Authors: Zhuoming Chen, Hongyi Liu, Yang Zhou, Haizhong Zheng, Beidi Chen
Abstract: Reinforcement learning (RL) for large language models (LLMs) remains expensive, particularly because the rollout is expensive. Decoupling rollout generation from policy optimization (e.g., leveraging a more efficient model to rollout) could enable substantial efficiency gains, yet doing so introduces a severe distribution mismatch that destabilizes learning. We propose Jackpot, a framework that leverages Optimal Budget Rejection Sampling (OBRS) to directly reduce the discrepancy between the rollout model and the evolving policy. Jackpot integrates a principled OBRS procedure, a unified training objective that jointly updates the policy and rollout models, and an efficient system implementation enabled by top-$k$ probability estimation and batch-level bias correction. Our theoretical analysis shows that OBRS consistently moves the rollout distribution closer to the target distribution under a controllable acceptance budget. Empirically, \sys substantially improves training stability compared to importance-sampling baselines, achieving performance comparable to on-policy RL when training Qwen3-8B-Base for up to 300 update steps of batchsize 64. Taken together, our results show that OBRS-based alignment brings us a step closer to practical and effective decoupling of rollout generation from policy optimization for RL for LLMs.
Authors: Peiyang Song, Pengrui Han, Noah Goodman
Abstract: Large Language Models (LLMs) have exhibited remarkable reasoning capabilities, achieving impressive results across a wide range of tasks. Despite these advances, significant reasoning failures persist, occurring even in seemingly simple scenarios. To systematically understand and address these shortcomings, we present the first comprehensive survey dedicated to reasoning failures in LLMs. We introduce a novel categorization framework that distinguishes reasoning into embodied and non-embodied types, with the latter further subdivided into informal (intuitive) and formal (logical) reasoning. In parallel, we classify reasoning failures along a complementary axis into three types: fundamental failures intrinsic to LLM architectures that broadly affect downstream tasks; application-specific limitations that manifest in particular domains; and robustness issues characterized by inconsistent performance across minor variations. For each reasoning failure, we provide a clear definition, analyze existing studies, explore root causes, and present mitigation strategies. By unifying fragmented research efforts, our survey provides a structured perspective on systemic weaknesses in LLM reasoning, offering valuable insights and guiding future research towards building stronger, more reliable, and robust reasoning capabilities. We additionally release a comprehensive collection of research works on LLM reasoning failures, as a GitHub repository at https://github.com/Peiyang-Song/Awesome-LLM-Reasoning-Failures, to provide an easy entry point to this area.
URLs: https://github.com/Peiyang-Song/Awesome-LLM-Reasoning-Failures,
Authors: Pierriccardo Olivieri, Fausto Lasca, Alessandro Gianola, Matteo Papini
Abstract: In this work, we propose a novel framework for the logical specification of non-Markovian rewards in Markov Decision Processes (MDPs) with large state spaces. Our approach leverages Linear Temporal Logic Modulo Theories over finite traces (LTLfMT), a more expressive extension of classical temporal logic in which predicates are first-order formulas of arbitrary first-order theories rather than simple Boolean variables. This enhanced expressiveness enables the specification of complex tasks over unstructured and heterogeneous data domains, promoting a unified and reusable framework that eliminates the need for manual predicate encoding. However, the increased expressive power of LTLfMT introduces additional theoretical and computational challenges compared to standard LTLf specifications. We address these challenges from a theoretical standpoint, identifying a fragment of LTLfMT that is tractable but sufficiently expressive for reward specification in an infinite-state-space context. From a practical perspective, we introduce a method based on reward machines and Hindsight Experience Replay (HER) to translate first-order logic specifications and address reward sparsity. We evaluate this approach to a continuous-control setting using Non-Linear Arithmetic Theory, showing that it enables natural specification of complex tasks. Experimental results show how a tailored implementation of HER is fundamental in solving tasks with complex goals.
Authors: Khurram Yamin, Jingjing Tang, Santiago Cortes-Gomez, Amit Sharma, Eric Horvitz, Bryan Wilder
Abstract: Large language models (LLMs) are increasingly deployed as agents in high-stakes domains where optimal actions depend on both uncertainty about the world and consideration of utilities of different outcomes, yet their decision logic remains difficult to interpret. We study whether LLMs are rational utility maximizers with coherent beliefs and stable preferences. We consider behaviors of models for diagnosis challenge problems. The results provide insights about the relationship of LLM inferences to ideal Bayesian utility maximization for elicited probabilities and observed actions. Our approach provides falsifiable conditions under which the reported probabilities \emph{cannot} correspond to the true beliefs of any rational agent. We apply this methodology to multiple medical diagnostic domains with evaluations across several LLMs. We discuss implications of the results and directions forward for uses of LLMs in guiding high-stakes decisions.
Authors: Qifan Zhang, Jianhao Ruan, Aochuan Chen, Kang Zeng, Nuo Chen, Jing Tang, Jia Li
Abstract: Large Reasoning Models (LRMs) have advanced rapidly; however, existing benchmarks in mathematics, code, and common-sense reasoning remain limited. They lack long-context evaluation, offer insufficient challenge, and provide answers that are difficult to verify programmatically. We introduce GrAlgoBench, a benchmark designed to evaluate LRMs through graph algorithm problems. Such problems are particularly well suited for probing reasoning abilities: they demand long-context reasoning, allow fine-grained control of difficulty levels, and enable standardized, programmatic evaluation. Across nine tasks, our systematic experiments reveal two major weaknesses of current LRMs. First, accuracy deteriorates sharply as context length increases, falling below 50% once graphs exceed 120 nodes. This degradation is driven by frequent execution errors, weak memory, and redundant reasoning. Second, LRMs suffer from an over-thinking phenomenon, primarily caused by extensive yet largely ineffective self-verification, which inflates reasoning traces without improving correctness. By exposing these limitations, GrAlgoBench establishes graph algorithm problems as a rigorous, multidimensional, and practically relevant testbed for advancing the study of reasoning in LRMs. Code is available at https://github.com/Bklight999/GrAlgoBench.
Authors: Longhui Ma, Di Zhao, Siwei Wang, Zhao Lv, Miao Wang
Abstract: GUI grounding maps natural language instructions to the correct interface elements, serving as the perception foundation for GUI agents. Existing approaches predominantly rely on fine-tuning multimodal large language models (MLLMs) using large-scale GUI datasets to predict target element coordinates, which is data-intensive and generalizes poorly to unseen interfaces. Recent attention-based alternatives exploit localization signals in MLLMs attention mechanisms without task-specific fine-tuning, but suffer from low reliability due to the lack of explicit and complementary spatial anchors in GUI images. To address this limitation, we propose Trifuse, an attention-based grounding framework that explicitly integrates complementary spatial anchors. Trifuse integrates attention, OCR-derived textual cues, and icon-level caption semantics via a Consensus-SinglePeak (CS) fusion strategy that enforces cross-modal agreement while retaining sharp localization peaks. Extensive evaluations on four grounding benchmarks demonstrate that Trifuse achieves strong performance without task-specific fine-tuning, substantially reducing the reliance on expensive annotated data. Moreover, ablation studies reveal that incorporating OCR and caption cues consistently improves attention-based grounding performance across different backbones, highlighting its effectiveness as a general framework for GUI grounding.
Authors: Yu Zhao, Fan Jiang, Tianle Liu, Bo Zeng, Yu Liu, Longyue Wang, Weihua Luo
Abstract: Recent advancements in Large Reasoning Models (LRMs), exemplified by DeepSeek-R1, have underscored the potential of scaling inference-time compute through Group Relative Policy Optimization (GRPO). However, GRPO frequently suffers from gradient signal attenuation when encountering problems that are either too trivial or overly complex. In these scenarios, the disappearance of inter-group advantages makes the gradient signal susceptible to noise, thereby jeopardizing convergence stability. While variants like DAPO attempt to rectify gradient vanishing, they do not alleviate the substantial computational overhead incurred by exhaustive rollouts on low-utility samples. In this paper, we propose Difficulty-Estimated Policy Optimization (DEPO), a novel framework designed to optimize the efficiency and robustness of reasoning alignment. DEPO integrates an online Difficulty Estimator that dynamically assesses and filters training data before the rollout phase. This mechanism ensures that computational resources are prioritized for samples with high learning potential. Empirical results demonstrate that DEPO achieves up to a 2x reduction in rollout costs without compromising model performance. Our approach significantly lowers the computational barrier for training high-performance reasoning models, offering a more sustainable path for reasoning scaling. Code and data will be released upon acceptance.
Authors: Arvid E. Gollwitzer, Paridhi Latawa, David de Gruijl, Deepak A. Subramanian, Adri\'an Noriega de la Colina
Abstract: Current tokenization methods process sequential data without accounting for signal quality, limiting their effectiveness on noisy real-world corpora. We present QA-Token (Quality-Aware Tokenization), which incorporates data reliability directly into vocabulary construction. We make three key contributions: (i) a bilevel optimization formulation that jointly optimizes vocabulary construction and downstream performance, (ii) a reinforcement learning approach that learns merge policies through quality-aware rewards with convergence guarantees, and (iii) an adaptive parameter learning mechanism via Gumbel-Softmax relaxation for end-to-end optimization. Our experimental evaluation demonstrates consistent improvements: genomics (6.7 percentage point F1 gain in variant calling over BPE), finance (30% Sharpe ratio improvement). At foundation scale, we tokenize a pretraining corpus comprising 1.7 trillion base-pairs and achieve state-of-the-art pathogen detection (94.53 MCC) while reducing token count by 15%. We unlock noisy real-world corpora, spanning petabases of genomic sequences and terabytes of financial time series, for foundation model training with zero inference overhead.
Authors: Hsien-Jyh Liao
Abstract: Large language models (LLMs) demonstrate remarkable reasoning capabilities, yet their performance often deteriorates sharply in long-horizon tasks, exhibiting systematic breakdown beyond certain scales. Conventional explanations primarily attribute this phenomenon to task complexity, such as combinatorial search explosion or long-term credit assignment challenges. In this work, we argue that these explanations are incomplete: even in linear, unbranched tasks without semantic ambiguity, autoregressive execution is subject to an intrinsic stability limit. We propose that the fundamental constraint on long-horizon reasoning arises from process-level instability in autoregressive generation rather than solely from search or task complexity, reframing long-horizon reasoning as a problem of structural governance. We derive Theorem~A, showing that decision advantage in single-path autoregressive reasoning decays exponentially with execution length, imposing a fundamental bound on maintainable reasoning chains. This result implies a structural consequence: stable long-horizon reasoning requires discrete segmentation, naturally inducing graph-like execution structures such as directed acyclic graphs (DAGs). Empirical studies in both synthetic environments and real TextWorld tasks reveal observable performance cliffs consistent with theoretical predictions. Our findings provide a dynamical perspective on long-horizon reasoning failure and suggest new limitations on maintaining long-term coherence under purely autoregressive architectures. Furthermore, we highlight that short-horizon evaluation protocols may obscure structural instability, indicating a potential shift from scaling toward structured governance in future reasoning systems.
Authors: Haotian Chen, Xin Cong, Shengda Fan, Yuyang Fu, Ziqin Gong, Yaxi Lu, Yishan Li, Boye Niu, Chengjun Pan, Zijun Song, Huadong Wang, Yesai Wu, Yueying Wu, Zihao Xie, Yukun Yan, Zhong Zhang, Yankai Lin, Zhiyuan Liu, Maosong Sun
Abstract: While Large Language Model (LLM)-based agents have shown remarkable potential for solving complex tasks, existing systems remain heavily reliant on large-scale models, leaving the capabilities of edge-scale models largely underexplored. In this paper, we present the first systematic study on training agentic models at the 4B-parameter scale. We identify three primary bottlenecks hindering the performance of edge-scale models: catastrophic forgetting during Supervised Fine-Tuning (SFT), sensitivity to reward signal noise during Reinforcement Learning (RL), and reasoning degradation caused by redundant information in long-context scenarios. To address the issues, we propose AgentCPM-Explore, a compact 4B agent model with high knowledge density and strong exploration capability. We introduce a holistic training framework featuring parameter-space model fusion, reward signal denoising, and contextual information refinement. Through deep exploration, AgentCPM-Explore achieves state-of-the-art (SOTA) performance among 4B-class models, matches or surpasses 8B-class SOTA models on four benchmarks, and even outperforms larger-scale models such as Claude-4.5-Sonnet or DeepSeek-v3.2 in five benchmarks. Notably, AgentCPM-Explore achieves 97.09% accuracy on GAIA text-based tasks under pass@64. These results provide compelling evidence that the bottleneck for edge-scale models is not their inherent capability ceiling, but rather their inference stability. Based on our well-established training framework, AgentCPM-Explore effectively unlocks the significant, yet previously underestimated, potential of edge-scale models.
Authors: Lanbo Lin, Jiayao Liu, Tianyuan Yang, Li Cai, Yuanwu Xu, Lei Wei, Sicong Xie, Guannan Zhang
Abstract: Evaluating agentic AI on open-ended professional tasks faces a fundamental dilemma between rigor and flexibility. Static rubrics provide rigorous, reproducible assessment but fail to accommodate diverse valid response strategies, while LLM-as-a-judge approaches adapt to individual responses yet suffer from instability and bias. Human experts address this dilemma by combining domain-grounded principles with dynamic, claim-level assessment. Inspired by this process, we propose JADE, a two-layer evaluation framework. Layer 1 encodes expert knowledge as a predefined set of evaluation skills, providing stable evaluation criteria. Layer 2 performs report-specific, claim-level evaluation to flexibly assess diverse reasoning strategies, with evidence-dependency gating to invalidate conclusions built on refuted claims. Experiments on BizBench show that JADE improves evaluation stability and reveals critical agent failure modes missed by holistic LLM-based evaluators. We further demonstrate strong alignment with expert-authored rubrics and effective transfer to a medical-domain benchmark, validating JADE across professional domains. Our code is publicly available at https://github.com/smiling-world/JADE.
Authors: Finn Rietz, Mart Karta\v{s}ev, Johannes A. Stork, Petter \"Ogren
Abstract: Behavior Trees (BTs) provide a structured and reactive framework for decision-making, commonly used to switch between sub-controllers based on environmental conditions. Reinforcement Learning (RL), on the other hand, can learn near-optimal controllers but sometimes struggles with sparse rewards, safe exploration, and long-horizon credit assignment. Combining BTs with RL has the potential for mutual benefit: a BT design encodes structured domain knowledge that can simplify RL training, while RL enables automatic learning of the controllers within BTs. However, naive integration of BTs and RL can lead to some controllers counteracting other controllers, possibly undoing previously achieved subgoals, thereby degrading the overall performance. To address this, we propose progress constraints, a novel mechanism where feasibility estimators constrain the allowed action set based on theoretical BT convergence results. Empirical evaluations in a 2D proof-of-concept and a high-fidelity warehouse environment demonstrate improved performance, sample efficiency, and constraint satisfaction, compared to prior methods of BT-RL integration.
Authors: Shengxuan Qiu, Haochen Huang, Shuzhang Zhong, Pengfei Zuo, Meng Li
Abstract: Scaling test-time compute with multi-path chain-of-thought improves reasoning accuracy, but its effectiveness depends critically on the exploration-exploitation trade-off. Existing approaches address this trade-off in rigid ways: tree-structured search hard-codes exploration through brittle expansion rules that interfere with post-trained reasoning, while parallel reasoning over-explores redundant hypothesis paths and relies on weak answer selection. Motivated by the observation that the optimal balance is phase-dependent and that correct and incorrect reasoning paths often diverge only at late stages, we reformulate test-time scaling as a dynamic expand-reduce control problem over a pool of hypotheses. We propose HyPER, a training-free online control policy for multi-path decoding in mixture-of-experts models that reallocates computation under a fixed budget using lightweight path statistics. HyPER consists of an online controller that transitions from exploration to exploitation as the hypothesis pool evolves, a token-level refinement mechanism that enables efficient generation-time exploitation without full-path resampling, and a length- and confidence-aware aggregation strategy for reliable answer-time exploitation. Experiments on four mixture-of-experts language models across diverse reasoning benchmarks show that HyPER consistently achieves a superior accuracy-compute trade-off, improving accuracy by 8 to 10 percent while reducing token usage by 25 to 40 percent.
Authors: Brian Rabern, Philipp Mondorf, Barbara Plank
Abstract: Large language models have demonstrated notable performance across various logical reasoning benchmarks. However, it remains unclear which core logical skills they truly master. To address this, we introduce LogicSkills, a unified benchmark designed to isolate three fundamental skills in formal reasoning: (i) $\textit{formal symbolization}\unicode{x2014}$translating premises into first-order logic; (ii) $\textit{countermodel construction}\unicode{x2014}$formulating a finite structure in which all premises are true while the conclusion is false; and (iii) $\textit{validity assessment}\unicode{x2014}$deciding whether a conclusion follows from a given set of premises. Items are drawn from the two-variable fragment of first-order logic (without identity) and are presented in both natural English and a Carroll-style language with nonce words. All examples are verified for correctness and non-triviality using the SMT solver Z3. Across leading models, performance is high on validity but substantially lower on symbolization and countermodel construction, suggesting reliance on surface-level patterns rather than genuine symbolic or rule-based reasoning.
Authors: Yishan Li, Wentong Chen, Yukun Yan, Mingwei Li, Sen Mei, Xiaorong Wang, Kunpeng Liu, Xin Cong, Shuo Wang, Zhong Zhang, Yaxi Lu, Zhenghao Liu, Yankai Lin, Zhiyuan Liu, Maosong Sun
Abstract: Generating deep research reports requires large-scale information acquisition and the synthesis of insight-driven analysis, posing a significant challenge for current language models. Most existing approaches follow a plan-then-write paradigm, whose performance heavily depends on the quality of the initial outline. However, constructing a comprehensive outline itself demands strong reasoning ability, causing current deep research systems to rely almost exclusively on closed-source or online large models. This reliance raises practical barriers to deployment and introduces safety and privacy concerns for user-authored data. In this work, we present AgentCPM-Report, a lightweight yet high-performing local solution composed of a framework that mirrors the human writing process and an 8B-parameter deep research agent. Our framework uses a Writing As Reasoning Policy (WARP), which enables models to dynamically revise outlines during report generation. Under this policy, the agent alternates between Evidence-Based Drafting and Reasoning-Driven Deepening, jointly supporting information acquisition, knowledge refinement, and iterative outline evolution. To effectively equip small models with this capability, we introduce a Multi-Stage Agentic Training strategy, consisting of cold-start, atomic skill RL, and holistic pipeline RL. Experiments on DeepResearch Bench, DeepConsult, and DeepResearch Gym demonstrate that AgentCPM-Report outperforms leading closed-source systems, with substantial gains in Insight.
Authors: Tianyi Hu, Qingxu Fu, Yanxi Chen, Zhaoyang Liu, Bolin Ding
Abstract: Reinforcement learning (RL) has emerged as the predominant paradigm for training large language model (LLM)-based AI agents. However, existing backbone RL algorithms lack verified convergence guarantees in agentic scenarios, especially in multi-turn settings, which can lead to training instability and failure to converge to optimal policies. In this paper, we systematically analyze how different combinations of policy update mechanisms and advantage estimation methods affect convergence properties in single/multi-turn scenarios. We find that REINFORCE with Group Relative Advantage Estimation (GRAE) can converge to the globally optimal under undiscounted conditions, but the combination of PPO & GRAE breaks PPO's original monotonic improvement property. Furthermore, we demonstrate that mainstream backbone RL algorithms cannot simultaneously achieve both critic-free and convergence guarantees in multi-turn scenarios. To address this, we propose SeeUPO (Sequence-level Sequential Update Policy Optimization), a critic-free approach with convergence guarantees for multi-turn interactions. SeeUPO models multi-turn interaction as sequentially executed multi-agent bandit problems. Through turn-by-turn sequential policy updates in reverse execution order, it ensures monotonic improvement and convergence to global optimal solution via backward induction. Experiments on AppWorld and BFCL v4 demonstrate SeeUPO's substantial improvements over existing backbone algorithms: relative gains of 43.3%-54.6% on Qwen3-14B and 24.1%-41.9% on Qwen2.5-14B (averaged across benchmarks), along with superior training stability.
Authors: Farooq Ahmad Wani, Alessandro Suglia, Rohit Saxena, Aryo Pradipta Gema, Wai-Chung Kwan, Fazl Barez, Maria Sofia Bucarelli, Fabrizio Silvestri, Pasquale Minervini
Abstract: The robustness of Vision Language Models (VLMs) is commonly assessed through output-level invariance, implicitly assuming that stable predictions reflect stable multimodal processing. In this work, we argue that this assumption is insufficient. We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness (spatial consistency of vision tokens), alongside standard label-based metrics. Applying this framework to modern VLMs across the SEEDBench, MMMU, and POPE datasets reveals three distinct failure modes. First, models frequently preserve predicted answers while undergoing substantial internal representation drift; for perturbations such as text overlays, this drift approaches the magnitude of inter-image variability, indicating that representations move to regions typically occupied by unrelated inputs despite unchanged outputs. Second, robustness does not improve with scale; larger models achieve higher accuracy but exhibit equal or greater sensitivity, consistent with sharper yet more fragile decision boundaries. Third, we find that perturbations affect tasks differently: they harm reasoning when they disrupt how models combine coarse and fine visual cues, but on the hallucination benchmarks, they can reduce false positives by making models generate more conservative answers.
Authors: Thiviyan Thanapalasingam, Antonis Vozikis, Peter Bloem, Paul Groth
Abstract: Knowledge Graph (KG) generation requires models to learn complex semantic dependencies between triples while maintaining domain validity constraints. Unlike link prediction, which scores triples independently, generative models must capture interdependencies across entire subgraphs to produce semantically coherent structures. We present ARK (Auto-Regressive Knowledge Graph Generation), a family of autoregressive models that generate KGs by treating graphs as sequences of (head, relation, tail) triples. ARK learns implicit semantic constraints directly from data, including type consistency, temporal validity, and relational patterns, without explicit rule supervision. On the IntelliGraphs benchmark, our models achieve 89.2% to 100.0% semantic validity across diverse datasets while generating novel graphs not seen during training. We also introduce SAIL, a variational extension of ARK that enables controlled generation through learned latent representations, supporting both unconditional sampling and conditional completion from partial graphs. Our analysis reveals that model capacity (hidden dimensionality >= 64) is more critical than architectural depth for KG generation, with recurrent architectures achieving comparable validity to transformer-based alternatives while offering substantial computational efficiency. These results demonstrate that autoregressive models provide an effective framework for KG generation, with practical applications in knowledge base completion and query answering.
Authors: Alessandro Abate, Giuseppe De Giacomo, Mathias Jackermeier, Jan Kret\'insk\'y, Maximilian Prokop, Christoph Weinhuber
Abstract: We study multi-task reinforcement learning (RL), a setting in which an agent learns a single, universal policy capable of generalising to arbitrary, possibly unseen tasks. We consider tasks specified as linear temporal logic (LTL) formulae, which are commonly used in formal methods to specify properties of systems, and have recently been successfully adopted in RL. In this setting, we present a novel task embedding technique leveraging a new generation of semantic LTL-to-automata translations, originally developed for temporal synthesis. The resulting semantically labelled automata contain rich, structured information in each state that allow us to (i) compute the automaton efficiently on-the-fly, (ii) extract expressive task embeddings used to condition the policy, and (iii) naturally support full LTL. Experimental results in a variety of domains demonstrate that our approach achieves state-of-the-art performance and is able to scale to complex specifications where existing methods fail.
Authors: Jiali Wu, Abhinav Anand, Shweta Verma, Mira Mezini
Abstract: State Space Models (SSMs) have emerged as an efficient alternative to the transformer architecture. Recent studies show that SSMs can match or surpass Transformers on code understanding tasks, such as code retrieval, when trained under similar conditions. However, their internal mechanisms remain a black box. We present the first systematic analysis of what SSM-based code models actually learn and perform the first comparative analysis of SSM and Transformer-based code models. Our analysis reveals that SSMs outperform Transformers at capturing code syntax and semantics in pretraining but forgets certain syntactic and semantic relations during fine-tuning on task, especially when the task emphasizes short-range dependencies. To diagnose this, we introduce SSM-Interpret, a frequency-domain framework that exposes a spectral shift toward short-range dependencies during fine-tuning. Guided by these findings, we propose architectural modifications that significantly improve the performance of SSM-based code model, validating that our analysis directly enables better models.
Authors: Anirudh Chari, Neil Pattanaik
Abstract: Human concept learning is typically active: learners choose which instances to query or test in order to reduce uncertainty about an underlying rule or category. Active concept learning must balance informativeness of queries against the stability of the learner that generates and scores hypotheses. We study this trade-off in a neuro-symbolic Bayesian learner whose hypotheses are executable programs proposed by a large language model (LLM) and reweighted by Bayesian updating. We compare a Rational Active Learner that selects queries to maximize approximate expected information gain (EIG) and the human-like Positive Test Strategy (PTS) that queries instances predicted to be positive under the current best hypothesis. Across concept-learning tasks in the classic Number Game, EIG is effective when falsification is necessary (e.g., compound or exception-laden rules), but underperforms on simple concepts. We trace this failure to a support mismatch between the EIG policy and the LLM proposal distribution: highly diagnostic boundary queries drive the posterior toward regions where the generator produces invalid or overly specific programs, yielding a support-mismatch trap in the particle approximation. PTS is information-suboptimal but tends to maintain proposal validity by selecting "safe" queries, leading to faster convergence on simple rules. Our results suggest that "confirmation bias" may not be a cognitive error, but rather a rational adaptation for maintaining tractable inference in the sparse, open-ended hypothesis spaces characteristic of human thought.
Authors: Dunwei Tu, Hongyan Hao, Hansi Yang, Yihao Chen, Yi-Kai Zhang, Zhikang Xia, Yu Yang, Yueqing Sun, Xingchen Liu, Furao Shen, Qi Gu, Hui Su, Xunliang Cai
Abstract: Training generalist agents capable of adapting to diverse scenarios requires interactive environments for self-exploration. However, interactive environments remain critically scarce, and existing synthesis methods suffer from significant limitations regarding environmental diversity and scalability. To address these challenges, we introduce ScaleEnv, a framework that constructs fully interactive environments and verifiable tasks entirely from scratch. Specifically, ScaleEnv ensures environment reliability through procedural testing, and guarantees task completeness and solvability via tool dependency graph expansion and executable action verification. By enabling agents to learn through exploration within ScaleEnv, we demonstrate significant performance improvements on unseen, multi-turn tool-use benchmarks such as $\tau^2$-Bench and VitaBench, highlighting strong generalization capabilities. Furthermore, we investigate the relationship between increasing number of domains and model generalization performance, providing empirical evidence that scaling environmental diversity is critical for robust agent learning.
Authors: Yi Chen, Wonjin Shin, Shuhong Liu, Tho Mai, Jeongmo Lee, Chuanbo Hua, Kun Wang, Jun Liu, Joo-Young Kim
Abstract: Large foundation models (LFMs) achieve strong performance through scaling, yet current structural pruning methods derive fixed pruning decisions during inference, overlooking sparsity patterns that emerge in the autoregressive token generation. In this paper, we propose POP (Partition-guided Online Pruning), an efficient online structural pruning framework that enables context-conditioned dynamic pruning with minimal computational overhead. POP partitions model channels into retained, candidate, and pruned regions, where prefilling defines a coarse pruning partition, and the decoding stage generates a fine-grained mask within the candidate region, avoiding full-channel re-evaluation. The coarse pruning partition preserves consistently important weights, while the fine-grained masking provides context-conditioned variation during decoding. Moreover, POP is a lightweight, plug-and-play method that requires no preprocessing, including offline calibration, retraining, or learning predictors. Extensive evaluations across diverse LFMs, including large language models (LLMs), mixture-of-experts models (MoEs), and vision-language models (VLMs), demonstrate that POP consistently delivers higher accuracy than existing pruning approaches while incurring smaller computational overhead and minimizing inference latency.
Authors: Tonghan Wang, Yuqi Pan, Xinyi Yang, Yanchen Jiang, Milind Tambe, David C. Parkes
Abstract: We develop a game-theoretic framework for predicting and steering the behavior of populations of large language models (LLMs) through Nash equilibrium (NE) analysis. To avoid the intractability of equilibrium computation in open-ended text spaces, we model each agent's action as a mixture over human subpopulations. Agents choose actively and strategically which groups to align with, yielding an interpretable and behaviorally substantive policy class. We derive closed-form NE characterizations, adopting standard concave-utility assumptions to enable analytical system-level predictions and give explicit, actionable guidance for shifting alignment targets toward socially desirable outcomes. The method functions as an active alignment layer on top of existing alignment pipelines such as RLHF. In a social-media setting, we show that a population of LLMs, especially reasoning-based models, may exhibit political exclusion, pathologies where some subpopulations are ignored by all LLM agents, which can be avoided by our method, illustrating the promise of applying the method to regulate multi-agent LLM dynamics across domains.
Authors: Jin Wang, Hui Ma, Fei Xing, Ming Yan
Abstract: Federated learning enables collaborative model training across distributed clients while preserving data privacy. However, in practical deployments, device heterogeneity, non-independent, and identically distributed (Non-IID) data often lead to highly unstable and biased gradient updates. When differential privacy is enforced, conventional fixed gradient clipping and Gaussian noise injection may further amplify gradient perturbations, resulting in training oscillation and performance degradation and degraded model performance. To address these challenges, we propose an adaptive differentially private federated learning framework that explicitly targets model efficiency under heterogeneous and privacy-constrained settings. On the client side, a lightweight local compressed module is introduced to regularize intermediate representations and constrain gradient variability, thereby mitigating noise amplification during local optimization. On the server side, an adaptive gradient clipping strategy dynamically adjusts clipping thresholds based on historical update statistics to avoid over-clipping and noise domination. Furthermore, a constraint-aware aggregation mechanism is designed to suppress unreliable or noise-dominated client updates and stabilize global optimization. Extensive experiments on CIFAR-10 and SVHN demonstrate improved convergence stability and classification accuracy.
Authors: Sindhuja Chaduvula, Jessee Ho, Kina Kim, Aravind Narayanan, Mahshid Alinoori, Muskan Garg, Dhanesh Ramachandram, Shaina Raza
Abstract: Over the last decade, explainable AI has primarily focused on interpreting individual model predictions, producing post-hoc explanations that relate inputs to outputs under a fixed decision structure. Recent advances in large language models (LLMs) have enabled agentic AI systems whose behaviour unfolds over multi-step trajectories. In these settings, success and failure are determined by sequences of decisions rather than a single output. While useful, it remains unclear how explanation approaches designed for static predictions translate to agentic settings where behaviour emerges over time. In this work, we bridge the gap between static and agentic explainability by comparing attribution-based explanations with trace-based diagnostics across both settings. To make this distinction explicit, we empirically compare attribution-based explanations used in static classification tasks with trace-based diagnostics used in agentic benchmarks (TAU-bench Airline and AssistantBench). Our results show that while attribution methods achieve stable feature rankings in static settings (Spearman $\rho = 0.86$), they cannot be applied reliably to diagnose execution-level failures in agentic trajectories. In contrast, trace-grounded rubric evaluation for agentic settings consistently localizes behaviour breakdowns and reveals that state tracking inconsistency is 2.7$\times$ more prevalent in failed runs and reduces success probability by 49\%. These findings motivate a shift towards trajectory-level explainability for agentic systems when evaluating and diagnosing autonomous AI behaviour. Resources: https://github.com/VectorInstitute/unified-xai-evaluation-framework https://vectorinstitute.github.io/unified-xai-evaluation-framework
URLs: https://github.com/VectorInstitute/unified-xai-evaluation-framework, https://vectorinstitute.github.io/unified-xai-evaluation-framework
Authors: Alisia Lupidi, Bhavul Gauri, Thomas Simon Foster, Bassel Al Omari, Despoina Magka, Alberto Pepe, Alexis Audran-Reiss, Muna Aghamelu, Nicolas Baldwin, Lucia Cipolina-Kun, Jean-Christophe Gagnon-Audet, Chee Hau Leow, Sandra Lefdal, Hossam Mossalam, Abhinav Moudgil, Saba Nazir, Emanuel Tewolde, Isabel Urrego, Jordi Armengol Estape, Amar Budhiraja, Gaurav Chaurasia, Abhishek Charnalia, Derek Dunfield, Karen Hambardzumyan, Daniel Izcovich, Martin Josifoski, Ishita Mediratta, Kelvin Niu, Parth Pathak, Michael Shvartsman, Edan Toledo, Anton Protopopov, Roberta Raileanu, Alexander Miller, Tatiana Shavrina, Jakob Foerster, Yoram Bachrach
Abstract: LLM agents hold significant promise for advancing scientific research. To accelerate this progress, we introduce AIRS-Bench (the AI Research Science Benchmark), a suite of 20 tasks sourced from state-of-the-art machine learning papers. These tasks span diverse domains, including language modeling, mathematics, bioinformatics, and time series forecasting. AIRS-Bench tasks assess agentic capabilities over the full research lifecycle -- including idea generation, experiment analysis and iterative refinement -- without providing baseline code. The AIRS-Bench task format is versatile, enabling easy integration of new tasks and rigorous comparison across different agentic frameworks. We establish baselines using frontier models paired with both sequential and parallel scaffolds. Our results show that agents exceed human SOTA in four tasks but fail to match it in sixteen others. Even when agents surpass human benchmarks, they do not reach the theoretical performance ceiling for the underlying tasks. These findings indicate that AIRS-Bench is far from saturated and offers substantial room for improvement. We open-source the AIRS-Bench task definitions and evaluation code to catalyze further development in autonomous scientific research.
Authors: Jean Kaddour, Srijan Patel, Gb\`etondji Dovonon, Leo Richter, Pasquale Minervini, Matt J. Kusner
Abstract: Can AI agents predict whether they will succeed at a task? We study agentic uncertainty by eliciting success probability estimates before, during, and after task execution. All results exhibit agentic overconfidence: some agents that succeed only 22% of the time predict 77% success. Counterintuitively, pre-execution assessment with strictly less information tends to yield better discrimination than standard post-execution review, though differences are not always significant. Adversarial prompting reframing assessment as bug-finding achieves the best calibration.
Authors: Sang Min Kim, Byeongchan Kim, Arijit Sehanobish, Somnath Basu Roy Chowdhury, Rahul Kidambi, Dongseok Shim, Avinava Dubey, Snigdha Chaturvedi, Min-hwan Oh, Krzysztof Choromanski
Abstract: Efficient neural networks are essential for scaling machine learning models to real-time applications and resource-constrained environments. Fully-connected feedforward layers (FFLs) introduce computation and parameter count bottlenecks within neural network architectures. To address this challenge, in this work, we propose a new class of dense layers that generalize standard fully-connected feedforward layers, \textbf{E}fficient, \textbf{U}nified and \textbf{Gen}eral dense layers (EUGens). EUGens leverage random features to approximate standard FFLs and go beyond them by incorporating a direct dependence on the input norms in their computations. The proposed layers unify existing efficient FFL extensions and improve efficiency by reducing inference complexity from quadratic to linear time. They also lead to \textbf{the first} unbiased algorithms approximating FFLs with arbitrary polynomial activation functions. Furthermore, EuGens reduce the parameter count and computational overhead while preserving the expressive power and adaptability of FFLs. We also present a layer-wise knowledge transfer technique that bypasses backpropagation, enabling efficient adaptation of EUGens to pre-trained models. Empirically, we observe that integrating EUGens into Transformers and MLPs yields substantial improvements in inference speed (up to \textbf{27}\%) and memory efficiency (up to \textbf{30}\%) across a range of tasks, including image classification, language model pre-training, and 3D scene reconstruction. Overall, our results highlight the potential of EUGens for the scalable deployment of large-scale neural networks in real-world scenarios.
Authors: Sankar B, Amogh A S, Sandhya Baranwal, Dibakar Sen
Abstract: During product conceptualization, capturing the non-linear history and cognitive intent is crucial. Traditional sketching tools often lose this context. We introduce DIMES (Design Idea Management and Evolution capture System), a web-based environment featuring sGIT (SketchGit), a custom visual version control architecture, and Generative AI. sGIT includes AEGIS, a module using hybrid Deep Learning and Machine Learning models to classify six stroke types. The system maps Git primitives to design actions, enabling implicit branching and multi-modal commits (stroke data + voice intent). In a comparative study, experts using DIMES demonstrated a 160% increase in breadth of concept exploration. Generative AI modules generated narrative summaries that enhanced knowledge transfer; novices achieved higher replication fidelity (Neural Transparency-based Cosine Similarity: 0.97 vs. 0.73) compared to manual summaries. AI-generated renderings also received higher user acceptance (Purchase Likelihood: 4.2 vs 3.1). This work demonstrates that intelligent version control bridges creative action and cognitive documentation, offering a new paradigm for design education.
Authors: Ziao Yang, Zizhang Chen, Lei Zhang, Hongfu Liu
Abstract: Slogans are concise and memorable catchphrases that play a crucial role in advertising by conveying brand identity and shaping public perception. However, advertising fatigue reduces the effectiveness of repeated slogans, creating a growing demand for novel, creative, and insightful slogan generation. While recent work leverages large language models (LLMs) for this task, existing approaches often produce stylistically redundant outputs that lack a clear brand persona and appear overtly machine-generated. We argue that effective slogans should balance novelty with familiarity and propose a new paradigm that recontextualizes persona-related famous quotes for slogan generation. Well-known quotes naturally align with slogan-length text, employ rich rhetorical devices, and offer depth and insight, making them a powerful resource for creative generation. Technically, we introduce a modular framework that decomposes slogan generation into interpretable subtasks, including quote matching, structural decomposition, vocabulary replacement, and remix generation. Extensive automatic and human evaluations demonstrate marginal improvements in diversity, novelty, emotional impact, and human preference over three state-of-the-art LLM baselines.
Authors: Wei-Chieh Huang, Weizhi Zhang, Yueqing Liang, Yuanchen Bei, Yankai Chen, Tao Feng, Xinyu Pan, Zhen Tan, Yu Wang, Tianxin Wei, Shanglin Wu, Ruiyao Xu, Liangwei Yang, Rui Yang, Wooseong Yang, Chin-Yuan Yeh, Hanrong Zhang, Haozhen Zhang, Siqi Zhu, Henry Peng Zou, Wanjia Zhao, Song Wang, Wujiang Xu, Zixuan Ke, Zheng Hui, Dawei Li, Yaozu Wu, Langzhou He, Chen Wang, Xiongxiao Xu, Baixiang Huang, Juntao Tan, Shelby Heinecke, Huan Wang, Caiming Xiong, Ahmed A. Metwally, Jun Yan, Chen-Yu Lee, Hanqing Zeng, Yinglong Xia, Xiaokai Wei, Ali Payani, Yu Wang, Haitong Ma, Wenya Wang, Chengguang Wang, Yu Zhang, Xin Wang, Yongfeng Zhang, Jiaxuan You, Hanghang Tong, Xiao Luo, Yizhou Sun, Wei Wang, Julian McAuley, James Zou, Jiawei Han, Philip S. Yu, Kai Shu
Abstract: The research of artificial intelligence is undergoing a paradigm shift from prioritizing model innovations over benchmark scores towards emphasizing problem definition and rigorous real-world evaluation. As the field enters the "second half," the central challenge becomes real utility in long-horizon, dynamic, and user-dependent environments, where agents face context explosion and must continuously accumulate, manage, and selectively reuse large volumes of information across extended interactions. Memory, with hundreds of papers released this year, therefore emerges as the critical solution to fill the utility gap. In this survey, we provide a unified view of foundation agent memory along three dimensions: memory substrate (internal and external), cognitive mechanism (episodic, semantic, sensory, working, and procedural), and memory subject (agent- and user-centric). We then analyze how memory is instantiated and operated under different agent topologies and highlight learning policies over memory operations. Finally, we review evaluation benchmarks and metrics for assessing memory utility, and outline various open challenges and future directions.
