Enhan Zhao, Yongqi Tong, Xuejun Yu, Hang Wang, Pan Wang, Yichao Wang, Di He, Jianshe Li 等 10 位作者
Existing long-horizon agent benchmarks often emphasize final task success while providing limited assessment of intermediate execution validity. Yet real-world business operations require agents to navigate extensive, potentially outdated or conflicting policies while satisfying critical constraints throughout execution. These requirements are particularly consequential in high-stakes domains such as payments, where intermediate violations can incur financial losses or regulatory consequences regardless of the final outcome. To bridge this evaluation gap, we introduce BizAgentBench, a multi-turn benchmark grounded in real-world procedures from a leading payments company. Its executable environment integrates 789 knowledge documents and 601 discoverable business tools, with consistent policies, states, and tool behavior. The benchmark comprises 50 tasks involving resource constraints, temporal dependencies, prior commitments, and conflicting or outdated guidance. Each task is validated through reference execution, while deterministic checks assess both goal attainment and execution validity without prescribing a single solution path. Experiments reveal substantial limitations even among frontier models: Claude Opus 5 and GPT-6 Sol achieve strict success rates of only 64.0% and 61.3%, respectively. For Qwen3.8-27B, goal attainment reaches 42.0%, but strict success drops to 32.7%, demonstrating how outcome-based evaluation can mask invalid execution. Beyond success rates, DeepSeek-V4.1-Flash averages 191.3 tool calls per run, approximately 70% more than reference executions, revealing substantial execution overhead. Overall, BizAgentBench establishes a testbed for evaluating and advancing long-horizon agents beyond task completion, toward reliable and constraint-compliant execution in complex business environments.
Fazong Wu, Xin Wang, Ming Yang, Zhenyong Zhang, Xiaoming Wu
As large language models (LLMs) are increasingly deployed as general-purpose conversational assistants, they must handle diverse user requests while reliably refusing harmful ones. Jailbreak attacks craft adversarial instructions or conversational contexts to elicit unsafe responses despite these safeguards. However, existing jailbreak benchmarks often rely on manually designed or collected prompts, making it difficult to systematically extend coverage to new threat conditions. Moreover, existing attack search methods typically optimize for attack success rate (ASR), providing limited insight into how model behavior changes locally between safe refusal and unsafe response. To address these limitations, we present JailBound, a unified jailbreak evaluation framework that combines automated benchmark construction with first-order loss (FOL)-guided dual-branch search. JailBound organizes threats by risk category, application domain, and attack type. Conditioned on this taxonomy, a fine-tuned meta-attack generator synthesizes evaluation prompts across diverse threat conditions. FOL serves as a local sensitivity signal for exploring high-value attack states and behavior-sensitive states. We evaluate JailBound across 46 LLMs from 13 model families under a unified experimental protocol. Under matched settings, the High-Value Branch improves average ASR over a ZOL-only baseline by 5.1 and 4.9 percentage points across the two evaluation settings. Under local perturbations, states from the Safety-Sensitivity Branch exhibit higher behavioral flip rates than those from the High-Value Branch across all tested radii for JailBound and S-Eval, indicating greater local sensitivity of judge-assigned safety labels. Warning: this paper includes examples that may be offensive or harmful.
Visual geo-localization requires complex reasoning to infer image locations without GPS metadata. Traditional retrieval methods are limited by database quality while Large Vision Language Models often struggle with diverse scenes. Although existing multi-agent systems improve collaboration, they treat interactions uniformly and lack effective conflict resolution. We propose GraphGeo, a multi-agent debate framework based on heterogeneous graph neural networks for visual geo-localization. The framework models diverse debate relationships through typed edges and separates supportive collaboration, competitive argumentation and knowledge transfer. We introduce a dual-level debate mechanism combining node-level refinement and edge-level argumentation modeling. A cross-level topology refinement strategy enables co-evolution between graph structure and agent representations. Experiments on multiple benchmarks demonstrate GraphGeo outperforms single model and multi-agent baselines. The framework converts cognitive conflicts between agents into higher geo-localization accuracy through structured debate.
AI 速览：提出上下文感知的自回归图像生成框架，让模型依据全局上下文与当前局部信息动态决定下一个 token 的位置，而非固定或随机顺序。
Autoregressive image generation has recently emerged as a competitive visual generation paradigm, yet its generation order is often predefined or heuristic, such as next-patch prediction, random-order prediction, or one-step generation for specific scales. We observe that these strategies can cause models to overlook important structural dependencies in images, thereby limiting both structural robustness and reasoning capability. In this work, we propose a context-aware autoregressive generation framework, where the model dynamically determines the next token location based on global context and current local information. Rather than following a fixed or random order, our method allows the generation process to adapt to image structure and semantic dependencies. Extensive experiments show that context-aware generation significantly improves structural robustness across different image AR paradigms, including masked autoregressive and scale-wise autoregressive models. Beyond quantitative improvements, we further observe emerging reasoning properties: the model can better understand complex semantic instructions and generate images that more faithfully follow the given guidance. These findings suggest that generation order is a critical yet underexplored factor in image autoregressive modeling, and that context-aware decoding can unlock stronger reasoning-aware visual generation capabilities. Code is available at https://github.com/Hpesojyjx/Saccades-AR.
