces Quantized Evolution Strategies (QES), an optimization paradigm that performs full-parameter fine-tuning directly in the quantized space. QES is based on two innovations: (1) it integrates accumulated error feedback to preserve high-precision gradient signals, and (2) it utilizes a stateless seed replay to reduce memory usage to low-precision inference levels. QES significantly outperforms the state-of-the-art zeroth-order fine-tuning method on arithmetic reasoning tasks, making direct fine-tuning for quantized models possible. It therefore opens up the possibility for scaling up LLMs entirely in the quantized space. The source code is available at https://github.com/dibbla/Quantized-Evolution-Strategies .

Method: Contrastive Concept-Tree Search (CCTS) extracts hierarchical concept representations from generated programs, learns a contrastive concept model that guides parent selection, and reweights parents using likelihood-ratio scores between high- and low-performing solutions to bias search toward useful concept combinations.

Result: CCTS improves search efficiency over fitness-based baselines and produces interpretable, task-specific concept trees across Erdős-type combinatorics problems. Gains are largely driven by learning which concepts to avoid. Controlled synthetic environment validates these findings.

Conclusion: Explicit concept hierarchy extraction and contrastive learning significantly improve LLM-assisted algorithm discovery by providing better guidance than algorithm lineage alone, with interpretable concept trees offering insights into successful search strategies.

Abstract: Large language Model (LLM)-assisted algorithm discovery is an iterative, black-box optimization process over programs to approximatively solve a target task, where an LLM proposes candidate programs and an external evaluator provides task feedback. Despite intense recent research on the topic and promising results, how can the LLM internal representation of the space of possible programs be maximally exploited to improve performance is an open question. Here, we introduce Contrastive Concept-Tree Search (CCTS), which extracts a hierarchical concept representation from the generated programs and learns a contrastive concept model that guides parent selection. By reweighting parents using a likelihood-ratio score between high- and low-performing solutions, CCTS biases search toward useful concept combinations and away from misleading ones, providing guidance through an explicit concept hierarchy rather than the algorithm lineage constructed by the LLM. We show that CCTS improves search efficiency over fitness-based baselines and produces interpretable, task-specific concept trees across a benchmark of open Erdős-type combinatorics problems. Our analysis indicates that the gains are driven largely by learning which concepts to avoid. We further validate these findings in a controlled synthetic algorithm-discovery environment, which reproduces qualitatively the search dynamics observed with the LLMs.

Method: Novel deep learning-based ensemble framework that leverages historical arrival patterns and real-time parcel status updates to forecast hub workloads.

Result: Empirical tests show the ensemble method outperforms traditional forecasting techniques and standalone deep learning models in a major city case study.

Conclusion: The method has significant potential to improve operational efficiency in logistics hubs and warrants broader adoption.

Abstract: The rapid expansion of online shopping has increased the demand for timely parcel delivery, compelling logistics service providers to enhance the efficiency, agility, and predictability of their hub networks. In order to solve the problem, we propose a novel deep learning-based ensemble framework that leverages historical arrival patterns and real-time parcel status updates to forecast upcoming workloads at logistic hubs. This approach not only facilitates the generation of short-term forecasts, but also improves the accuracy of future hub workload predictions for more strategic planning and resource management. Empirical tests of the algorithm, conducted through a case study of a major city’s parcel logistics, demonstrate the ensemble method’s superiority over both traditional forecasting techniques and standalone deep learning models. Our findings highlight the significant potential of this method to improve operational efficiency in logistics hubs and advocate for its broader adoption.

Method: Proposes a subspace-aware semidefinite programming formulation of nuclear norm that explicitly informs reconstruction with prior singular-subspace information. Also studies a lightweight implicit subspace-alignment strategy via concatenation of matrices from consecutive days to encourage alignment of spatial/temporal singular directions.

Result: Consistently outperforms standard matrix completion methods (SoftImpute, IterativeSVD), statistical, and deep learning baselines at high occlusion levels across Beijing and Shanghai datasets. Explicit SDP approach is markedly more robust than implicit strategy when large fractions of entries are missing.

Conclusion: SATORIS-N provides accurate traffic-density reconstruction essential for intelligent vehicles and V2X systems. The framework generalizes to other spatiotemporal settings where singular subspaces evolve slowly over time.