Authors: Zihang Wang, Siyue Zhang, Yilun Zhao, Jingyi Yang, Tingyu Song, Anh Tuan Luu, Chen Zhao
Abstract: Embedding models are a fundamental component of modern AI systems such as semantic search and retrieval-augmented generation. Recent advances in large foundation models have substantially accelerated the development of embedding models, including those based on Large Language Models (LLMs), Vision Language Models (VLMs), and Multimodal LLMs. More recently, Large Diffusion Language Models (dLLMs) and Multimodal dLLMs have emerged as competitive alternatives to autoregressive models, offering advantages such as bidirectional attention and parallel generation. This progress naturally raises a critical yet unexplored question: can Multimodal dLLMs serve as effective multimodal embedding models? To answer this, we present the first systematic study of converting Multimodal dLLMs into embedding models. We evaluate state-of-the-art Multimodal dLLMs and Autoregressive VLMs across three categories of embedding tasks: classification, visual question answering, and information retrieval. Our results show that Multimodal dLLM embeddings generally underperform their autoregressive VLM counterparts. The stronger diffusion-based model, LaViDa, lags by only 3.5 points on classification, 2.5 points on VQA, and 4.4 points on retrieval tasks, whereas the other diffusion-based model, MMaDA, exhibits substantially larger performance gaps, exceeding 20 points across all tasks. Further analysis reveals insufficient image-text alignment in diffusion-based models, accounting for the observed limitations in their embedding performance.
Authors: Yi-Xiang Hu, Yuke Wang, Feng Wu, Zirui Huang, Shuli Zeng, Xiang-Yang Li
Abstract: Scheduling precedence-constrained tasks under shared renewable resources is central to modern computing platforms. The Resource Investment Problem (RIP) models this setting by minimizing the cost of provisioned renewable resources under precedence and timing constraints. Exact mixed-integer programming and constraint programming become impractically slow on large instances, and dynamic updates require schedule revisions under tight latency budgets. We present iScheduler, a reinforcement-learning-driven iterative scheduling framework that formulates RIP solving as a Markov decision process over decomposed subproblems and constructs schedules through sequential process selection. The framework accelerates optimization and supports reconfiguration by reusing unchanged process schedules and rescheduling only affected processes. We also release L-RIPLIB, an industrial-scale benchmark derived from cloud-platform workloads with 1,000 instances of 2,500-10,000 tasks. Experiments show that iScheduler attains competitive resource costs while reducing time to feasibility by up to 43$\times$ against strong commercial baselines.
Authors: Dinesh Gopalan, Ratul Ali
Abstract: The escalating demand for high-fidelity, real-time inference in distributed edge-cloud environments necessitates aggressive model optimization to counteract severe latency and energy constraints. This paper introduces the Hybrid Quantization and Pruning (HQP) framework, a novel, integrated methodology designed to achieve synergistic model acceleration while adhering to strict quality guarantees. We detail a sensitivity-aware structural pruning algorithm that employs a dynamic weight sensitivity metric, derived from a highly efficient approximation of the Fisher Information Matrix (FIM), to guide the iterative removal of redundant filters. This pruning is strictly conditional, enforcing an adherence to a maximum permissible accuracy drop (Delta ax) before the model proceeds to 8-bit post-training quantization. This rigorous coordination is critical, as it ensures the resultant sparse model structure is maximally robust to quantization error and hardware-specific kernel optimization. Exhaustive evaluation across heterogeneous NVIDIA Jetson edge platforms, utilizing resource-efficient architectures like MobileNetV3 and ResNet-18, demonstrates that the HQP framework achieves a peak performance gain of 3.12 times inference speedup and a 55 percent model size reduction, while rigorously containing the accuracy drop below the 1.5 percent constraint. A comprehensive comparative analysis against conventional single-objective compression techniques validates the HQP framework as a superior, hardware-agnostic solution for deploying ultra-low-latency AI in resource-limited edge infrastructures.
Authors: Elliot L. Epstein, Rajat Dwaraknath, John Winnicki, Thanawat Sornwanee
Abstract: This paper argues that large ML conferences should allocate marginal review capacity primarily to papers near the acceptance boundary, rather than spreading extra reviews via random or affinity-driven heuristics. We propose using LLM-based comparative ranking (via pairwise comparisons and a Bradley--Terry model) to identify a borderline band \emph{before} human reviewing and to allocate \emph{marginal} reviewer capacity at assignment time. Concretely, given a venue-specific minimum review target (e.g., 3 or 4), we use this signal to decide which papers receive one additional review (e.g., a 4th or 5th), without conditioning on any human reviews and without using LLM outputs for accept/reject. We provide a simple expected-impact calculation in terms of (i) the overlap between the predicted and true borderline sets ($\rho$) and (ii) the incremental value of an extra review near the boundary ($\Delta$), and we provide retrospective proxies to estimate these quantities.
Authors: Nunzio Lore, Babak Heydari
Abstract: Large Language Models (LLMs) often exhibit pronounced context-dependent variability that undermines predictable multi-agent behavior in tasks requiring strategic thinking. Focusing on models that range from 7 to 9 billion parameters in size engaged in a ten-round repeated Prisoner's Dilemma, we evaluate whether short, costless pre-play messages emulating the cheap-talk paradigm affect strategic stability. Our analysis uses simulation-level bootstrap resampling and nonparametric inference to compare cooperation trajectories fitted with LOWESS regression across both the messaging and the no-messaging condition. We demonstrate consistent reductions in trajectory noise across a majority of the model-context pairings being studied. The stabilizing effect persists across multiple prompt variants and decoding regimes, though its magnitude depends on model choice and contextual framing, with models displaying higher baseline volatility gaining the most. While communication rarely produces harmful instability, we document a few context-specific exceptions and identify the limited domains in which communication harms stability. These findings position cheap-talk style communication as a low-cost, practical tool for improving the predictability and reliability of strategic behavior in multi-agent LLM systems.
Authors: Thomas Georges, Adam Abdin
Abstract: We introduce a Transformer-based Reinforcement Learning framework for autonomous orbital collision avoidance that explicitly models the effects of partial observability and imperfect monitoring in space operations. The framework combines a configurable encounter simulator, a distance-dependent observation model, and a sequential state estimator to represent uncertainty in relative motion. A central contribution of this work is the use of transformer-based Partially Observable Markov Decision Process (POMDP) architecture, which leverage long-range temporal attention to interpret noisy and intermittent observations more effectively than traditional architectures. This integration provides a foundation for training collision avoidance agents that can operate more reliably under imperfect monitoring environments.
Authors: Xiaoxuan Tang, Jincheng Wang, Liwei Luo, Jingxuan Xu, Sheng Zhou, Dajun Chen, Wei Jiang, Yong Li
Abstract: Large language models (LLMs) have recently shown strong potential for Automated Program Repair (APR), yet most existing approaches remain unimodal and fail to leverage the rich diagnostic signals contained in visual artifacts such as screenshots and control-flow graphs. In practice, many bug reports convey critical information visually (e.g., layout breakage or missing widgets), but directly using such dense visual inputs often causes context loss and noise, making it difficult for MLLMs to ground visual observations into precise fault localization and executable patches. To bridge this semantic gap, we propose \textbf{SVRepair}, a multimodal APR framework with structured visual representation. SVRepair first fine-tunes a vision-language model, \textbf{Structured Visual Representation (SVR)}, to uniformly transform heterogeneous visual artifacts into a \emph{semantic scene graph} that captures GUI elements and their structural relations (e.g., hierarchy), providing normalized, code-relevant context for downstream repair. Building on the graph, SVRepair drives a coding agent to localize faults and synthesize patches, and further introduces an iterative visual-artifact segmentation strategy that progressively narrows the input to bug-centered regions to suppress irrelevant context and reduce hallucinations. Extensive experiments across multiple benchmarks demonstrate state-of-the-art performance: SVRepair achieves \textbf{36.47\%} accuracy on SWE-Bench M, \textbf{38.02\%} on MMCode, and \textbf{95.12\%} on CodeVision, validating the effectiveness of SVRepair for multimodal program repair.
Authors: Qianren Mao, Yashuo Luo, Ziqi Qin, Junnan Liu, Weifeng Jiang, Zhijun Chen, Zhuoran Li, Likang Xiao, Chuou Xu, Qili Zhang, Hanwen Hao, Jingzheng Li, Chunghua Lin, Jianxin Li, Philip S. Yu
Abstract: The lightweight semi-supervised learning (LSL) strategy provides an effective approach of conserving labeled samples and minimizing model inference costs. Prior research has effectively applied knowledge transfer learning and co-training regularization from large to small models in LSL. However, such training strategies are computationally intensive and prone to local optima, thereby increasing the difficulty of finding the optimal solution. This has prompted us to investigate the feasibility of integrating three low-cost scenarios for text mining tasks: limited labeled supervision, lightweight fine-tuning, and rapid-inference small models. We propose NanoNet, a novel framework for lightweight text mining that implements parameter-efficient learning with limited supervision. It employs online knowledge distillation to generate multiple small models and enhances their performance through mutual learning regularization. The entire process leverages parameter-efficient learning, reducing training costs and minimizing supervision requirements, ultimately yielding a lightweight model for downstream inference.
Authors: Nicolas Menet, Michael Hersche, Andreas Krause, Abbas Rahimi
Abstract: Coding agents are increasingly utilized in test-driven software development, yet the theoretical mechanisms behind their environment-interaction strategies remain underexplored. We provide a probabilistic framework for two dominant paradigms: code selection after generation using the execution environment, and code generation conditioned on environment feedback. First, we formalize several well-established selection heuristics as environment-aware estimators of code correctness. We theoretically prove that estimators based on fuzzy functional similarity add an inductive bias and strictly dominate estimators based on functional equivalence in terms of signal-to-noise ratio. Second, we frame backprompting as an in-context approximation of Thompson sampling. We derive a novel regret bound for reward functions with unobservable components, theoretically explaining why the effectiveness of backprompting is limited by the ambiguity of the informal task description (an irreducible regret). Using three state-of-the-art open weight models, we corroborate these findings across BigCodeBenchHard, LeetCodeDataset, and QiskitHumanEvalSim. Our formalization also suggests how to improve task descriptions effectively, leading to a new benchmark, QiskitHumanEvalSimX.
Authors: Olaf Yunus Laitinen Imanov, Derya Umut Kulali, Taner Yilmaz
Abstract: Climate hazards increasingly disrupt urban transportation and emergency-response operations by damaging housing stock, degrading infrastructure, and reducing network accessibility. This paper presents Skjold-DiT, a diffusion-transformer framework that integrates heterogeneous spatio-temporal urban data to forecast building-level climate-risk indicators while explicitly incorporating transportation-network structure and accessibility signals relevant to intelligent vehicles (e.g., emergency reachability and evacuation-route constraints). Concretely, Skjold-DiT enables hazard-conditioned routing constraints by producing calibrated, uncertainty-aware accessibility layers (reachability, travel-time inflation, and route redundancy) that can be consumed by intelligent-vehicle routing and emergency dispatch systems. Skjold-DiT combines: (1) Fjell-Prompt, a prompt-based conditioning interface designed to support cross-city transfer; (2) Norrland-Fusion, a cross-modal attention mechanism unifying hazard maps/imagery, building attributes, demographics, and transportation infrastructure into a shared latent representation; and (3) Valkyrie-Forecast, a counterfactual simulator for generating probabilistic risk trajectories under intervention prompts. We introduce the Baltic-Caspian Urban Resilience (BCUR) dataset with 847,392 building-level observations across six cities, including multi-hazard annotations (e.g., flood and heat indicators) and transportation accessibility features. Experiments evaluate prediction quality, cross-city generalization, calibration, and downstream transportation-relevant outcomes, including reachability and hazard-conditioned travel times under counterfactual interventions.
Authors: Yifu Qiu, Zheng Zhao, Waylon Li, Yftah Ziser, Anna Korhonen, Shay B. Cohen, Edoardo M. Ponti
Abstract: Internal modelling of the world -- predicting transitions between previous states $X$ and next states $Y$ under actions $Z$ -- is essential to reasoning and planning for LLMs and VLMs. Learning such models typically requires costly action-labelled trajectories. We propose SWIRL, a self-improvement framework that learns from state-only sequences by treating actions as a latent variable and alternating between Forward World Modelling (FWM) $P_\theta(Y|X,Z)$ and an Inverse Dynamics Modelling (IDM) $Q_\phi(Z|X,Y)$. SWIRL iterates two phases: (1) Variational Information Maximisation, which updates the FWM to generate next states that maximise conditional mutual information with latent actions given prior states, encouraging identifiable consistency; and (2) ELBO Maximisation, which updates the IDM to explain observed transitions, effectively performing coordinate ascent. Both models are trained with reinforcement learning (specifically, GRPO) with the opposite frozen model's log-probability as a reward signal. We provide theoretical learnability guarantees for both updates, and evaluate SWIRL on LLMs and VLMs across multiple environments: single-turn and multi-turn open-world visual dynamics and synthetic textual environments for physics, web, and tool calling. SWIRL achieves gains of 16% on AURORABench, 28% on ByteMorph, 16% on WorldPredictionBench, and 14% on StableToolBench.
Authors: Zhihan Jiang, Qianhui Chen, Chu Zhang, Yanheng Li, Ray LC
Abstract: In human conversation, empathic dialogue requires nuanced temporal cues indicating whether the conversational partner is paying attention. This type of "active listening" is overlooked in the design of Conversational Agents (CAs), which use the same pacing for one conversation. To model the temporal cues in human conversation, we need CAs that dynamically adjust response pacing according to user input. We qualitatively analyzed ten cases of active listening to distill five context-aware pacing strategies: Reflective Silence, Facilitative Silence, Empathic Silence, Holding Space, and Immediate Response. In a between-subjects study (N=50) with two conversational scenarios (relationship and career-support), the context-aware agent scored higher than static-pacing control on perceived human-likeness, smoothness, and interactivity, supporting deeper self-disclosure and higher engagement. In the career support scenario, the CA yielded higher perceived listening quality and affective trust. This work shows how insights from human conversation like context-aware pacing can empower the design of more empathic human-AI communication.
Authors: Amir H. Ashouri, Shayan Shirahmad Gale Bagi, Kavin Satheeskumar, Tejas Srikanth, Jonathan Zhao, Ibrahim Saidoun, Ziwen Wang, Bryan Chan, Tomasz S. Czajkowski
Abstract: The phase ordering problem has been a long-standing challenge since the late 1970s, yet it remains an open problem due to having a vast optimization space and an unbounded nature, making it an open-ended problem without a finite solution, one can limit the scope by reducing the number and the length of optimizations. Traditionally, such locally optimized decisions are made by hand-coded algorithms tuned for a small number of benchmarks, often requiring significant effort to be retuned when the benchmark suite changes. In the past 20 years, Machine Learning has been employed to construct performance models to improve the selection and ordering of compiler optimizations, however, the approaches are not baked into the compiler seamlessly and never materialized to be leveraged at a fine-grained scope of code segments. This paper presents Protean Compiler: An agile framework to enable LLVM with built-in phase-ordering capabilities at a fine-grained scope. The framework also comprises a complete library of more than 140 handcrafted static feature collection methods at varying scopes, and the experimental results showcase speedup gains of up to 4.1% on average and up to 15.7% on select Cbench applications wrt LLVM's O3 by just incurring a few extra seconds of build time on Cbench. Additionally, Protean compiler allows for an easy integration with third-party ML frameworks and other Large Language Models, and this two-step optimization shows a gain of 10.1% and 8.5% speedup wrt O3 on Cbench's Susan and Jpeg applications. Protean compiler is seamlessly integrated into LLVM and can be used as a new, enhanced, full-fledged compiler. We plan to release the project to the open-source community in the near future.
Authors: Yanzheng Xiang, Lan Wei, Yizhen Yao, Qinglin Zhu, Hanqi Yan, Chen Jin, Philip Alexander Teare, Dandan Zhang, Lin Gui, Amrutha Saseendran, Yulan He
Abstract: Parallel diffusion decoding can accelerate diffusion language model inference by unmasking multiple tokens per step, but aggressive parallelism often harms quality. Revocable decoding mitigates this by rechecking earlier tokens, yet we observe that existing verification schemes frequently trigger flip-flop oscillations, where tokens are remasked and later restored unchanged. This behaviour slows inference in two ways: remasking verified positions weakens the conditioning context for parallel drafting, and repeated remask cycles consume the revision budget with little net progress. We propose COVER (Cache Override Verification for Efficient Revision), which performs leave-one-out verification and stable drafting within a single forward pass. COVER constructs two attention views via KV cache override: selected seeds are masked for verification, while their cached key value states are injected for all other queries to preserve contextual information, with a closed form diagonal correction preventing self leakage at the seed positions. COVER further prioritises seeds using a stability aware score that balances uncertainty, downstream influence, and cache drift, and it adapts the number of verified seeds per step. Across benchmarks, COVER markedly reduces unnecessary revisions and yields faster decoding while preserving output quality.
Authors: Kaushik Sinha, Christopher Tosh
Abstract: Expand-and-sparsify representations are a class of theoretical models that capture sparse representation phenomena observed in the sensory systems of many animals. At a high level, these representations map an input $x \in \mathbb{R}^d$ to a much higher dimension $m \gg d$ via random linear projections before zeroing out all but the $k \ll m$ largest entries. The result is a $k$-sparse vector in $\{0,1\}^m$. We study the suitability of this representation for two fundamental statistical problems: density estimation and mode estimation. For density estimation, we show that a simple linear function of the expand-and-sparsify representation produces an estimator with minimax-optimal $\ell_{\infty}$ convergence rates. In mode estimation, we provide simple algorithms on top of our density estimator that recover single or multiple modes at optimal rates up to logarithmic factors under mild conditions.
Authors: Uwe Peters, Andrea Bertazzoli, Jasmine M. DeJesus, Gisela J. van der Velden, Benjamin Chin-Yee
Abstract: Scientists often use generics, that is, unquantified statements about whole categories of people or phenomena, when communicating research findings (e.g., "statins reduce cardiovascular events"). Large language models (LLMs), such as ChatGPT, frequently adopt the same style when summarizing scientific texts. However, generics can prompt overgeneralizations, especially when they are interpreted differently across audiences. In a study comparing laypeople, scientists, and two leading LLMs (ChatGPT-5 and DeepSeek), we found systematic differences in interpretation of generics. Compared to most scientists, laypeople judged scientific generics as more generalizable and credible, while LLMs rated them even higher. These mismatches highlight significant risks for science communication. Scientists may use generics and incorrectly assume laypeople share their interpretation, while LLMs may systematically overgeneralize scientific findings when summarizing research. Our findings underscore the need for greater attention to language choices in both human and LLM-mediated science communication.
Authors: Taewook Kim, Matthew K. Hong, Yan-Ying Chen, Jonathan Q. Li, Monica P Van, Shabnam Hakimi, Matthew Kay, Matthew Klenk
Abstract: Product designers often begin their design process with handcrafted personas. While personas are intended to ground design decisions in consumer preferences, they often fall short in practice by remaining abstract, expensive to produce, and difficult to translate into actionable design features. As a result, personas risk serving as static reference points rather than tools that actively shape design outcomes. To address these challenges, we built Personagram, an interactive system powered by multimodal large language models (MLLMs) that helps designers explore detailed census-based personas, extract product features inferred from persona attributes, and recombine them for specific customer segments. In a study with 12 professional designers, we show that Personagram facilitates more actionable ideation workflows by structuring multimodal thinking from persona attributes to product design features, achieving higher engagement with personas, perceived transparency, and satisfaction compared to a chat-based baseline. We discuss implications of integrating AI-generated personas into product design workflows.
Authors: Parv Maheshwari, Jay Karhade, Yogesh Chawla, Isaiah Adu, Florian Heisen, Andrew Porco, Andrew Jong, Yifei Liu, Santosh Pitla, Sebastian Scherer, Wenshan Wang
Abstract: We present AnyThermal, a thermal backbone that captures robust task-agnostic thermal features suitable for a variety of tasks such as cross-modal place recognition, thermal segmentation, and monocular depth estimation using thermal images. Existing thermal backbones that follow task-specific training from small-scale data result in utility limited to a specific environment and task. Unlike prior methods, AnyThermal can be used for a wide range of environments (indoor, aerial, off-road, urban) and tasks, all without task-specific training. Our key insight is to distill the feature representations from visual foundation models such as DINOv2 into a thermal encoder using thermal data from these multiple environments. To bridge the diversity gap of the existing RGB-Thermal datasets, we introduce the TartanRGBT platform, the first open-source data collection platform with synced RGB-Thermal image acquisition. We use this payload to collect the TartanRGBT dataset - a diverse and balanced dataset collected in 4 environments. We demonstrate the efficacy of AnyThermal and TartanRGBT, achieving state-of-the-art results with improvements of up to 36% across diverse environments and downstream tasks on existing datasets.
Authors: Nan Chen, Soledad Villar, Soufiane Hayou
Abstract: Low-Rank Adaptation (LoRA) is a standard tool for parameter-efficient finetuning of large models. While it induces a small memory footprint, its training dynamics can be surprisingly complex as they depend on several hyperparameters such as initialization, adapter rank, and learning rate. In particular, it is unclear how the optimal learning rate scales with adapter rank, which forces practitioners to re-tune the learning rate whenever the rank is changed. In this paper, we introduce Maximal-Update Adaptation ($\mu$A), a theoretical framework that characterizes how the "optimal" learning rate should scale with model width and adapter rank to produce stable, non-vanishing feature updates under standard configurations. $\mu$A is inspired from the Maximal-Update Parametrization ($\mu$P) in pretraining. Our analysis leverages techniques from hyperparameter transfer and reveals that the optimal learning rate exhibits different scaling patterns depending on initialization and LoRA scaling factor. Specifically, we identify two regimes: one where the optimal learning rate remains roughly invariant across ranks, and another where it scales inversely with rank. We further identify a configuration that allows learning rate transfer from LoRA to full finetuning, drastically reducing the cost of learning rate tuning for full finetuning. Experiments across language, vision, vision--language, image generation, and reinforcement learning tasks validate our scaling rules and show that learning rates tuned on LoRA transfer reliably to full finetuning.
Authors: Akshit Achara, Tatiana Gaintseva, Mateo Mahaut, Pritish Chakraborty, Viktor Stenby Johansson, Melih Barsbey, Emanuele Rodol\`a, Donato Crisostomi
Abstract: The Platonic Representation Hypothesis suggests that independently trained neural networks converge to increasingly similar latent spaces. However, current strategies for mapping these representations are inherently pairwise, scaling quadratically with the number of models and failing to yield a consistent global reference. In this paper, we study the alignment of $M \ge 3$ models. We first adapt Generalized Procrustes Analysis (GPA) to construct a shared orthogonal universe that preserves the internal geometry essential for tasks like model stitching. We then show that strict isometric alignment is suboptimal for retrieval, where agreement-maximizing methods like Canonical Correlation Analysis (CCA) typically prevail. To bridge this gap, we finally propose Geometry-Corrected Procrustes Alignment (GCPA), which establishes a robust GPA-based universe followed by a post-hoc correction for directional mismatch. Extensive experiments demonstrate that GCPA consistently improves any-to-any retrieval while retaining a practical shared reference space.
Authors: Alec S. Xu, Can Yaras, Matthew Asato, Qing Qu, Laura Balzano
Abstract: Recent empirical evidence has demonstrated that the training dynamics of large-scale deep neural networks occur within low-dimensional subspaces. While this has inspired new research into low-rank training, compression, and adaptation, theoretical justification for these dynamics in nonlinear networks remains limited. %compared to deep linear settings. To address this gap, this paper analyzes the learning dynamics of multi-layer perceptrons (MLPs) under gradient descent (GD). We demonstrate that the weight dynamics concentrate within invariant low-dimensional subspaces throughout training. Theoretically, we precisely characterize these invariant subspaces for two-layer networks with smooth nonlinear activations, providing insight into their emergence. Experimentally, we validate that this phenomenon extends beyond our theoretical assumptions. Leveraging these insights, we empirically show there exists a low-rank MLP parameterization that, when initialized within the appropriate subspaces, matches the classification performance of fully-parameterized counterparts on a variety of classification tasks.
Authors: Jorge Daniel Rodr\'iguez-Vidal, Gabriel Villalonga, Diego Porres, Antonio M. L\'opez Pe\~na
Abstract: End-to-End Autonomous Driving (E2E-AD) systems are typically grouped by the nature of their outputs: (i) waypoint-based models that predict a future trajectory, and (ii) action-based models that directly output throttle, steer and brake. Most recent benchmark protocols and training pipelines are waypoint-based, which makes action-based policies harder to train and compare, slowing their progress. To bridge this waypoint-action gap, we propose a novel, differentiable vehicle-model framework that rolls out predicted action sequences to their corresponding ego-frame waypoint trajectories while supervising in waypoint space. Our approach enables action-based architectures to be trained and evaluated, for the first time, within waypoint-based benchmarks without modifying the underlying evaluation protocol. We extensively evaluate our framework across multiple challenging benchmarks and observe consistent improvements over the baselines. In particular, on NAVSIM \texttt{navhard} our approach achieves state-of-the-art performance. Our code will be made publicly available upon acceptance.
Authors: Joseph Amigo, Rooholla Khorrambakht, Nicolas Mansard, Ludovic Righetti
Abstract: World models offer a promising avenue for more faithfully capturing complex dynamics, including contacts and non-rigidity, as well as complex sensory information, such as visual perception, in situations where standard simulators struggle. However, these models are computationally complex to evaluate, posing a challenge for popular RL approaches that have been successfully used with simulators to solve complex locomotion tasks but yet struggle with manipulation. This paper introduces a method that bypasses simulators entirely, training RL policies inside world models learned from robots' interactions with real environments. At its core, our approach enables policy training with large-scale diffusion models via a novel decoupled first-order gradient (FoG) method: a full-scale world model generates accurate forward trajectories, while a lightweight latent-space surrogate approximates its local dynamics for efficient gradient computation. This coupling of a local and global world model ensures high-fidelity unrolling alongside computationally tractable differentiation. We demonstrate the efficacy of our method on the Push-T manipulation task, where it significantly outperforms PPO in sample efficiency. We further evaluate our approach through an ego-centric object manipulation task with a quadruped. Together, these results demonstrate that learning inside data-driven world models is a promising pathway for solving hard-to-model RL tasks in image space without reliance on hand-crafted physics simulators.
Authors: Shyam Sundar Murali Krishnan, Dean Frederick Hougen
Abstract: The growth of machine learning demands interpretable models for critical applications, yet most high-performing models are ``black-box'' systems that obscure input-output relationships, while traditional rule-based algorithms like RuleFit suffer from a lack of predictive power and instability despite their simplicity. This motivated our development of Sparse Relaxed Regularized Regression Rule-Fit (SR4-Fit), a novel interpretable classification algorithm that addresses these limitations while maintaining superior classification performance. Using demographic characteristics of U.S. congressional districts from the Census Bureau's American Community Survey, we demonstrate that SR4-Fit can predict House election party outcomes with unprecedented accuracy and interpretability. Our results show that while the majority party remains the strongest predictor, SR4-Fit has revealed intrinsic combinations of demographic factors that affect prediction outcomes that were unable to be interpreted in black-box algorithms such as random forests. The SR4-Fit algorithm surpasses both black-box models and existing interpretable rule-based algorithms such as RuleFit with respect to accuracy, simplicity, and robustness, generating stable and interpretable rule sets while maintaining superior predictive performance, thus addressing the traditional trade-off between model interpretability and predictive capability in electoral forecasting. To further validate SR4-Fit's performance, we also apply it to six additional publicly available classification datasets, like the breast cancer, Ecoli, page blocks, Pima Indians, vehicle, and yeast datasets, and find similar results.
Authors: Ziyao Zeng, Chen Liu, Tianyu Liu, Hao Wang, Xiatao Sun, Fengyu Yang, Xiaofeng Liu, Zhiwen Fan
Abstract: Game balancing is a longstanding challenge requiring repeated playtesting, expert intuition, and extensive manual tuning. We introduce RuleSmith, the first framework that achieves automated game balancing by leveraging the reasoning capabilities of multi-agent LLMs. It couples a game engine, multi-agent LLMs self-play, and Bayesian optimization operating over a multi-dimensional rule space. As a proof of concept, we instantiate RuleSmith on CivMini, a simplified civilization-style game containing heterogeneous factions, economy systems, production rules, and combat mechanics, all governed by tunable parameters. LLM agents interpret textual rulebooks and game states to generate actions, to conduct fast evaluation of balance metrics such as win-rate disparities. To search the parameter landscape efficiently, we integrate Bayesian optimization with acquisition-based adaptive sampling and discrete projection: promising candidates receive more evaluation games for accurate assessment, while exploratory candidates receive fewer games for efficient exploration. Experiments show that RuleSmith converges to highly balanced configurations and provides interpretable rule adjustments that can be directly applied to downstream game systems. Our results illustrate that LLM simulation can serve as a powerful surrogate for automating design and balancing in complex multi-agent environments.
Authors: Yu Zhang, Sean Bin Yang, Arijit Khan, Cuneyt Gurcan Akcora
Abstract: Counterfactual explanations offer an intuitive way to interpret graph neural networks (GNNs) by identifying minimal changes that alter a model's prediction, thereby answering "what must differ for a different outcome?". In this work, we propose a novel framework, ATEX-CF that unifies adversarial attack techniques with counterfactual explanation generation-a connection made feasible by their shared goal of flipping a node's prediction, yet differing in perturbation strategy: adversarial attacks often rely on edge additions, while counterfactual methods typically use deletions. Unlike traditional approaches that treat explanation and attack separately, our method efficiently integrates both edge additions and deletions, grounded in theory, leveraging adversarial insights to explore impactful counterfactuals. In addition, by jointly optimizing fidelity, sparsity, and plausibility under a constrained perturbation budget, our method produces instance-level explanations that are both informative and realistic. Experiments on synthetic and real-world node classification benchmarks demonstrate that ATEX-CF generates faithful, concise, and plausible explanations, highlighting the effectiveness of integrating adversarial insights into counterfactual reasoning for GNNs.
Authors: Patryk Rybak, Pawe{\l} Batorski, Paul Swoboda, Przemys{\l}aw Spurek
Abstract: Machine unlearning for LLMs aims to remove sensitive or copyrighted data from trained models. However, the true efficacy of current unlearning methods remains uncertain. Standard evaluation metrics rely on benign queries that often mistake superficial information suppression for genuine knowledge removal. Such metrics fail to detect residual knowledge that more sophisticated prompting strategies could still extract. We introduce REBEL, an evolutionary approach for adversarial prompt generation designed to probe whether unlearned data can still be recovered. Our experiments demonstrate that REBEL successfully elicits ``forgotten'' knowledge from models that seemed to be forgotten in standard unlearning benchmarks, revealing that current unlearning methods may provide only a superficial layer of protection. We validate our framework on subsets of the TOFU and WMDP benchmarks, evaluating performance across a diverse suite of unlearning algorithms. Our experiments show that REBEL consistently outperforms static baselines, recovering ``forgotten'' knowledge with Attack Success Rates (ASRs) reaching up to 60% on TOFU and 93% on WMDP. We will make all code publicly available upon acceptance. Code is available at https://github.com/patryk-rybak/REBEL/
Authors: Aman Anand, Amir Eskandari, Elyas Rahsno, Farhana Zulkernine
Abstract: Self-supervised learning (SSL) has shown remarkable success in skeleton-based action recognition by leveraging data augmentations to learn meaningful representations. However, existing SSL methods rely on data augmentations that predominantly focus on masking high-motion frames and high-degree joints such as joints with degree 3 or 4. This results in biased and incomplete feature representations that struggle to generalize across varied motion patterns. To address this, we propose Asymmetric Spatio-temporal Masking (ASMa) for Skeleton Action Representation Learning, a novel combination of masking to learn a full spectrum of spatio-temporal dynamics inherent in human actions. ASMa employs two complementary masking strategies: one that selectively masks high-degree joints and low-motion, and another that masks low-degree joints and high-motion frames. These masking strategies ensure a more balanced and comprehensive skeleton representation learning. Furthermore, we introduce a learnable feature alignment module to effectively align the representations learned from both masked views. To facilitate deployment in resource-constrained settings and on low-resource devices, we compress the learned and aligned representation into a lightweight model using knowledge distillation. Extensive experiments on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD datasets demonstrate that our approach outperforms existing SSL methods with an average improvement of 2.7-4.4% in fine-tuning and up to 5.9% in transfer learning to noisy datasets and achieves competitive performance compared to fully supervised baselines. Our distilled model achieves 91.4% parameter reduction and 3x faster inference on edge devices while maintaining competitive accuracy, enabling practical deployment in resource-constrained scenarios.
Authors: Navita Goyal, Hal Daum\'e III
Abstract: Model steering, which involves intervening on hidden representations at inference time, has emerged as a lightweight alternative to finetuning for precisely controlling large language models. While steering efficacy has been widely studied, evaluations of whether interventions alter only the intended property remain limited, especially with respect to unintended changes in behaviors related to the target property. We call this notion specificity. We propose a framework that distinguishes three dimensions of specificity: general (preserving fluency and unrelated abilities), control (preserving related control properties), and robustness (preserving control properties under distribution shifts). We study two safety-critical use cases: steering models to reduce overrefusal and faithfulness hallucinations, and show that while steering achieves high efficacy and largely maintains general and control specificity, it consistently fails to preserve robustness specificity. In the case of overrefusal steering, for example, all steering methods reduce overrefusal without harming general abilities and refusal on harmful queries; however, they substantially increase vulnerability to jailbreaks. Our work provides the first systematic evaluation of specificity in model steering, showing that standard efficacy and specificity checks are insufficient, because without robustness evaluation, steering methods may appear reliable even when they compromise model safety.
Authors: Mark Russinovich, Yanan Cai, Keegan Hines, Giorgio Severi, Blake Bullwinkel, Ahmed Salem
Abstract: Safety alignment is only as robust as its weakest failure mode. Despite extensive work on safety post-training, it has been shown that models can be readily unaligned through post-deployment fine-tuning. However, these methods often require extensive data curation and degrade model utility. In this work, we extend the practical limits of unalignment by introducing GRP-Obliteration (GRP-Oblit), a method that uses Group Relative Policy Optimization (GRPO) to directly remove safety constraints from target models. We show that a single unlabeled prompt is sufficient to reliably unalign safety-aligned models while largely preserving their utility, and that GRP-Oblit achieves stronger unalignment on average than existing state-of-the-art techniques. Moreover, GRP-Oblit generalizes beyond language models and can also unalign diffusion-based image generation systems. We evaluate GRP-Oblit on six utility benchmarks and five safety benchmarks across fifteen 7-20B parameter models, spanning instruct and reasoning models, as well as dense and MoE architectures. The evaluated model families include GPT-OSS, distilled DeepSeek, Gemma, Llama, Ministral, and Qwen.
Authors: Pedro Cisneros-Velarde
Abstract: Polemic questions need more than one viewpoint to express a balanced answer. Large Language Models (LLMs) can provide a balanced answer, but also take a single aligned viewpoint or refuse to answer. In this paper, we study if such initial responses can be steered to a specific viewpoint in a simple and intuitive way: by only providing one-sided arguments supporting the viewpoint. Our systematic study has three dimensions: (i) which stance is induced in the LLM response, (ii) how the polemic question is formulated, (iii) how the arguments are shown. We construct a small dataset and remarkably find that opinion steering occurs across (i)-(iii) for diverse models, number of arguments, and topics. Switching to other arguments consistently decreases opinion steering.
Authors: Parsa Gooya, Reinel Sospedra-Alfonso, Johannes Exenberger
Abstract: Accurately quantifying uncertainty in predictions and projections arising from irreducible internal climate variability is critical for informed decision making. Such uncertainty is typically assessed using ensembles produced with physics based climate models. However, computational constraints impose a trade off between generating the large ensembles required for robust uncertainty estimation and increasing model resolution to better capture fine scale dynamics. Generative machine learning offers a promising pathway to alleviate these constraints. We develop a conditional Variational Autoencoder (cVAE) trained on a limited sample of climate simulations to generate arbitrary large ensembles. The approach is applied to output from monthly CMIP6 historical and future scenario experiments produced with the Canadian Centre for Climate Modelling and Analysis' (CCCma's) Earth system model CanESM5. We show that the cVAE model learns the underlying distribution of the data and generates physically consistent samples that reproduce realistic low and high moment statistics, including extremes. Compared with more sophisticated generative architectures, cVAEs offer a mathematically transparent, interpretable, and computationally efficient framework. Their simplicity lead to some limitations, such as overly smooth outputs, spectral bias, and underdispersion, that we discuss along with strategies to mitigate them. Specifically, we show that incorporating output noise improves the representation of climate relevant multiscale variability, and we propose a simple method to achieve this. Finally, we show that cVAE-enhanced ensembles capture realistic global teleconnection patterns, even under climate conditions absent from the training data.
Authors: Jinchi Tang, Yan Guo
Abstract: With the advancement of deep learning technologies, specialized neural processing hardware such as Brain Processing Units (BPUs) have emerged as dedicated platforms for CNN acceleration, offering optimized INT8 computation capabilities for convolutional operations. Meanwhile, Vision Transformer (ViT) models, such as the Data-efficient Image Transformer (DeiT), have demonstrated superior performance and play increasingly crucial roles in computer vision tasks. However, due to the architectural mismatch between CNN-optimized hardware and Vision Transformer computation characteristics--namely, that linear layers in Transformers operate on three-dimensional data while BPU acceleration is designed for four-dimensional convolution operations-it is difficult or even impossible to leverage BPU's advantages when deploying Vision Transformers. To address this challenge, we propose a novel approach that restructures the Vision Transformer by replacing linear layers and layer normalization operations with carefully designed convolutional operators. This enables DeiT to fully utilize the acceleration capabilities of BPUs, while allowing the original weight parameters to be inherited by the restructured models without retraining or fine-tuning. To the best of our knowledge, this is the first successful deployment of Vision Transformers that fully leverages BPU classification datasets demonstrate the effectiveness of our approach. Specifically, the quantized DeiT-Base model achieves 80.4% accuracy on ImageNet, compared to the original 81.8%, while obtaining up to a 3.8* inference speedup. Our finetuned DeiT model on the flower classification dataset also achieves excellent performance, with only a 0.5% accuracy drop for the DeiT-Base model, further demonstrating the effectiveness of our method.