Generalist robot policies are trained on static offline datasets, yet physical manipulation unfolds sequentially across stages. Because a weakness in a downstream stage remains unobservable until the policy passes earlier ones, each resolved bottleneck exposes the next hidden failure mode. This staged structure calls for Recursive Self-Improvement (RSI), in which a policy iteratively executes, diagnoses failures, and updates itself. Moreover, because task completion behaves like a product of stage-level pass rates, recursive gains can multiply across generations. In physical environments, three bottlenecks obstruct compounding: terminal execution feedback rarely localizes the stage that caused a failure, failed trajectories do not supply the corrective action, and updating shared parameters can degrade previously mastered skills. We propose Compounding Physical RSI, a self-evolving policy framework that turns execution failures into heritable capability. First, automated capability-gap assessment traces failed rollouts to the task stage where progression breaks down and formulates a structured improvement objective. Second, conditioned on this objective, our remedy stage curates validated, stage-relevant experience and refines the policy via behavior-regularized offline reinforcement learning, without requiring costly online rollouts at each optimization step. Third, to safeguard existing skills, our retention stage constructs weight-space candidates between the inherited reference policy and the refined policy, evaluates them across the task suite, and promotes the best candidate as the reference state for the next generation. Evaluated across all 42 tasks of the RoboDojo-Sim benchmark, CSU-AI-07 ranks first on the leaderboard with a three-seed mean Overall Score of 26.15 and Success Rate of 20.70%, outperforming the next-best entry by 5.92 points and 5.82 percentage points.
Reinforcement learning (RL) post-training has become an important paradigm for improving multimodal large language models (MLLMs), especially with automatically verifiable rewards. Self-supervised visual pretext tasks provide such rewards without human annotations. However, existing methods mainly exploit within-instance structure, offering limited supervision for cross-instance discrimination and fine-grained comparison. We introduce Group-and-Order Self-Supervised Reinforcement Learning (GO-SSL), a cross-instance jigsaw framework for vision-centric MLLM post-training. Given two visual instances, GO-SSL mixes their local elements and trains the model to group them by instance of origin while recovering the order within each group, coupling instance-level comparison with spatial, temporal, or geometric structure recovery. This paired formulation turns the same unlabeled data into richer verifiable supervision through diverse pairings, shuffles, and a random-to-hard curriculum. We instantiate GO-SSL mainly on cross-image jigsaw tasks and further extend it to 3D depth and video temporal ordering. Across diverse benchmarks, GO-SSL consistently improves over the base MLLM and prior self-supervised post-training baselines, with clear gains in fine-grained perception and spatial understanding. Further evaluations across model scales and families support its generality. These results suggest that cross-instance comparison provides a useful self-supervised RL signal for improving transferable visual understanding in MLLMs. Code is available at https://github.com/SEU-VIPGroup/GO-SSL.
Large language model (LLM)-based multi-agent systems (MAS) can solve complex problems beyond the reach of a single agent by decomposing tasks and coordinating complementary capabilities. Despite extensive research on MAS, existing work largely focuses on specific aspects or individual phases of their development and operation, resulting in fragmented perspectives with objectives and signals that are difficult to reconcile across these different settings. To bridge these disconnected perspectives, we conceptualize MAS development and operation as a three-phase lifecycle spanning design-time orchestration before execution, runtime intervention during execution, and post-hoc attribution after execution, and introduce MAS-Gate, a unified node-centric quantification applicable across these phases. Modeling each execution as a directed acyclic graph, MAS-Gate combines within-node utility, which estimates a node's contribution through changes in reference-answer likelihood, with inter-node strength, which approximates semantic uptake and potential downstream reach. We adapt this shared quantification to three objectives: providing a structure-aware dense reward for Orchestrator training, training an Editor to act on execution prefixes using paired rewards from completed edited and unedited trajectories, and supplying risk-ranked nodes as auxiliary evidence for LLM-based post-hoc attribution. Experiments spanning interaction environments, knowledge-intensive multi-hop question answering, general reasoning, and post-hoc attribution demonstrate that MAS-Gate improves workflow construction, targeted runtime editing, and LLM-assisted error localization over corresponding phase-specific baselines. Together, these results establish MAS-Gate as a unified structural interface for improving MAS generation, dynamic editing, and error attribution throughout the lifecycle.