Abstract: Traffic-density matrices from different days exhibit both low rank and stable correlations in their singular-vector subspaces. Leveraging this, we introduce SATORIS-N, a framework for imputing partially observed traffic-density by informed subspace priors from neighboring days. Our contribution is a subspace-aware semidefinite programming (SDP)} formulation of nuclear norm that explicitly informs the reconstruction with prior singular-subspace information. This convex formulation jointly enforces low rank and subspace alignment, providing a single global optimum and substantially improving accuracy under medium and high occlusion. We also study a lightweight implicit subspace-alignment} strategy in which matrices from consecutive days are concatenated to encourage alignment of spatial or temporal singular directions. Although this heuristic offers modest gains when missing rates are low, the explicit SDP approach is markedly more robust when large fractions of entries are missing. Across two real-world datasets (Beijing and Shanghai), SATORIS-N consistently outperforms standard matrix-completion methods such as SoftImpute, IterativeSVD, statistical, and even deep learning baselines at high occlusion levels. The framework generalizes to other spatiotemporal settings in which singular subspaces evolve slowly over time. In the context of intelligent vehicles and vehicle-to-everything (V2X) systems, accurate traffic-density reconstruction enables critical applications including cooperative perception, predictive routing, and vehicle-to-infrastructure (V2I) communication optimization. When infrastructure sensors or vehicle-reported observations are incomplete - due to communication dropouts, sensor occlusions, or sparse connected vehicle penetration-reliable imputation becomes essential for safe and efficient autonomous navigation.

Method: Reformulates TSF as experience-conditioned reasoning task with hierarchical memory: historical patterns from predictions, reasoning wisdom from inference trajectories, and general laws from temporal features. Uses dynamic confidence adaptation for continual evolution without test set distribution leakage.

Result: Extensive experiments on multiple datasets show MemCast consistently outperforms previous methods, validating the effectiveness of the hierarchical memory and continual evolution approach.

Conclusion: MemCast successfully addresses the limitations of existing LLM-based forecasters by introducing explicit experience accumulation and continual evolution through hierarchical memory organization and dynamic confidence adaptation.

Abstract: Time series forecasting (TSF) plays a critical role in decision-making for many real-world applications. Recently, LLM-based forecasters have made promising advancements. Despite their effectiveness, existing methods often lack explicit experience accumulation and continual evolution. In this work, we propose MemCast, a learning-to-memory framework that reformulates TSF as an experience-conditioned reasoning task. Specifically, we learn experience from the training set and organize it into a hierarchical memory. This is achieved by summarizing prediction results into historical patterns, distilling inference trajectories into reasoning wisdom, and inducing extracted temporal features into general laws. Furthermore, during inference, we leverage historical patterns to guide the reasoning process and utilize reasoning wisdom to select better trajectories, while general laws serve as criteria for reflective iteration. Additionally, to enable continual evolution, we design a dynamic confidence adaptation strategy that updates the confidence of individual entries without leaking the test set distribution. Extensive experiments on multiple datasets demonstrate that MemCast consistently outperforms previous methods, validating the effectiveness of our approach. Our code is available at https://github.com/Xiaoyu-Tao/MemCast-TS.

Method: Used pretrained vision models (both convolutional and transformer-based) to embed novel visual categories along a one-dimensional morph continuum. Tested various subset selection strategies emphasizing prototypicality, joint representativeness, and diversity. Had adult participants select 1-3 exemplars for teaching, then compared human judgments with model predictions.

Result: Strategies based on joint representativeness, or its combination with diversity, best captured human exemplar selection. Purely prototypical or diversity-based strategies performed worse. Transformer-based representations consistently aligned more closely with human behavior than convolutional networks.

Conclusion: Dataset distillation methods from machine learning can serve as computational models for human teaching behavior, with transformer representations providing better alignment with human exemplar selection strategies.

Abstract: Teaching requires distilling a rich category distribution into a small set of informative exemplars. Although prior work shows that humans consider both representativeness and diversity when teaching, the computational principles underlying these tradeoffs remain unclear. We address this gap by modeling human exemplar selection using neural network feature representations and principled subset selection criteria. Novel visual categories were embedded along a one-dimensional morph continuum using pretrained vision models, and selection strategies varied in their emphasis on prototypicality, joint representativeness, and diversity. Adult participants selected one to three exemplars to teach a learner. Model-human comparisons revealed that strategies based on joint representativeness, or its combination with diversity, best captured human judgments, whereas purely prototypical or diversity-based strategies performed worse. Moreover, transformer-based representations consistently aligned more closely with human behavior than convolutional networks. These results highlight the potential utility of dataset distillation methods in machine learning as computational models for teaching.

Method: RLPT decouples strategic decision-making from token generation by: 1) Using base model’s semantic priors to identify a dynamic set of promising tokens, 2) Constraining policy optimization exclusively to this refined subset via masking, 3) Reducing gradient variance by focusing on relevant action space.

Result: RLPT outperforms standard RL baselines on math, coding, and telecom reasoning tasks. It reduces gradient variance, stabilizes training, improves sample efficiency, and works effectively across various model sizes (4B and 8B) and RL algorithms (GRPO and DAPO).