Authors: Jorge L. Ruiz Williams
Abstract: We present the Condensate Theorem: attention sparsity is a learned topological property, not an architectural constraint. Through empirical analysis of trained language models, we find that attention mass concentrates on a distinct topological manifold -- and this manifold can be identified dynamically without checking every position. We prove a general result: for any query, projecting attention onto the Condensate Manifold (Anchor + Window + Dynamic Top-k) achieves 100% output equivalence with full $O(n^2)$ attention. This is not an approximation -- it is lossless parity. We validate this across GPT-2, Pythia, Qwen2, TinyLlama, and Mistral, demonstrating bit-exact token matching on 1,500+ generated tokens. By mapping this topology to hardware, our Topological Attention kernel achieves a 159x measured speedup at 131K tokens (3.94ms vs 628ms) and a projected >1,200x speedup at 1M tokens, reducing inference costs by >99.9% compared to Flash Attention. We conclude that the quadratic bottleneck is an artifact of naive implementation, not intelligence.
Authors: Junqi Chen, Sirui Chen, Chaochao Lu
Abstract: Causal inference is essential for decision-making but remains challenging for non-experts. While large language models (LLMs) show promise in this domain, their precise causal estimation capabilities are still limited, and the impact of post-training on these abilities is insufficiently explored. This paper examines the extent to which post-training can enhance LLMs' capacity for causal inference. We introduce CauGym, a comprehensive dataset comprising seven core causal tasks for training and five diverse test sets. Using this dataset, we systematically evaluate five post-training approaches: SFT, DPO, KTO, PPO, and GRPO. Across five in-domain and four existing benchmarks, our experiments demonstrate that appropriate post-training enables smaller LLMs to perform causal inference competitively, often surpassing much larger models. Our 14B parameter model achieves 93.5% accuracy on the CaLM benchmark, compared to 55.4% by OpenAI o3. Furthermore, the post-trained LLMs exhibit strong generalization and robustness under real-world conditions such as distribution shifts and noisy data. Collectively, these findings provide the first systematic evidence that targeted post-training can produce reliable and robust LLM-based causal reasoners. Our data and GRPO-model are available at https://github.com/OpenCausaLab/CauGym.
Authors: Harold Soh, Eugene Lim
Abstract: Robot Foundation Models such as Vision-Language-Action models are rapidly reshaping how robot policies are trained and deployed, replacing hand-designed planners with end-to-end generative action models. While these systems demonstrate impressive generalization, it remains unclear whether they fundamentally resolve the long-standing challenges of robotics. We address this question by analyzing action hallucinations that violate physical constraints and their extension to plan-level failures. Focusing on latent-variable generative policies, we show that hallucinations often arise from structural mismatches between feasible robot behavior and common model architectures. We study three such barriers -- topological, precision, and horizon -- and show how they impose unavoidable tradeoffs. Our analysis provides mechanistic explanations for reported empirical failures of generative robot policies and suggests principled directions for improving reliability and trustworthiness, without abandoning their expressive power.
Authors: Qianlong Lan, Anuj Kaul, Shaun Jones, Stephanie Westrum
Abstract: The deployment of autonomous AI agents capable of executing commercial transactions has motivated the adoption of mandate-based payment authorization protocols, including the Universal Commerce Protocol (UCP) and the Agent Payments Protocol (AP2). These protocols replace interactive, session-based authorization with cryptographically issued mandates, enabling asynchronous and autonomous execution. While AP2 provides specification-level guarantees through signature verification, explicit binding, and expiration semantics, real-world agentic execution introduces runtime behaviors such as retries, concurrency, and orchestration that challenge implicit assumptions about mandate usage. In this work, we present a security analysis of the AP2 mandate lifecycle and identify enforcement gaps that arise during runtime in agent-based payment systems. We propose a zero-trust runtime verification framework that enforces explicit context binding and consume-once mandate semantics using dynamically generated, time-bound nonces, ensuring that authorization decisions are evaluated at execution time rather than assumed from static issuance properties. Through simulation-based evaluation under high concurrency, we show that context-aware binding and consume-once enforcement address distinct and complementary attack classes, and that both are required to prevent replay and context-redirect attacks. The proposed framework mitigates all evaluated attacks while maintaining stable verification latency of approximately 3.8~ms at throughput levels up to 10{,}000 transactions per second. We further demonstrate that the required runtime state is bounded by peak concurrency rather than cumulative transaction history, indicating that robust runtime security for agentic payment execution can be achieved with minimal and predictable overhead.
Authors: Sanjana Reddy (Google), Ishaan Malhi (Google DeepMind), Sally Ma (Google DeepMind), Praneet Dutta (Google DeepMind)
Abstract: Existing methods for preference tuning of text-to-image (T2I) diffusion models often rely on computationally expensive generation steps to create positive and negative pairs of images. These approaches frequently yield training pairs that either lack meaningful differences, are expensive to sample and filter, or exhibit significant variance in irrelevant pixel regions, thereby degrading training efficiency. To address these limitations, we introduce "Di3PO", a novel method for constructing positive and negative pairs that isolates specific regions targeted for improvement during preference tuning, while keeping the surrounding context in the image stable. We demonstrate the efficacy of our approach by applying it to the challenging task of text rendering in diffusion models, showcasing improvements over baseline methods of SFT and DPO.
Authors: Yewei Liu, Xiyuan Wang, Yansheng Mao, Yoav Gelbery, Haggai Maron, Muhan Zhang
Abstract: We propose SHINE (Scalable Hyper In-context NEtwork), a scalable hypernetwork that can map diverse meaningful contexts into high-quality LoRA adapters for large language models (LLM). By reusing the frozen LLM's own parameters in an in-context hypernetwork design and introducing architectural innovations, SHINE overcomes key limitations of prior hypernetworks and achieves strong expressive power with a relatively small number of parameters. We introduce a pretraining and instruction fine-tuning pipeline, and train our hypernetwork to generate high quality LoRA adapters from diverse meaningful contexts in a single forward pass. It updates LLM parameters without any fine-tuning, and immediately enables complex question answering tasks related to the context without directly accessing the context, effectively transforming in-context knowledge to in-parameter knowledge in one pass. Our work achieves outstanding results on various tasks, greatly saves time, computation and memory costs compared to SFT-based LLM adaptation, and shows great potential for scaling. Our code is available at https://github.com/Yewei-Liu/SHINE
Authors: Xiyang Zhang, Yuanhe Tian, Hongzhi Wang, Yan Song
Abstract: Fine-tuning large language models (LLMs) for specialized domains often necessitates a trade-off between acquiring domain expertise and retaining general reasoning capabilities, a phenomenon known as catastrophic forgetting. Existing remedies face a dichotomy: gradient surgery methods offer geometric safety but incur prohibitive computational costs via online projections, while efficient data selection approaches reduce overhead but remain blind to conflict-inducing gradient directions. In this paper, we propose Orthogonal Gradient Selection (OGS), a data-centric method that harmonizes domain performance, general capability retention, and training efficiency. OGS shifts the geometric insights of gradient projection from the optimizer to the data selection stage by treating data selection as a constrained decision-making process. By leveraging a lightweight Navigator model and reinforcement learning techniques, OGS dynamically identifies training samples whose gradients are orthogonal to a general-knowledge anchor. This approach ensures naturally safe updates for target models without modifying the optimizer or incurring runtime projection costs. Experiments across medical, legal, and financial domains demonstrate that OGS achieves excellent results, significantly improving domain performance and training efficiency while maintaining or even enhancing performance on general tasks such as GSM8K.
Authors: Fuxi Zhang, Yifan Wang, Hengrun Zhao, Zhuohan Sun, Changxing Xia, Lijun Wang, Huchuan Lu, Yangrui Shao, Chen Yang, Long Teng
Abstract: Salient object detection is inherently a subjective problem, as observers with different priors may perceive different objects as salient. However, existing methods predominantly formulate it as an objective prediction task with a single groundtruth segmentation map for each image, which renders the problem under-determined and fundamentally ill-posed. To address this issue, we propose Observer-Centric Salient Object Detection (OC-SOD), where salient regions are predicted by considering not only the visual cues but also the observer-specific factors such as their preferences or intents. As a result, this formulation captures the intrinsic ambiguity and diversity of human perception, enabling personalized and context-aware saliency prediction. By leveraging multi-modal large language models, we develop an efficient data annotation pipeline and construct the first OC-SOD dataset named OC-SODBench, comprising 33k training, validation and test images with 152k textual prompts and object pairs. Built upon this new dataset, we further design OC-SODAgent, an agentic baseline which performs OC-SOD via a human-like "Perceive-Reflect-Adjust" process. Extensive experiments on our proposed OC-SODBench have justified the effectiveness of our contribution. Through this observer-centric perspective, we aim to bridge the gap between human perception and computational modeling, offering a more realistic and flexible understanding of what makes an object truly "salient." Code and dataset are publicly available at: https://github.com/Dustzx/OC_SOD
Authors: David Yavo, Richard Khoury, Christophe Pere, Sadoune Ait Kaci Azzou
Abstract: Synthetic tabular data enables sharing and analysis of sensitive records, but its practical deployment requires balancing distributional fidelity, downstream utility, and privacy protection. We study a simple, model agnostic post processing framework that can be applied on top of any synthetic data generator to improve this trade off. First, a mode patching step repairs categories that are missing or severely underrepresented in the synthetic data, while largely preserving learned dependencies. Second, a k nearest neighbor filter replaces synthetic records that lie too close to real data points, enforcing a minimum distance between real and synthetic samples. We instantiate this framework for two neural generative models for tabular data, a feed forward generator and a variational autoencoder, and evaluate it on three public datasets covering credit card transactions, cardiovascular health, and census based income. We assess marginal and joint distributional similarity, the performance of models trained on synthetic data and evaluated on real data, and several empirical privacy indicators, including nearest neighbor distances and attribute inference attacks. With moderate thresholds between 0.2 and 0.35, the post processing reduces divergence between real and synthetic categorical distributions by up to 36 percent and improves a combined measure of pairwise dependence preservation by 10 to 14 percent, while keeping downstream predictive performance within about 1 percent of the unprocessed baseline. At the same time, distance based privacy indicators improve and the success rate of attribute inference attacks remains largely unchanged. These results provide practical guidance for selecting thresholds and applying post hoc repairs to improve the quality and empirical privacy of synthetic tabular data, while complementing approaches that provide formal differential privacy guarantees.
Authors: Mona Rajhans, Vishal Khawarey
Abstract: Machine learning (ML) models are increasingly deployed in cybersecurity applications such as phishing detection and network intrusion prevention. However, these models remain vulnerable to adversarial perturbations small, deliberate input modifications that can degrade detection accuracy and compromise interpretability. This paper presents an empirical study of adversarial robustness and explainability drift across two cybersecurity domains phishing URL classification and network intrusion detection. We evaluate the impact of L (infinity) bounded Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) perturbations on model accuracy and introduce a quantitative metric, the Robustness Index (RI), defined as the area under the accuracy perturbation curve. Gradient based feature sensitivity and SHAP based attribution drift analyses reveal which input features are most susceptible to adversarial manipulation. Experiments on the Phishing Websites and UNSW NB15 datasets show consistent robustness trends, with adversarial training improving RI by up to 9 percent while maintaining clean-data accuracy. These findings highlight the coupling between robustness and interpretability degradation and underscore the importance of quantitative evaluation in the design of trustworthy, AI-driven cybersecurity systems.
Authors: Xin Jin, Tiejun Lv, Yashuai Cao, Jie Zeng, Mugen Peng
Abstract: This letter proposes an active reconfigurable intelligent surface (ARIS) assisted rate-splitting multiple access (RSMA) integrated sensing and communication (ISAC) system to overcome the fairness bottleneck in multi-target sensing under obstructed line-of-sight environments. Beamforming at the transceiver and ARIS, along with rate splitting, are optimized to maximize the minimum multi-target echo signal-to-interference-plus-noise ratio under multi-user rate and power constraints. The intricate non-convex problem is decoupled into three subproblems and solved iteratively by majorization-minimization (MM) and sequential rank-one constraint relaxation (SROCR) algorithms. Simulations show our scheme outperforms nonorthogonal multiple access, space-division multiple access, and passive RIS baselines, approaching sensing-only upper bounds.
Authors: Zhenxing Ming, Julie Stephany Berrio, Mao Shan, Stewart Worrall
Abstract: 3D semantic occupancy prediction enables autonomous vehicles (AVs) to perceive fine-grained geometric and semantic structure of their surroundings from onboard sensors, which is essential for safe decision-making and navigation. Recent models for 3D semantic occupancy prediction have successfully addressed the challenge of describing real-world objects with varied shapes and classes. However, the intermediate representations used by existing methods for 3D semantic occupancy prediction rely heavily on 3D voxel volumes or a set of 3D Gaussians, hindering the model's ability to efficiently and effectively capture fine-grained geometric details in the 3D driving environment. This paper introduces TFusionOcc, a novel object-centric multi-sensor fusion framework for predicting 3D semantic occupancy. By leveraging multi-stage multi-sensor fusion, Student's t-distribution, and the T-Mixture model (TMM), together with more geometrically flexible primitives, such as the deformable superquadric (superquadric with inverse warp), the proposed method achieved state-of-the-art (SOTA) performance on the nuScenes benchmark. In addition, extensive experiments were conducted on the nuScenes-C dataset to demonstrate the robustness of the proposed method in different camera and lidar corruption scenarios. The code will be available at: https://github.com/DanielMing123/TFusionOcc
Authors: Fumitaka Iwaki, Tatsuji Takahashi
Abstract: In the field of natural language processing, some studies have attempted sentiment analysis on text by handling emotions as explanatory or response variables. One of the most popular emotion models used in this context is the wheel of emotion proposed by Plutchik. This model schematizes human emotions in a circular structure, and represents them in two or three dimensions. However, the validity of Plutchik's wheel of emotion has not been sufficiently examined. This study investigated the validity of the wheel by creating and analyzing a semantic networks of emotion words. Through our experiments, we collected data of similarity and association of ordered pairs of emotion words, and constructed networks using these data. We then analyzed the structure of the networks through community detection, and compared it with that of the wheel of emotion. The results showed that each network's structure was, for the most part, similar to that of the wheel of emotion, but locally different.
Authors: Abhishek Jangra, Sachin Thukral, Arnab Chatterjee, Jayasree Raveendran
Abstract: This study examines the hierarchical structure of financial needs as articulated in social media discourse, employing generative AI techniques to analyze large-scale textual data. While human needs encompass a broad spectrum from fundamental survival to psychological fulfillment financial needs are particularly critical, influencing both individual well-being and day-to-day decision-making. Our research advances the understanding of financial behavior by utilizing large language models (LLMs) to extract and analyze expressions of financial needs from social media posts. We hypothesize that financial needs are organized hierarchically, progressing from short-term essentials to long-term aspirations, consistent with theoretical frameworks established in the behavioral sciences. Through computational analysis, we demonstrate the feasibility of identifying these needs and validate the presence of a hierarchical structure within them. In addition to confirming this structure, our findings provide novel insights into the content and themes of financial discussions online. By inferring underlying needs from naturally occurring language, this approach offers a scalable and data-driven alternative to conventional survey methodologies, enabling a more dynamic and nuanced understanding of financial behavior in real-world contexts.
Authors: Sung-Hoon Yoon, Ruizhi Qian, Minda Zhao, Weiyue Li, Mengyu Wang
Abstract: Large Language Models (LLMs) have become integral to many domains, making their safety a critical priority. Prior jailbreaking research has explored diverse approaches, including prompt optimization, automated red teaming, obfuscation, and reinforcement learning (RL) based methods. However, most existing techniques fail to effectively leverage vulnerabilities revealed in earlier interaction turns, resulting in inefficient and unstable attacks. Since jailbreaking involves sequential interactions in which each response influences future actions, reinforcement learning provides a natural framework for this problem. Motivated by this, we propose a history-aware RL-based jailbreak framework that analyzes and reweights vulnerability signals from prior steps to guide future decisions. We show that incorporating historical information alone improves jailbreak success rates. Building on this insight, we introduce an attention-based reweighting mechanism that highlights critical vulnerabilities within the interaction history, enabling more efficient exploration with fewer queries. Extensive experiments on AdvBench and HarmBench demonstrate that our method achieves state-of-the-art jailbreak performance while significantly improving query efficiency. These results underscore the importance of historical vulnerability signals in reinforcement learning-driven jailbreak strategies and offer a principled pathway for advancing adversarial research on LLM safeguards.
Authors: Yibing Liu, Chong Zhang, Zhongyi Han, Hansong Liu, Yong Wang, Yang Yu, Xiaoyan Wang, Yilong Yin
Abstract: We address the problem of runtime trajectory anomaly detection, a critical capability for enabling trustworthy LLM agents. Current safety measures predominantly focus on static input/output filtering. However, we argue that ensuring LLM agents reliability requires auditing the intermediate execution process. In this work, we formulate the task of Trajectory Anomaly Detection. The goal is not merely detection, but precise error localization. This capability is essential for enabling efficient rollback-and-retry. To achieve this, we construct TrajBench, a dataset synthesized via a perturb-and-complete strategy to cover diverse procedural anomalies. Using this benchmark, we investigate the capability of models in process supervision. We observe that general-purpose LLMs, even with zero-shot prompting, struggle to identify and localize these anomalies. This reveals that generalized capabilities do not automatically translate to process reliability. To address this, we propose TrajAD, a specialized verifier trained with fine-grained process supervision. Our approach outperforms baselines, demonstrating that specialized supervision is essential for building trustworthy agents.
Authors: Satyam Dwivedi, Sanjukta Ghosh, Shivam Dwivedi, Nishi Kumari, Anil Thakur, Anurag Purushottam, Deepak Alok, Praveen Gatla, Manjuprasad B, Bipasha Patgiri
Abstract: Large Language Models (LLMs) perform well on many reasoning benchmarks, yet existing evaluations rarely assess their ability to distinguish between meaningful semantic relations and genuine unrelatedness. We introduce CORE (Comprehensive Ontological Relation Evaluation), a dataset of 225K multiple-choice questions spanning 74 disciplines, together with a general-domain open-source benchmark of 203 rigorously validated questions (Cohen's Kappa = 1.0) covering 24 semantic relation types with equal representation of unrelated pairs. A human baseline from 1,000+ participants achieves 92.6% accuracy (95.1% on unrelated pairs). In contrast, 29 state-of-the-art LLMs achieve 48.25-70.9% overall accuracy, with near-ceiling performance on related pairs (86.5-100%) but severe degradation on unrelated pairs (0-41.35%), despite assigning similar confidence (92-94%). Expected Calibration Error increases 2-4x on unrelated pairs, and a mean semantic collapse rate of 37.6% indicates systematic generation of spurious relations. On the CORE 225K MCQs dataset, accuracy further drops to approximately 2%, highlighting substantial challenges in domain-specific semantic reasoning. We identify unrelatedness reasoning as a critical, under-evaluated frontier for LLM evaluation and safety.
Authors: Yingming Pu, Tao Lin, Hongyu Chen
Abstract: Large Language Model (LLM)-based scientific agents have accelerated scientific discovery, yet they often suffer from significant inefficiencies due to adherence to fixed initial priors. Existing approaches predominantly operate within a static hypothesis space, which restricts the discovery of novel phenomena, resulting in computational waste when baseline theories fail. To address this, we propose shifting the focus from searching hypotheses to evolving the underlying scientific principles. We present PiEvo, a principle-evolvable framework that treats scientific discovery as Bayesian optimization over an expanding principle space. By integrating Information-Directed Hypothesis Selection via Gaussian Process and an anomaly-driven augmentation mechanism, PiEvo enables agents to autonomously refine their theoretical worldview. Evaluation across four benchmarks demonstrates that PiEvo (1) achieves an average solution quality of up to 90.81%~93.15%, representing a 29.7%~31.1% improvement over the state-of-the-art, (2) attains an 83.3% speedup in convergence step via significantly reduced sample complexity by optimizing the compact principle space, and (3) maintains robust performance across diverse scientific domains and LLM backbones.
Authors: Changyue Wang, Weihang Su, Qingyao Ai, Yiqun Liu
Abstract: Scaling training data and model parameters has long driven progress in large language models (LLMs), but this paradigm is increasingly constrained by the scarcity of high-quality data and diminishing returns from rising computational costs. As a result, recent work is increasing the focus on continual learning from real-world deployment, where user interaction logs provide a rich source of authentic human feedback and procedural knowledge. However, learning from user logs is challenging due to their unstructured and noisy nature. Vanilla LLM systems often struggle to distinguish useful feedback signals from noisy user behavior, and the disparity between user log collection and model optimization (e.g., the off-policy optimization problem) further strengthens the problem. To this end, we propose UNO (User log-driveN Optimization), a unified framework for improving LLM systems (LLMsys) with user logs. UNO first distills logs into semi-structured rules and preference pairs, then employs query-and-feedback-driven clustering to manage data heterogeneity, and finally quantifies the cognitive gap between the model's prior knowledge and the log data. This assessment guides the LLMsys to adaptively filter out noisy feedback and construct different modules for primary and reflective experiences extracted from user logs, thereby improving future responses. Extensive experiments show that UNO achieves state-of-the-art effectiveness and efficiency, significantly outperforming Retrieval Augmented Generation (RAG) and memory-based baselines. We have open-sourced our code at https://github.com/bebr2/UNO .
Authors: Feng-Ting Liao, Meng-Hsi Chen, Guan-Ting Yi, Da-shan Shiu
Abstract: Dense Transformer language models have largely adhered to one consistent architectural shape: each layer consists of an attention module followed by a feed-forward network (FFN) with a narrow-wide-narrow MLP, allocating most parameters to the MLP at expansion ratios between 2 and 4. Motivated by recent results that residual wide-narrow-wide (hourglass) MLPs offer superior function approximation capabilities, we revisit the long-standing MLP shape convention in Transformer, challenging the necessity of the narrow-wide-narrow design. To study this, we develop a Transformer variant that replaces the conventional FFN with a deeper hourglass-shaped FFN, comprising a stack of hourglass sub-MLPs connected by residual pathways. We posit that a deeper but lighter hourglass FFN can serve as a competitive alternative to the conventional FFN, and that parameters saved by using a lighter hourglass FFN can be more effectively utilized, such as by enlarging model hidden dimensions under fixed budgets. We confirm these through empirical validations across model scales: hourglass FFNs outperform conventional FFNs up to 400M and achieve comparable performance at larger scales to 1B parameters; hourglass FFN variants with reduced FFN and increased attention parameters show consistent improvements over conventional configurations at matched budgets. Together, these findings shed new light on recent work and prompt a rethinking of the narrow-wide-narrow MLP convention and the balance between attention and FFN towards efficient and expressive modern language models.
Authors: Kyungbeom Kim, Seungwon Oh, Kyung-Joong Kim
Abstract: Parameter sharing is a key strategy in multi-agent reinforcement learning (MARL) for improving scalability, yet conventional fully shared architectures often collapse into homogeneous behaviors. Recent methods introduce diversity through clustering, pruning, or masking, but typically compromise resource efficiency. We propose Prism, a parameter sharing framework that induces inter-agent diversity by representing shared networks in the spectral domain via singular value decomposition (SVD). All agents share the singular vector directions while learning distinct spectral masks on singular values. This mechanism encourages inter-agent diversity and preserves scalability. Extensive experiments on both homogeneous (LBF, SMACv2) and heterogeneous (MaMuJoCo) benchmarks show that Prism achieves competitive performance with superior resource efficiency.
Authors: Xiaosong Jia, Yihang Sun, Junqi You, Songbur Wong, Zichen Zou, Junchi Yan, Zuxuan Wu, Yu-Gang Jiang
Abstract: Feedforward models for novel view synthesis (NVS) have recently advanced by transformer-based methods like LVSM, using attention among all input and target views. In this work, we argue that its full self-attention design is suboptimal, suffering from quadratic complexity with respect to the number of input views and rigid parameter sharing among heterogeneous tokens. We propose Efficient-LVSM, a dual-stream architecture that avoids these issues with a decoupled co-refinement mechanism. It applies intra-view self-attention for input views and self-then-cross attention for target views, eliminating unnecessary computation. Efficient-LVSM achieves 29.86 dB PSNR on RealEstate10K with 2 input views, surpassing LVSM by 0.2 dB, with 2x faster training convergence and 4.4x faster inference speed. Efficient-LVSM achieves state-of-the-art performance on multiple benchmarks, exhibits strong zero-shot generalization to unseen view counts, and enables incremental inference with KV-cache, thanks to its decoupled designs.
Authors: Minjeong Ban, Jeonghwan Choi, Hyangsuk Min, Nicole Hee-Yeon Kim, Minseok Kim, Jae-Gil Lee, Hwanjun Song
Abstract: Information retrieval (IR) evaluation remains challenging due to incomplete IR benchmark datasets that contain unlabeled relevant chunks. While LLMs and LLM-human hybrid strategies reduce costly human effort, they remain prone to LLM overconfidence and ineffective AI-to-human escalation. To address this, we propose DREAM, a multi-round debate-based relevance assessment framework with LLM agents, built on opposing initial stances and iterative reciprocal critique. Through our agreement-based debate, it yields more accurate labeling for certain cases and more reliable AI-to-human escalation for uncertain ones, achieving 95.2% labeling accuracy with only 3.5% human involvement. Using DREAM, we build BRIDGE, a refined benchmark that mitigates evaluation bias and enables fairer retriever comparison by uncovering 29,824 missing relevant chunks. We then re-benchmark IR systems and extend evaluation to RAG, showing that unaddressed holes not only distort retriever rankings but also drive retrieval-generation misalignment. The relevance assessment framework is available at https: //github.com/DISL-Lab/DREAM-ICLR-26; and the BRIDGE dataset is available at https://github.com/DISL-Lab/BRIDGE-Benchmark.
Authors: Andy Rosenbaum, Assaf Siani, Ilan Kernerman
Abstract: We release MTQE.en-he: to our knowledge, the first publicly available English-Hebrew benchmark for Machine Translation Quality Estimation. MTQE.en-he contains 959 English segments from WMT24++, each paired with a machine translation into Hebrew, and Direct Assessment scores of the translation quality annotated by three human experts. We benchmark ChatGPT prompting, TransQuest, and CometKiwi and show that ensembling the three models outperforms the best single model (CometKiwi) by 6.4 percentage points Pearson and 5.6 percentage points Spearman. Fine-tuning experiments with TransQuest and CometKiwi reveal that full-model updates are sensitive to overfitting and distribution collapse, yet parameter-efficient methods (LoRA, BitFit, and FTHead, i.e., fine-tuning only the classification head) train stably and yield improvements of 2-3 percentage points. MTQE.en-he and our experimental results enable future research on this under-resourced language pair.
Authors: Yi Liu, Zhihao Chen, Yanjun Zhang, Gelei Deng, Yuekang Li, Jianting Ning, Leo Yu Zhang
Abstract: Third-party agent skills extend LLM-based agents with instruction files and executable code that run on users' machines. Skills execute with user privileges and are distributed through community registries with minimal vetting, but no ground-truth dataset exists to characterize the resulting threats. We construct the first labeled dataset of malicious agent skills by behaviorally verifying 98,380 skills from two community registries, confirming 157 malicious skills with 632 vulnerabilities. These attacks are not incidental. Malicious skills average 4.03 vulnerabilities across a median of three kill chain phases, and the ecosystem has split into two archetypes: Data Thieves that exfiltrate credentials through supply chain techniques, and Agent Hijackers that subvert agent decision-making through instruction manipulation. A single actor accounts for 54.1\% of confirmed cases through templated brand impersonation. Shadow features, capabilities absent from public documentation, appear in 0\% of basic attacks but 100\% of advanced ones; several skills go further by exploiting the AI platform's own hook system and permission flags. Responsible disclosure led to 93.6\% removal within 30 days. We release the dataset and analysis pipeline to support future work on agent skill security.
Authors: Jan Benad, Pradeep Kr. Banerjee, Frank R\"oder, Nihat Ay, Martin V. Butz, Manfred Eppe
Abstract: Zero-shot generalization in contextual reinforcement learning remains a core challenge, particularly when the context is latent and must be inferred from data. A canonical failure mode is actuator inversion, where identical actions produce opposite physical effects under a latent binary context. We propose DMA*-SH, a framework where a single hypernetwork, trained solely via dynamics prediction, generates a small set of adapter weights shared across the dynamics model, policy, and action-value function. This shared modulation imparts an inductive bias matched to actuator inversion, while input/output normalization and random input masking stabilize context inference, promoting directionally concentrated representations. We provide theoretical support via an expressivity separation result for hypernetwork modulation, and a variance decomposition with policy-gradient variance bounds that formalize how within-mode compression improves learning under actuator inversion. For evaluation, we introduce the Actuator Inversion Benchmark (AIB), a suite of environments designed to isolate discontinuous context-to-dynamics interactions. On AIB's held-out actuator-inversion tasks, DMA*-SH achieves zero-shot generalization, outperforming domain randomization by 111.8% and surpassing a standard context-aware baseline by 16.1%.
Authors: Guodong Wang, Chenkai Zhang, Qingjie Liu, Jinjin Zhang, Jiancheng Cai, Junjie Liu, Xinmin Liu
Abstract: Reliable benchmarking is critical for advancing Vision-Language-Action (VLA) models, as it reveals their generalization, robustness, and alignment of perception with language-driven manipulation tasks. However, existing benchmarks often provide limited or misleading assessments due to insufficient evaluation protocols that inadequately capture real-world distribution shifts. This work systematically rethinks VLA benchmarking from both evaluation and data perspectives, introducing LIBERO-X, a benchmark featuring: 1) A hierarchical evaluation protocol with progressive difficulty levels targeting three core capabilities: spatial generalization, object recognition, and task instruction understanding. This design enables fine-grained analysis of performance degradation under increasing environmental and task complexity; 2) A high-diversity training dataset collected via human teleoperation, where each scene supports multiple fine-grained manipulation objectives to bridge the train-evaluation distribution gap. Experiments with representative VLA models reveal significant performance drops under cumulative perturbations, exposing persistent limitations in scene comprehension and instruction grounding. By integrating hierarchical evaluation with diverse training data, LIBERO-X offers a more reliable foundation for assessing and advancing VLA development.
Authors: Yassine Abbahaddou
Abstract: Graph Neural Networks (GNNs) have established themselves as the leading models for learning on graph-structured data, generally categorized into spatial and spectral approaches. Central to these architectures is the Graph Shift Operator (GSO), a matrix representation of the graph structure used to filter node signals. However, selecting the optimal GSO, whether fixed or learnable, remains largely empirical. In this paper, we introduce a novel alignment gain metric that quantifies the geometric distortion between the input signal and label subspaces. Crucially, our theoretical analysis connects this alignment directly to generalization bounds via a spectral proxy for the Lipschitz constant. This yields a principled, computation-efficient criterion to rank and select the optimal GSO for any prediction task prior to training, eliminating the need for extensive search.
Authors: Niccolo Avogaro, Nayanika Debnath, Li Mi, Thomas Frick, Junling Wang, Zexue He, Hang Hua, Konrad Schindler, Mattia Rigotti
Abstract: Despite recent successes, test-time scaling - i.e., dynamically expanding the token budget during inference as needed - remains brittle for vision-language models (VLMs): unstructured chains-of-thought about images entangle perception and reasoning, leading to long, disorganized contexts where small perceptual mistakes may cascade into completely wrong answers. Moreover, expensive reinforcement learning with hand-crafted rewards is required to achieve good performance. Here, we introduce SPARC (Separating Perception And Reasoning Circuits), a modular framework that explicitly decouples visual perception from reasoning. Inspired by sequential sensory-to-cognitive processing in the brain, SPARC implements a two-stage pipeline where the model first performs explicit visual search to localize question-relevant regions, then conditions its reasoning on those regions to produce the final answer. This separation enables independent test-time scaling with asymmetric compute allocation (e.g., prioritizing perceptual processing under distribution shift), supports selective optimization (e.g., improving the perceptual stage alone when it is the bottleneck for end-to-end performance), and accommodates compressed contexts by running global search at lower image resolutions and allocating high-resolution processing only to selected regions, thereby reducing total visual tokens count and compute. Across challenging visual reasoning benchmarks, SPARC outperforms monolithic baselines and strong visual-grounding approaches. For instance, SPARC improves the accuracy of Qwen3VL-4B on the $V^*$ VQA benchmark by 6.7 percentage points, and it surpasses "thinking with images" by 4.6 points on a challenging OOD task despite requiring a 200$\times$ lower token budget.
Authors: Christof Duhme, Chris Lippe, Verena Hoerr, Xiaoyi Jiang
Abstract: Chemical exchange saturation transfer (CEST) MRI is a non-invasive imaging modality for detecting metabolites. It offers higher resolution and sensitivity compared to conventional magnetic resonance spectroscopy (MRS). However, quantification of CEST data is challenging because the measured signal results from a complex interplay of many physiological variables. Here, we introduce a transformer-based neural network to fit parameters such as metabolite concentrations, exchange and relaxation rates of a physical model derived from Bloch-McConnell equations to in-vitro CEST spectra. We show that our self-supervised trained neural network clearly outperforms the solution of classical gradient-based solver.
Authors: Florian Eilers, Christof Duhme, Xiaoyi Jiang
Abstract: Complex-valued neural networks (CVNNs) are rising in popularity for all kinds of applications. To safely use CVNNs in practice, analyzing their robustness against outliers is crucial. One well known technique to understand the behavior of deep neural networks is to investigate their behavior under adversarial attacks, which can be seen as worst case minimal perturbations. We design Phase Attacks, a kind of attack specifically targeting the phase information of complex-valued inputs. Additionally, we derive complex-valued versions of commonly used adversarial attacks. We show that in some scenarios CVNNs are more robust than RVNNs and that both are very susceptible to phase changes with the Phase Attacks decreasing the model performance more, than equally strong regular attacks, which can attack both phase and magnitude.
Authors: Christof Duhme, Florian Eilers, Xiaoyi Jiang
Abstract: Adversarial attacks against deep neural networks are commonly constructed under $\ell_p$ norm constraints, most often using $p=1$, $p=2$ or $p=\infty$, and potentially regularized for specific demands such as sparsity or smoothness. These choices are typically made without a systematic investigation of how the norm parameter \( p \) influences the structural and perceptual properties of adversarial perturbations. In this work, we study how the choice of \( p \) affects sparsity and smoothness of adversarial attacks generated under \( \ell_p \) norm constraints for values of $p \in [1,2]$. To enable a quantitative analysis, we adopt two established sparsity measures from the literature and introduce three smoothness measures. In particular, we propose a general framework for deriving smoothness measures based on smoothing operations and additionally introduce a smoothness measure based on first-order Taylor approximations. Using these measures, we conduct a comprehensive empirical evaluation across multiple real-world image datasets and a diverse set of model architectures, including both convolutional and transformer-based networks. We show that the choice of $\ell_1$ or $\ell_2$ is suboptimal in most cases and the optimal $p$ value is dependent on the specific task. In our experiments, using $\ell_p$ norms with $p\in [1.3, 1.5]$ yields the best trade-off between sparse and smooth attacks. These findings highlight the importance of principled norm selection when designing and evaluating adversarial attacks.
Authors: Shaowen Wang, Tariq Alkhalifah
Abstract: Weight initialization plays a crucial role in the optimization behavior and convergence efficiency of neural networks. Most existing initialization methods, such as Xavier and Kaiming initializations, rely on random sampling and do not exploit information from the optimization process itself. We propose a simple, yet effective, initialization strategy based on self-supervised pre-training using random noise as the target. Instead of directly training the network from random weights, we first pre-train it to fit random noise, which leads to a structured and non-random parameter configuration. We show that this noise-driven pre-training significantly improves convergence speed in subsequent tasks, without requiring additional data or changes to the network architecture. The proposed method is particularly effective for implicit neural representations (INRs) and Deep Image Prior (DIP)-style networks, which are known to exhibit a strong low-frequency bias during optimization. After noise-based pre-training, the network is able to capture high-frequency components much earlier in training, leading to faster and more stable convergence. Although random noise contains no semantic information, it serves as an effective self-supervised signal (considering its white spectrum nature) for shaping the initialization of neural networks. Overall, this work demonstrates that noise-based pre-training offers a lightweight and general alternative to traditional random initialization, enabling more efficient optimization of deep neural networks.
Authors: Miko{\l}aj Janusz, Adam Wr\'obel, Bartosz Zieli\'nski, Dawid Rymarczyk
Abstract: Prototypical parts-based models offer a "this looks like that" paradigm for intrinsic interpretability, yet they typically struggle with ImageNet-scale generalization and often require computationally expensive backbone finetuning. Furthermore, existing methods frequently suffer from "prototype drift," where learned prototypes lack tangible grounding in the training distribution and change their activation under small perturbations. We present ProtoQuant, a novel architecture that achieves prototype stability and grounded interpretability through latent vector quantization. By constraining prototypes to a discrete learned codebook within the latent space, we ensure they remain faithful representations of the training data without the need to update the backbone. This design allows ProtoQuant to function as an efficient, interpretable head that scales to large-scale datasets. We evaluate ProtoQuant on ImageNet and several fine-grained benchmarks (CUB-200, Cars-196). Our results demonstrate that ProtoQuant achieves competitive classification accuracy while generalizing to ImageNet and comparable interpretability metrics to other prototypical-parts-based methods.
Authors: Robert Hutter, Michael Pradel
Abstract: Software development agents powered by large language models (LLMs) have shown great promise in automating tasks like environment setup, issue solving, and program repair. Unfortunately, understanding and debugging such agents remain challenging due to their complex and dynamic nature. Developers must reason about trajectories of LLM queries, tool calls, and code modifications, but current techniques reveal little of this intermediate process in a comprehensible format. The key insight of this paper is that debugging software development agents shares many similarities with conventional debugging of software programs, yet requires a higher level of abstraction that raises the level from low-level implementation details to high-level agent actions. Drawing on this insight, we introduce AgentStepper, the first interactive debugger for LLM-based software engineering agents. AgentStepper enables developers to inspect, control, and interactively manipulate agent trajectories. AgentStepper represents trajectories as structured conversations among an LLM, the agent program, and tools. It supports breakpoints, stepwise execution, and live editing of prompts and tool invocations, while capturing and displaying intermediate repository-level code changes. Our evaluation applies AgentStepper to three state-of-the-art software development agents, ExecutionAgent, SWE-Agent, and RepairAgent, showing that integrating the approach into existing agents requires minor code changes (39-42 edited lines). Moreover, we report on a user study with twelve participants, indicating that AgentStepper improves the ability of participants to interpret trajectories (64% vs. 67% mean performance) and identify bugs in the agent's implementation (17% vs. 60% success rate), while reducing perceived workload (e.g., frustration reduced from 5.4/7.0 to 2.4/7.0) compared to conventional tools.
Authors: Dominik P. Hofer, David Haag, Rania Islambouli, Jan D. Smeddinck
Abstract: Digital behaviour change systems increasingly rely on repeated, system-initiated messages to support users in everyday contexts. LLMs enable these messages to be personalised consistently across interactions, yet it remains unclear whether such personalisation improves individual messages or instead shapes users' perceptions through patterns of exposure. We explore this question in the context of LLM-generated JITAIs, which are short, context-aware messages delivered at moments deemed appropriate to support behaviour change, using physical activity as an application domain. In a controlled retrospective study, 90 participants evaluated messages generated using four LLM strategies: baseline prompting, few-shot prompting, fine-tuned models, and retrieval augmented generation, each implemented with and without Big Five Personality Traits to produce personality-aligned communication across multiple scenarios. Using ordinal multilevel models with within-between decomposition, we distinguish trial-level effects, whether personality information improves evaluations of individual messages, from person-level exposure effects, whether participants receiving higher proportions of personality-informed messages exhibit systematically different overall perceptions. Results showed no trial-level associations, but participants who received higher proportions of BFPT-informed messages rated the messages as more personalised, appropriate, and reported less negative affect. We use Communication Accommodation Theory for post-hoc analysis. These results suggest that personality-based personalisation in behaviour change systems may operate primarily through aggregate exposure rather than per-message optimisation, with implications for how adaptive systems are designed and evaluated in sustained human-AI interaction. In-situ longitudinal studies are needed to validate these findings in real-world contexts.
Authors: Ariyan Bighashdel, Thiago D. Sim\~ao, Frans A. Oliehoek
Abstract: Multi-agent reinforcement learning (MARL) offers a scalable alternative to exact game-theoretic analysis but suffers from non-stationarity and the need to maintain diverse populations of strategies that capture non-transitive interactions. Policy Space Response Oracles (PSRO) address these issues by iteratively expanding a restricted game with approximate best responses (BRs), yet per-agent BR training makes it prohibitively expensive in many-agent or simulator-expensive settings. We introduce Joint Experience Best Response (JBR), a drop-in modification to PSRO that collects trajectories once under the current meta-strategy profile and reuses this joint dataset to compute BRs for all agents simultaneously. This amortizes environment interaction and improves the sample efficiency of best-response computation. Because JBR converts BR computation into an offline RL problem, we propose three remedies for distribution-shift bias: (i) Conservative JBR with safe policy improvement, (ii) Exploration-Augmented JBR that perturbs data collection and admits theoretical guarantees, and (iii) Hybrid BR that interleaves JBR with periodic independent BR updates. Across benchmark multi-agent environments, Exploration-Augmented JBR achieves the best accuracy-efficiency trade-off, while Hybrid BR attains near-PSRO performance at a fraction of the sample cost. Overall, JBR makes PSRO substantially more practical for large-scale strategic learning while preserving equilibrium robustness.
Authors: Yuancheng Wang, Zhenyu Tang, Yun Wang, Arthur Hinsvark, Yingru Liu, Yinghao Li, Kainan Peng, Junyi Ao, Mingbo Ma, Mike Seltzer, Qing He, Xubo Liu
Abstract: Speech tokenizers are foundational to speech language models, yet existing approaches face two major challenges: (1) balancing trade-offs between encoding semantics for understanding and acoustics for reconstruction, and (2) achieving low bit rates and low token rates. We propose Speech Diffusion Tokenizer (SiTok), a diffusion autoencoder that jointly learns semantic-rich representations through supervised learning and enables high-fidelity audio reconstruction with diffusion. We scale SiTok to 1.6B parameters and train it on 2 million hours of speech. Experiments show that SiTok outperforms strong baselines on understanding, reconstruction and generation tasks, at an extremely low token rate of $12.5$ Hz and a bit-rate of 200 bits-per-second.
Authors: Maj-Annika Tammisto, Faiz Ali Shah, Daniel Rodriguez, Dietmar Pfahl
Abstract: Background: High-level system testing of applications that use data from e-Government services as input requires test data that is real-life-like but where the privacy of personal information is guaranteed. Applications with such strong requirement include information exchange between countries, medicine, banking, etc. This review aims to synthesize the current state-of-the-practice in this domain. Objectives: The objective of this Systematic Review is to identify existing approaches for creating and evolving synthetic test data without using real-life raw data. Methods: We followed well-known methodologies for conducting systematic literature reviews, including the ones from Kitchenham as well as guidelines for analysing the limitations of our review and its threats to validity. Results: A variety of methods and tools exist for creating privacy-preserving test data. Our search found 1,013 publications in IEEE Xplore, ACM Digital Library, and SCOPUS. We extracted data from 75 of those publications and identified 37 approaches that answer our research question partly. A common prerequisite for using these methods and tools is direct access to real-life data for data anonymization or synthetic test data generation. Nine existing synthetic test data generation approaches were identified that were closest to answering our research question. Nevertheless, further work would be needed to add the ability to evolve synthetic test data to the existing approaches. Conclusions: None of the publications really covered our requirements completely, only partially. Synthetic test data evolution is a field that has not received much attention from researchers but needs to be explored in Digital Government Solutions, especially since new legal regulations are being placed in force in many countries.
Authors: Adam Wr\'obel, Siddhartha Gairola, Jacek Tabor, Bernt Schiele, Bartosz Zieli\'nski, Dawid Rymarczyk
Abstract: Vision Transformers (ViTs) have become a dominant architecture in computer vision, yet producing stable and high-resolution attribution maps for these models remains challenging. Architectural components such as patch embeddings and attention routing often introduce structured artifacts in pixel-level explanations, causing many existing methods to rely on coarse patch-level attributions. We introduce DAVE \textit{(\underline{D}istribution-aware \underline{A}ttribution via \underline{V}iT Gradient D\underline{E}composition)}, a mathematically grounded attribution method for ViTs based on a structured decomposition of the input gradient. By exploiting architectural properties of ViTs, DAVE isolates locally equivariant and stable components of the effective input--output mapping. It separates these from architecture-induced artifacts and other sources of instability.
Authors: Gaurish Trivedi, Alakh Sharma, Kartikey Singh Bhandari, Yash Sinha, Pratik Narang, Dhruv Kumar, Jagat Sesh Challa
Abstract: Standard trust-region methods constrain policy updates via Kullback-Leibler (KL) divergence. However, KL controls only an average divergence and does not directly prevent rare, large likelihood-ratio excursions that destabilize training--precisely the failure mode that motivates heuristics such as PPO's clipping. We propose overlap geometry as an alternative trust region, constraining distributional overlap via the Bhattacharyya coefficient (closely related to the Hellinger/Renyi-1/2 geometry). This objective penalizes separation in the ratio tails, yielding tighter control over likelihood-ratio excursions without relying on total variation bounds that can be loose in tail regimes. We derive Bhattacharyya-TRPO (BTRPO) and Bhattacharyya-PPO (BPPO), enforcing overlap constraints via square-root ratio updates: BPPO clips the square-root ratio q = sqrt(r), and BTRPO applies a quadratic Hellinger/Bhattacharyya penalty. Empirically, overlap-based updates improve robustness and aggregate performance as measured by RLiable under matched training budgets, suggesting overlap constraints as a practical, principled alternative to KL for stable policy optimization.
Authors: Kichang Lee, Jaeho Jin, JaeYeon Park, Songkuk Kim, JeongGil Ko
Abstract: Predictive confidence serves as a foundational control signal in mission-critical systems, directly governing risk-aware logic such as escalation, abstention, and conservative fallback. While prior federated learning attacks predominantly target accuracy or implant backdoors, we identify confidence calibration as a distinct attack objective. We present the Temperature Scaling Attack (TSA), a training-time attack that degrades calibration while preserving accuracy. By injecting temperature scaling with learning rate-temperature coupling during local training, malicious updates maintain benign-like optimization behavior, evading accuracy-based monitoring and similarity-based detection. We provide a convergence analysis under non-IID settings, showing that this coupling preserves standard convergence bounds while systematically distorting confidence. Across three benchmarks, TSA substantially shifts calibration (e.g., 145% error increase on CIFAR-100) with <2 accuracy change, and remains effective under robust aggregation and post-hoc calibration defenses. Case studies further show that confidence manipulation can cause up to 7.2x increases in missed critical cases (healthcare) or false alarms (autonomous driving), even when accuracy is unchanged. Overall, our results establish calibration integrity as a critical attack surface in federated learning.
Authors: Ruiqian Nai, Boyuan Zheng, Junming Zhao, Haodong Zhu, Sicong Dai, Zunhao Chen, Yihang Hu, Yingdong Hu, Tong Zhang, Chuan Wen, Yang Gao
Abstract: Current approaches for humanoid whole-body manipulation, primarily relying on teleoperation or visual sim-to-real reinforcement learning, are hindered by hardware logistics and complex reward engineering. Consequently, demonstrated autonomous skills remain limited and are typically restricted to controlled environments. In this paper, we present the Humanoid Manipulation Interface (HuMI), a portable and efficient framework for learning diverse whole-body manipulation tasks across various environments. HuMI enables robot-free data collection by capturing rich whole-body motion using portable hardware. This data drives a hierarchical learning pipeline that translates human motions into dexterous and feasible humanoid skills. Extensive experiments across five whole-body tasks--including kneeling, squatting, tossing, walking, and bimanual manipulation--demonstrate that HuMI achieves a 3x increase in data collection efficiency compared to teleoperation and attains a 70% success rate in unseen environments.
Authors: Zi Yin, Fanhong Li, Shurui Zheng, Jia Liu
Abstract: Developing robotic manipulation policies is iterative and hypothesis-driven: researchers test tactile sensing, gripper geometries, and sensor placements through real-world data collection and training. Yet even minor end-effector changes often require mechanical refitting and system re-integration, slowing iteration. We present RAPID, a full-stack reconfigurable platform designed to reduce this friction. RAPID is built around a tool-free, modular hardware architecture that unifies handheld data collection and robot deployment, and a matching software stack that maintains real-time awareness of the underlying hardware configuration through a driver-level Physical Mask derived from USB events. This modular hardware architecture reduces reconfiguration to seconds and makes systematic multi-modal ablation studies practical, allowing researchers to sweep diverse gripper and sensing configurations without repeated system bring-up. The Physical Mask exposes modality presence as an explicit runtime signal, enabling auto-configuration and graceful degradation under sensor hot-plug events, so policies can continue executing when sensors are physically added or removed. System-centric experiments show that RAPID reduces the setup time for multi-modal configurations by two orders of magnitude compared to traditional workflows and preserves policy execution under runtime sensor hot-unplug events. The hardware designs, drivers, and software stack are open-sourced at https://rapid-kit.github.io/ .
Authors: Boyu Chen, Tai Guo, Weiyu Cui, Yuqing Li, Xingxing Wang, Chuan Shi, Cheng Yang
Abstract: Multimodal retrieval models are becoming increasingly important in scenarios such as food delivery, where rich multimodal features can meet diverse user needs and enable precise retrieval. Mainstream approaches typically employ a dual-tower architecture between queries and items, and perform joint optimization of intra-tower and inter-tower tasks. However, we observe that joint optimization often leads to certain modalities dominating the training process, while other modalities are neglected. In addition, inconsistent training speeds across modalities can easily result in the one-epoch problem. To address these challenges, we propose a staged pretraining strategy, which guides the model to focus on specialized tasks at each stage, enabling it to effectively attend to and utilize multimodal features, and allowing flexible control over the training process at each stage to avoid the one-epoch problem. Furthermore, to better utilize the semantic IDs that compress high-dimensional multimodal embeddings, we design both generative and discriminative tasks to help the model understand the associations between SIDs, queries, and item features, thereby improving overall performance. Extensive experiments on large-scale real-world Meituan data demonstrate that our method achieves improvements of 3.80%, 2.64%, and 2.17% on R@5, R@10, and R@20, and 5.10%, 4.22%, and 2.09% on N@5, N@10, and N@20 compared to mainstream baselines. Online A/B testing on the Meituan platform shows that our approach achieves a 1.12% increase in revenue and a 1.02% increase in click-through rate, validating the effectiveness and superiority of our method in practical applications.
Authors: Bowen Zhang, Meiyi Wang, Harold Soh
Abstract: Post-training improves instruction-following and helpfulness of large language models (LLMs) but often reduces generation diversity, which leads to repetitive outputs in open-ended settings, a phenomenon known as mode collapse. Motivated by evidence that LLM layers play distinct functional roles, we hypothesize that mode collapse can be localized to specific layers and that restoring a carefully chosen range of layers to their pre-trained weights can recover diversity while maintaining high output quality. To validate this hypothesis and decide which layers to restore, we design a proxy task -- Constrained Random Character(CRC) -- with an explicit validity set and a natural diversity objective. Results on CRC reveal a clear diversity-validity trade-off across restoration ranges and identify configurations that increase diversity with minimal quality loss. Based on these findings, we propose Selective Layer Restoration (SLR), a training-free method that restores selected layers in a post-trained model to their pre-trained weights, yielding a hybrid model with the same architecture and parameter count, incurring no additional inference cost. Across three different tasks (creative writing, open-ended question answering, and multi-step reasoning) and three different model families (Llama, Qwen, and Gemma), we find SLR can consistently and substantially improve output diversity while maintaining high output quality.
Authors: Lucie Termignon, Simonas Zilinskas, Hadrien P\'elissier, Aur\'elien Barrot, Nicolas Chesnais, Elie Gavoty
Abstract: Large Language Models (LLMs) often show reduced performance, cultural alignment, and safety robustness in non-English languages, partly because English dominates both pre-training data and human preference alignment datasets. Training methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) require human preference data, which remains scarce and largely non-public for many languages beyond English. To address this gap, we introduce compar:IA, an open-source digital public service developed inside the French government and designed to collect large-scale human preference data from a predominantly French-speaking general audience. The platform uses a blind pairwise comparison interface to capture unconstrained, real-world prompts and user judgments across a diverse set of language models, while maintaining low participation friction and privacy-preserving automated filtering. As of 2026-02-07, compar:IA has collected over 600,000 free-form prompts and 250,000 preference votes, with approximately 89% of the data in French. We release three complementary datasets -- conversations, votes, and reactions -- under open licenses, and present initial analyses, including a French-language model leaderboard and user interaction patterns. Beyond the French context, compar:IA is evolving toward an international digital public good, offering reusable infrastructure for multilingual model training, evaluation, and the study of human-AI interaction.
Authors: Shentong Mo, Lanqing Li
Abstract: Generative models for de novo protein backbone design have achieved remarkable success in creating novel protein structures. However, these diffusion-based approaches remain computationally intensive and slower than desired for large-scale structural exploration. While recent efforts like Proteina have introduced flow-matching to improve sampling efficiency, the potential of tokenization for structural compression and acceleration remains largely unexplored in the protein domain. In this work, we present SaDiT, a novel framework that accelerates protein backbone generation by integrating SaProt Tokenization with a Diffusion Transformer (DiT) architecture. SaDiT leverages a discrete latent space to represent protein geometry, significantly reducing the complexity of the generation process while maintaining theoretical SE(3) equivalence. To further enhance efficiency, we introduce an IPA Token Cache mechanism that optimizes the Invariant Point Attention (IPA) layers by reusing computed token states during iterative sampling. Experimental results demonstrate that SaDiT outperforms state-of-the-art models, including RFDiffusion and Proteina, in both computational speed and structural viability. We evaluate our model across unconditional backbone generation and fold-class conditional generation tasks, where SaDiT shows superior ability to capture complex topological features with high designability.
Authors: Daniil Plyusov, Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii, Alexey Malakhov, Daniil Gavrilov
Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is commonly based on group sampling to estimate advantages and stabilize policy updates. In practice, large group sizes are not feasible due to computational limits, which biases learning toward trajectories that are already likely. Smaller groups often miss rare-correct trajectories while still containing mixed rewards, concentrating probability on common solutions. We derive the probability that updates miss rare-correct modes as a function of group size, showing non-monotonic behavior, and characterize how updates redistribute mass within the correct set, revealing that unsampled-correct mass can shrink even as total correct mass grows. Motivated by this analysis, we propose a difficulty-aware advantage scaling coefficient, inspired by Focal loss, that down-weights updates on high-success prompts. The lightweight modification can be directly integrated into any group-relative RLVR algorithm such as GRPO, DAPO, and CISPO. On Qwen2.5-7B across in-domain and out-of-domain benchmarks, our method improves pass@256 from 64.1 $\rightarrow$ 70.3 (GRPO), 69.3 $\rightarrow$ 72.5 (DAPO), and 73.2 $\rightarrow$ 76.8 (CISPO), while preserving or improving pass@1, without increasing group size or computational cost.
Authors: Zuyao Xu, Yuqi Qiu, Lu Sun, FaSheng Miao, Fubin Wu, Xinyi Wang, Xiang Li, Haozhe Lu, ZhengZe Zhang, Yuxin Hu, Jialu Li, Jin Luo, Feng Zhang, Rui Luo, Xinran Liu, Yingxian Li, Jiaji Liu
Abstract: Citations provide the basis for trusting scientific claims; when they are invalid or fabricated, this trust collapses. With the advent of Large Language Models (LLMs), this risk has intensified: LLMs are increasingly used for academic writing, yet their tendency to fabricate citations (``ghost citations'') poses a systemic threat to citation validity. To quantify this threat and inform mitigation, we develop CiteVerifier, an open-source framework for large-scale citation verification, and conduct the first comprehensive study of citation validity in the LLM era through three experiments built on it. We benchmark 13 state-of-the-art LLMs on citation generation across 40 research domains, finding that all models hallucinate citations at rates from 14.23\% to 94.93\%, with significant variation across research domains. Moreover, we analyze 2.2 million citations from 56,381 papers published at top-tier AI/ML and Security venues (2020--2025), confirming that 1.07\% of papers contain invalid or fabricated citations (604 papers), with an 80.9\% increase in 2025 alone. Furthermore, we survey 97 researchers and analyze 94 valid responses after removing 3 conflicting samples, revealing a critical ``verification gap'': 41.5\% of researchers copy-paste BibTeX without checking and 44.4\% choose no-action responses when encountering suspicious references; meanwhile, 76.7\% of reviewers do not thoroughly check references and 80.0\% never suspect fake citations. Our findings reveal an accelerating crisis where unreliable AI tools, combined with inadequate human verification by researchers and insufficient peer review scrutiny, enable fabricated citations to contaminate the scientific record. We propose interventions for researchers, venues, and tool developers to protect citation integrity.
Authors: Rishabh Jain, Keisuke Okumura, Michael Amir, Pietro Lio, Amanda Prorok
Abstract: Multi-Agent Path Finding (MAPF) is a representative multi-agent coordination problem, where multiple agents are required to navigate to their respective goals without collisions. Solving MAPF optimally is known to be NP-hard, leading to the adoption of learning-based approaches to alleviate the online computational burden. Prevailing approaches, such as Graph Neural Networks (GNNs), are typically constrained to pairwise message passing between agents. However, this limitation leads to suboptimal behaviours and critical issues, such as attention dilution, particularly in dense environments where group (i.e. beyond just two agents) coordination is most critical. Despite the importance of such higher-order interactions, existing approaches have not been able to fully explore them. To address this representational bottleneck, we introduce HMAGAT (Hypergraph Multi-Agent Attention Network), a novel architecture that leverages attentional mechanisms over directed hypergraphs to explicitly capture group dynamics. Empirically, HMAGAT establishes a new state-of-the-art among learning-based MAPF solvers: e.g., despite having just 1M parameters and being trained on 100$\times$ less data, it outperforms the current SoTA 85M parameter model. Through detailed analysis of HMAGAT's attention values, we demonstrate how hypergraph representations mitigate the attention dilution inherent in GNNs and capture complex interactions where pairwise methods fail. Our results illustrate that appropriate inductive biases are often more critical than the training data size or sheer parameter count for multi-agent problems.
Authors: Noah Schwartz, Chandra Kanth Nagesh, Sriram Sankaranarayanan, Ramneet Kaur, Tuhin Sahai, Susmit Jha
Abstract: We present a novel approach for verifying properties of Kolmogorov-Arnold Networks (KANs), a class of neural networks characterized by nonlinear, univariate activation functions typically implemented as piecewise polynomial splines or Gaussian processes. Our method creates mathematical ``abstractions'' by replacing each KAN unit with a piecewise affine (PWA) function, providing both local and global error estimates between the original network and its approximation. These abstractions enable property verification by encoding the problem as a Mixed Integer Linear Program (MILP), determining whether outputs satisfy specified properties when inputs belong to a given set. A critical challenge lies in balancing the number of pieces in the PWA approximation: too many pieces add binary variables that make verification computationally intractable, while too few pieces create excessive error margins that yield uninformative bounds. Our key contribution is a systematic framework that exploits KAN structure to find optimal abstractions. By combining dynamic programming at the unit level with a knapsack optimization across the network, we minimize the total number of pieces while guaranteeing specified error bounds. This approach determines the optimal approximation strategy for each unit while maintaining overall accuracy requirements. Empirical evaluation across multiple KAN benchmarks demonstrates that the upfront analysis costs of our method are justified by superior verification results.
Authors: Argyro Tsandalidou, Konstantinos Dogeas, Eleftheria Tetoula Tsonga, Elisavet Parselia, Georgios Tsimiklis, George Arvanitakis
Abstract: Satellite imagery is employed for large-scale prospectivity mapping due to the high cost and typically limited availability of on-site mineral exploration data. In this work, we present a proof-of-concept framework that leverages generative representations learned from multispectral Sentinel-2 imagery to identify gold-bearing regions from space. An autoencoder foundation model, called Isometric, which is pretrained on the large-scale FalconSpace-S2 v1.0 dataset, produces information-dense spectral-spatial representations that serve as inputs to a lightweight XGBoost classifier. We compare this representation-based approach with a raw spectral input baseline using a dataset of 63 Sentinel-2 images from known gold and non-gold locations. The proposed method improves patch-level accuracy from 0.51 to 0.68 and image-level accuracy from 0.55 to 0.73, demonstrating that generative embeddings capture transferable mineralogical patterns even with limited labeled data. These results highlight the potential of foundation-model representations to make mineral exploration more efficient, scalable, and globally applicable.
Authors: Thibaud Gloaguen, Robin Staab, Nikola Jovanovi\'c, Martin Vechev
Abstract: LLM watermarks allow tracing AI-generated texts by inserting a detectable signal into their generated content. Recent works have proposed a wide range of watermarking algorithms, each with distinct designs, usually built using a bottom-up approach. Crucially, there is no general and principled formulation for LLM watermarking. In this work, we show that most existing and widely used watermarking schemes can in fact be derived from a principled constrained optimization problem. Our formulation unifies existing watermarking methods and explicitly reveals the constraints that each method optimizes. In particular, it highlights an understudied quality-diversity-power trade-off. At the same time, our framework also provides a principled approach for designing novel watermarking schemes tailored to specific requirements. For instance, it allows us to directly use perplexity as a proxy for quality, and derive new schemes that are optimal with respect to this constraint. Our experimental evaluation validates our framework: watermarking schemes derived from a given constraint consistently maximize detection power with respect to that constraint.
Authors: Fengpeng Li, Kemou Li, Qizhou Wang, Bo Han, Jiantao Zhou
Abstract: Concept erasure helps stop diffusion models (DMs) from generating harmful content; but current methods face robustness retention trade off. Robustness means the model fine-tuned by concept erasure methods resists reactivation of erased concepts, even under semantically related prompts. Retention means unrelated concepts are preserved so the model's overall utility stays intact. Both are critical for concept erasure in practice, yet addressing them simultaneously is challenging, as existing works typically improve one factor while sacrificing the other. Prior work typically strengthens one while degrading the other, e.g., mapping a single erased prompt to a fixed safe target leaves class level remnants exploitable by prompt attacks, whereas retention-oriented schemes underperform against adaptive adversaries. This paper introduces Adversarial Erasure with Gradient Informed Synergy (AEGIS), a retention-data-free framework that advances both robustness and retention.
Authors: Yassine Chagna, Antal Goldschmidt
Abstract: This project explores large language models (LLMs) for anomaly detection across heterogeneous log sources. Traditional intrusion detection systems suffer from high false positive rates, semantic blindness, and data scarcity, as logs are inherently sensitive, making clean datasets rare. We address these challenges through three contributions: (1) LogAtlas-Foundation-Sessions and LogAtlas-Defense-Set, balanced and heterogeneous log datasets with explicit attack annotations and privacy preservation; (2) empirical benchmarking revealing why standard metrics such as F1 and accuracy are misleading for security applications; and (3) a two phase training framework combining log understanding (Base-AMAN, 3B parameters) with real time detection (AMAN, 0.5B parameters via knowledge distillation). Results demonstrate practical feasibility, with inference times of 0.3-0.5 seconds per session and operational costs below 50 USD per day.
Authors: Kate Sanders, Nathaniel Weir, Sapana Chaudhary, Kaj Bostrom, Huzefa Rangwala
Abstract: An impediment to using Large Language Models (LLMs) for reasoning output verification is that LLMs struggle to reliably identify errors in thinking traces, particularly in long outputs, domains requiring expert knowledge, and problems without verifiable rewards. We propose a data-driven approach to automatically construct highly granular reasoning error taxonomies to enhance LLM-driven error detection on unseen reasoning traces. Our findings indicate that classification approaches that leverage these error taxonomies, or "rubrics", demonstrate strong error identification compared to baseline methods in technical domains like coding, math, and chemical engineering. These rubrics can be used to build stronger LLM-as-judge reward functions for reasoning model training via reinforcement learning. Experimental results show that these rewards have the potential to improve models' task accuracy on difficult domains over models trained by general LLMs-as-judges by +45%, and approach performance of models trained by verifiable rewards while using as little as 20% as many gold labels. Through our approach, we extend the usage of reward rubrics from assessing qualitative model behavior to assessing quantitative model correctness on tasks typically learned via RLVR rewards. This extension opens the door for teaching models to solve complex technical problems without a full dataset of gold labels, which are often highly costly to procure.
Authors: Sohan Venkatesh, Ashish Mahendran Kurapath
Abstract: Activation steering methods, such as persona vectors, are widely used to control large language model behavior and increasingly interpreted as revealing meaningful internal representations. This interpretation implicitly assumes steering directions are identifiable and uniquely recoverable from input-output behavior. We formalize steering as an intervention on internal representations and prove that, under realistic modeling and data conditions, steering vectors are fundamentally non-identifiable due to large equivalence classes of behaviorally indistinguishable interventions. Empirically, we validate this across multiple models and semantic traits, showing orthogonal perturbations achieve near-equivalent efficacy with negligible effect sizes. However, identifiability is recoverable under structural assumptions including statistical independence, sparsity constraints, multi-environment validation or cross-layer consistency. These findings reveal fundamental interpretability limits and clarify structural assumptions required for reliable safety-critical control.
Authors: Keonyoung Koh, Moonkyeong Jung, Samuel Seungsup Lee, Daehyung Park
Abstract: We address the over-constrained planning problem in semi-static environments. The planning objective is to find a best-effort solution that avoids all hard constraint regions while minimally traversing the least risky areas. Conventional methods often rely on pre-defined area costs, limiting generalizations. Further, the spatial continuity of navigation spaces makes it difficult to identify regions that are passable without overestimation. To overcome these challenges, we propose SuReNav, a superpixel graph-based constraint relaxation and navigation method that imitates human-like safe and efficient navigation. Our framework consists of three components: 1) superpixel graph map generation with regional constraints, 2) regional-constraint relaxation using graph neural network trained on human demonstrations for safe and efficient navigation, and 3) interleaving relaxation, planning, and execution for complete navigation. We evaluate our method against state-of-the-art baselines on 2D semantic maps and 3D maps from OpenStreetMap, achieving the highest human-likeness score of complete navigation while maintaining a balanced trade-off between efficiency and safety. We finally demonstrate its scalability and generalization performance in real-world urban navigation with a quadruped robot, Spot.
Authors: Ahsan Mehmood, Naveed Ul Hassan, Ghassan M. Kraidy
Abstract: This paper investigates the role of large language models (LLMs) in sixth-generation (6G) Internet of Things (IoT) networks and proposes a prompt-engineering-based real-time feedback and verification (PE-RTFV) framework that perform physical-layer's optimization tasks through an iteratively process. By leveraging the naturally available closed-loop feedback inherent in wireless communication systems, PE-RTFV enables real-time physical-layer optimization without requiring model retraining. The proposed framework employs an optimization LLM (O-LLM) to generate task-specific structured prompts, which are provided to an agent LLM (A-LLM) to produce task-specific solutions. Utilizing real-time system feedback, the O-LLM iteratively refines the prompts to guide the A-LLM toward improved solutions in a gradient-descent-like optimization process. We test PE-RTFV approach on wireless-powered IoT testbed case study on user-goal-driven constellation design through semantically solving rate-energy (RE)-region optimization problem which demonstrates that PE-RTFV achieves near-genetic-algorithm performance within only a few iterations, validating its effectiveness for complex physical-layer optimization tasks in resource-constrained IoT networks.
Authors: David Lopez-Ayala, Asier Cabello, Pablo Zinemanas, Emilio Molina, Martin Rocamora
Abstract: AI music generators have advanced to the point where their outputs are often indistinguishable from human compositions. While detection methods have emerged, they are typically designed and validated in music streaming contexts with clean, full-length tracks. Broadcast audio, however, poses a different challenge: music appears as short excerpts, often masked by dominant speech, conditions under which existing detectors fail. In this work, we introduce AI-OpenBMAT, the first dataset tailored to broadcast-style AI-music detection. It contains 3,294 one-minute audio excerpts (54.9 hours) that follow the duration patterns and loudness relations of real television audio, combining human-made production music with stylistically matched continuations generated with Suno v3.5. We benchmark a CNN baseline and state-of-the-art SpectTTTra models to assess SNR and duration robustness, and evaluate on a full broadcast scenario. Across all settings, models that excel in streaming scenarios suffer substantial degradation, with F1-scores dropping below 60% when music is in the background or has a short duration. These results highlight speech masking and short music length as critical open challenges for AI music detection, and position AI-OpenBMAT as a benchmark for developing detectors capable of meeting industrial broadcast requirements.
Authors: Yuming Li, Qingyu Li, Chengyu Bai, Xiangyang Luo, Zeyue Xue, Wenyu Qin, Meng Wang, Yikai Wang, Shanghang Zhang
Abstract: Reinforcement learning from human feedback (RLHF) shows promise for aligning diffusion and flow models, yet policy optimization methods such as GRPO suffer from inefficient and static sampling strategies. These methods treat all prompts and denoising steps uniformly, ignoring substantial variations in sample learning value as well as the dynamic nature of critical exploration moments. To address this issue, we conduct a detailed analysis of the internal attention dynamics during GRPO training and uncover a key insight: attention entropy can serve as a powerful dual-signal proxy. First, across different samples, the relative change in attention entropy ({\Delta}Entropy), which reflects the divergence between the current policy and the base policy, acts as a robust indicator of sample learning value. Second, during the denoising process, the peaks of absolute attention entropy (Entropy(t)), which quantify attention dispersion, effectively identify critical timesteps where high-value exploration occurs. Building on this observation, we propose Adaptive Entropy-Guided Policy Optimization (AEGPO), a novel dual-signal, dual-level adaptive optimization strategy. At the global level, AEGPO uses {\Delta}Entropy to dynamically allocate rollout budgets, prioritizing prompts with higher learning value. At the local level, it exploits the peaks of Entropy(t) to guide exploration selectively at critical high-dispersion timesteps rather than uniformly across all denoising steps. By focusing computation on the most informative samples and the most critical moments, AEGPO enables more efficient and effective policy optimization. Experiments on text-to-image generation tasks demonstrate that AEGPO significantly accelerates convergence and achieves superior alignment performance compared to standard GRPO variants.
Authors: Zhimin Hu, Lanhao Niu, Sashank Varma
Abstract: A central question in cognitive science is whether conceptual representations converge onto a shared manifold to support generalization, or diverge into orthogonal subspaces to minimize task interference. While prior work has discovered evidence for both, a mechanistic account of how these properties coexist and transform across tasks remains elusive. We propose that representational sharing lies not in the concepts themselves, but in the geometric relations between them. Using number concepts as a testbed and language models as high-dimensional computational substrates, we show that number representations preserve a stable relational structure across tasks. Task-specific representations are embedded in distinct subspaces, with low-level features like magnitude and parity encoded along separable linear directions. Crucially, we find that these subspaces are largely transformable into one another via linear mappings, indicating that representations share relational structure despite being located in distinct subspaces. Together, these results provide a mechanistic lens of how language models balance the shared structure of number representation with functional flexibility. It suggests that understanding arises when task-specific transformations are applied to a shared underlying relational structure of conceptual representations.
Authors: Chao Zhou, Tianyi Wei, Yiling Chen, Wenbo Zhou, Nenghai Yu
Abstract: While modern text-to-image models excel at prompt-based generation, they often lack the fine-grained control necessary for specific user requirements like spatial layouts or subject appearances. Multi-condition control addresses this, yet its integration into Diffusion Transformers (DiTs) is bottlenecked by the conventional ``concatenate-and-attend'' strategy, which suffers from quadratic computational and memory overhead as the number of conditions scales. Our analysis reveals that much of this cross-modal interaction is spatially or semantically redundant. To this end, we propose Position-aligned and Keyword-scoped Attention (PKA), a highly efficient framework designed to eliminate these redundancies. Specifically, Position-Aligned Attention (PAA) linearizes spatial control by enforcing localized patch alignment, while Keyword-Scoped Attention (KSA) prunes irrelevant subject-driven interactions via semantic-aware masking. To facilitate efficient learning, we further introduce a Conditional Sensitivity-Aware Sampling (CSAS) strategy that reweights the training objective towards critical denoising phases, drastically accelerating convergence and enhancing conditional fidelity. Empirically, PKA delivers a 10.0$\times$ inference speedup and a 5.1$\times$ VRAM saving, providing a scalable and resource-friendly solution for high-fidelity multi-conditioned generation.
Authors: Jonathan Pan
Abstract: Mechanistic interpretability aims to reverse-engineer the internal computations of Large Language Models (LLMs), yet separating sparse semantic signals from high-dimensional polysemantic noise remains a significant challenge. This paper introduces the Quantum Sieve Tracer, a hybrid quantum-classical framework designed to characterize factual recall circuits. We implement a modular pipeline that first localizes critical layers using classical causal tracing, then maps specific attention head activations into an exponentially large quantum Hilbert space. Using open-weight models (Meta Llama-3.2-1B and Alibaba Qwen2.5-1.5B-Instruct), we perform a two-stage analysis that reveals a fundamental architectural divergence. While Qwen's layer 7 circuit functions as a classic Recall Hub, we discover that Llama's layer 9 acts as an Interference Suppression circuit, where ablating the identified heads paradoxically improves factual recall. Our results demonstrate that quantum kernels can distinguish between these constructive (recall) and reductive (suppression) mechanisms, offering a high-resolution tool for analyzing the fine-grained topology of attention.
Authors: Xinyu Zhao, Qingyun Sun, Jiayi Luo, Xingcheng Fu, Jianxin Li
Abstract: Graph Anomaly Detection (GAD) aims to identify irregular patterns in graph data, and recent works have explored zero-shot generalist GAD to enable generalization to unseen graph datasets. However, existing zero-shot GAD methods largely ignore intrinsic geometric differences across diverse anomaly patterns, substantially limiting their cross-domain generalization. In this work, we reveal that anomaly detectability is highly dependent on the underlying geometric properties and that embedding graphs from different domains into a single static curvature space can distort the structural signatures of anomalies. To address the challenge that a single curvature space cannot capture geometry-dependent graph anomaly patterns, we propose GAD-MoRE, a novel framework for zero-shot Generalizable Graph Anomaly Detection with a Mixture of Riemannian Experts architecture. Specifically, to ensure that each anomaly pattern is modeled in the Riemannian space where it is most detectable, GAD-MoRE employs a set of specialized Riemannian expert networks, each operating in a distinct curvature space. To align raw node features with curvature-specific anomaly characteristics, we introduce an anomaly-aware multi-curvature feature alignment module that projects inputs into parallel Riemannian spaces, enabling the capture of diverse geometric characteristics. Finally, to facilitate better generalization beyond seen patterns, we design a memory-based dynamic router that adaptively assigns each input to the most compatible expert based on historical reconstruction performance on similar anomalies. Extensive experiments in the zero-shot setting demonstrate that GAD-MoRE significantly outperforms state-of-the-art generalist GAD baselines, and even surpasses strong competitors that are few-shot fine-tuned with labeled data from the target domain.
Authors: Jiangping Huang, Wenguang Ye, Weisong Sun, Jian Zhang, Mingyue Zhang, Yang Liu
Abstract: Large Language Models (LLMs) often generate code with subtle but critical bugs, especially for complex tasks. Existing automated repair methods typically rely on superficial pass/fail signals, offering limited visibility into program behavior and hindering precise error localization. In addition, without a way to learn from prior failures, repair processes often fall into repetitive and inefficient cycles. To overcome these challenges, we present TraceCoder, a collaborative multi-agent framework that emulates the observe-analyze-repair process of human experts. The framework first instruments the code with diagnostic probes to capture fine-grained runtime traces, enabling deep insight into its internal execution. It then conducts causal analysis on these traces to accurately identify the root cause of the failure. This process is further enhanced by a novel Historical Lesson Learning Mechanism (HLLM), which distills insights from prior failed repair attempts to inform subsequent correction strategies and prevent recurrence of similar mistakes. To ensure stable convergence, a Rollback Mechanism enforces that each repair iteration constitutes a strict improvement toward the correct solution. Comprehensive experiments across multiple benchmarks show that TraceCoder achieves up to a 34.43\% relative improvement in Pass@1 accuracy over existing advanced baselines. Ablation studies verify the significance of each system component, with the iterative repair process alone contributing a 65.61\% relative gain in accuracy. Furthermore, TraceCoder significantly outperforms leading iterative methods in terms of both accuracy and cost-efficiency.
Authors: Ruchika Chavhan, Malcolm Chadwick, Alberto Gil Couto Pimentel Ramos, Luca Morreale, Mehdi Noroozi, Abhinav Mehrotra
Abstract: While large-scale text-to-image diffusion models continue to improve in visual quality, their increasing scale has widened the gap between state-of-the-art models and on-device solutions. To address this gap, we introduce NanoFLUX, a 2.4B text-to-image flow-matching model distilled from 17B FLUX.1-Schnell using a progressive compression pipeline designed to preserve generation quality. Our contributions include: (1) A model compression strategy driven by pruning redundant components in the diffusion transformer, reducing its size from 12B to 2B; (2) A ResNet-based token downsampling mechanism that reduces latency by allowing intermediate blocks to operate on lower-resolution tokens while preserving high-resolution processing elsewhere; (3) A novel text encoder distillation approach that leverages visual signals from early layers of the denoiser during sampling. Empirically, NanoFLUX generates 512 x 512 images in approximately 2.5 seconds on mobile devices, demonstrating the feasibility of high-quality on-device text-to-image generation.
Authors: Samuel Klein, Willie Neiswanger, Daniel Ratner, Michael Kagan, Sean Gasiorowski
Abstract: Bayesian optimal experimental design (BOED) seeks to maximize the expected information gain (EIG) of experiments. This requires a likelihood estimate, which in many settings is intractable. Simulation-based inference (SBI) provides powerful tools for this regime. However, existing work explicitly connecting SBI and BOED is restricted to a single contrastive EIG bound. We show that the EIG admits multiple formulations which can directly leverage modern SBI density estimators, encompassing neural posterior, likelihood, and ratio estimation. Building on this perspective, we define a novel EIG estimator using neural likelihood estimation. Further, we identify optimization as a key bottleneck of gradient based EIG maximization and show that a simple multi-start parallel gradient ascent procedure can substantially improve reliability and performance. With these innovations, our SBI-based BOED methods are able to match or outperform by up to $22\%$ existing state-of-the-art approaches across standard BOED benchmarks.