Conclusion: By focusing RL optimization on promising tokens rather than the full vocabulary, RLPT addresses the action space complexity problem in LLM alignment, leading to more stable and efficient training while maintaining performance across diverse reasoning tasks.

Abstract: Reinforcement learning (RL) has emerged as a key paradigm for aligning and optimizing large language models (LLMs). Standard approaches treat the LLM as the policy and apply RL directly over the full vocabulary space. However, this formulation includes the massive tail of contextually irrelevant tokens in the action space, which could distract the policy from focusing on decision-making among the truly reasonable tokens. In this work, we verify that valid reasoning paths could inherently concentrate within a low-rank subspace. Based on this insight, we introduce Reinforcement Learning with Promising Tokens (RLPT), a framework that mitigates the action space issue by decoupling strategic decision-making from token generation. Specifically, RLPT leverages the semantic priors of the base model to identify a dynamic set of \emph{promising tokens} and constrains policy optimization exclusively to this refined subset via masking. Theoretical analysis and empirical results demonstrate that RLPT effectively reduces gradient variance, stabilizes the training process, and improves sample efficiency. Experiment results on math, coding, and telecom reasoning show that RLPT outperforms standard RL baselines and integrates effectively across various model sizes (4B and 8B) and RL algorithms (GRPO and DAPO).

Method: Introduces Psychological Regret Model (PRM) that computes regret signals after each decision step based on the difference between expected optimal action value and actual action value. This transforms sparse rewards into dense feedback through step-wise scoring.

Result: PRM achieves stable performance approximately 36% faster than traditional PPO in benchmark environments like Lunar Lander. It’s particularly effective in continuous control tasks and environments with delayed feedback.

Conclusion: PRM bridges behavioral economics and RL by formalizing human-inspired counterfactual thinking as computable regret signals, making it suitable for real-world applications requiring rapid policy adaptation.

Abstract: Reinforcement learning algorithms often suffer from slow convergence due to sparse reward signals, particularly in complex environments where feedback is delayed or infrequent. This paper introduces the Psychological Regret Model (PRM), a novel approach that accelerates learning by incorporating regret-based feedback signals after each decision step. Rather than waiting for terminal rewards, PRM computes a regret signal based on the difference between the expected value of the optimal action and the value of the action taken in each state. This transforms sparse rewards into dense feedback signals through a step-wise scoring framework, enabling faster convergence. We demonstrate that PRM achieves stable performance approximately 36% faster than traditional Proximal Policy Optimization (PPO) in benchmark environments such as Lunar Lander. Our results indicate that PRM is particularly effective in continuous control tasks and environments with delayed feedback, making it suitable for real-world applications such as robotics, finance, and adaptive education where rapid policy adaptation is critical. The approach formalizes human-inspired counterfactual thinking as a computable regret signal, bridging behavioral economics and reinforcement learning.

Method: Proposes computing Expected Future Reward (EFR) using only marginal samples from pre-trained diffusion models, detaching neural dependency between current state and EFR. Introduces lookahead sampling to collect marginal samples efficiently, and uses an accurate solver to guide particles toward high-reward lookahead samples (LiDAR sampling).

Result: LiDAR achieves substantial performance improvements with only three samples and 3-step lookahead, showing steep performance gains with increased lookahead accuracy and sample count. Reaches same GenEval performance as latest gradient guidance method for SDXL with 9.5x speedup.

Conclusion: LiDAR sampling provides an efficient test-time scaling method for aligning diffusion model outputs with human intent, significantly reducing computational cost while maintaining or improving performance.

Abstract: Diffusion models have demonstrated strong generative performance; however, generated samples often fail to fully align with human intent. This paper studies a test-time scaling method that enables sampling from regions with higher human-aligned reward values. Existing gradient guidance methods approximate the expected future reward (EFR) at an intermediate particle $\mathbf{x}_t$ using a Taylor approximation, but this approximation at each time step incurs high computational cost due to sequential neural backpropagation. We show that the EFR at any $\mathbf{x}_t$ can be computed using only marginal samples from a pre-trained diffusion model. The proposed EFR formulation detaches the neural dependency between $\mathbf{x}_t$ and the EFR, enabling closed-form guidance computation without neural backpropagation. To further improve efficiency, we introduce lookahead sampling to collect marginal samples. For final sample generation, we use an accurate solver that guides particles toward high-reward lookahead samples. We refer to this sampling scheme as LiDAR sampling. LiDAR achieves substantial performance improvements using only three samples with a 3-step lookahead solver, exhibiting steep performance gains as lookahead accuracy and sample count increase; notably, it reaches the same GenEval performance as the latest gradient guidance method for SDXL with a 9.5x speedup.