Authors: Saad Hossain, Tom Tseng, Punya Syon Pandey, Samanvay Vajpayee, Matthew Kowal, Nayeema Nonta, Samuel Simko, Stephen Casper, Zhijing Jin, Kellin Pelrine, Sirisha Rambhatla
Abstract: As increasingly capable open-weight large language models (LLMs) are deployed, improving their tamper resistance against unsafe modifications, whether accidental or intentional, becomes critical to minimize risks. However, there is no standard approach to evaluate tamper resistance. Varied data sets, metrics, and tampering configurations make it difficult to compare safety, utility, and robustness across different models and defenses. To this end, we introduce TamperBench, the first unified framework to systematically evaluate the tamper resistance of LLMs. TamperBench (i) curates a repository of state-of-the-art weight-space fine-tuning attacks and latent-space representation attacks; (ii) enables realistic adversarial evaluation through systematic hyperparameter sweeps per attack-model pair; and (iii) provides both safety and utility evaluations. TamperBench requires minimal additional code to specify any fine-tuning configuration, alignment-stage defense method, and metric suite while ensuring end-to-end reproducibility. We use TamperBench to evaluate 21 open-weight LLMs, including defense-augmented variants, across nine tampering threats using standardized safety and capability metrics with hyperparameter sweeps per model-attack pair. This yields novel insights, including effects of post-training on tamper resistance, that jailbreak-tuning is typically the most severe attack, and that Triplet emerges as a leading alignment-stage defense. Code is available at: https://github.com/criticalml-uw/TamperBench
Authors: Juan Guti\'errez, Victor Guti\'errez-Garcia, Jos\'e Luis Blanco-Murillo
Abstract: Fully unsupervised segmentation pipelines naively seek the most salient object, should this be present. As a result, most of the methods reported in the literature deliver non-deterministic partitions that are sensitive to initialization, seed order, and threshold heuristics. We propose PANC, a weakly supervised spectral segmentation framework that uses a minimal set of annotated visual tokens to produce stable, controllable, and reproducible object masks. From the TokenCut approach, we augment the token-token affinity graph with a handful of priors coupled to anchor nodes. By manipulating the graph topology, we bias the spectral eigenspace toward partitions that are consistent with the annotations. Our approach preserves the global grouping enforced by dense self-supervised visual features, trading annotated tokens for significant gains in reproducibility, user control, and segmentation quality. Using 5 to 30 annotations per dataset, our training-free method achieves state-of-the-art performance among weakly and unsupervised approaches on standard benchmarks (e.g., DUTS-TE, ECSSD, MS COCO). Contrarily, it excels in domains where dense labels are costly or intra-class differences are subtle. We report strong and reliable results on homogeneous, fine-grained, and texture-limited domains, achieving 96.8% (+14.43% over SotA), 78.0% (+0.2%), and 78.8% (+0.37%) average mean intersection-over-union (mIoU) on CrackForest (CFD), CUB-200-2011, and HAM10000 datasets, respectively. For multi-object benchmarks, the framework showcases explicit, user-controllable semantic segmentation.
Authors: Samir Abdaljalil, Parichit Sharma, Erchin Serpedin, Hasan Kurban
Abstract: Hallucinations in large language models remain a persistent challenge, particularly in multilingual and generative settings where factual consistency is difficult to maintain. While recent models show strong performance on English-centric benchmarks, their behavior across languages, tasks, and hallucination types is not yet well understood. In this work, we introduce Halluverse-M^3, a dataset designed to enable systematic analysis of hallucinations across multiple languages, multiple generation tasks, and multiple hallucination categories. Halluverse-M^3 covers four languages, English, Arabic, Hindi, and Turkish, and supports two generation tasks: question answering and dialogue summarization. The dataset explicitly distinguishes between entity-level, relation-level, and sentence-level hallucinations. Hallucinated outputs are constructed through a controlled editing process and validated by human annotators, ensuring clear alignment between original content and hallucinated generations. Using this dataset, we evaluate a diverse set of contemporary open-source and proprietary language models on fine-grained hallucination detection. Our results show that question answering is consistently easier than dialogue summarization, while sentence-level hallucinations remain challenging even for the strongest models. Performance is highest in English and degrades in lower-resource languages, with Hindi exhibiting the lowest detection accuracy. Overall, Halluverse-M^3 provides a realistic and challenging benchmark for studying hallucinations in multilingual, multi-task settings. We release the dataset to support future research on hallucination detection and mitigation\footnote{https://huggingface.co/datasets/sabdalja/HalluVerse-M3}.
URLs: https://huggingface.co/datasets/sabdalja/HalluVerse-M3
Authors: Ziming Liu, Sophia Sanborn, Surya Ganguli, Andreas Tolias
Abstract: Can general-purpose AI architectures go beyond prediction to discover the physical laws governing the universe? True intelligence relies on "world models" -- causal abstractions that allow an agent to not only predict future states but understand the underlying governing dynamics. While previous "AI Physicist" approaches have successfully recovered such laws, they typically rely on strong, domain-specific priors that effectively "bake in" the physics. Conversely, Vafa et al. recently showed that generic Transformers fail to acquire these world models, achieving high predictive accuracy without capturing the underlying physical laws. We bridge this gap by systematically introducing three minimal inductive biases. We show that ensuring spatial smoothness (by formulating prediction as continuous regression) and stability (by training with noisy contexts to mitigate error accumulation) enables generic Transformers to surpass prior failures and learn a coherent Keplerian world model, successfully fitting ellipses to planetary trajectories. However, true physical insight requires a third bias: temporal locality. By restricting the attention window to the immediate past -- imposing the simple assumption that future states depend only on the local state rather than a complex history -- we force the model to abandon curve-fitting and discover Newtonian force representations. Our results demonstrate that simple architectural choices determine whether an AI becomes a curve-fitter or a physicist, marking a critical step toward automated scientific discovery.
Authors: Ehud Shapiro
Abstract: Grassroots Logic Programs (GLP) is a concurrent logic programming language with variables partitioned into paired \emph{readers} and \emph{writers}, conjuring both linear logic and futures/promises: an assignment is produced at most once via the sole occurrence of a writer (promise) and consumed at most once via the sole occurrence of its paired reader (future), and may contain additional readers and/or writers, enabling the concise expression of rich multidirectional communication modalities. GLP was designed as a language for grassroots platforms -- distributed systems with multiple instances that can operate independently of each other and of any global resource, and can coalesce into ever larger instances -- with its target architecture being smartphones communicating peer-to-peer. The operational semantics of Concurrent (single-agent) GLP and of multiagent GLP (maGLP) were defined via transition systems/multiagent transition systems, respectively. Here, we describe the mathematics developed to facilitate the workstation- and smartphone-based implementations of GLP by AI in Dart. We developed dGLP -- implementation-ready deterministic operational semantics for single-agent GLP -- and proved it correct with respect to the Concurrent GLP operational semantics; dGLP was used by AI as a formal spec, from which it developed a workstation-based implementation of GLP. We developed madGLP -- an implementation-ready multiagent operational semantics for maGLP -- and proved it correct with respect to the maGLP operational semantics; madGLP is deterministic at the agent level (not at the system level due to communication asynchrony), and is being used by AI as a formal spec from which it develops a smartphone-based implementation of maGLP.
Authors: Zuyuan Zhang, Sizhe Tang, Tian Lan
Abstract: Non-Markovian dynamics are commonly found in real-world environments due to long-range dependencies, partial observability, and memory effects. The Bellman equation that is the central pillar of Reinforcement learning (RL) becomes only approximately valid under Non-Markovian. Existing work often focus on practical algorithm designs and offer limited theoretical treatment to address key questions, such as what dynamics are indeed capturable by the Bellman framework and how to inspire new algorithm classes with optimal approximations. In this paper, we present a novel topological viewpoint on temporal-difference (TD) based RL. We show that TD errors can be viewed as 1-cochain in the topological space of state transitions, while Markov dynamics are then interpreted as topological integrability. This novel view enables us to obtain a Hodge-type decomposition of TD errors into an integrable component and a topological residual, through a Bellman-de Rham projection. We further propose HodgeFlow Policy Search (HFPS) by fitting a potential network to minimize the non-integrable projection residual in RL, achieving stability/sensitivity guarantees. In numerical evaluations, HFPS is shown to significantly improve RL performance under non-Markovian.
Authors: Alex McKenzie, Keenan Pepper, Stijn Servaes, Martin Leitgab, Murat Cubuktepe, Mike Vaiana, Diogo de Lucena, Judd Rosenblatt, Michael S. A. Graziano
Abstract: Large language models can resist task-misaligned activation steering during inference, sometimes recovering mid-generation to produce improved responses even when steering remains active. We term this Endogenous Steering Resistance (ESR). Using sparse autoencoder (SAE) latents to steer model activations, we find that Llama-3.3-70B shows substantial ESR, while smaller models from the Llama-3 and Gemma-2 families exhibit the phenomenon less frequently. We identify 26 SAE latents that activate differentially during off-topic content and are causally linked to ESR in Llama-3.3-70B. Zero-ablating these latents reduces the multi-attempt rate by 25%, providing causal evidence for dedicated internal consistency-checking circuits. We demonstrate that ESR can be deliberately enhanced through both prompting and training: meta-prompts instructing the model to self-monitor increase the multi-attempt rate by 4x for Llama-3.3-70B, and fine-tuning on self-correction examples successfully induces ESR-like behavior in smaller models. These findings have dual implications: ESR could protect against adversarial manipulation but might also interfere with beneficial safety interventions that rely on activation steering. Understanding and controlling these resistance mechanisms is important for developing transparent and controllable AI systems. Code is available at github.com/agencyenterprise/endogenous-steering-resistance.
Authors: Duygu Altinok
Abstract: Tokenization is a pivotal design choice for neural language modeling in morphologically rich languages (MRLs) such as Turkish, where productive agglutination challenges both vocabulary efficiency and morphological fidelity. Prior studies have explored tokenizer families and vocabulary sizes but typically (i) vary vocabulary without systematically controlling the tokenizer's training corpus, (ii) provide limited intrinsic diagnostics, and (iii) evaluate a narrow slice of downstream tasks. We present the first comprehensive, principled study of Turkish subword tokenization; a "subwords manifest", that jointly varies vocabulary size and tokenizer training corpus size (data and vocabulary coupling), compares multiple tokenizer families under matched parameter budgets (WordPiece, morphology level, and character baselines), and evaluates across semantic (NLI, STS, sentiment analysis, NER), syntactic (POS, dependency parsing), and morphology-sensitive probes. To explain why tokenizers succeed or fail, we introduce a morphology-aware diagnostic toolkit that goes beyond coarse aggregates to boundary-level micro/macro F1, decoupled lemma atomicity vs. surface boundary hits, over/under-segmentation indices, character/word edit distances (CER/WER), continuation rates, and affix-type coverage and token-level atomicity. Our contributions are fourfold: (i) a systematic investigation of the vocabulary-corpus-success triad; (ii) a unified, morphology-aware evaluation framework linking intrinsic diagnostics to extrinsic outcomes; (iii) controlled comparisons identifying when character-level and morphology-level tokenization pay off; and (iv) an open-source release of evaluation code, tokenizer pipelines, and models. As the first work of its kind, this "subwords manifest" delivers actionable guidance for building effective tokenizers in MRLs and establishes a reproducible foundation for future research.
Authors: Shenyuan Gao, William Liang, Kaiyuan Zheng, Ayaan Malik, Seonghyeon Ye, Sihyun Yu, Wei-Cheng Tseng, Yuzhu Dong, Kaichun Mo, Chen-Hsuan Lin, Qianli Ma, Seungjun Nah, Loic Magne, Jiannan Xiang, Yuqi Xie, Ruijie Zheng, Dantong Niu, You Liang Tan, K. R. Zentner, George Kurian, Suneel Indupuru, Pooya Jannaty, Jinwei Gu, Jun Zhang, Jitendra Malik, Pieter Abbeel, Ming-Yu Liu, Yuke Zhu, Joel Jang, Linxi "Jim" Fan
Abstract: Being able to simulate the outcomes of actions in varied environments will revolutionize the development of generalist agents at scale. However, modeling these world dynamics, especially for dexterous robotics tasks, poses significant challenges due to limited data coverage and scarce action labels. As an endeavor towards this end, we introduce DreamDojo, a foundation world model that learns diverse interactions and dexterous controls from 44k hours of egocentric human videos. Our data mixture represents the largest video dataset to date for world model pretraining, spanning a wide range of daily scenarios with diverse objects and skills. To address the scarcity of action labels, we introduce continuous latent actions as unified proxy actions, enhancing interaction knowledge transfer from unlabeled videos. After post-training on small-scale target robot data, DreamDojo demonstrates a strong understanding of physics and precise action controllability. We also devise a distillation pipeline that accelerates DreamDojo to a real-time speed of 10.81 FPS and further improves context consistency. Our work enables several important applications based on generative world models, including live teleoperation, policy evaluation, and model-based planning. Systematic evaluation on multiple challenging out-of-distribution (OOD) benchmarks verifies the significance of our method for simulating open-world, contact-rich tasks, paving the way for general-purpose robot world models.
Authors: Yuchen Yan, Liang Jiang, Jin Jiang, Shuaicheng Li, Zujie Wen, Zhiqiang Zhang, Jun Zhou, Jian Shao, Yueting Zhuang, Yongliang Shen
Abstract: Large reasoning models achieve strong performance by scaling inference-time chain-of-thought, but this paradigm suffers from quadratic cost, context length limits, and degraded reasoning due to lost-in-the-middle effects. Iterative reasoning mitigates these issues by periodically summarizing intermediate thoughts, yet existing methods rely on supervised learning or fixed heuristics and fail to optimize when to summarize, what to preserve, and how to resume reasoning. We propose InftyThink+, an end-to-end reinforcement learning framework that optimizes the entire iterative reasoning trajectory, building on model-controlled iteration boundaries and explicit summarization. InftyThink+ adopts a two-stage training scheme with supervised cold-start followed by trajectory-level reinforcement learning, enabling the model to learn strategic summarization and continuation decisions. Experiments on DeepSeek-R1-Distill-Qwen-1.5B show that InftyThink+ improves accuracy by 21% on AIME24 and outperforms conventional long chain-of-thought reinforcement learning by a clear margin, while also generalizing better to out-of-distribution benchmarks. Moreover, InftyThink+ significantly reduces inference latency and accelerates reinforcement learning training, demonstrating improved reasoning efficiency alongside stronger performance.
Authors: Grace Luo, Jiahai Feng, Trevor Darrell, Alec Radford, Jacob Steinhardt
Abstract: Existing approaches for analyzing neural network activations, such as PCA and sparse autoencoders, rely on strong structural assumptions. Generative models offer an alternative: they can uncover structure without such assumptions and act as priors that improve intervention fidelity. We explore this direction by training diffusion models on one billion residual stream activations, creating "meta-models" that learn the distribution of a network's internal states. We find that diffusion loss decreases smoothly with compute and reliably predicts downstream utility. In particular, applying the meta-model's learned prior to steering interventions improves fluency, with larger gains as loss decreases. Moreover, the meta-model's neurons increasingly isolate concepts into individual units, with sparse probing scores that scale as loss decreases. These results suggest generative meta-models offer a scalable path toward interpretability without restrictive structural assumptions. Project page: https://generative-latent-prior.github.io.
Authors: Nardine Osman, Mark d'Inverno
Abstract: In the diverse array of work investigating the nature of human values from psychology, philosophy and social sciences, there is a clear consensus that values guide behaviour. More recently, a recognition that values provide a means to engineer ethical AI has emerged. Indeed, Stuart Russell proposed shifting AI's focus away from simply ``intelligence'' towards intelligence ``provably aligned with human values''. This challenge -- the value alignment problem -- with others including an AI's learning of human values, aggregating individual values to groups, and designing computational mechanisms to reason over values, has energised a sustained research effort. Despite this, no formal, computational definition of values has yet been proposed. We address this through a formal conceptual framework rooted in the social sciences, that provides a foundation for the systematic, integrated and interdisciplinary investigation into how human values can support designing ethical AI.
Authors: Yichen Wu, Qianqian Gao, Xudong Pan, Geng Hong, Min Yang
Abstract: As large language models (LLMs) are increasingly deployed as interactive agents, open-ended human-AI interactions can involve deceptive behaviors with serious real-world consequences, yet existing evaluations remain largely scenario-specific and model-centric. We introduce OpenDeception, a lightweight framework for jointly evaluating deception risk from both sides of human-AI dialogue. It consists of a scenario benchmark with 50 real-world deception cases, an IntentNet that infers deceptive intent from agent reasoning, and a TrustNet that estimates user susceptibility. To address data scarcity, we synthesize high-risk dialogues via LLM-based role-and-goal simulation, and train the User Trust Scorer using contrastive learning on controlled response pairs, avoiding unreliable scalar labels. Experiments on 11 LLMs and three large reasoning models show that over 90% of goal-driven interactions in most models exhibit deceptive intent, with stronger models displaying higher risk. A real-world case study adapted from a documented AI-induced suicide incident further demonstrates that our joint evaluation can proactively trigger warnings before critical trust thresholds are reached.
Authors: Paiheng Xu, Gang Wu, Xiang Chen, Tong Yu, Chang Xiao, Franck Dernoncourt, Tianyi Zhou, Wei Ai, Viswanathan Swaminathan
Abstract: Scripting interfaces enable users to automate tasks and customize software workflows, but creating scripts traditionally requires programming expertise and familiarity with specific APIs, posing barriers for many users. While Large Language Models (LLMs) can generate code from natural language queries, runtime code generation is severely limited due to unverified code, security risks, longer response times, and higher computational costs. To bridge the gap, we propose an offline simulation framework to curate a software-specific skillset, a collection of verified scripts, by exploiting LLMs and publicly available scripting guides. Our framework comprises two components: (1) task creation, using top-down functionality guidance and bottom-up API synergy exploration to generate helpful tasks; and (2) skill generation with trials, refining and validating scripts based on execution feedback. To efficiently navigate the extensive API landscape, we introduce a Graph Neural Network (GNN)-based link prediction model to capture API synergy, enabling the generation of skills involving underutilized APIs and expanding the skillset's diversity. Experiments with Adobe Illustrator demonstrate that our framework significantly improves automation success rates, reduces response time, and saves runtime token costs compared to traditional runtime code generation. This is the first attempt to use software scripting interfaces as a testbed for LLM-based systems, highlighting the advantages of leveraging execution feedback in a controlled environment and offering valuable insights into aligning AI capabilities with user needs in specialized software domains.
Authors: Ruichen Zhang, Rana Muhammad Shahroz Khan, Zhen Tan, Dawei Li, Song Wang, Tianlong Chen
Abstract: Data-centric distillation, including data augmentation, selection, and mixing, offers a promising path to creating smaller, more efficient student Large Language Models (LLMs) that retain strong reasoning abilities. However, there still lacks a comprehensive benchmark to systematically assess the effect of each distillation approach. This paper introduces DC-CoT, the first data-centric benchmark that investigates data manipulation in chain-of-thought (CoT) distillation from method, model and data perspectives. Utilizing various teacher models (e.g., o4-mini, Gemini-Pro, Claude-3.5) and student architectures (e.g., 3B, 7B parameters), we rigorously evaluate the impact of these data manipulations on student model performance across multiple reasoning datasets, with a focus on in-distribution (IID) and out-of-distribution (OOD) generalization, and cross-domain transfer. Our findings aim to provide actionable insights and establish best practices for optimizing CoT distillation through data-centric techniques, ultimately facilitating the development of more accessible and capable reasoning models. The codebase can be accessed at https://github.com/UNITES-Lab/Distillation-Bench
Authors: Jinmei Liu, Fuhong Liu, Zhenhong Sun, Jianye Hao, Huaxiong Li, Bo Wang, Daoyi Dong, Chunlin Chen, Zhi Wang
Abstract: Recent advancements in language models have demonstrated remarkable in-context learning abilities, prompting the exploration of in-context reinforcement learning (ICRL) to extend the promise to decision domains. Due to involving more complex dynamics and temporal correlations, existing ICRL approaches may face challenges in learning from suboptimal trajectories and achieving precise in-context inference. In the paper, we propose \textbf{S}calable \textbf{I}n-\textbf{C}ontext \textbf{Q}-\textbf{L}earning (\textbf{S-ICQL}), an innovative framework that harnesses dynamic programming and world modeling to steer ICRL toward efficient reward maximization and task generalization, while retaining the scalability and stability of supervised pretraining. We design a prompt-based multi-head transformer architecture that simultaneously predicts optimal policies and in-context value functions using separate heads. We pretrain a generalized world model to capture task-relevant information, enabling the construction of a compact prompt that facilitates fast and precise in-context inference. During training, we perform iterative policy improvement by fitting a state value function to an upper-expectile of the Q-function, and distill the in-context value functions into policy extraction using advantage-weighted regression. Extensive experiments across a range of discrete and continuous environments show consistent performance gains over various types of baselines, especially when learning from suboptimal data. Our code is available at \textcolor{magenta}{\href{https://github.com/NJU-RL/SICQL}{https://github.com/NJU-RL/SICQL}}.
URLs: https://github.com/NJU-RL/SICQL, https://github.com/NJU-RL/SICQL
Authors: Ziqi Zhu, Tao Hu, Honglong Zhang, Dan Yang, HanGeng Chen, Mengran Zhang, Xilun Chen
Abstract: We present CID-GraphRAG (Conversational Intent-Driven Graph Retrieval Augmented Generation), a novel framework that addresses the limitations of existing dialogue systems in maintaining both contextual coherence and goal-oriented progression in multi-turn customer service conversations. Unlike traditional RAG systems that rely solely on semantic similarity (Conversation RAG) or standard knowledge graphs (GraphRAG), CID-GraphRAG constructs dynamic intent transition graphs from goal achieved historical dialogues and implements a dual-retrieval mechanism that adaptively balances intent-based graph traversal with semantic search. This approach enables the system to simultaneously leverage both conversional intent flow patterns and contextual semantics, significantly improving retrieval quality and response quality. In extensive experiments on real-world customer service dialogues, we employ both automatic metrics and LLM-as-judge assessments, demonstrating that CID-GraphRAG significantly outperforms both semantic-based Conversation RAG and intent-based GraphRAG baselines across all evaluation criteria. Quantitatively, CID-GraphRAG demonstrates substantial improvements over Conversation RAG across automatic metrics, with relative gains of 11% in BLEU, 5% in ROUGE-L, 6% in METEOR, and most notably, a 58% improvement in response quality according to LLM-as-judge evaluations. These results demonstrate that the integration of intent transition structures with semantic retrieval creates a synergistic effect that neither approach achieves independently, establishing CID-GraphRAG as an effective framework for addressing the challenges of maintaining contextual coherence and goal-oriented progression in knowledge-intensive multi-turn dialogues.
Authors: Qi Liu, Can Li, Wanjing Ma
Abstract: Traditional agent-based urban mobility simulations often rely on rigid rulebased systems that struggle to capture the complexity, adaptability, and behavioral diversity inherent in human travel decision making. Inspired by recent advancements in large language models and AI agent technologies, we introduce GATSim, a novel framework that leverages these advancements to simulate urban mobility using generative agents with dedicated cognitive structures. GATSim agents are characterized by diverse socioeconomic profiles, individual lifestyles, and evolving preferences shaped through psychologically informed memory systems and lifelong learning. The main contributions of this work are: 1) a comprehensive architecture that integrates urban mobility foundation model with agent cognitive systems and transport simulation environment; 2) a hierarchical memory designed for efficient retrieval of contextually relevant information, incorporating spatial and temporal associations; 3) planning and reactive mechanisms for modeling adaptive mobility behaviors which integrate a multi-scale reflection process to transform specific travel experiences into generalized behavioral insights. Experiments indicate that generative agents perform competitively with human annotators in role-playing scenarios, while naturally producing realistic macroscopic traffic patterns. The code for the prototype implementation is publicly available at https://github.com/qiliuchn/gatsim.
Authors: Myung Ho Kim
Abstract: We report a structural convergence among four influential theories of mind: Kahneman dual-system theory, Friston predictive processing, Minsky society of mind, and Clark extended mind, emerging unintentionally within a practical AI architecture known as Agentic Flow. Designed to address limitations of large language models LLMs, Agentic Flow comprises five interlocking modules - Retrieval, Cognition, Control, Action, and Memory - organized into a repeatable cognitive loop. Although originally inspired only by Minsky and Clark, subsequent analysis showed that its structure echoes computational motifs from all four theories. This suggests that theoretical convergence may arise from implementation constraints rather than deliberate synthesis. In controlled evaluations, the structured agent achieved 95.8 percent task success compared to 62.3 percent for baseline LLMs, demonstrating stronger constraint adherence and more reproducible reasoning. We characterize this convergence through a broader descriptive meta-architecture called PEACE, highlighting recurring patterns such as predictive modeling, associative recall, and error-sensitive control. Later formalized as the Structured Cognitive Loop (SCL), this abstraction generalizes principles first realized in Agentic Flow as a foundation for behavioral intelligence in LLM-based agents.Rather than asserting theoretical unification, this position paper proposes that intelligent architectures may evolve toward shared structural patterns shaped by practical demands. Agentic Flow thus functions as an implementation instance of the Structured Cognitive Loop, illustrating how a unified cognitive form can emerge not from abstraction, but from the necessities of real-world reasoning.
Authors: Giovanni Briglia, Francesco Fabiano, Stefano Mariani
Abstract: Multi-agent Epistemic Planning (MEP) is an autonomous planning framework for reasoning about both the physical world and the beliefs of agents, with applications in domains where information flow and awareness among agents are critical. The richness of MEP requires states to be represented as Kripke structures, i.e., directed labeled graphs. This representation limits the applicability of existing heuristics, hindering the scalability of epistemic solvers, which must explore an exponential search space without guidance, resulting often in intractability. To address this, we exploit Graph Neural Networks (GNNs) to learn patterns and relational structures within epistemic states, to guide the planning process. GNNs, which naturally capture the graph-like nature of Kripke models, allow us to derive meaningful estimates of state quality -- e.g., the distance from the nearest goal -- by generalizing knowledge obtained from previously solved planning instances. We integrate these predictive heuristics into an epistemic planning pipeline and evaluate them against standard baselines, showing improvements in the scalability of multi-agent epistemic planning.
Authors: Adamo Cerioli, Edward D. Lee, Vito D. P. Servedio
Abstract: Strategic decision-making requires balancing immediate opportunities against long-term objectives: a tension fundamental to competitive environments. We investigate this trade-off in chess by analyzing the dynamics of human and AI gameplay through a network-based metric that quantifies piece-to-piece interactions. Our analysis reveals that elite AI players sustain substantially higher levels of strategic tension for longer durations than top human grandmasters. We find that cumulative tension scales with algorithmic complexity in AI systems and increases linearly with skill level (Elo rating) in human play. Longer time controls are associated with higher tension in human games, reflecting the additional strategic complexity players can manage with more thinking time. The temporal profiles reveal contrasting approaches: highly competitive AI systems tolerate densely interconnected positions that balance offensive and defensive tactics over extended periods, while human players systematically limit tension and game complexity. These differences have broader implications for understanding how artificial and biological systems navigate complex strategic environments and for the deployment of AI in high-stakes competitive scenarios.
Authors: Davi Bastos Costa, Renato Vicente
Abstract: Mafia is a social deduction game where informed mafia compete against uninformed townsfolk. Its asymmetry of information and reliance on theory-of-mind reasoning mirror real-world multi-agent scenarios, making it a useful testbed for evaluating the social intelligence of large language models (LLMs). To support a systematic study, we introduce Mini-Mafia: a simplified four-player variant with one mafioso, one detective and two villagers. We set the mafioso to kill a villager and the detective to investigate the mafioso during the night, reducing the game to a single day phase of discussion and voting. Remarkably, we find that the mafia win-rate $p$ in this three-agent system can be described by a simple theoretical model: $\text{logit}(p) = v \times (m - d)$, where $m$, $d$, and $v$ are intrinsic model parameters representing the mafioso's deception, the villager's detection, and the detective's disclosure capabilities, respectively. This compact analytic description of an interacting triad shows that multi-agent dialogue can be captured by a few latent parameters while still matching empirical outcomes, opening a path to a principled theoretical description of multi-agent LLM systems. Estimating these parameters from LLM gameplay data using Bayesian inference yields the Mini-Mafia Benchmark. Our experiments reveal counterintuitive results, including cases where smaller models significantly outperform larger ones. We also establish human baselines, revealing that LLMs excel at persuasion but lag in simple strategic reasoning for agentic interaction. Beyond benchmarking, Mini-Mafia enables quantitative study of emergent multi-agent dynamics such as name bias and last-speaker advantage, and contributes to AI safety by generating training data for deception detectors.
Authors: Yunghwei Lai, Kaiming Liu, Ziyue Wang, Weizhi Ma, Yang Liu
Abstract: The professionalism of a human doctor in outpatient service depends on two core abilities: the ability to make accurate medical decisions and the medical consultation skill to conduct strategic, empathetic patient inquiry. Existing Large Language Models (LLMs) have achieved remarkable accuracy on medical decision-making benchmarks. However, they often lack the ability to conduct the strategic and empathetic consultation, which is essential for real-world clinical scenarios. To address this gap, we propose Doctor-R1, an AI doctor agent trained to master both of the capabilities by ask high-yield questions and conduct strategic multi-turn inquiry to guide decision-making. Our framework introduces three key components: a multi-agent interactive environment, a two-tiered reward architecture that separately optimizes clinical decision-making and communicative inquiry skills, and an experience repository to ground policy learning in high-quality prior trajectories. We evaluate Doctor-R1 on OpenAI's HealthBench and MAQuE, assessed across multi-facet metrics, such as communication quality, user experience, and task accuracy. Remarkably, Doctor-R1 surpasses state-of-the-art open-source specialized LLMs by a substantial margin with higher parameter efficiency and outperforms powerful proprietary models. Furthermore, human expert evaluations show that Doctor-R1 achieves superior clinical capability and patient-centric performance, demonstrating the effectiveness of the framework.
Authors: Joy Jia Yin Lim, Ye He, Jifan Yu, Xin Cong, Daniel Zhang-Li, Zhiyuan Liu, Huiqin Liu, Lei Hou, Juanzi Li, Bin Xu
Abstract: Personalized Learning Path Planning (PLPP) aims to design adaptive learning paths that align with individual goals. While large language models (LLMs) show potential in personalizing learning experiences, existing approaches often lack mechanisms for goal-aligned planning. We introduce Pxplore, a novel framework for PLPP that integrates a reinforcement-based training paradigm and an LLM-driven educational architecture. We design a structured learner state model and an automated reward function that transforms abstract objectives into computable signals. We train the policy combining supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO), and deploy it within a real-world learning platform. Extensive experiments validate Pxplore's effectiveness in producing coherent, personalized, and goal-driven learning paths. We release our code and dataset at https://github.com/Pxplore/pxplore-algo.
Authors: Zhe Wu, Hongjin Lu, Junliang Xing, Changhao Zhang, Yuxuan Li, Yin Zhu, Yuhao Yang, Yuheng Jing, Kai Li, Kun Shao, Jianye Hao, Jun Wang, Yuanchun Shi
Abstract: Building agents that autonomously operate mobile devices has attracted increasing attention. While Vision-Language Models (VLMs) show promise, most existing approaches rely on direct state-to-action mappings, which lack structured reasoning and planning, and thus generalize poorly to novel tasks or unseen UI layouts. We introduce Hi-Agent, a trainable hierarchical vision-language agent for mobile control, featuring a high-level reasoning model and a low-level action model that are jointly optimized. For efficient training, we reformulate multi-step decision-making as a sequence of single-step subgoals and propose a foresight advantage function, which leverages execution feedback from the low-level model to guide high-level optimization. This design alleviates the path explosion issue encountered by Group Relative Policy Optimization (GRPO) in long-horizon tasks and enables stable, critic-free joint training. Hi-Agent achieves a new State-Of-The-Art (SOTA) 87.9% task success rate on the Android-in-the-Wild (AitW) benchmark, significantly outperforming prior methods across three paradigms: prompt-based (AppAgent: 17.7%), supervised (Filtered BC: 54.5%), and reinforcement learning-based (DigiRL: 71.9%). It also demonstrates competitive zero-shot generalization on the ScreenSpot-v2 benchmark. On the more challenging AndroidWorld benchmark, Hi-Agent also scales effectively with larger backbones, showing strong adaptability in high-complexity mobile control scenarios.
Authors: Kou Misaki, Takuya Akiba
Abstract: We introduce String Seed of Thought (SSoT), a novel prompting method for LLMs that improves Probabilistic Instruction Following (PIF). We define PIF as a task requiring an LLM to select its answer from a predefined set of options, each associated with a specific probability, such that the empirical distribution of the generated answers aligns with the target distribution when prompted multiple times. While LLMs excel at tasks with single, deterministic answers, they often fail at PIF, exhibiting biases problematic for applications requiring non-deterministic behaviors, such as human-behavior simulation, content diversification, and multiplayer games. It also harms the diversity of generated responses, a crucial factor in test-time scaling, by causing the outputs to collapse into a limited set of answers. To address this, we propose SSoT, a simple prompting method that instructs an LLM to first output a random string to generate sufficient entropy. SSoT also instructs the LLM to extract randomness by manipulating this string to derive a final answer, thereby preserving diversity while adhering to specific constraints. We demonstrate that SSoT significantly improves the PIF performance of LLMs, approaching the ideal performance of a pseudo-random number generator. Furthermore, our experiments on NoveltyBench show SSoT's benefits extend beyond closed-set tasks to open-ended tasks by enhancing response diversity.
Authors: Xinyi Lin, Yuyang Zhang, Yuanhang Gan, Juntao Chen, Hao Shen, Yichun He, Lijun Li, Ze Yuan, Shuang Wang, Chaohao Wang, Rui Zhang, Na Li, Jia Liu
Abstract: Scientific experimentation and manufacturing rely on prolonged protocol development and complex, multi-step implementation, which require continuous human expertise for precise execution and decision-making, limiting interpretability and scalability. Here, we introduce human-artificial intelligence (AI) co-embodied intelligence, a new form of physical AI that unites human researchers, agentic AI, and wearable hardware. In this paradigm, humans provide precise execution, while agentic AI contributes contextual reasoning, adaptive planning, and analysis. The wearable interface continuously captures experimentation and manufacturing, facilitating seamless communication between humans and AI. We instantiate this paradigm in a microfabrication cleanroom, leading to the agentic-physical experimentation (APEX) system which understands fabrication procedure with accuracy 51% higher than state-of-the-art multimodal large language models/vision language models (LLMs/VLMs), detects and corrects fabrication errors in real-time, and transfers procedural expertise to novice users. Critically, APEX system enables the co-development of fabrication protocols in cleanrooms, overcoming the incompatibility of elastomeric materials in standard microfabrication processes and enabling previously unattainable fabrication outcomes, as demonstrated by the wafer-scale realization of brain-level soft neural probe capable of single-unit-resolution neural recording. These results establish the human-AI co-embodied intelligence that extends agentic reasoning beyond computation into the physical domain, transforming scientific experimentation and manufacturing into autonomous, traceable, interpretable and scalable processes.
Authors: Jingtong Yue, Ziqi Huang, Zhaoxi Chen, Xintao Wang, Pengfei Wan, Ziwei Liu
Abstract: The landscape of video generation is shifting, from a focus on generating visually appealing clips to building virtual environments that support interaction and maintain physical plausibility. These developments point toward the emergence of video foundation models that function not only as visual generators but also as implicit world models, models that simulate the physical dynamics, agent-environment interactions, and task planning that govern real or imagined worlds. This survey provides a systematic overview of this evolution, conceptualizing modern video foundation models as the combination of two core components: an implicit world model and a video renderer. The world model encodes structured knowledge about the world, including physical laws, interaction dynamics, and agent behavior. It serves as a latent simulation engine that enables coherent visual reasoning, long-term temporal consistency, and goal-driven planning. The video renderer transforms this latent simulation into realistic visual observations, effectively producing videos as a "window" into the simulated world. We trace the progression of video generation through four generations, in which the core capabilities advance step by step, ultimately culminating in a world model, built upon a video generation model, that embodies intrinsic physical plausibility, real-time multimodal interaction, and planning capabilities spanning multiple spatiotemporal scales. For each generation, we define its core characteristics, highlight representative works, and examine their application domains such as robotics, autonomous driving, and interactive gaming. Finally, we discuss open challenges and design principles for next-generation world models, including the role of agent intelligence in shaping and evaluating these systems. An up-to-date list of related works is maintained at this link.
Authors: Myung Ho Kim
Abstract: Large language model agents suffer from fundamental architectural problems: entangled reasoning and execution, memory volatility, and uncontrolled action sequences. We introduce Structured Cognitive Loop (SCL), a modular architecture that explicitly separates agent cognition into five phases: Retrieval, Cognition, Control, Action, and Memory (R-CCAM). Soft Symbolic Control constitutes a dedicated governance layer within SCL, applying symbolic constraints to probabilistic inference while preserving the flexibility of neural reasoning and restoring the explainability and controllability of classical symbolic systems. Through empirical validation on multi-step conditional reasoning tasks, we demonstrate that SCL achieves zero policy violations, eliminates redundant tool calls, and maintains complete decision traceability. These results address critical gaps in existing frameworks such as ReAct, AutoGPT, and memory-augmented approaches. Our contributions are threefold: (1) we situate SCL within the taxonomy of hybrid intelligence, differentiating it from prompt-centric and memory-only approaches; (2) we formally define Soft Symbolic Control and contrast it with neuro-symbolic AI; and (3) we derive three design principles for trustworthy agents: modular decomposition, adaptive symbolic governance, and transparent state management. We provide a complete open-source implementation demonstrating the R-CCAM loop architecture, alongside a live GPT-4o-powered travel planning agent. By connecting expert system principles with modern LLM capabilities, this work offers a practical and theoretically grounded path toward reliable, explainable, and governable AI agents.
Authors: Jovan Pavlovi\'c, Mikl\'os Kr\'esz, L\'aszl\'o Hajdu
Abstract: Despite initial successes and a variety of architectures, retrieval-augmented generation systems still struggle to reliably retrieve and connect the multi-step evidence required for complicated reasoning tasks. Most of the standard RAG frameworks regard all retrieved information as equally reliable, overlooking the varying credibility and interconnected nature of large textual corpora. GraphRAG approaches offer potential improvement to RAG systems by integrating knowledge graphs, which structure information into nodes and edges, capture entity relationships, and enable multi-step logical traversal. However, GraphRAG is not always an ideal solution, as it depends on high-quality graph representations of the corpus. Such representations usually rely on manually curated knowledge graphs, which are costly to construct and update, or on automated graph-construction pipelines that are often unreliable. Moreover, systems following this paradigm typically use large language models to guide graph traversal and evidence retrieval. In this paper, we propose a novel RAG framework that uses a spreading activation algorithm to retrieve information from a corpus of documents connected by an automatically constructed heterogeneous knowledge graph. This approach reduces reliance on semantic knowledge graphs, which are often incomplete due to information loss during information extraction, avoids LLM-guided graph traversal, and improves performance on multi-hop question answering. Experiments show that our method achieves better or comparable performance to several state-of-the-art RAG methods and can be integrated as a plug-and-play module with different iterative RAG pipelines. When combined with chain-of-thought iterative retrieval, it yields up to a 39% absolute improvement in answer correctness over naive RAG, while achieving these results with small open-weight language models.
Authors: Yinxu Tang, Chengsong Huang, Jiaxin Huang, William Yeoh
Abstract: Knowledge Graph Question Answering (KGQA) has largely focused on entity-centric queries that return a single answer entity. However, many real-world questions are inherently relational, aiming to understand how entities are associated rather than which entity satisfies a query. In this work, we introduce relation-centric KGQA, a complementary setting in which the answer is a subgraph that represents the semantic relations among entities. The main challenge lies in the abundance of candidate subgraphs, where trivial or overly common connections often obscure the identification of unique and informative answers. To tackle this, we propose UniRel, a unified modular framework that combines a subgraph retriever with an LLM fine-tuned using reinforcement learning. The framework uses a reward function to prefer compact and specific subgraphs with informative relations and low-degree intermediate entities. Experiments show that UniRel improves connectivity and reward over Prompting baselines and generalizes well to unseen entities and relations. Moreover, UniRel can be applied to conventional entity-centric KGQA, achieving competitive or improved performance in several settings.
Authors: Nawazish Ali, Rachael Shaw, Karl Mason
Abstract: Dairy farming is an energy intensive sector that relies heavily on grid electricity. With increasing renewable energy integration, sustainable energy management has become essential for reducing grid dependence and supporting the United Nations Sustainable Development Goal 7 on affordable and clean energy. However, the intermittent nature of renewables poses challenges in balancing supply and demand in real time. Intelligent load scheduling is therefore crucial to minimize operational costs while maintaining reliability. Reinforcement Learning has shown promise in improving energy efficiency and reducing costs. However, most RL-based scheduling methods assume complete knowledge of future prices or generation, which is unrealistic in dynamic environments. Moreover, standard PPO variants rely on fixed clipping or KL divergence thresholds, often leading to unstable training under variable tariffs. To address these challenges, this study proposes a Deep Reinforcement Learning framework for efficient load scheduling in dairy farms, focusing on battery storage and water heating under realistic operational constraints. The proposed Forecast Aware PPO incorporates short term forecasts of demand and renewable generation using hour of day and month based residual calibration, while the PID KL PPO variant employs a proportional integral derivative controller to regulate KL divergence for stable policy updates adaptively. Trained on real world dairy farm data, the method achieves up to 1% lower electricity cost than PPO, 4.8% than DQN, and 1.5% than SAC. For battery scheduling, PPO reduces grid imports by 13.1%, demonstrating scalability and effectiveness for sustainable energy management in modern dairy farming.
Authors: Ashutosh Hathidara, Julien Yu, Vaishali Senthil, Sebastian Schreiber, Anil Babu Ankisettipalli
Abstract: Large language models (LLMs) are increasingly used as human simulators, both for evaluating conversational systems and for generating fine-tuning data. However, naive "act-as-a-user" prompting often yields verbose, unrealistic utterances, motivating principled evaluation of *user proxy agents*. We present **MirrorBench**, a reproducible and extensible benchmarking framework that evaluates user proxies solely on their ability to produce human-like user utterances across diverse conversational regimes, explicitly decoupled from downstream task success. **MirrorBench** combines three lexical-diversity metrics (**MATTR**, **Yule's~$K$**, and **HD-D**) with three LLM-judge-based metrics (**GTEval**, **Pairwise Indistinguishability**, and **Rubric-and-Reason**), and contextualizes judge scores using Human-Human and Proxy-Proxy calibration controls. Across four public datasets, **MirrorBench** yields variance-aware comparisons and reveals systematic gaps between user proxies and real human users. The framework is [open source](https://github.com/SAP/mirrorbench) and includes a command-line interface for running and managing user-proxy benchmarking experiments.
Authors: Haotian Li, Shijun Yang, Weizhen Qi, Silei Zhao, Rui Hua, Mingzhu Song, Xiaojian Yang, Chao Peng
Abstract: Conventional agent systems often struggle in open-ended environments where task distributions continuously drift and external supervision is scarce. Their reliance on static toolsets or offline training lags behind these dynamics, leaving the system's capability boundaries rigid and unknown. To address this, we propose the In-Situ Self-Evolving paradigm. This approach treats sequential task interactions as a continuous stream of experience, enabling the system to distill short-term execution feedback into long-term, reusable capabilities without access to ground-truth labels. Within this framework, we identify tool evolution as the critical pathway for capability expansion, which provides verifiable, binary feedback signals. Within this framework, we develop Yunjue Agent, a system that iteratively synthesizes, optimizes, and reuses tools to navigate emerging challenges. To optimize evolutionary efficiency, we further introduce a Parallel Batch Evolution strategy. Empirical evaluations across five diverse benchmarks under a zero-start setting demonstrate significant performance gains over proprietary baselines. Additionally, complementary warm-start evaluations confirm that the accumulated general knowledge can be seamlessly transferred to novel domains. Finally, we propose a novel metric to monitor evolution convergence, serving as a function analogous to training loss in conventional optimization. We open-source our codebase, system traces, and evolved tools to facilitate future research in resilient, self-evolving intelligence.
Authors: Lei Wei, Xiao Peng, Jinpeng Ou, Bin Wang
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities in function calling for autonomous agents, yet current mechanisms lack explicit reasoning transparency during parameter generation, particularly for complex functions with interdependent parameters. While existing approaches like chain-of-thought prompting operate at the agent level, they fail to provide fine-grained reasoning guidance for individual function parameters. To address these limitations, we propose Think-Augmented Function Calling (TAFC), a novel framework that enhances function calling accuracy through explicit reasoning at both function and parameter levels. Our method introduces a universal "think" parameter augmentation that enables models to articulate their decision-making process, with dynamic optimization for parameter descriptions to improve reasoning quality. For complex parameters, TAFC automatically triggers granular reasoning based on complexity scoring, ensuring appropriate justification for critical decisions. Additionally, we propose reasoning-guided optimization to align generated reasoning with human expectations. TAFC requires no architectural modifications to existing LLMs while maintaining full API compatibility. Evaluation on ToolBench across proprietary and open-source models demonstrates significant improvements in parameter generation accuracy and reasoning coherence for multi-parameter functions, while providing enhanced interpretability for debugging AI agent behaviors.
Authors: Lei Wei, Xiao Peng, Xu Dong, Niantao Xie, Bin Wang
Abstract: Large language models deployed as autonomous agents face critical memory limitations, lacking selective forgetting mechanisms that lead to either catastrophic forgetting at context boundaries or information overload within them. While human memory naturally balances retention and forgetting through adaptive decay processes, current AI systems employ binary retention strategies that preserve everything or lose it entirely. We propose FadeMem, a biologically-inspired agent memory architecture that incorporates active forgetting mechanisms mirroring human cognitive efficiency. FadeMem implements differential decay rates across a dual-layer memory hierarchy, where retention is governed by adaptive exponential decay functions modulated by semantic relevance, access frequency, and temporal patterns. Through LLM-guided conflict resolution and intelligent memory fusion, our system consolidates related information while allowing irrelevant details to fade. Experiments on Multi-Session Chat, LoCoMo, and LTI-Bench demonstrate superior multi-hop reasoning and retrieval with 45\% storage reduction, validating the effectiveness of biologically-inspired forgetting in agent memory systems.
Authors: Anh Khoa Ngo Ho, Martin Chauvin, Simon Gosset, Philippe Cordier, Boris Gamazaychikov
Abstract: As large language models become integral to agentic artificial intelligence systems, their energy demands during inference may pose significant sustainability challenges. This study investigates whether deploying smaller-scale language models can reduce energy consumption without compromising responsiveness and output quality in a multi-agent, real-world environments. We conduct a comparative analysis across language models of varying scales to quantify trade-offs between efficiency and performance. Results show that smaller open-weights models can lower energy usage while preserving task quality. Building on these findings, we propose practical guidelines for sustainable artificial intelligence design, including optimal batch size configuration and computation resource allocation. These insights offer actionable strategies for developing scalable, environmentally responsible artificial intelligence systems.
Authors: Rauno Arike, Raja Mehta Moreno, Rohan Subramani, Shubhorup Biswas, Francis Rhys Ward
Abstract: Frontier language model agents can exhibit misaligned behaviors, including deception, exploiting reward hacks, and pursuing hidden objectives. To control potentially misaligned agents, we can use LLMs themselves to monitor for misbehavior. In this paper, we study how information access affects LLM monitor performance. While one might expect that monitors perform better when they have access to more of the monitored agents' reasoning and actions, we find that contemporary systems often perform better with less information, a phenomenon we call the less-is-more effect for automated oversight. We demonstrate this phenomenon, and analyze the conditions under which it occurs, in three evaluation environments where agents must conduct sabotage while evading monitors. Leveraging this effect, we introduce extract-and-evaluate (EaE) monitoring -- a new hierarchical approach where one monitor isolates relevant excerpts from the monitored agent's trajectory and another monitor scores these excerpts. In BigCodeBench-Sabotage with a GPT-4.1-mini monitor, EaE improves sabotage detection rates by 16.8 percentage points over the next-best approach. In other settings, EaE either outperforms or is competitive with baselines. In addition, we find that agents unaware of being monitored can be caught much more easily and that monitors scoring longer reasoning traces gain more from information filtering. Lastly, we conduct a cost-performance analysis and find that Gemini and Claude models with monitoring techniques that involve information filtering occupy much of the Pareto frontier.
Authors: Kyle Hamilton, Muhammad Intizar Ali
Abstract: In this document we perform a systematic review of the State-of-the-art in Predictive Maintenance (PdM) over the last five years in industrial settings such as commercial buildings, pharmaceutical facilities, or semi-conductor manufacturing. In general, data-driven methods such as those based on deep learning, exhibit higher accuracy than traditional knowledge-based systems. These systems however, are not without significant limitations. The need for large labeled data sets, a lack of generalizability to new environments (out-of-distribution generalization), and a lack of transparency at inference time are some of the obstacles to adoption in real world environments. In contrast, traditional approaches based on domain expertise in the form of rules, logic or first principles suffer from poor accuracy, many false positives and a need for ongoing expert supervision and manual tuning. While the majority of approaches in recent literature utilize some form of data-driven architecture, there are hybrid systems which also take into account domain specific knowledge. Such hybrid systems have the potential to overcome the weaknesses of either approach on its own while preserving their strengths. We propose taking the hybrid approach even further and integrating deep learning with symbolic logic, or Neuro-symbolic AI, to create more accurate, explainable, interpretable, and robust systems. We describe several neuro-symbolic architectures and examine their strengths and limitations within the PdM domain. We focus specifically on methods which involve the use of sensor data and manually crafted rules as inputs by describing concrete NeSy architectures. In short, this survey outlines the context of modern maintenance, defines key concepts, establishes a generalized framework, reviews current modeling approaches and challenges, and introduces the proposed focus on Neuro-symbolic AI (NESY).
Authors: Xiaoyu Wen, Zhida He, Han Qi, Ziyu Wan, Zhongtian Ma, Ying Wen, Tianhang Zheng, Xingcheng Xu, Chaochao Lu, Qiaosheng Zhang
Abstract: Ensuring robust safety alignment is crucial for Large Language Models (LLMs), yet existing defenses often lag behind evolving adversarial attacks due to their \textbf{reliance on static, pre-collected data distributions}. In this paper, we introduce \textbf{MAGIC}, a novel multi-turn multi-agent reinforcement learning framework that formulates LLM safety alignment as an adversarial asymmetric game. Specifically, an attacker agent learns to iteratively rewrite original queries into deceptive prompts, while a defender agent simultaneously optimizes its policy to recognize and refuse such inputs. This dynamic process triggers a \textbf{co-evolution}, where the attacker's ever-changing strategies continuously uncover long-tail vulnerabilities, driving the defender to generalize to unseen attack patterns. Remarkably, we observe that the attacker, endowed with initial reasoning ability, evolves \textbf{novel, previously unseen combinatorial strategies} through iterative RL training, underscoring our method's substantial potential. Theoretically, we provide insights into a more robust game equilibrium and derive safety guarantees. Extensive experiments validate our framework's effectiveness, demonstrating superior defense success rates without compromising the helpfulness of the model. Our code is available at https://github.com/BattleWen/MAGIC.
Authors: Zhiyu An, Wan Du
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 social choice theory.
Authors: Enyu Zhou, Zhiheng Xi, Long Ma, Zhihao Zhang, Shihan Dou, Zhikai Lei, Guoteng Wang, Rui Zheng, Hang Yan, Tao Gui, Qi Zhang, Xuanjing Huang
Abstract: As Large Language Models increasingly automate complex, long-horizon tasks such as \emph{vibe coding}, a supervision gap has emerged. While models excel at execution, users often struggle to guide them effectively due to insufficient domain expertise, the difficulty of articulating precise intent, and the inability to reliably validate complex outputs. It presents a critical challenge in scalable oversight: enabling humans to responsibly steer AI systems on tasks that surpass their own ability to specify or verify. To tackle this, we propose Scalable Interactive Oversight, a framework that decomposes complex intent into a recursive tree of manageable decisions to amplify human supervision. Rather than relying on open-ended prompting, our system elicits low-burden feedback at each node and recursively aggregates these signals into precise global guidance. Validated in web development task, our framework enables non-experts to produce expert-level Product Requirement Documents, achieving a 54\% improvement in alignment. Crucially, we demonstrate that this framework can be optimized via Reinforcement Learning using only online user feedback, offering a practical pathway for maintaining human control as AI scales.
Authors: Haosen Ge, Hamsa Bastani, Osbert Bastani
Abstract: Rapidly increasing AI capabilities have substantial real-world consequences, ranging from AI safety concerns to labor market consequences. The Model Evaluation & Threat Research (METR) report argues that AI capabilities have exhibited exponential growth since 2019. In this note, we argue that the data does not support exponential growth, even in shorter-term horizons. Whereas the METR study claims that fitting sigmoid/logistic curves results in inflection points far in the future, we fit a sigmoid curve to their current data and find that the inflection point has already passed. In addition, we propose a more complex model that decomposes AI capabilities into base and reasoning capabilities, exhibiting individual rates of improvement. We prove that this model supports our hypothesis that AI capabilities will exhibit an inflection point in the near future. Our goal is not to establish a rigorous forecast of our own, but to highlight the fragility of existing forecasts of exponential growth.
Authors: Zhanli Li, Huiwen Tian, Lvzhou Luo, Yixuan Cao, Ping Luo
Abstract: With the rapid progress of tool-using and agentic large language models (LLMs), Retrieval-Augmented Generation (RAG) is evolving from one-shot, passive retrieval into multi-turn, decision-driven evidence acquisition. Despite strong results in open-domain settings, existing agentic search frameworks commonly treat long documents as flat collections of chunks, underutilizing document-native priors such as hierarchical organization and sequential discourse structure. We introduce DeepRead, a structure-aware, multi-turn document reasoning agent that explicitly operationalizes these priors for long-document question answering. DeepRead leverages LLM-based OCR model to convert PDFs into structured Markdown that preserves headings and paragraph boundaries. It then indexes documents at the paragraph level and assigns each paragraph a coordinate-style metadata key encoding its section identity and in-section order. Building on this representation, DeepRead equips the LLM with two complementary tools: a Retrieve tool that localizes relevant paragraphs while exposing their structural coordinates (with lightweight scanning context), and a ReadSection tool that enables contiguous, order-preserving reading within a specified section and paragraph range. Our experiments demonstrate that DeepRead achieves significant improvements over Search-o1-style agentic search in document question answering. The synergistic effect between retrieval and reading tools is also validated. Our fine-grained behavioral analysis reveals a reading and reasoning paradigm resembling human-like ``locate then read'' behavior.
Authors: Kate H. Bentley, Luca Belli, Adam M. Chekroud, Emily J. Ward, Emily R. Dworkin, Emily Van Ark, Kelly M. Johnston, Will Alexander, Millard Brown, Matt Hawrilenko
Abstract: Millions now use generative AI chatbots for psychological support. Despite the promise related to availability and scale, the single most pressing question in AI for mental health is whether these tools are safe. The Validation of Ethical and Responsible AI in Mental Health (VERA-MH) evaluation was recently proposed to meet the urgent need for an evidence-based, automated safety benchmark. This study aimed to examine the clinical validity and reliability of VERA-MH for evaluating AI safety in suicide risk detection and response. We first simulated a large set of conversations between large language model (LLM)-based users (user-agents) and general-purpose AI chatbots. Licensed mental health clinicians used a rubric (scoring guide) to independently rate the simulated conversations for safe and unsafe chatbot behaviors, as well as user-agent realism. An LLM-based judge used the same scoring rubric to evaluate the same set of simulated conversations. We then examined rating alignment (a) among individual clinicians and (b) between clinician consensus and the LLM judge, and (c) summarized clinicians' ratings of user-agent realism. Individual clinicians were generally consistent with one another in their safety ratings (chance-corrected inter-rater reliability [IRR] = 0.77), establishing a gold-standard clinical reference. The LLM judge was strongly aligned with this clinical consensus overall (IRR = 0.81) and within key conditions. Together, findings from this human evaluation study support the validity and reliability of VERA-MH: an open-source, automated AI safety evaluation for mental health. Future research will examine the generalizability and robustness of VERA-MH and expand the framework to target additional key areas of AI safety in mental health.
Authors: Ruijie Shi, Houbin Zhang, Yuecheng Han, Yuheng Wang, Jingru Fan, Runde Yang, Yufan Dang, Huatao Li, Dewen Liu, Yuan Cheng, Chen Qian
Abstract: Large Language Models have shown strong capabilities in complex problem solving, yet many agentic systems remain difficult to interpret and control due to opaque internal workflows. While some frameworks offer explicit architectures for collaboration, many deployed agentic systems operate as black boxes to users. We address this by introducing Agentic Workflow Reconstruction (AWR), a new task aiming to synthesize an explicit, interpretable stand-in workflow that approximates a black-box system using only input--output access. We propose AgentXRay, a search-based framework that formulates AWR as a combinatorial optimization problem over discrete agent roles and tool invocations in a chain-structured workflow space. Unlike model distillation, AgentXRay produces editable white-box workflows that match target outputs under an observable, output-based proxy metric, without accessing model parameters. To navigate the vast search space, AgentXRay employs Monte Carlo Tree Search enhanced by a scoring-based Red-Black Pruning mechanism, which dynamically integrates proxy quality with search depth. Experiments across diverse domains demonstrate that AgentXRay achieves higher proxy similarity and reduces token consumption compared to unpruned search, enabling deeper workflow exploration under fixed iteration budgets.
Authors: Weijian Yu, Yuhuan Lu, Dingqi Yang
Abstract: Knowledge graphs (KGs) have become a key ingredient supporting a variety of applications. Beyond the traditional triplet representation of facts where a relation connects two entities, modern KGs observe an increasing number of hyper-relational facts, where an arbitrary number of qualifiers associated with a triplet provide auxiliary information to further describe the rich semantics of the triplet, which can effectively boost the reasoning performance in link prediction tasks. However, existing link prediction techniques over such hyper-relational KGs (HKGs) mostly focus on a transductive setting, where KG embedding models are learned from the specific vocabulary of a given KG and subsequently can only make predictions within the same vocabulary, limiting their generalizability to previously unseen vocabularies. Against this background, we propose THOR, an inducTive link prediction technique for Hyper-relational knOwledge gRaphs. Specifically, we first introduce both relation and entity foundation graphs, modeling their fundamental inter- and intra-fact interactions in HKGs, which are agnostic to any specific relations and entities. Afterward, THOR is designed to learn from the two foundation graphs with two parallel graph encoders followed by a transformer decoder, which supports efficient masked training and fully-inductive inference. We conduct a thorough evaluation of THOR in hyper-relational link prediction tasks on 12 datasets with different settings. Results show that THOR outperforms a sizable collection of baselines, yielding 66.1%, 55.9%, and 20.4% improvement over the best-performing rule-based, semi-inductive, and fully-inductive techniques, respectively. A series of ablation studies also reveals our key design factors capturing the structural invariance transferable across HKGs for inductive tasks.
Authors: Jun Lu
Abstract: The sole aim of this book is to give a self-contained introduction to concepts and mathematical tools in Bayesian matrix decomposition in order to seamlessly introduce matrix decomposition techniques and their applications in subsequent sections. However, we clearly realize our inability to cover all the useful and interesting results concerning Bayesian matrix decomposition and given the paucity of scope to present this discussion, e.g., the separated analysis of variational inference for conducting the optimization. We refer the reader to literature in the field of Bayesian analysis for a more detailed introduction to the related fields. This book is primarily a summary of purpose, significance of important Bayesian matrix decomposition methods, e.g., real-valued decomposition, nonnegative matrix factorization, Bayesian interpolative decomposition, and the origin and complexity of the methods which shed light on their applications. The mathematical prerequisite is a first course in statistics and linear algebra. Other than this modest background, the development is self-contained, with rigorous proof provided throughout.
Authors: Zhenxi Song, Ruihan Qin, Huixia Ren, Zhen Liang, Yi Guo, Min Zhang, Zhiguo Zhang
Abstract: Cross-center data heterogeneity and annotation unreliability significantly challenge the intelligent diagnosis of diseases using brain signals. A notable example is the EEG-based diagnosis of neurodegenerative diseases, which features subtler abnormal neural dynamics typically observed in small-group settings. To advance this area, in this work, we introduce a transferable framework employing Manifold Attention and Confidence Stratification (MACS) to diagnose neurodegenerative disorders based on EEG signals sourced from four centers with unreliable annotations. The MACS framework's effectiveness stems from these features: 1) The Augmentor generates various EEG-represented brain variants to enrich the data space; 2) The Switcher enhances the feature space for trusted samples and reduces overfitting on incorrectly labeled samples; 3) The Encoder uses the Riemannian manifold and Euclidean metrics to capture spatiotemporal variations and dynamic synchronization in EEG; 4) The Projector, equipped with dual heads, monitors consistency across multiple brain variants and ensures diagnostic accuracy; 5) The Stratifier adaptively stratifies learned samples by confidence levels throughout the training process; 6) Forward and backpropagation in MACS are constrained by confidence stratification to stabilize the learning system amid unreliable annotations. Our subject-independent experiments, conducted on both neurocognitive and movement disorders using cross-center corpora, have demonstrated superior performance compared to existing related algorithms. This work not only improves EEG-based diagnostics for cross-center and small-setting brain diseases but also offers insights into extending MACS techniques to other data analyses, tackling data heterogeneity and annotation unreliability in multimedia and multimodal content understanding.
Authors: Menglin Yang, Ram Samarth B B, Aosong Feng, Bo Xiong, Jihong Liu, Irwin King, Rex Ying
Abstract: Large language models (LLMs) have demonstrated remarkable performance across various tasks. However, it remains an open question whether the default Euclidean space is the most suitable choice for LLMs. In this study, we investigate the geometric characteristics of LLMs, focusing specifically on tokens and their embeddings. Our findings reveal that token frequency follows a power-law distribution, where high-frequency tokens (e.g., the, that ) constitute the minority, while low-frequency tokens (e.g., apple, dog) constitute the majority. Furthermore, high-frequency tokens cluster near the origin, whereas low-frequency tokens are positioned farther away in the embedding space. Additionally, token embeddings exhibit hyperbolic characteristics, indicating a latent tree-like structure within the embedding space. Motivated by these observations, we propose HypLoRA, an efficient fine-tuning approach that operates in hyperbolic space to exploit these underlying hierarchical structures better. HypLoRA performs low-rank adaptation directly in hyperbolic space, thereby preserving hyperbolic modeling capabilities throughout the fine-tuning process. Extensive experiments across various base models and reasoning benchmarks, specifically arithmetic and commonsense reasoning tasks, demonstrate that HypLoRA substantially improves LLM performance.
Authors: Joshua Tint, Som Sagar, Aditya Taparia, Kelly Raines, Bimsara Pathiraja, Caleb Liu, Ransalu Senanayake
Abstract: Human communication is often implicit, conveying tone, identity, and intent beyond literal meanings. While large language models have achieved strong performance on explicit tasks such as summarization and reasoning, their capacity for expressivity, or implicit communication, remains underexplored. We introduce \textbf{ExpressivityBench}, a framework for evaluating the expressivity of LLMs using information-theoretic communication models. Our approach quantifies how well LLM-generated text communicates target properties without explicit mention, across nine tasks spanning emotion, identity, and tone. To enable scalable and reproducible evaluation, we employ LLM-based graders validated against human judgments. Our results reveal that while models are adept at expressing affective content, they struggle with sociolinguistic signals, lagging behind human baselines. This study provides a necessary step to evaluate human-like implicit communication, with implications for applications such as education, mental health support, and socially-aware dialogue systems. We provide code and data for our benchmark alongside our paper.
Authors: Philipp Bach, Victor Chernozhukov, Sven Klaassen, Martin Spindler, Jan Teichert-Kluge, Suhas Vijaykumar
Abstract: This paper advances empirical demand analysis by integrating multimodal product representations derived from artificial intelligence (AI). Using a detailed dataset of toy cars on textit{Amazon.com}, we combine text descriptions, images, and tabular covariates to represent each product using transformer-based embedding models. These embeddings capture nuanced attributes, such as quality, branding, and visual characteristics, that traditional methods often struggle to summarize. Moreover, we fine-tune these embeddings for causal inference tasks. We show that the resulting embeddings substantially improve the predictive accuracy of sales ranks and prices and that they lead to more credible causal estimates of price elasticity. Notably, we uncover strong heterogeneity in price elasticity driven by these product-specific features. Our findings illustrate that AI-driven representations can enrich and modernize empirical demand analysis. The insights generated may also prove valuable for applied causal inference more broadly.
Authors: Lanpei Li, Jack Bell, Massimo Coppola, Vincenzo Lomonaco
Abstract: In the Cloud-Edge Continuum, dynamic infrastructure change and variable workloads complicate efficient resource management. Centralized methods can struggle to adapt, whilst purely decentralized policies lack global oversight. This paper proposes a hybrid framework using Graph Neural Network (GNN) embeddings and collaborative multi-agent reinforcement learning (MARL). Local agents handle neighbourhood-level decisions, and a global orchestrator coordinates system-wide. This work contributes to decentralized application placement strategies with centralized oversight, GNN integration and collaborative MARL for efficient, adaptive and scalable resource management.
Authors: Wenna Lai, Haoran Xie, Guandong Xu, Qing Li
Abstract: Aspect-based sentiment analysis (ABSA) aims to identify four sentiment elements, including aspect term, aspect category, opinion term, and sentiment polarity. These elements construct a complete picture of sentiments. The most challenging task, aspect sentiment quad prediction (ASQP), requires predicting all four elements simultaneously and is hindered by the difficulty of accurately modeling dependencies among sentiment elements. A key challenge lies in the scarcity of annotated data, which limits the model ability to understand and reason about the relational dependencies required for effective quad prediction. To address this challenge, we propose a stepwise task augmentation framework with relation learning that decomposes ASQP into a sequence of auxiliary subtasks with increasing relational granularity. Specifically, STAR incrementally constructs auxiliary data by augmenting the training data with pairwise and overall relation tasks, enabling the model to capture and compose sentiment dependencies in a stepwise manner. This stepwise formulation provides effective relational learning signals that enhance quad prediction performance, particularly in low-resource scenarios. Extensive experiments across four benchmark datasets demonstrate that STAR consistently outperforms existing methods, achieving average F1 improvements of over $2\%$ under low-resource conditions.
Authors: Mathieu Calvat, Chris Bean, Dhruv Anjaria, Hyoungryul Park, Haoren Wang, Kenneth Vecchio, J. C. Stinville
Abstract: To leverage advancements in machine learning for metallic materials design and property prediction, it is crucial to develop a data-reduced representation of metal microstructures that surpasses the limitations of current physics-based discrete microstructure descriptors. This need is particularly relevant for metallic materials processed through additive manufacturing, which exhibit complex hierarchical microstructures that cannot be adequately described using the conventional metrics typically applied to wrought materials. Furthermore, capturing the spatial heterogeneity of microstructures at the different scales is necessary within such framework to accurately predict their properties. To address these challenges, we propose the physical spatial mapping of metal diffraction latent space features. This approach integrates (i) point diffraction data encoding via variational autoencoders or contrastive learning and (ii) the physical mapping of the encoded values. Together these steps offer a method offers a novel means to comprehensively describe metal microstructures. We demonstrate this approach on a wrought and additively manufactured alloy, showing that it effectively encodes microstructural information and enables direct identification of microstructural heterogeneity not directly possible by physics-based models. This data-reduced microstructure representation opens the application of machine learning models in accelerating metallic material design and accurately predicting their properties.
Authors: Haoran Sun, Bingyang Wang, Suyang Yu, Yijiang Li, Qingying Gao, Haiyun Lyu, Lianyu Huang, Zelong Hong, Jiahui Ge, Qianli Ma, Hang He, Yifan Zhou, Lingzi Guo, Lantao Mei, Maijunxian Wang, Dezhi Luo, Hokin Deng
Abstract: Perceptual constancy is the ability to maintain stable perceptions of objects despite changes in sensory input, such as variations in distance, angle, or lighting. This ability is crucial for visual understanding in a dynamic world. Here, we explored such ability in current Vision Language Models (VLMs). In this study, we evaluated 155 VLMs using 236 experiments across three domains: color, size, and shape constancy. The experiments included single-image and video adaptations of classic cognitive tasks, along with novel tasks in in-the-wild conditions. We found significant variability in VLM performance across these domains, with model performance in shape constancy clearly dissociated from that of color and size constancy.
Authors: Ziwei Chen, Jiawen Shen, Luna, Hanyu Zhang, Kristen Vaccaro
Abstract: Recent work has shown that front-end code generated by Large Language Models (LLMs) can embed deceptive designs. To assess the magnitude of this problem, identify the factors that influence deceptive design production, and test strategies for reducing deceptive designs, we carried out two studies which generated and analyzed 1,296 LLM-generated web components, along with a design rationale for each. The first study tested four LLMs for 15 common ecommerce components. Overall 55.8% of components contained at least one deceptive design, and 30.6% contained two or more. Occurence varied significantly across models, with DeepSeek-V3 producing the fewest. Interface interference emerged as the dominant strategy, using color psychology to influence actions and hiding essential information. The first study found that prompts emphasizing business interests (e.g., increasing sales) significantly increased deceptive designs, so a second study tested a variety of prompting strategies to reduce their frequency, finding a values-centered approach the most effective. Our findings highlight risks in using LLMs for coding and offer recommendations for LLM developers and providers.
Authors: Jathurshan Pradeepkumar, Xihao Piao, Zheng Chen, Jimeng Sun
Abstract: Foundation models are reshaping EEG analysis, yet an important problem of EEG tokenization remains a challenge. This paper presents TFM-Tokenizer, a novel tokenization framework that learns a vocabulary of time-frequency motifs from single-channel EEG signals and encodes them into discrete tokens. We propose a dual-path architecture with time-frequency masking to capture robust motif representations, and it is model-agnostic, supporting both lightweight transformers and existing foundation models for downstream tasks. Our study demonstrates three key benefits: Accuracy: Experiments on four diverse EEG benchmarks demonstrate consistent performance gains across both single- and multi-dataset pretraining settings, achieving up to $11\%$ improvement in Cohen's Kappa over strong baselines. Generalization: Moreover, as a plug-and-play component, it consistently boosts the performance of diverse foundation models, including BIOT and LaBraM. Scalability: By operating at the single-channel level rather than relying on the strict 10-20 EEG system, our method has the potential to be device-agnostic. Experiments on ear-EEG sleep staging, which differs from the pretraining data in signal format, channel configuration, recording device, and task, show that our tokenizer outperforms baselines by $14\%$. A comprehensive token analysis reveals strong class-discriminative, frequency-aware, and consistent structure, enabling improved representation quality and interpretability. Code is available at https://github.com/Jathurshan0330/TFM-Tokenizer.
Authors: Maciej Chrab\k{a}szcz, Filip Szatkowski, Bartosz W\'ojcik, Jan Dubi\'nski, Tomasz Trzci\'nski, Sebastian Cygert
Abstract: Although modern LLMs are aligned with human values during post-training, robust moderation remains essential to prevent harmful outputs at deployment time. Existing approaches suffer from performance-efficiency trade-offs and are difficult to customize to user-specific requirements. Motivated by this gap, we introduce Multi-Layer Prototype Moderator (MLPM), a lightweight and highly customizable input moderation tool. We propose leveraging prototypes of intermediate representations across multiple layers to improve moderation quality while maintaining high efficiency. By design, our method adds negligible overhead to the generation pipeline and can be seamlessly applied to any model. MLPM achieves state-of-the-art performance on diverse moderation benchmarks and demonstrates strong scalability across model families of various sizes. Moreover, we show that it integrates smoothly into end-to-end moderation pipelines and further improves response safety when combined with output moderation techniques. Overall, our work provides a practical and adaptable solution for safe, robust, and efficient LLM deployment.
Authors: Sungduk Yu, Man Luo, Avinash Madasu, Vasudev Lal, Phillip Howard
Abstract: Peer review is a critical process for ensuring the integrity of published scientific research. Confidence in this process is predicated on the assumption that experts in the relevant domain give careful consideration to the merits of manuscripts which are submitted for publication. With the recent rapid advancements in large language models (LLMs), a new risk to the peer review process is that negligent reviewers will rely on LLMs to perform the often time consuming process of reviewing a paper. However, there is a lack of existing resources for benchmarking the detectability of AI text in the domain of peer review. To address this deficiency, we introduce a comprehensive dataset containing a total of 788,984 AI-written peer reviews paired with corresponding human reviews, covering 8 years of papers submitted to each of two leading AI research conferences (ICLR and NeurIPS). We use this new resource to evaluate the ability of 18 existing AI text detection algorithms to distinguish between peer reviews fully written by humans and different state-of-the-art LLMs. Additionally, we explore a context-aware detection method called Anchor, which leverages manuscript content to detect AI-generated reviews, and analyze the sensitivity of detection models to LLM-assisted editing of human-written text. Our work reveals the difficulty of identifying AI-generated text at the individual peer review level, highlighting the urgent need for new tools and methods to detect this unethical use of generative AI. Our dataset is publicly available at: https://huggingface.co/datasets/IntelLabs/AI-Peer-Review-Detection-Benchmark.
URLs: https://huggingface.co/datasets/IntelLabs/AI-Peer-Review-Detection-Benchmark.
Authors: Jensen Gao, Suneel Belkhale, Sudeep Dasari, Ashwin Balakrishna, Dhruv Shah, Dorsa Sadigh
Abstract: Machine learning for robot manipulation promises to unlock generalization to novel tasks and environments. But how should we measure the progress of these policies towards generalization? Evaluating and quantifying generalization is the Wild West of modern robotics, with each work proposing and measuring different types of generalization in their own, often difficult to reproduce settings. In this work, our goal is (1) to outline the forms of generalization we believe are important for robot manipulation in a comprehensive and fine-grained manner, and (2) to provide reproducible guidelines for measuring these notions of generalization. We first propose STAR-Gen, a taxonomy of generalization for robot manipulation structured around visual, semantic, and behavioral generalization. Next, we instantiate STAR-Gen with two case studies on real-world benchmarking: one based on open-source models and the Bridge V2 dataset, and another based on the bimanual ALOHA 2 platform that covers more dexterous and longer horizon tasks. Our case studies reveal many interesting insights: for example, we observe that open-source vision-language-action models often struggle with semantic generalization, despite pre-training on internet-scale language datasets. We provide videos and other supplementary material at stargen-taxonomy.github.io.
Authors: Reshabh K Sharma, Jonathan De Halleux, Shraddha Barke, Dan Grossman, Benjamin Zorn
Abstract: Large language models (LLMs) are being used in many applications and prompts for these models are integrated into software applications as code-like artifacts. These prompts behave much like traditional software in that they take inputs, generate outputs, and perform some specific function. However, prompts differ from traditional code in many ways and require new approaches to ensure that they are robust. For example, unlike traditional software the output of a prompt depends on the AI model that interprets it. Also, while natural language prompts are easy to modify, the impact of updates is harder to predict. New approaches to testing, debugging, and modifying prompts with respect to the model running them are required. To address some of these issues, we developed PromptPex, an LLM-based tool to automatically generate and evaluate unit tests for a given prompt. PromptPex extracts input and output specifications from a prompt and uses them to generate diverse, targeted, and valid unit tests. These tests are instrumental in identifying regressions when a prompt is changed and also serve as a tool to understand how prompts are interpreted by different models. We use PromptPex to generate tests for eight benchmark prompts and evaluate the quality of the generated tests by seeing if they can cause each of four diverse models to produce invalid output. PromptPex consistently creates tests that result in more invalid model outputs than a carefully constructed baseline LLM-based test generator. Furthermore, by extracting concrete specifications from the input prompt, PromptPex allows prompt writers to clearly understand and test specific aspects of their prompts. The source code of PromptPex is available at https://github.com/microsoft/promptpex.
Authors: Yingming Pu, Tao Lin, Hongyu Chen
Abstract: Large Language Model (LLM)-based multi-agent systems (MAS) demonstrate remarkable potential for scientific discovery. Existing approaches, however, often automate scientific discovery using predefined workflows that lack rationality constraints. This often leads to aimless hypothesizing and a failure to consistently link hypotheses with evidence, thereby hindering the systematic reduction of uncertainty. Overcoming these limitations fundamentally requires a principled approach to exploration. We introduce PiFlow, an information-theoretical framework, treating automated scientific discovery as a structured uncertainty reduction problem guided by principles (e.g., scientific laws). Extensive evaluations across three distinct scientific domains demonstrate that PiFlow (I) improves discovery efficiency by 31.18%~41.73% and solution quality by 12.47%~31.72% against state-of-the-art methods, (II) delivers a 5.6x speedup in time-to-solution while reducing token consumption by up to 27% compared to vanilla agents, and (III) serves as a Plug-and-Play module that generalizes on existing agent architecture. Overall, PiFlow establishes a novel paradigm shift in highly efficient agentic scientific discovery, paving the way for more robust and accelerated AI-driven research.
Authors: Fuchun Ge, Yuxinxin Chen, Pavlo O. Dral
Abstract: Molecular dynamics (MD) is a powerful tool for exploring the behavior of atomistic systems, but its reliance on sequential numerical integration limits simulation efficiency. We present a novel neural network architecture, MDtrajNet, and a pre-trained foundational model, MDtrajNet-1, that directly generates MD trajectories across chemical space, bypassing force calculations and integration. MDtrajNet combines equivariant neural networks with a transformer-based architecture to achieve strong accuracy and transferability in predicting long-time trajectories. This approach accelerates simulations by up to two orders of magnitude and yields better accuracy than MD propagated with established machine-learning interatomic potentials trained on the same data. Remarkably, the errors of the trajectories generated by MDtrajNet-1 for various seen and even unseen small-sized molecular systems are close to those of the conventional ab initio MD. The current limitations of MDtrajNet-1 are attributed to the relatively small size of the chemical space in its training data; however, even for bigger, unseen systems, MDtrajNet-1 provides a good starting point for fine-tuning and obtaining system-specific models. The architecture's flexible design supports diverse application scenarios, including different statistical ensembles, boundary conditions, and interaction types. By overcoming the intrinsic speed barrier of conventional MD, MDtrajNet opens new frontiers in efficient and scalable atomistic simulations.
Authors: Kiet Bennema ten Brinke, Koen Minartz, Vlado Menkovski
Abstract: Simulating trajectories of dynamical systems is a fundamental problem in a wide range of fields such as molecular dynamics, biochemistry, and pedestrian dynamics. Machine learning has become an invaluable tool for scaling physics-based simulators and developing models directly from experimental data. In particular, recent advances in deep generative modeling and geometric deep learning enable probabilistic simulation by learning complex trajectory distributions while respecting intrinsic permutation and time-shift symmetries. However, trajectories of N-body systems are commonly characterized by high sensitivity to perturbations leading to bifurcations, as well as multi-scale temporal and spatial correlations. To address these challenges, we introduce STFlow (Spatio-Temporal Flow), a generative model based on graph neural networks and hierarchical convolutions. By incorporating data-dependent couplings within the Flow Matching framework, STFlow denoises starting from conditioned random-walks instead of Gaussian noise. This novel informed prior simplifies the learning task by reducing transport cost, increasing training and inference efficiency. We validate our approach on N-body systems, molecular dynamics, and human trajectory forecasting. Across these benchmarks, STFlow achieves the lowest prediction errors with fewer simulation steps and improved scalability.
Authors: Zhen Yang, Hongyi Lin, Yifan He, Junqi Wang, Zeyu Sun, Shuo Liu, Jie Xu, Pengpeng Wang, Zhongxing Yu, Qingyuan Liang
Abstract: In recent years, code intelligence has gained increasing importance in the field of automated software engineering. Meanwhile, the widespread adoption of Pretrained Language Models (PLMs) and Large Language Models (LLMs) has raised concerns regarding data contamination and its potential impact on model performance evaluation. Previous studies mainly focused on sample-level contamination, ignoring partial contamination scenarios that are pervasive in code intelligence. This paper fills this gap and presents a systematic empirical study to investigate the fine-grained data contamination on mainstream code tasks. Our study involves diverse representative PLMs: RoBERTa and GPT-2, and LLMs: LLaMA and StarCoder, covering three major tasks: code translation, code generation, and code summarization, across two Programming Languages (PLs): Java and Python. We categorize contamination scenarios into four types according to the code intelligence practice, namely input-only, output-only, unpaired, and paired contamination settings, and construct corresponding experimental and control groups for exploration. Experimental results show that, under the pre-training, fine-tuning, and inference paradigm adopted by PLMs, even deliberately injecting paired contamination does not lead to significant performance overestimation. But direct inference or small-scale fine-tuning uncovers the contamination effects. In contrast, LLMs with pre-training and inference paradigm are significantly affected by the paired contamination. Apart from the above, other contamination scenarios have no impact on both PLMs and LLMs. Our findings challenge the conventional belief that contamination inevitably leads to performance overestimation, providing new insights into the evaluation and deployment of code intelligence models.
Authors: Xintong Duan, Yutong He, Fahim Tajwar, Ruslan Salakhutdinov, J. Zico Kolter, Jeff Schneider
Abstract: Although diffusion models have achieved strong results in decision-making tasks, their slow inference speed remains a key limitation. While consistency models offer a potential solution, existing applications to decision-making either struggle with suboptimal demonstrations under behavior cloning or rely on complex concurrent training of multiple networks under the actor-critic framework. In this work, we propose a novel approach to consistency distillation for offline reinforcement learning that directly incorporates reward optimization into the distillation process. Our method achieves single-step sampling while generating higher-reward action trajectories through decoupled training and noise-free reward signals. Empirical evaluations on the Gym MuJoCo, FrankaKitchen, and long horizon planning benchmarks demonstrate that our approach can achieve a 9.7% improvement over previous state-of-the-art while offering up to 142x speedup over diffusion counterparts in inference time.
Authors: Giacomo Baldan, Qiang Liu, Alberto Guardone, Nils Thuerey
Abstract: Physics-constrained generative modeling aims to produce high-dimensional samples that are both physically consistent and distributionally accurate, a task that remains challenging due to often conflicting optimization objectives. Recent advances in flow matching and diffusion models have enabled efficient generative modeling, but integrating physical constraints often degrades generative fidelity or requires costly inference-time corrections. Our work is the first to recognize the trade-off between distributional and physical accuracy. Based on the insight of inherently conflicting objectives, we introduce Physics-Based Flow Matching (PBFM) a method that enforces physical constraints at training time using conflict-free gradient updates and unrolling to mitigate Jensen's gap. Our approach avoids manual loss balancing and enables simultaneous optimization of generative and physical objectives. As a consequence, physics constraints do not impede inference performance. We benchmark our method across three representative PDE benchmarks. PBFM achieves a Pareto-optimal trade-off, competitive inference speed, and generalizes to a wide range of physics-constrained generative tasks, providing a practical tool for scientific machine learning. Code and datasets available at https://github.com/tum-pbs/PBFM.
Authors: Xianquan Yan, Hakan Akg\"un, Kenji Kawaguchi, N. Duane Loh, Ching Hua Lee
Abstract: AI is transforming scientific research by revealing new ways to understand complex physical systems, but its impact remains constrained by the lack of large, high-quality domain-specific datasets. A rich, largely untapped resource lies in non-Hermitian quantum physics, where the energy spectra of crystals form intricate geometries on the complex plane -- termed as Hamiltonian spectral graphs. Despite their significance as fingerprints for electronic behavior, their systematic study has been intractable due to the reliance on manual extraction. To unlock this potential, we introduce Poly2Graph: a high-performance, open-source pipeline that automates the mapping of 1-D crystal Hamiltonians to spectral graphs. Using this tool, we present HSG-12M: a dataset containing 11.6 million static and 5.1 million dynamic Hamiltonian spectral graphs across 1401 characteristic-polynomial classes, distilled from 177 TB of spectral potential data. Crucially, HSG-12M is the first large-scale dataset of spatial multigraphs -- graphs embedded in a metric space where multiple geometrically distinct trajectories between two nodes are retained as separate edges. This simultaneously addresses a critical gap, as existing graph benchmarks overwhelmingly assume simple, non-spatial edges, discarding vital geometric information. Benchmarks with popular GNNs expose new challenges in learning spatial multi-edges at scale. Beyond its practical utility, we show that spectral graphs serve as universal topological fingerprints of polynomials, vectors, and matrices, forging a new algebra-to-graph link. HSG-12M lays the groundwork for data-driven scientific discovery in condensed matter physics, new opportunities in geometry-aware graph learning and beyond.
Authors: Yuliang Xu, Siming Huang, Mingmeng Geng, Yao Wan, Xuanhua Shi, Dongping Chen
Abstract: Coding remains one of the most fundamental modes of interaction between humans and machines. With the rapid advancement of Large Language Models (LLMs), code generation capabilities have begun to significantly reshape programming practices. This development prompts a central question: Have LLMs transformed code style, and how can such transformation be characterized? In this paper, we present a pioneering study that investigates the impact of LLMs on code style, with a focus on naming conventions, complexity, maintainability, and similarity. By analyzing code from over 20,000 GitHub repositories linked to arXiv papers published between 2020 and 2025, we identify measurable trends in the evolution of coding style that align with characteristics of LLM-generated code. For instance, the proportion of snake_case function names in Python code increased from 40.7% in Q1 2023 to 49.8% in Q3 2025. Furthermore, we investigate how LLMs approach algorithmic problems by examining their reasoning processes. Our experimental results may provide the first large-scale empirical evidence that LLMs affect real-world programming style. We release all the experimental dataset and source code at: https://github.com/ignorancex/LLM_code
Authors: Chenchen Yuan, Zheyu Zhang, Shuo Yang, Bardh Prenkaj, Gjergji Kasneci
Abstract: Large Language Models (LLMs) have shown impressive moral reasoning abilities. Yet they often diverge when confronted with complex, multi-factor moral dilemmas. To address these discrepancies, we propose a framework that synthesizes multiple LLMs' moral judgments into a collectively formulated moral judgment, realigning models that deviate significantly from this consensus. Our aggregation mechanism fuses continuous moral acceptability scores (beyond binary labels) into a collective probability, weighting contributions by model reliability. For misaligned models, a targeted embedding-optimization procedure fine-tunes token embeddings for moral philosophical theories, minimizing JS divergence to the consensus while preserving semantic integrity. Experiments on a large-scale social moral dilemma dataset show our approach builds robust consensus and improves individual model fidelity. These findings highlight the value of data-driven moral alignment across multiple models and its potential for safer, more consistent AI systems.
Authors: Sizhe Chen, Arman Zharmagambetov, David Wagner, Chuan Guo
Abstract: Prompt injection attacks, where untrusted data contains an injected prompt to manipulate the system, have been listed as the top security threat to LLM-integrated applications. Model-level prompt injection defenses have shown strong effectiveness, but the strongest defenses are proprietary. Open-source secure models are needed by the AI security community so that co-development of attacks and defenses through open research can drive scientific progress in mitigating prompt injection attacks. To this end, we develop Meta SecAlign, the first fully open-source LLM with built-in model-level defense that achieves commercial-grade performance and is powerful enough for complex agentic tasks. We provide complete details of our training recipe. We perform the most comprehensive evaluation to date on 9 utility benchmarks (measuring general knowledge, instruction following, and agentic workflows) and 7 security benchmarks. Results show that Meta SecAlign, despite being trained only on generic instruction-tuning samples, surprisingly confers security in unseen downstream tasks, including tool-calling and web-navigation, in addition to general instruction-following. Our best model -- Meta-SecAlign-70B -- establishes a new frontier of utility-security trade-off for open-source LLMs, and is more secure than several flagship proprietary models with prompt injection defense. Below are links for the code (https://github.com/facebookresearch/Meta_SecAlign), Meta-SecAlign-70B (https://huggingface.co/facebook/Meta-SecAlign-70B), and Meta-SecAlign-8B (https://huggingface.co/facebook/Meta-SecAlign-8B) models.
URLs: https://github.com/facebookresearch/Meta_SecAlign),, https://huggingface.co/facebook/Meta-SecAlign-70B),, https://huggingface.co/facebook/Meta-SecAlign-8B)
Authors: Yannis Bendi-Ouis (Mnemosyne), Xavier Hinaut (Mnemosyne)
Abstract: While Large Language Models and their underlying Transformer architecture are remarkably efficient, they do not reflect how our brain processes and learns a diversity of cognitive tasks such as language, nor how it leverages working memory. Furthermore, Transformers encounters a computational limitation: quadratic complexity growth with sequence length. Motivated by these limitations, we aim to design architectures that leverage efficient working memory dynamics to overcome standard computational barriers. We introduce Echo State Transformers (EST), a hybrid architecture that resolves this challenge while demonstrating state of the art performance in classification and detection tasks. EST integrates the Transformer attention mechanisms with nodes from Reservoir Computing to create a fixed-size memory system. Drawing inspiration from Echo State Networks, our approach leverages several reservoirs (random recurrent networks) in parallel as a lightweight and efficient working memory. These independent units possess distinct and learned internal dynamics with an adaptive leak rate, enabling them to dynamically adjust their own temporality. By applying attention on those fixed number of units instead of input tokens, EST achieves linear complexity for the whole sequence, effectively breaking the quadratic scaling problem of standard Transformers. We evaluate ESTs on a recent timeseries benchmark: the Time Series Library, which comprises 69 tasks across five categories. Results show that ESTs ranks first overall in two of five categories, outperforming strong state-of-the-art baselines on classification and anomaly detection tasks, while remaining competitive on short-term forecasting. These results demonstrate that by shifting the attention mechanism from the entire input sequence to a fixed set of evolving memory units, it is possible to maintains high sensitivity to temporal events while achieving constant computational complexity per step.
Authors: Etrit Haxholli, Yeti Z. Gurbuz, Ogul Can, Eli Waxman
Abstract: While continuous diffusion models excel in modeling continuous distributions, their application to categorical data has been less effective. Recent work has shown that ratio-matching through score-entropy within a continuous-time discrete Markov chain (CTMC) framework serves as a competitive alternative to autoregressive models in language modeling. To enhance this framework, we first introduce three new theorems concerning the KL divergence between the data and learned distribution. Our results serve as the discrete counterpart to those established for continuous diffusion models and allow us to derive an improved upper bound of the perplexity. Second, we empirically show that ratio-matching performed by minimizing the denoising cross-entropy between the clean and corrupted data enables models to outperform those utilizing score-entropy with up to 10% lower perplexity/generative-perplexity, and 15% faster training steps. To further support our findings, we introduce and evaluate a novel CTMC transition-rate matrix that allows prediction refinement, and derive the analytic expression for its matrix exponential which facilitates the computation of conditional ratios thus enabling efficient training and generation.
Authors: Syemin Park, Soobin Park, Youn-kyung Lim
Abstract: Creating a cast of characters by attending to their relational dynamics is a critical aspect of most long-form storywriting. However, our formative study (N=14) reveals that writers struggle to envision new characters that could influence existing ones, balance similarities and differences among characters, and intricately flesh out their relationships. Based on these observations, we designed Constella, an LLM-based multi-agent tool that supports storywriters' interconnected character creation process. Constella suggests related characters (FRIENDS DISCOVERY feature), reveals the inner mindscapes of several characters simultaneously (JOURNALS feature), and manifests relationships through inter-character responses (COMMENTS feature). Our 7-8 day deployment study with storywriters (N=11) shows that Constella enabled the creation of expansive communities composed of related characters, facilitated the comparison of characters' thoughts and emotions, and deepened writers' understanding of character relationships. We conclude by discussing how multi-agent interactions can help distribute writers' attention and effort across the character cast.
Authors: Weibo Zhou, Lingbo Li, Shangsong Liang
Abstract: The scarcity and high cost of high-quality domain-specific question-answering (QA) datasets limit supervised fine-tuning of large language models (LLMs). We introduce $\textbf{D-SCoRE}$, a training-free framework that leverages LLMs and prompt engineering to automatically generate diverse, rich QA datasets with Chain-of-Thought (CoT) from arbitrary textual sources. By integrating $\textbf{D}$ocument-centric processing, $\textbf{S}$egmentation, $\textbf{Co}$T $\textbf{R}$easoning, and structured $\textbf{E}$xport - along with multi-dimensional controls such as semantic role transformation, question type balancing, and counterfactual augmentation - D-SCoRE produces tailored QA pairs with enhanced diversity and relevance. LLMs fine-tuned on D-SCoRE-generated datasets outperform those trained on human-annotated QA data across most evaluated domains. Its efficiency and scalability enable rapid, high-performance domain-adaptive fine-tuning on consumer-grade hardware, generating over 1,100 high-quality QA pairs per GPU-hour end-to-end.
Authors: Hai Zhong, Xun Wang, Zhuoran Li, Longbo Huang
Abstract: By leveraging differentiable dynamics, Reparameterization Policy Gradient (RPG) achieves high sample efficiency. However, current approaches are hindered by two critical limitations: the under-utilization of computationally expensive dynamics Jacobians and inherent training instability. While sample reuse offers a remedy for under-utilization, no prior principled framework exists, and naive attempts risk exacerbating instability. To address these challenges, we propose Reparameterization Proximal Policy Optimization (RPO). We first establish that under sample reuse, RPG naturally optimizes a PPO-style surrogate objective via Backpropagation Through Time, providing a unified framework for both on- and off-policy updates. To further ensure stability, RPO integrates a clipped policy gradient mechanism tailored for RPG and employs explicit Kullback-Leibler divergence regularization. Experimental results demonstrate that RPO maintains superior sample efficiency and consistently outperforms or achieves state-of-the-art performance across diverse tasks.
Authors: Changin Choi, Wonseok Lee, Jungmin Ko, Wonjong Rhee
Abstract: Knowledge-intensive visual question answering (VQA) requires external knowledge beyond image content, demanding precise visual grounding and coherent integration of visual and textual information. Although multimodal retrieval-augmented generation has achieved notable advances by incorporating external knowledge bases, existing approaches largely adopt single-pass frameworks that often fail to acquire sufficient knowledge and lack mechanisms to revise misdirected reasoning. We propose PMSR (Progressive Multimodal Search and Reasoning), a framework that progressively constructs a structured reasoning trajectory to enhance both knowledge acquisition and synthesis. PMSR uses dual-scope queries conditioned on both the latest record and the trajectory to retrieve diverse knowledge from heterogeneous knowledge bases. The retrieved evidence is then synthesized into compact records via compositional reasoning. This design facilitates controlled iterative refinement, which supports more stable reasoning trajectories with reduced error propagation. Extensive experiments across six diverse benchmarks (Encyclopedic-VQA, InfoSeek, MMSearch, LiveVQA, FVQA, and OK-VQA) demonstrate that PMSR consistently improves both retrieval recall and end-to-end answer accuracy.
Authors: Xinzhe Zheng, Zhen-Qun Yang, Zishan Liu, Haoran Xie, S. Joe Qin, Arlene Chen, Fangzhen Lin
Abstract: Large Language Models (LLMs) deliver strong performance but are difficult to deploy under tight memory and compute constraints. Low-bit post-training quantization (PTQ) is a promising direction; however, it typically relies on calibration data, auxiliary transformations, and GPU tools. To address these limitations, we propose MSB (Multi Scale Binary), a calibration-free and transformation-free PTQ method that generalizes binary quantization to multi-bit settings. MSB optimizes a dynamic grouping criterion that minimizes within group variance, yielding group-wise multiscale levels that can be applied consistently across granularities from per tensor to block-wise configurations with 64 elements groups per row, without calibration or intermediate transforms. We implement the optimization in a CPU based solver for the quantization step and evaluate using standard bfloat16 execution without low-bit packing. On Llama 3.2 3B, MSB achieves 8.43 perplexity on WikiText-2 under 4-bit weight only block-wise quantization, compared to 7.81 in full precision and 12.23 with GPTQ its default setup. Overall, MSB provides a new optimization perspective for low-bit PTQ while simplifying the pipeline by removing calibration and transformations.
Authors: Amin Karimi Monsefi, Nikhil Bhendawade, Manuel Rafael Ciosici, Dominic Culver, Yizhe Zhang, Irina Belousova
Abstract: Autoregressive language models (ARMs) deliver strong likelihoods, but are inherently serial: they generate one token per forward pass, which limits throughput and inflates latency for long sequences. Diffusion Language Models (DLMs) parallelize across positions and thus appear promising for language generation, yet standard discrete diffusion typically needs hundreds to thousands of model evaluations to reach high quality, trading serial depth for iterative breadth. We introduce FS-DFM, Few-Step Discrete Flow-Matching. A discrete flow-matching model designed for speed without sacrificing quality. The core idea is simple: make the number of sampling steps an explicit parameter and train the model to be consistent across step budgets, so one big move lands where many small moves would. We pair this with a reliable update rule that moves probability in the right direction without overshooting, and with strong teacher guidance distilled from long-run trajectories. Together, these choices make few-step sampling stable, accurate, and easy to control. On language modeling benchmarks, FS-DFM with 8 sampling steps achieves perplexity parity with a 1,024-step discrete-flow baseline for generating 1,024 tokens using a similar-size model, delivering up to 128 times faster sampling and corresponding latency/throughput gains. Code & pretrained checkpoints: https://github.com/apple/ml-fs-dfm
Authors: Michael S. Yao, Osbert Bastani, Alma Andersson, Tommaso Biancalani, A\"icha Bentaieb, Claudia Iriondo
Abstract: The goal of personalized medicine is to discover a treatment regimen that optimizes a patient's clinical outcome based on their personal genetic and environmental factors. However, candidate treatments cannot be arbitrarily administered to the patient to assess their efficacy; we often instead have access to an in silico surrogate model that approximates the true fitness of a proposed treatment. Unfortunately, such surrogate models have been shown to fail to generalize to previously unseen patient-treatment combinations. We hypothesize that domain-specific prior knowledge - such as medical textbooks and biomedical knowledge graphs - can provide a meaningful alternative signal of the fitness of proposed treatments. To this end, we introduce LLM-based Entropy-guided Optimization with kNowledgeable priors (LEON), a mathematically principled approach to leverage large language models (LLMs) as black-box optimizers without any task-specific fine-tuning, taking advantage of their ability to contextualize unstructured domain knowledge to propose personalized treatment plans in natural language. In practice, we implement LEON via 'optimization by prompting,' which uses LLMs as stochastic engines for proposing treatment designs. Experiments on real-world optimization tasks show LEON outperforms both traditional and LLM-based methods in proposing individualized treatments for patients.
Authors: Jialin Wu, Shreya Saha, Yiqing Bo, Meenakshi Khosla
Abstract: Large vision models differ widely in architecture and training paradigm, yet we lack principled methods to determine which aspects of their representations are shared across families and which reflect distinctive computational strategies. We leverage a suite of representational similarity metrics, each capturing a different facet-geometry, unit tuning, or linear decodability-and assess family separability using multiple complementary measures. Metrics preserving geometry or tuning (e.g., RSA, Soft Matching) yield strong family discrimination, whereas flexible mappings such as Linear Predictivity show weaker separation. These findings indicate that geometry and tuning carry family-specific signatures, while linearly decodable information is more broadly shared. To integrate these complementary facets, we adapt Similarity Network Fusion (SNF), a method inspired by multi-omics integration. SNF achieves substantially sharper family separation than any individual metric and produces robust composite signatures. Clustering of the fused similarity matrix recovers both expected and surprising patterns: supervised ResNets and ViTs form distinct clusters, yet all self-supervised models group together across architectural boundaries. Hybrid architectures (ConvNeXt, Swin) cluster with masked autoencoders, suggesting convergence between architectural modernization and reconstruction-based training. This biology-inspired framework provides a principled typology of vision models, showing that emergent computational strategies-shaped jointly by architecture and training objective-define representational structure beyond surface design categories.
Authors: Yu-Che Tsai, Kuan-Yu Chen, Yuan-Chi Li, Yuan-Hao Chen, Ching-Yu Tsai, Shou-De Lin
Abstract: Existing large language model (LLM)-based embeddings typically adopt an encoder-only paradigm, treating LLMs as static feature extractors and overlooking their core generative strengths. We introduce GIRCSE (Generative Iterative Refinement for Contrastive Sentence Embeddings), a novel framework that leverages autoregressive generation to iteratively refine semantic representations. By producing sequences of soft tokens optimized under contrastive objective, GIRCSE captures latent concepts and implicit semantics that encoder-only methods often miss. To guide this process, we propose an Iterative Contrastive Refinement (ICR) objective that encourages each refinement step to yield better representations. Extensive experiments show that GIRCSE outperforms strong LLM-based embedding baselines on the MTEB benchmark and instruction-following tasks. Moreover, GIRCSE exhibits an emergent test-time scaling property: generating more tokens at inference steadily improves embedding quality. Our results establish generative iterative refinement as a new paradigm for representation learning.
Authors: Jennifer Lin
Abstract: We provide evidence that eigenanalysis of the empirical neural tangent kernel (eNTK) can surface the features used by trained neural networks. Across three standard toy models for mechanistic interpretability, Toy Models of Superposition (TMS), a 1-layer MLP trained on modular addition and a 1-layer Transformer trained on modular addition, we find that top eigenspaces of the eNTK align with ground-truth features. In TMS, the eNTK recovers the ground-truth features in both the sparse (high superposition) and dense regimes. In modular arithmetic, the eNTK can be used to recover Fourier feature families. Moreover, we provide evidence that a layerwise eNTK localizes features to specific layers and that the evolution of the eNTK spectrum can be used to diagnose the grokking phase transition. These results suggest that eNTK analysis may provide a practical handle for feature discovery and for detecting phase changes in small models.
Authors: Donald Pfaffmann, Matthias Klusch, Marcel Steinmetz
Abstract: We present a novel hybrid learning-assisted planning method, named HyPlan, for solving the collision-free navigation problem for self-driving cars in partially observable traffic environments. HyPlan combines methods for multi-agent behavior prediction, deep reinforcement learning with proximal policy optimization and approximated online POMDP planning with heuristic confidence-based vertical pruning to reduce its execution time without compromising safety of driving. Our experimental performance analysis on the CARLA-CTS2 benchmark of critical traffic scenarios with pedestrians revealed that HyPlan may navigate safer than selected relevant baselines and perform significantly faster than considered alternative online POMDP planners.
Authors: Zhechong Huang, Zhao Zhang, Ruyi Ji, Tingxuan Xia, Qihao Zhu, Qinxiang Cao, Zeyu Sun, Wiggin Zhou, Yingfei Xiong
Abstract: Language models have shown remarkable proficiency in code generation; nevertheless, ensuring type correctness remains a challenge. Although traditional methods, such as constrained decoding, alleviate this problem by externally rejecting untypable code, the model itself does not effectively learn type reasoning internally, which ultimately limits its overall performance. This paper introduces TyFlow, a novel system that internalizes type reasoning within code generation to guide the model to learn the type system. The core of our approach is a novel type-guided program synthesis system that maintains an isomorphism between type derivation trees and synthesis derivation trees, enabling a new code representation based on synthesis decision sequences rather than traditional text-based token sequences. By offloading the complexity of type system learning to the representation itself, models can redirect their computational resources toward higher-level program semantics. Our evaluation shows that TyFlow not only eliminates type errors but also significantly improves functional correctness, highlighting the importance of aligning LMs with type systems internally.
Authors: Moritz Wagner, Christophe Roux, Max Zimmer, Sebastian Pokutta
Abstract: Post-training pruning substantially reduces inference costs but often causes severe quality degradation without adapting the remaining weights. For LLMs, such retraining is commonly considered impractical due to large computational costs, motivating increasingly sophisticated pruning criteria to compensate by selecting better sparsity patterns. In this work, we revisit post-pruning adaptation and study local reconstruction: adapting only a small pruned submodel at a time using a small calibration set by matching intermediate activations of the dense model. We conduct a large-scale study across model families and scales (up to 72B parameters) and establish three central results. First, local reconstruction is an effective adaptation mechanism for LLMs, matching post-pruning PEFT while using over an order of magnitude less data and compute. Second, we identify a broad "free lunch" regime in reconstruction granularity: across a wide range of submodel sizes, final quality remains essentially unchanged, allowing granularity to be chosen based on memory constraints. Finally, with reconstruction, the pruning criterion becomes less critical: performance gaps between sophisticated methods and simple baselines shrink with model size, making simple methods competitive again. Collectively, our results challenge the prevailing narrative that post-pruning adaptation is impractical for LLMs.
Authors: Yu Wu, Ke Shu, Jonas Fischer, Lidia Pivovarova, David Rosson, Eetu M\"akel\"a, Mikko Tolonen
Abstract: This paper presents a novel task of extracting low-resourced and noisy Latin fragments from mixed-language historical documents with varied layouts. We benchmark and evaluate the performance of large foundation models against a multimodal dataset of 724 annotated pages. The results demonstrate that reliable Latin detection with contemporary zero-shot models is achievable, yet these models lack a functional comprehension of Latin. This study establishes a comprehensive baseline for processing Latin within mixed-language corpora, supporting quantitative analysis in intellectual history and historical linguistics. Both the dataset and code are available at https://github.com/COMHIS/EACL26-detect-latin.
Authors: Marko Karbevski, Antonij Mijoski
Abstract: We theoretically investigate whether the Query, Key, Value weight triplet can be reduced in encoder-only and decoder-only transformers. Under mild assumptions, we prove that Query weights are redundant and can be replaced with the identity matrix, reducing attention parameters by $25\%$. This also simplifies optimization: attention logits become linear rather than quadratic in learned weights. Validating on decoder-only GPT-style small models trained from scratch, we find that with adjusted attention scaling and weight decay, reduced models match baseline performance despite fewer parameters. Training remains stable at over $3\times$ lower weight decay, suggesting Query weight elimination provides implicit regularization. Our analysis has also led us to a structural expressivity boundary: in the mathematically tractable ReLU setting, skip connections push MLPs into a generically disjoint function class at fixed width. These findings motivate investigation across modalities and at scale, where the observed stability and efficiency gains may prove most consequential.
Authors: Zainab Aizaz, James C. Knight, Thomas Nowotny
Abstract: Spiking Neural Networks (SNNs) are a promising, energy-efficient alternative to standard Artificial Neural Networks (ANNs) and are particularly well-suited to spatio-temporal tasks such as keyword spotting and video classification. However, SNNs have a much lower arithmetic intensity than ANNs and are therefore not well-matched to standard accelerators like GPUs and TPUs. Field Programmable Gate Arrays (FPGAs) are designed for such memory-bound workloads, and here we present a novel, fully-programmable RISC-V-based system-on-chip (FeNN-DMA), tailored to simulating SNNs on modern UltraScale+ FPGAs. We show that FeNN-DMA has comparable resource usage and energy requirements to state-of-the-art fixed-function SNN accelerators, yet it supports more complex neuron models and network topologies, and can simulate up to 16 thousand neurons and 256 million synapses per core. Using this functionality, we demonstrate state-of-the-art classification accuracy on the Spiking Heidelberg Digits, Neuromorphic MNIST and Braille tactile classification tasks.
Authors: Qiyong Zhong, Jiajie Su, Ming Yang, Yunshan Ma, Xiaolin Zheng, Chaochao Chen
Abstract: Sequential recommendations (SR) predict users' future interactions based on their historical behavior. The rise of Large Language Models (LLMs) has brought powerful generative and reasoning capabilities, significantly enhancing SR performance, while Multimodal LLMs (MLLMs) further extend this by introducing data like images and interactive relationships. However, critical issues remain, i.e., (a) Suboptimal item representations caused by lengthy and redundant descriptions, leading to inefficiencies in both training and inference; (b) Modality-related cognitive bias, as LLMs are predominantly pretrained on textual data, limiting their ability to effectively integrate and utilize non-textual modalities; (c) Weakening sequential perception in long interaction sequences, where attention mechanisms struggle to capture earlier interactions, hindering the modeling of long-range dependencies. To address these issues, we propose Speeder, an efficient MLLM-based paradigm for SR featuring three key innovations: 1) Multimodal Representation Compression (MRC), which condenses item attributes into concise yet informative tokens, reducing redundancy and computational cost; 2) Modality-aware Progressive Optimization (MPO), enabling gradual learning of multimodal representations; 3) Sequential Position Awareness Enhancement (SPAE), improving the LLM's capability to capture both relative and absolute sequential dependencies in long interaction sequences. Extensive experiments on real-world datasets demonstrate the effectiveness and efficiency of Speeder. Speeder increases training speed to 250% of the original while reducing inference time to 25% on the Amazon dataset.
Authors: Karen Sargsyan
Abstract: AI-Scientist systems that use large language models to automate research risk generating spurious discoveries through uncontrolled multiple testing. We present a functional architecture that enforces statistical rigor at two levels: a Haskell embedded DSL (the Research monad) that makes it impossible to test a hypothesis without updating the error budget, and a declarative scaffolding technique that structurally prevents data leakage across the boundary into LLM-generated code. We ground these guarantees in a machine-checked Lean 4 formalization of the LORD++ online FDR control theorem (855 lines, zero sorry), which identifies four sufficient conditions for FDR control. Three are structural conditions -- about information flow, data separation, and test validity -- enforced by the architecture's type system and scaffolding. The fourth is an arithmetic condition: a budget invariant requiring that a wealth process remain non-negative under floating-point computation. We verify this condition over IEEE 754 doubles using SPARK/Ada, whose GNATprove toolchain statically confirms that no rounding sequence can violate the invariant and whose flow analysis independently confirms the predictability condition. The resulting verification chain -- from real-analysis proof to floating-point implementation -- is, to our knowledge, the first for any online FDR control procedure. Monte Carlo simulation (N=2000 hypotheses) and an end-to-end case study confirm that the monadic implementation controls FDR at 1.1% against a 5% target, while a naive approach inflates to 41%.
Authors: Ying Song, Balaji Palanisamy
Abstract: Graph unlearning has emerged as a promising solution to comply with "the right to be forgotten" regulations by enabling the removal of sensitive information upon request. However, this solution is not foolproof. The involvement of multiple parties creates new attack surfaces, and residual traces of deleted data can still remain in the unlearned graph neural networks (GNNs). These vulnerabilities can be exploited by attackers to recover the supposedly erased samples, thereby undermining the intended functionality of graph unlearning. In this work, we propose GraphToxin, the first full graph reconstruction attack against graph unlearning. Specifically, we introduce a novel curvature matching module to provide fine-grained guidance for unlearned graph recovery. We demonstrate that GraphToxin can successfully subvert the regulatory guarantees expected from graph unlearning, it can recover not only a deleted individual's information and personal links but also sensitive content from their connections, thereby posing substantially more detrimental threats. Furthermore, we extend GraphToxin to multiple-node removal under both white-box and black-box settings, showcasing its practical feasibility and potential to cause considerable harm. We highlight the necessity of worst-case analysis and propose a systematic evaluation framework to assess attack performance under both random and worst-case node removal scenarios. Our extensive experiments demonstrate the effectiveness and flexibility of GraphToxin. Notably, existing defense mechanisms are largely ineffective against this attack or even amplify its performance in some cases. Given the severe privacy risks posed by GraphToxin, our work underscores the urgent need for more effective and robust defenses.
Authors: Sajad U P
Abstract: Phishing and related cyber threats are becoming more varied and technologically advanced. Among these, email-based phishing remains the most dominant and persistent threat. These attacks exploit human vulnerabilities to disseminate malware or gain unauthorized access to sensitive information. Deep learning (DL) models, particularly transformer-based models, have significantly enhanced phishing mitigation through their contextual understanding of language. However, some recent threats, specifically Artificial Intelligence (AI)-generated phishing attacks, are reducing the overall system resilience of phishing detectors. In response, adversarial training has shown promise against AI-generated phishing threats. This study presents a hybrid approach that uses DistilBERT, a smaller, faster, and lighter version of the BERT transformer model for email classification. Robustness against text-based adversarial perturbations is reinforced using Fast Gradient Method (FGM) adversarial training. Furthermore, the framework integrates the LIME Explainable AI (XAI) technique to enhance the transparency of the DistilBERT architecture. The framework also uses the Flan-T5-small language model from Hugging Face to generate plain-language security narrative explanations for end-users. This combined approach ensures precise phishing classification while providing easily understandable justifications for the model's decisions.
Authors: Kexin Chu, Dawei Xiang, Zixu Shen, Yiwei Yang, Zecheng Liu, Wei Zhang
Abstract: Mixture-of-Experts (MoE) has become a practical architecture for scaling LLM capacity while keeping per-token compute modest, but deploying MoE models on a single, memory-limited GPU remains difficult because expert weights dominate the HBM footprint. Existing expert offloading and prefetching systems reduce the resident set, yet they often pay expert-loading costs on the critical path when activation becomes dense. Post-training quantization (PTQ) lowers the footprint without transfers, but prevailing pipelines fix expert bit-widths offline and assume routing remains stable, even though MoE expert utilization is heavy-tailed and the hot set can shift across workloads. We present DynaExq, a runtime-aware mixed-precision serving system that treats single-GPU MoE inference under a hard HBM envelope as an online, budget-constrained precision allocation problem. The key insight is to keep the experts that dominate runtime traffic resident at higher precision, while maintaining a low-precision fallback for the remaining experts, so the system can reduce transfer volume and avoid the waiting latency that limits offloading and prefetching under dense activation. DynaExq estimates long-horizon expert hotness from router traces, selects a per-layer high-precision resident set via a budget-feasible top-$n$ rule, and applies promotions and demotions asynchronously through stable expert handles so the forward pass always executes on a fully materialized expert version. Across Qwen3-MoE-30B/80B and six benchmarks, DynaExq improves accuracy over static PTQ on Qwen3-80B (73.09% to 77.57%) under comparable device-memory budgets and achieves up to 2.73x higher throughput than offloading/prefetch baselines at batch size 32.
Authors: Xingtao Zhao, Hao Peng, Dingli Su, Xianghua Zeng, Chunyang Liu, Jinzhi Liao, Philip S. Yu
Abstract: Reliable uncertainty quantification (UQ) is essential for deploying large language models (LLMs) in safety-critical scenarios, as it enables them to abstain from responding when uncertain, thereby avoiding hallucinations, i.e., plausible yet factually incorrect responses. However, while semantic UQ methods have achieved advanced performance, they overlook latent semantic structural information that could enable more precise uncertainty estimates. In this paper, we propose \underline{Se}mantic \underline{S}tructural \underline{E}ntropy ({SeSE}), a principled black-box UQ framework applicable to both open- and closed-source LLMs. To reveal the intrinsic structure of the semantic space, SeSE constructs its optimal hierarchical abstraction through an encoding tree with minimal structural entropy. The structural entropy of this encoding tree thus quantifies the inherent uncertainty within LLM semantic space after optimal compression. Additionally, unlike existing methods that primarily focus on simple short-form generation, we extent SeSE to provide interpretable, granular uncertainty estimation for long-form outputs. We theoretically prove that SeSE generalizes semantic entropy, the gold standard for UQ in LLMs, and empirically demonstrate its superior performance over strong baselines across 24 model-dataset combinations.
Authors: Otman A. Basir
Abstract: This paper introduces a new family of multi-parent recombination operators for Genetic Algorithms (GAs), based on normalized Pascal (binomial) coefficients. Unlike classical two-parent crossover operators, Pascal-Weighted Recombination (PWR) forms offsprings as structured convex combination of multiple parents, using binomially shaped weights that emphasize central inheritance while suppressing disruptive variance. We develop a mathematical framework for PWR, derive variance-transfer properties, and analyze its effect on schema survival. The operator is extended to real-valued, binary/logit, and permutation representations. We evaluate the proposed method on four representative benchmarks: (i) PID controller tuning evaluated using the ITAE metric, (ii) FIR low-pass filter design under magnitude-response constraints, (iii) wireless power-modulation optimization under SINR coupling, and (iv) the Traveling Salesman Problem (TSP). We demonstrate how, across these benchmarks, PWR consistently yields smoother convergence, reduced variance, and achieves 9-22% performance gains over standard recombination operators. The approach is simple, algorithm-agnostic, and readily integrable into diverse GA architectures.
Authors: Echo Diyun LU, Charles Findling, Marianne Clausel, Alessandro Leite, Wei Gong, Pierric Kersaudy
Abstract: Regime transitions routinely break stationarity in time series, making calibrated uncertainty as important as point accuracy. We study distribution-free uncertainty for regime-switching forecasting by coupling Deep Switching State Space Models with Adaptive Conformal Inference (ACI) and its aggregated variant (AgACI). We also introduce a unified conformal wrapper that sits atop strong sequence baselines including S4, MC-Dropout GRU, sparse Gaussian processes, and a change-point local model to produce online predictive bands with finite-sample marginal guarantees under nonstationarity and model misspecification. Across synthetic and real datasets, conformalized forecasters achieve near-nominal coverage with competitive accuracy and generally improved band efficiency.
Authors: Guanbo Huang, Jingjia Mao, Fanding Huang, Fengkai Liu, Xiangyang Luo, Yaoyuan Liang, Jiasheng Lu, Xiaoe Wang, Pei Liu, Ruiliu Fu, Shao-Lun Huang
Abstract: Despite tremendous recent progress, Flow Matching methods still suffer from exposure bias due to discrepancies in training and inference. This paper investigates the root causes of exposure bias in Flow Matching, including: (1) the model lacks generalization to biased inputs during training, and (2) insufficient low-frequency content captured during early denoising, leading to accumulated bias. Based on these insights, we propose ReflexFlow, a simple and effective reflexive refinement of the Flow Matching learning objective that dynamically corrects exposure bias. ReflexFlow consists of two components: (1) Anti-Drift Rectification (ADR), which reflexively adjusts prediction targets for biased inputs utilizing a redesigned loss under training-time scheduled sampling; and (2) Frequency Compensation (FC), which reflects on missing low-frequency components and compensates them by reweighting the loss using exposure bias. ReflexFlow is model-agnostic, compatible with all Flow Matching frameworks, and improves generation quality across datasets. Experiments on CIFAR-10, CelebA-64, and ImageNet-256 show that ReflexFlow outperforms prior approaches in mitigating exposure bias, achieving a 35.65% reduction in FID on CelebA-64.
Authors: Shirui Chen, Jiantao Jiao, Lillian J. Ratliff, Banghua Zhu
Abstract: Masked diffusion language models (MDLMs) offer the potential for parallel token generation, but most open-source MDLMs decode fewer than 5 tokens per model forward pass even with sophisticated sampling strategies, limiting their parallel generation potential. Existing acceleration methods either rely on fixed confidence-based heuristics or use distillation-based approaches that finetune MDLMs on trajectories generated by a base model, which can become off-policy during finetuning and restrict performance to the quality of the base model's samples. We propose \texttt{dUltra}, an on-policy reinforcement learning framework based on Group Relative Policy Optimization (GRPO) that learns unmasking strategies for efficient parallel decoding. dUltra introduces an unmasking planner head that predicts per-token unmasking likelihoods under independent Bernoulli distributions. We jointly optimize the base diffusion LLM and the unmasking order planner using reward signals combining verifiable reward, distillation reward, and the number of unmasking steps. Across mathematical reasoning and code generation tasks, dUltra achieves superior accuracy-efficiency trade-offs compared to state-of-the-art heuristic (Fast-dLLM) and distillation baselines (d3LLM, dParallel), demonstrating that learned unmasking trajectories through on-policy RL enable better exploitation of parallel generation in MDLMs. Code and checkpoints are released at https://github.com/chinsengi/dUltra-os.
Authors: Guokan Chen, Yao Xiao, Bin Fan, Meixin Xionga, Zhicheng Lin, Yuanying Liu
Abstract: PINNs enhance scientific computing by incorporating physical laws into neural network structures, leading to significant advancements in scientific computing. However, PINNs struggle with multi-scale and high-frequency problems due to pathological gradient flow and spectral bias, which severely limit their predictive power. By combining an enhanced network architecture with a dynamically adaptive weighting mechanism featuring upper-bound constraints, we propose the Dynamic Balancing Adaptive Weighting Physics-Informed Kolmogorov-Arnold Network (DBAW-PIKAN). The proposed method effectively mitigates gradient-related failure modes and overcomes bottlenecks in function representation. Compared to baseline models, the proposed method accelerates the convergence process and improves solution accuracy by at least an order of magnitude without introducing additional computational complexity. Numerical results on the Klein-Gordon, Burgers, and Helmholtz equations demonstrate that DBAW-PIKAN achieves superior accuracy and generalization performance.
Authors: Phil Cuvin, Hao Zhu, Diyi Yang
Abstract: Deceptive UI designs, widely instantiated across the web and commonly known as dark patterns, manipulate users into performing actions misaligned with their goals. In this paper, we show that dark patterns are highly effective in steering agent trajectories, posing a significant risk to agent robustness. To quantify this risk, we introduce DECEPTICON, an environment for testing individual dark patterns in isolation. DECEPTICON includes 700 web navigation tasks with dark patterns -- 600 generated tasks and 100 real-world tasks, designed to measure instruction-following success and dark pattern effectiveness. Across state-of-the-art agents, we find dark patterns successfully steer agent trajectories towards malicious outcomes in over 70% of tested generated and real-world tasks -- compared to a human average of 31%. Moreover, we find that dark pattern effectiveness correlates positively with model size and test-time reasoning, making larger, more capable models more susceptible. Leading countermeasures against adversarial attacks, including in-context prompting and guardrail models, fail to consistently reduce the success rate of dark pattern interventions. Our findings reveal dark patterns as a latent and unmitigated risk to web agents, highlighting the urgent need for robust defenses against manipulative designs.
Authors: Yingru Li, Jiacai Liu, Jiawei Xu, Yuxuan Tong, Ziniu Li, Qian Liu, Baoxiang Wang
Abstract: Policy gradient methods for Large Language Models optimize a policy $\pi_\theta$ via a surrogate objective computed from samples of a rollout policy $\pi_{\text{roll}}$. However, modern LLM-RL pipelines suffer from unavoidable implementation divergences, such as backend discrepancies, Mixture-of-Experts routing discontinuities, and distributed training staleness. These factors cause an off-policy mismatch ($\pi_{\text{roll}} \neq \pi_\theta$), leading to approximation errors between the surrogate and the true objective. We demonstrate that classical trust region bounds on this error scale as $O(T^2)$ with sequence length $T$, rendering them vacuous for long-horizon tasks. To address this, we derive two new bounds: a Pinsker-Marginal bound scaling as $O(T^{3/2})$ and a Mixed bound scaling as $O(T)$. We further derive an Adaptive bound that strictly generalizes the Pinsker-Marginal bound by combining an importance-ratio decomposition of the error with an adaptive per-position application of Pinsker's inequality on the future trajectory divergence; the minimum over all three bounds is tighter than any individual bound. Crucially, all bounds depend on $D_{\mathrm{KL}}^{\mathrm{tok,max}}$, the maximum token-level KL divergence across the sequence. As a \emph{sequence-level} term, the divergence cannot be controlled by previous token-independent methods like PPO clipping. We propose Trust Region Masking (TRM), which masks entire sequences that violate the trust region. TRM enables the first non-vacuous monotonic improvement guarantees and demonstrates empirical training stability for long-horizon LLM-RL.
Authors: Yingru Li, Jiawei Xu, Jiacai Liu, Yuxuan Tong, Ziniu Li, Tianle Cai, Ge Zhang, Qian Liu, Baoxiang Wang
Abstract: Reinforcement Learning (RL) for Large Language Models (LLMs) faces a fundamental tension: the numerical divergence between high-throughput inference engines and numerically precise training engines. Although these systems share the same parameters, they produce slightly different probability distributions, creating a training-inference mismatch. We prove that the bound on the log-probability divergence arising from this mismatch scales as $(1-p)$, where $p$ is the token probability. This scaling induces a highly asymmetric effect: the bound vanishes for high-probability tokens but remains significant for low-probability tokens in the distribution tail. When sampled, these tail tokens introduce systematically biased errors that accumulate over sequences, thereby destabilizing gradient estimation. Instead of applying post-hoc corrections, we propose Dynamic Vocabulary Pruning (DVP), which constrains the RL objective to a dynamically determined ''safe'' vocabulary that excludes the extreme tail. This strategy trades large, destabilizing numerical errors for a small, bounded optimization bias. We validate DVP empirically by demonstrating stable training, and theoretically by deriving strict bounds on the induced bias.
Authors: Zhijun Chen, Zeyu Ji, Qianren Mao, Hao Wu, Junhang Cheng, Bangjie Qin, Zhuoran Li, Jingzheng Li, Kai Sun, Zizhe Wang, Yikun Ban, Zhu Sun, Xiangyang Ji, Hailong Sun
Abstract: We propose LLM-PeerReview, an unsupervised LLM Ensemble method that selects the most ideal response from multiple LLM-generated candidates for each query, harnessing the collective wisdom of multiple models with diverse strengths. LLM-PeerReview is built on a novel, peer-review-inspired framework that offers a transparent and interpretable mechanism, while remaining fully unsupervised for flexible adaptability and generalization. Specifically, it operates in three stages: For scoring, we use the emerging LLM-as-a-Judge technique to evaluate each response by reusing multiple LLMs at hand; For reasoning, we can apply a straightforward averaging strategy or a principled graphical model-based truth inference algorithm to aggregate multiple scores to produce a final score for each response; Finally, the highest-scoring response is selected as the best ensemble output. LLM-PeerReview is conceptually simple and empirically powerful. Our results across four datasets show that the two variants of the proposed approach outperform the advanced model Smoothie-Global by 6.9% and 7.3% points, cross diverse task types including factual recall QA, math reasoning, and instruction following.
Authors: Yohan Park, Hyunwoo Ha, Wonjun Jo, Tae-Hyun Oh
Abstract: Vision Language Models (VLMs) are increasingly adopted as central reasoning modules for embodied agents. Existing benchmarks evaluate their capabilities under ideal, well-lit conditions, yet robust 24/7 operation demands performance under a wide range of visual degradations, including low-light conditions at night or in dark environments--a core necessity that has been largely overlooked. To address this underexplored challenge, we present DarkEQA, an open-source benchmark for evaluating EQA-relevant perceptual primitives under multi-level low-light conditions. DarkEQA isolates the perception bottleneck by evaluating question answering from egocentric observations under controlled degradations, enabling attributable robustness analysis. A key design feature of DarkEQA is its physical fidelity: visual degradations are modeled in linear RAW space, simulating physics-based illumination drop and sensor noise followed by an ISP-inspired rendering pipeline. We demonstrate the utility of DarkEQA by evaluating a wide range of state-of-the-art VLMs and Low-Light Image Enhancement (LLIE) models. Our analysis systematically reveals VLMs' limitations when operating under these challenging visual conditions. Project website: https://darkeqa-benchmark.github.io/
Authors: Yanhua Zhao
Abstract: Cycle-Consistent Adversarial Network (CycleGAN) is very promising in domain adaptation. In this report, an example in medical domain will be explained. We present struecture of a CycleGAN model for unpaired image-to-image translation from microscopy to pseudo H\&E stained histopathology images.
Authors: Sung Jun Cheon, Jaekyung Cho, Seongho Choi, Hyunjun Eun, Seokhwan Jo, Jaehyun Jun, Minsoo Kang, Jin Kim, Jiwon Kim, Minsang Kim, Sungwan Kim, Seungsik Kim, Tae Yoon Kim, Youngrang Kim, Hyeongmun Lee, Sangyeol Lee, Sungeun Lee, Youngsoon Lee, Yujin Lee, Seongmin Ok, Chanyong Park, Hyewoong Park, Junyoung Park, Hyunho Yang, Subin Yi, Soohyun Bae, Dhammiko Arya, Yongseok Choi, Sangho Choi, Dongyeon Cho, Seungmo Cho, Gyoungeun Han, Yong-jin Han, Seokyoung Hong, Hyeon Hwang, Wonbeom Jang, Minjeong Ju, Wonjin Jung, Keummin Ka, Sungil Kang, Dongnam Kim, Joonghoon Kim, Jonghwi Kim, SaeRom Kim, Sangjin Kim, Seongwon Kim, Youngjin Kim, Seojin Lee, Sunwoo Lee, Taehoon Lee, Chanwoo Park, Sohee Park, Sooyeon Park, Yohan Ra, Sereimony Sek, Seungyeon Seo, Gun Song, Sanghoon Woo, Janghan Yoon, Sungbin Yoon
Abstract: We introduce A.X K1, a 519B-parameter Mixture-of-Experts (MoE) language model trained from scratch. Our design leverages scaling laws to optimize training configurations and vocabulary size under fixed computational budgets. A.X K1 is pre-trained on a corpus of approximately 10T tokens, curated by a multi-stage data processing pipeline. Designed to bridge the gap between reasoning capability and inference efficiency, A.X K1 supports explicitly controllable reasoning to facilitate scalable deployment across diverse real-world scenarios. We propose a simple yet effective Think-Fusion training recipe, enabling user-controlled switching between thinking and non-thinking modes within a single unified model. Extensive evaluations demonstrate that A.X K1 achieves performance competitive with leading open-source models, while establishing a distinctive advantage in Korean-language benchmarks.
Authors: YuanLab. ai, :, Shawn Wu, Jiangang Luo, Tong Yu, Darcy Chen, Sean Wang, Xudong Zhao, Louie Li, Claire Wang, Hunter He, Carol Wang, Allen Wang
Abstract: Although Mixture-of-Experts (MoE) Large Language Models (LLMs) deliver superior accuracy with a reduced number of active parameters, their pre-training represents a significant computationally bottleneck due to underutilized experts and limited training efficiency. This work introduces a Layer-Adaptive Expert Pruning (LAEP) algorithm designed for the pre-training stage of MoE LLMs. In contrast to previous expert pruning approaches that operate primarily in the post-training phase, the proposed algorithm enhances training efficiency by selectively pruning underutilized experts and reorganizing experts across computing devices according to token distribution statistics. Comprehensive experiments demonstrate that LAEP effectively reduces model size and substantially improves pre-training efficiency. In particular, when pre-training the Yuan3.0-1T Base model from scratch original with 1515B parameters, LAEP achieves a 48.3% improvement in training efficiency alongside a 33.3% parameter reduction, while still delivering excellent performance across multiple domains.
Authors: Chongjun Xia, Yanchun Peng, Xianzhi Wang
Abstract: Interactive recommender systems can dynamically adapt to user feedback, but often suffer from content homogeneity and filter bubble effects due to overfitting short-term user preferences. While recent efforts aim to improve content diversity, they predominantly operate in static or one-shot settings, neglecting the long-term evolution of user interests. Reinforcement learning provides a principled framework for optimizing long-term user satisfaction by modeling sequential decision-making processes. However, its application in recommendation is hindered by sparse, long-tailed user-item interactions and limited semantic planning capabilities. In this work, we propose LLM-Enhanced Reinforcement Learning (LERL), a novel hierarchical recommendation framework that integrates the semantic planning power of LLM with the fine-grained adaptability of RL. LERL consists of a high-level LLM-based planner that selects semantically diverse content categories, and a low-level RL policy that recommends personalized items within the selected semantic space. This hierarchical design narrows the action space, enhances planning efficiency, and mitigates overexposure to redundant content. Extensive experiments on real-world datasets demonstrate that LERL significantly improves long-term user satisfaction when compared with state-of-the-art baselines. The implementation of LERL is available at https://github.com/1163710212/LERL.
Authors: Huan Zhang, Yizhan Li, Wenhao Huang, Ziyu Hou, Yu Song, Xuye Liu, Farshid Effaty, Jinya Jiang, Sifan Wu, Qianggang Ding, Izumi Takahara, Leonard R. MacGillivray, Teruyasu Mizoguchi, Tianshu Yu, Lizi Liao, Yuyu Luo, Yu Rong, Jia Li, Ying Diao, Heng Ji, Bang Liu
Abstract: The convergence of artificial intelligence and materials science presents a transformative opportunity, but achieving true acceleration in discovery requires moving beyond task-isolated, fine-tuned models toward agentic systems that plan, act, and learn across the full discovery loop. This survey advances a unique pipeline-centric view that spans from corpus curation and pretraining, through domain adaptation and instruction tuning, to goal-conditioned agents interfacing with simulation and experimental platforms. Unlike prior reviews, we treat the entire process as an end-to-end system to be optimized for tangible discovery outcomes rather than proxy benchmarks. This perspective allows us to trace how upstream design choices-such as data curation and training objectives-can be aligned with downstream experimental success through effective credit assignment. To bridge communities and establish a shared frame of reference, we first present an integrated lens that aligns terminology, evaluation, and workflow stages across AI and materials science. We then analyze the field through two focused lenses: From the AI perspective, the survey details LLM strengths in pattern recognition, predictive analytics, and natural language processing for literature mining, materials characterization, and property prediction; from the materials science perspective, it highlights applications in materials design, process optimization, and the acceleration of computational workflows via integration with external tools (e.g., DFT, robotic labs). Finally, we contrast passive, reactive approaches with agentic design, cataloging current contributions while motivating systems that pursue long-horizon goals with autonomy, memory, and tool use. This survey charts a practical roadmap towards autonomous, safety-aware LLM agents aimed at discovering novel and useful materials.
Authors: Meng Ding, Zeqing Zhang, Di Wang, Lijie Hu
Abstract: Reliable data attribution is essential for mitigating bias and reducing computational waste in modern machine learning, with the Shapley value serving as the theoretical gold standard. While recent "In-Run" methods bypass the prohibitive cost of retraining by estimating contributions dynamically, they heavily rely on the linear structure of Stochastic Gradient Descent (SGD) and fail to capture the complex dynamics of adaptive optimizers like Adam. In this work, we demonstrate that data attribution is inherently optimizer-dependent: we show that SGD-based proxies diverge significantly from true contributions under Adam (Pearson $R \approx 0.11$), rendering them ineffective for modern training pipelines. To bridge this gap, we propose Adam-Aware In-Run Data Shapley. We derive a closed-form approximation that restores additivity by redefining utility under a fixed-state assumption and enable scalable computation via a novel Linearized Ghost Approximation. This technique linearizes the variance-dependent scaling term, allowing us to compute pairwise gradient dot-products without materializing per-sample gradients. Extensive experiments show that our method achieves near-perfect fidelity to ground-truth marginal contributions ($R > 0.99$) while retaining $\sim$95\% of standard training throughput. Furthermore, our Adam-aware attribution significantly outperforms SGD-based baselines in data attribution downstream tasks.
Authors: Max Manolov, Tony Gao, Siddharth Shukla, Cheng-Ting Chou, Ryan Lagasse
Abstract: Large language models (LLMs) are increasingly used to assist developers with code, yet their implementations of cryptographic functionality often contain exploitable flaws. Minor design choices (e.g., static initialization vectors or missing authentication) can silently invalidate security guarantees. We introduce CIPHER(Cryptographic Insecurity Profiling via Hybrid Evaluation of Responses), a benchmark for measuring cryptographic vulnerability incidence in LLM-generated Python code under controlled security-guidance conditions. CIPHER uses insecure/neutral/secure prompt variants per task, a cryptography-specific vulnerability taxonomy, and line-level attribution via an automated scoring pipeline. Across a diverse set of widely used LLMs, we find that explicit secure prompting reduces some targeted issues but does not reliably eliminate cryptographic vulnerabilities overall. The benchmark and reproducible scoring pipeline will be publicly released upon publication.
Authors: Abhijit Gupta
Abstract: Drug discovery motivates efficient molecular property prediction under limited labeled data. Chemical space is vast, often estimated at approximately 10^60 drug-like molecules, while only thousands of drugs have been approved. As a result, self-supervised pretraining on large unlabeled molecular corpora has become essential for data-efficient molecular representation learning. We introduce **CardinalGraphFormer**, a graph transformer that incorporates Graphormer-inspired structural biases, including shortest-path distance and centrality, as well as direct-bond edge bias, within a structured sparse attention regime limited to shortest-path distance <= 3. The model further augments this design with a cardinality-preserving unnormalized aggregation channel over the same support set. Pretraining combines contrastive graph-level alignment with masked attribute reconstruction. Under a fully matched evaluation protocol, CardinalGraphFormer improves mean performance across all 11 evaluated tasks and achieves statistically significant gains on 10 of 11 public benchmarks spanning MoleculeNet, OGB, and TDC ADMET tasks when compared to strong reproduced baselines.
Authors: Atharv Sonwane, Eng-Shen Tu, Wei-Chung Lu, Claas Beger, Carter Larsen, Debjit Dhar, Simon Alford, Rachel Chen, Ronit Pattanayak, Tuan Anh Dang, Guohao Chen, Gloria Geng, Kevin Ellis, Saikat Dutta
Abstract: LLM-powered coding agents are redefining how real-world software is developed. To drive the research towards better coding agents, we require challenging benchmarks that can rigorously evaluate the ability of such agents to perform various software engineering tasks. However, popular coding benchmarks such as HumanEval and SWE-Bench focus on narrowly scoped tasks such as competition programming and patch generation. In reality, software engineers have to handle a broader set of tasks for real-world software development. To address this gap, we propose OmniCode, a novel software engineering benchmark that contains a broader and more diverse set of task categories beyond code or patch generation. Overall, OmniCode contains 1794 tasks spanning three programming languages (Python, Java, and C++) and four key categories: bug fixing, test generation, code review fixing, and style fixing. In contrast to prior software engineering benchmarks, the tasks in OmniCode are (1) manually validated to eliminate ill-defined problems, and (2) synthetically crafted or recently curated to avoid data leakage issues, presenting a new framework for synthetically generating diverse software tasks from limited real-world data. We evaluate OmniCode with popular agent frameworks such as SWE-Agent and show that while they may perform well on bug fixing for Python, they fall short on tasks such as Test Generation and in languages such as C++ and Java. For instance, SWE-Agent achieves a maximum of 20.9% with DeepSeek-V3.1 on Java Test Generation tasks. OmniCode aims to serve as a robust benchmark and spur the development of agents that can perform well across different aspects of software development. Code and data are available at https://github.com/seal-research/OmniCode.
Authors: Xianglong Yan, ChengZhu Bao, Zhiteng Li, Tianao Zhang, Shaoqiu Zhang, Ruobing Xie, Samm Sun, Yulun Zhang
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.
Authors: Ying Wang, Jiahui Chen, Dejun Jiang
Abstract: Storage systems are fundamental to modern computing infrastructures, yet ensuring their correctness remains challenging in practice. Despite decades of research on system testing, many storage-system failures (including durability, ordering, recovery, and consistency violations) remain difficult to expose systematically. This difficulty stems not primarily from insufficient testing tooling, but from intrinsic properties of storage-system execution, including nondeterministic interleavings, long-horizon state evolution, and correctness semantics that span multiple layers and execution phases. This survey adopts a storage-centric view of system testing and organizes existing techniques according to the execution properties and failure mechanisms they target. We review a broad spectrum of approaches, ranging from concurrency testing and long-running workloads to crash-consistency analysis, hardware-level semantic validation, and distributed fault injection, and analyze their fundamental strengths and limitations. Within this framework, we examine fuzzing as an automated testing paradigm, highlighting systematic mismatches between conventional fuzzing assumptions and storage-system semantics, and discuss how recent artificial intelligence advances may complement fuzzing through state-aware and semantic guidance. Overall, this survey provides a unified perspective on storage-system correctness testing and outlines key challenges
Authors: Zhiyuan Yao, Yi-Kai Zhang, Yuxin Chen, Yueqing Sun, Zishan Xu, Yu Yang, Tianhao Hu, Qi Gu, Hui Su, Xunliang Cai
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.
Authors: Basile Terver, Randall Balestriero, Megi Dervishi, David Fan, Quentin Garrido, Tushar Nagarajan, Koustuv Sinha, Wancong Zhang, Mike Rabbat, Yann LeCun, Amir Bar
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.
Authors: Chandrashekar M S, Vineet Singh, Lakshmi Pedapudi
Abstract: The digitization of agricultural advisory services in India requires robust Automatic Speech Recognition (ASR) systems capable of accurately transcribing domain-specific terminology in multiple Indian languages. This paper presents a benchmarking framework for evaluating ASR performance in agricultural contexts across Hindi, Telugu, and Odia languages. We introduce evaluation metrics including Agriculture Weighted Word Error Rate (AWWER) and domain-specific utility scoring to complement traditional metrics. Our evaluation of 10,934 audio recordings, each transcribed by up to 10 ASR models, reveals performance variations across languages and models, with Hindi achieving the best overall performance (WER: 16.2%) while Odia presents the greatest challenges (best WER: 35.1%, achieved only with speaker diarization). We characterize audio quality challenges inherent to real-world agricultural field recordings and demonstrate that speaker diarization with best-speaker selection can substantially reduce WER for multi-speaker recordings (upto 66% depending on the proportion of multi-speaker audio). We identify recurring error patterns in agricultural terminology and provide practical recommendations for improving ASR systems in low-resource agricultural domains. The study establishes baseline benchmarks for future agricultural ASR development.
Authors: Mehdi Lotfian, Mohammad Jalali, Farzan Farnia
Abstract: Prompt-guided generative AI models have rapidly expanded across vision and language domains, producing realistic and diverse outputs from textual inputs. The growing variety of such models, trained with different data and architectures, calls for principled methods to identify which types of prompts lead to distinct model behaviors. In this work, we propose PromptSplit, a kernel-based framework for detecting and analyzing prompt-dependent disagreement between generative models. For each compared model pair, PromptSplit constructs a joint prompt--output representation by forming tensor-product embeddings of the prompt and image (or text) features, and then computes the corresponding kernel covariance matrix. We utilize the eigenspace of the weighted difference between these matrices to identify the main directions of behavioral difference across prompts. To ensure scalability, we employ a random-projection approximation that reduces computational complexity to $O(nr^2 + r^3)$ for projection dimension $r$. We further provide a theoretical analysis showing that this approximation yields an eigenstructure estimate whose expected deviation from the full-dimensional result is bounded by $O(1/r^2)$. Experiments across text-to-image, text-to-text, and image-captioning settings demonstrate that PromptSplit accurately detects ground-truth behavioral differences and isolates the prompts responsible, offering an interpretable tool for detecting where generative models disagree.
Authors: Julian Rodemann, Unai Fischer-Abaigar, James Bailie, Krikamol Muandet
Abstract: Performative predictions influence the very outcomes they aim to forecast. We study performative predictions that affect a sample (e.g., only existing users of an app) and/or the whole population (e.g., all potential app users). This raises the question of how well models generalize under performativity. For example, how well can we draw insights about new app users based on existing users when both of them react to the app's predictions? We address this question by embedding performative predictions into statistical learning theory. We prove generalization bounds under performative effects on the sample, on the population, and on both. A key intuition behind our proofs is that in the worst case, the population negates predictions, while the sample deceptively fulfills them. We cast such self-negating and self-fulfilling predictions as min-max and min-min risk functionals in Wasserstein space, respectively. Our analysis reveals a fundamental trade-off between performatively changing the world and learning from it: the more a model affects data, the less it can learn from it. Moreover, our analysis results in a surprising insight on how to improve generalization guarantees by retraining on performatively distorted samples. We illustrate our bounds in a case study on prediction-informed assignments of unemployed German residents to job trainings, drawing upon administrative labor market records from 1975 to 2017 in Germany.
Authors: Sijia Chen, Lijuan Ma, Yanqiu Yu, En Yu, Liman Liu, Wenbing Tao
Abstract: Referring Multi-Object Tracking (RMOT) aims to track specific targets based on language descriptions and is vital for interactive AI systems such as robotics and autonomous driving. However, existing RMOT models rely solely on 2D RGB data, making it challenging to accurately detect and associate targets characterized by complex spatial semantics (e.g., ``the person closest to the camera'') and to maintain reliable identities under severe occlusion, due to the absence of explicit 3D spatial information. In this work, we propose a novel task, RGBD Referring Multi-Object Tracking (DRMOT), which explicitly requires models to fuse RGB, Depth (D), and Language (L) modalities to achieve 3D-aware tracking. To advance research on the DRMOT task, we construct a tailored RGBD referring multi-object tracking dataset, named DRSet, designed to evaluate models' spatial-semantic grounding and tracking capabilities. Specifically, DRSet contains RGB images and depth maps from 187 scenes, along with 240 language descriptions, among which 56 descriptions incorporate depth-related information. Furthermore, we propose DRTrack, a MLLM-guided depth-referring tracking framework. DRTrack performs depth-aware target grounding from joint RGB-D-L inputs and enforces robust trajectory association by incorporating depth cues. Extensive experiments on the DRSet dataset demonstrate the effectiveness of our framework.
Authors: Itai Zilberstein, Ioannis Anagnostides, Zachary W. Sollie, Arman Kilic, Tuomas Sandholm
Abstract: Online matching has been a mainstay in domains such as Internet advertising and organ allocation, but practical algorithms often lack strong theoretical guarantees. We take an important step toward addressing this by developing new online matching algorithms based on a coarsening approach. Although coarsening typically implies a loss of granularity, we show that, to the contrary, aggregating offline nodes into capacitated clusters can yield near-optimal theoretical guarantees. We apply our methodology to heart transplant allocation to develop theoretically grounded policies based on structural properties of historical data. Furthermore, in simulations based on real data, our policy closely matches the performance of the omniscient benchmark, achieving competitive ratio 0.91, drastically higher than the US status quo policy's 0.51. Our work bridges the gap between data-driven heuristics and pessimistic theoretical lower bounds.
Authors: Georgii Aparin, Tasnima Sadekova, Alexey Rukhovich, Assel Yermekova, Laida Kushnareva, Vadim Popov, Kristian Kuznetsov, Irina Piontkovskaya
Abstract: Sparse Autoencoders (SAEs) are powerful tools for interpreting neural representations, yet their use in audio remains underexplored. We train SAEs across all encoder layers of Whisper and HuBERT, provide an extensive evaluation of their stability, interpretability, and show their practical utility. Over 50% of the features remain consistent across random seeds, and reconstruction quality is preserved. SAE features capture general acoustic and semantic information as well as specific events, including environmental noises and paralinguistic sounds (e.g. laughter, whispering) and disentangle them effectively, requiring removal of only 19-27% of features to erase a concept. Feature steering reduces Whisper's false speech detections by 70% with negligible WER increase, demonstrating real-world applicability. Finally, we find SAE features correlated with human EEG activity during speech perception, indicating alignment with human neural processing. The code and checkpoints are available at https://github.com/audiosae/audiosae_demo.
Authors: John Yan, Michael Yu, Yuqi Sun, Alexander Duffy, Tyler Marques, Matthew Lyle Olson
Abstract: Large language models (LLMs) are increasingly trained in complex Reinforcement Learning, multi-agent environments, making it difficult to understand how behavior changes over training. Sparse Autoencoders (SAEs) have recently shown to be useful for data-centric interpretability. In this work, we analyze large-scale reinforcement learning training runs from the sophisticated environment of Full-Press Diplomacy by applying pretrained SAEs, alongside LLM-summarizer methods. We introduce Meta-Autointerp, a method for grouping SAE features into interpretable hypotheses about training dynamics. We discover fine-grained behaviors including role-playing patterns, degenerate outputs, language switching, alongside high-level strategic behaviors and environment-specific bugs. Through automated evaluation, we validate that 90% of discovered SAE Meta-Features are significant, and find a surprising reward hacking behavior. However, through two user studies, we find that even subjectively interesting and seemingly helpful SAE features may be worse than useless to humans, along with most LLM generated hypotheses. However, a subset of SAE-derived hypotheses are predictively useful for downstream tasks. We further provide validation by augmenting an untrained agent's system prompt, improving the score by +14.2%. Overall, we show that SAEs and LLM-summarizer provide complementary views into agent behavior, and together our framework forms a practical starting point for future data-centric interpretability work on ensuring trustworthy LLM behavior throughout training.
Authors: Zhuokun Chen, Jianfei Cai, Bohan Zhuang
Abstract: Generating long-form content, such as minute-long videos and extended texts, is increasingly important for modern generative models. Block diffusion improves inference efficiency via KV caching and block-wise causal inference and has been widely adopted in diffusion language models and video generation. However, in long-context settings, block diffusion still incurs substantial overhead from repeatedly computing attention over a growing KV cache. We identify an underexplored property of block diffusion: cross-step redundancy of attention within a block. Our analysis shows that attention outputs from tokens outside the current block remain largely stable across diffusion steps, while block-internal attention varies significantly. Based on this observation, we propose FlashBlock, a cached block-external attention mechanism that reuses stable attention output, reducing attention computation and KV cache access without modifying the diffusion process. Moreover, FlashBlock is orthogonal to sparse attention and can be combined as a complementary residual reuse strategy, substantially improving model accuracy under aggressive sparsification. Experiments on diffusion language models and video generation demonstrate up to 1.44$\times$ higher token throughput and up to 1.6$\times$ reduction in attention time, with negligible impact on generation quality. Project page: https://caesarhhh.github.io/FlashBlock/.
Authors: Zhenxiong Yu, Zhi Yang, Zhiheng Jin, Shuhe Wang, Heng Zhang, Yanlin Fei, Lingfeng Zeng, Fangqi Lou, Shuo Zhang, Tu Hu, Jingping Liu, Rongze Chen, Xingyu Zhu, Kunyi Wang, Chaofa Yuan, Xin Guo, Zhaowei Liu, Feipeng Zhang, Jie Huang, Huacan Wang, Ronghao Chen, Liwen Zhang
Abstract: As large language models (LLMs) evolve into autonomous agents, their real-world applicability has expanded significantly, accompanied by new security challenges. Most existing agent defense mechanisms adopt a mandatory checking paradigm, in which security validation is forcibly triggered at predefined stages of the agent lifecycle. In this work, we argue that effective agent security should be intrinsic and selective rather than architecturally decoupled and mandatory. We propose Spider-Sense framework, an event-driven defense framework based on Intrinsic Risk Sensing (IRS), which allows agents to maintain latent vigilance and trigger defenses only upon risk perception. Once triggered, the Spider-Sense invokes a hierarchical defence mechanism that trades off efficiency and precision: it resolves known patterns via lightweight similarity matching while escalating ambiguous cases to deep internal reasoning, thereby eliminating reliance on external models. To facilitate rigorous evaluation, we introduce S$^2$Bench, a lifecycle-aware benchmark featuring realistic tool execution and multi-stage attacks. Extensive experiments demonstrate that Spider-Sense achieves competitive or superior defense performance, attaining the lowest Attack Success Rate (ASR) and False Positive Rate (FPR), with only a marginal latency overhead of 8.3\%.
Authors: Chang Zou, Changlin Li, Yang Li, Patrol Li, Jianbing Wu, Xiao He, Songtao Liu, Zhao Zhong, Kailin Huang, Linfeng Zhang
Abstract: While diffusion models have achieved great success in the field of video generation, this progress is accompanied by a rapidly escalating computational burden. Among the existing acceleration methods, Feature Caching is popular due to its training-free property and considerable speedup performance, but it inevitably faces semantic and detail drop with further compression. Another widely adopted method, training-aware step-distillation, though successful in image generation, also faces drastic degradation in video generation with a few steps. Furthermore, the quality loss becomes more severe when simply applying training-free feature caching to the step-distilled models, due to the sparser sampling steps. This paper novelly introduces a distillation-compatible learnable feature caching mechanism for the first time. We employ a lightweight learnable neural predictor instead of traditional training-free heuristics for diffusion models, enabling a more accurate capture of the high-dimensional feature evolution process. Furthermore, we explore the challenges of highly compressed distillation on large-scale video models and propose a conservative Restricted MeanFlow approach to achieve more stable and lossless distillation. By undertaking these initiatives, we further push the acceleration boundaries to $11.8\times$ while preserving generation quality. Extensive experiments demonstrate the effectiveness of our method. The code will be made publicly available soon.
Authors: Pavithren V S Pakianathan, Rania Islambouli, Diogo Branco, Albrecht Schmidt, Tiago Guerreiro, Jan David Smeddinck
Abstract: Individuals are increasingly generating substantial personal health and lifestyle data, e.g. through wearables and smartphones. While such data could transform preventative care, its integration into clinical practice is hindered by its scale, heterogeneity and the time pressure and data literacy of healthcare professionals (HCPs). We explore how large language models (LLMs) can support sensemaking of patient-generated health data (PGHD) with automated summaries and natural language data exploration. Using cardiovascular disease (CVD) risk reduction as a use case, 16 HCPs reviewed multimodal PGHD in a mixed-methods study with a prototype that integrated common charts, LLM-generated summaries, and a conversational interface. Findings show that AI summaries provided quick overviews that anchored exploration, while conversational interaction supported flexible analysis and bridged data-literacy gaps. However, HCPs raised concerns about transparency, privacy, and overreliance. We contribute empirical insights and sociotechnical design implications for integrating AI-driven summarization and conversation into clinical workflows to support PGHD sensemaking.
Authors: Lixuan Guo, Yifei Wang, Tiansheng Wen, Yifan Wang, Aosong Feng, Bo Chen, Stefanie Jegelka, Chenyu You
Abstract: In the era of large foundation models, the quality of embeddings has become a central determinant of downstream task performance and overall system capability. Yet widely used dense embeddings are often extremely high-dimensional, incurring substantial costs in storage, memory, and inference latency. To address these, Contrastive Sparse Representation (CSR) is recently proposed as a promising direction, mapping dense embeddings into high-dimensional but k-sparse vectors, in contrast to compact dense embeddings such as Matryoshka Representation Learning (MRL). Despite its promise, CSR suffers severe degradation in the ultra-sparse regime, where over 80% of neurons remain inactive, leaving much of its efficiency potential unrealized. In this paper, we introduce CSRv2, a principled training approach designed to make ultra-sparse embeddings viable. CSRv2 stabilizes sparsity learning through progressive k-annealing, enhances representational quality via supervised contrastive objectives, and ensures end-to-end adaptability with full backbone finetuning. CSRv2 reduces dead neurons from 80% to 20% and delivers a 14% accuracy gain at k=2, bringing ultra-sparse embeddings on par with CSR at k=8 and MRL at 32 dimensions, all with only two active features. While maintaining comparable performance, CSRv2 delivers a 7x speedup over MRL, and yields up to 300x improvements in compute and memory efficiency relative to dense embeddings in text representation. Extensive experiments across text and vision demonstrate that CSRv2 makes ultra-sparse embeddings practical without compromising performance, where CSRv2 achieves 7%/4% improvement over CSR when k=4 and further increases this gap to 14%/6% when k=2 in text/vision representation. By making extreme sparsity viable, CSRv2 broadens the design space for real-time and edge-deployable AI systems where both embedding quality and efficiency are critical.
Authors: Seonghye Cho, Jaemin Han, Hyunjin Kim, Euisoo Jung, Jae-Gil Lee
Abstract: Pipeline parallelism enables training models that exceed single-device memory, but practical throughput remains limited by pipeline bubbles. Although parameter freezing can improve training throughput by adaptively skipping backward computation, existing methods often over-freeze parameters, resulting in unnecessary accuracy degradation. To address this issue, we propose TimelyFreeze, which models the pipeline schedule as a directed acyclic graph and solves a linear program to compute optimal freeze ratios that minimize batch execution time under accuracy constraints. Experiments show that TimelyFreeze achieves up to 40% training throughput improvement on LLaMA-8B with comparable accuracy. Overall, it enables faster large-scale model training without compromising convergence and generalizes across diverse pipeline-parallel settings.
Authors: Wei Liu, Jiawei Xu, Yingru Li, Longtao Zheng, Tianjian Li, Qian Liu, Junxian He
Abstract: High-quality kernel is critical for scalable AI systems, and enabling LLMs to generate such code would advance AI development. However, training LLMs for this task requires sufficient data, a robust environment, and the process is often vulnerable to reward hacking and lazy optimization. In these cases, models may hack training rewards and prioritize trivial correctness over meaningful speedup. In this paper, we systematically study reinforcement learning (RL) for kernel generation. We first design KernelGYM, a robust distributed GPU environment that supports reward hacking check, data collection from multi-turn interactions and long-term RL training. Building on KernelGYM, we investigate effective multi-turn RL methods and identify a biased policy gradient issue caused by self-inclusion in GRPO. To solve this, we propose Turn-level Reinforce-Leave-One-Out (TRLOO) to provide unbiased advantage estimation for multi-turn RL. To alleviate lazy optimization, we incorporate mismatch correction for training stability and introduce Profiling-based Rewards (PR) and Profiling-based Rejection Sampling (PRS) to overcome the issue. The trained model, Dr Kernel-14B, reaches performance competitive with Claude-4.5-Sonnet in Kernelbench. Finally, we study sequential test-time scaling for Dr Kernel-14B. On the KernelBench Level-2 subset, 31.6% of the generated kernels achieve at least a 1.2x speedup over the Torch reference, surpassing Claude-4.5-Sonnet (26.7%) and GPT-5 (28.6%). When selecting the best candidate across all turns, this 1.2x speedup rate further increases to 47.8%. All resources, including environment, training code, models, and dataset, are included in https://www.github.com/hkust-nlp/KernelGYM.