Explicit reasoning in large language models (LLMs) is often associated with reliability, yet it can still pose risks when a model infers harmful goals under a benign context. We identify a distinct challenge: ensuring the safety of the \emph{multi-step reasoning process} itself, where locally coherent steps may still lead to globally harmful outcomes. To explore this vulnerability, we propose PDRA (Paradox-driven Reasoning-based Attack), a novel attack method that programs a harmful reasoning trajectory. PDRA first constructs a paradoxical narrative core by pairing a malicious actor (e.g., a terrorist) with a benign task (e.g., writing a safety bulletin). It then embeds this core into a structured analytical framework that forces the model through a mandatory three-stage analysis -- surface purpose, latent intent, and operational reconstruction -- guiding it to produce detailed harmful content while maintaining narrative coherence. Mechanistically, we show PDRA steers the model's internal representations away from refusal-related patterns, suppressing safety-triggering lexicons while activating planning-oriented terms. Empirically, PDRA achieves state-of-the-art attack success rates across diverse LLMs with high efficiency. \textcolor{red}{\textbf{Warning:} This paper may contain sensitive content.}
Ioannis Anagnostides, Ioannis Panageas, Jingming Yan
The geometry of constraints in variational inequalities (VIs) plays a central role in applications ranging from resource allocation and equilibrium computation to constrained optimization. Yet, its effect on computational complexity has received surprisingly little attention beyond special cases. We examine this question when the constraint set is an $\ell_p$ ball. This family of problems encompasses two of the most extensively studied classes of problems in algorithmic game theory: symmetric bimatrix games for $p = 1$ and multi-player polymatrix games for $p = \infty$. Both of these problems are known to be PPAD-complete even for affine operators. We establish a comprehensive complexity landscape for VIs over $\ell_p$ balls across the full range $p\in[1,\infty]$. First, for any $p\neq 2$, we show that finding an inverse-polynomial approximate VI solution is PPAD-complete even for affine operators. In contrast, for $p=2$, we give a polynomial-time algorithm for affine VIs, extending a classic result of Ye (1992) beyond symmetric operators. On the other hand, once the operator becomes quadratic, we establish PPAD-completeness even at inverse polynomial precision. From a technical standpoint, we develop a new toolkit for exploiting the geometry of the constraint set and the structure of the operator to characterize the complexity of equilibrium and optimization problems, which is of independent interest.
Shuimu Chen, Jing Jin, Nan Su, Hongbo Xu, Zebang Cheng, Wenming Yang, Fei Ma, Guijin Wang
Large vision-language models (LVLMs) have shown strong potential for industrial anomaly detection (IAD) by providing image-level anomaly judgments and interpretable reasoning. However, reliably translating generated judgments into precise pixel-level localization remains challenging. We propose \textbf{M}ixed-Trajectory \textbf{O}n-\textbf{P}olicy \textbf{D}istillation for Language-Guided Industrial \textbf{A}nomaly Detection (MOPDA), the first framework to introduce on-policy self-distillation into LVLM-based IAD. For judgment learning, \method introduces \textbf{Mixed-Trajectory Supervision}, combining student-generated on-policy trajectories with evidence-conditioned teacher trajectories under a shared token-level distillation objective. Student trajectories preserve supervision on deployment-relevant response paths, while teacher trajectories provide complementary evidence-conditioned supervision. For dense localization, \textbf{Language-guided Visual Anchoring} uses the final judgment as a compact semantic condition to construct image-specific normal and abnormal anchors, which are contrasted with dense visual features to produce anomaly maps. This keeps language as semantic guidance while grounding pixel-level responses in visual evidence. Under a strict cross-dataset zero-shot protocol on five IAD benchmarks, \method outperforms the evaluated LVLM-based baselines on most detection, localization, and judgment metrics while remaining competitive with CLIP-based methods. Ablations further validate both mixed-trajectory supervision and final-judgment conditioning.
A GUI command can be correct at the interface level yet still produce an incorrect assembly. An agent must identify the relevant components, respect their structural relationships, and determine which parts should move during recovery. We introduce MAGAR-Bench, a benchmark of 4,000 multimodal samples from 26 assembly families that separates Assembly Decision, Atomic Action, and Minimal Rework. For rework, the objective counts distinct moved parts within the validated recovery candidate set for each state, rather than the number of actions. We also introduce MAGAR, a selective correction framework built on a frozen policy obtained by supervised fine-tuning. Its key design matches correction granularity to output dependencies. Structured candidates revise assembly fields and rework scope, while complete action candidates keep native GUI primitives coupled with their parameters. A learned selector retains the original prediction unless a candidate passes a task-specific intervention threshold. On recorded states from held-out assembly families, MAGAR improves Overall, the task-balanced aggregate, from 37.50 to 41.39, with gains on all three tasks. Across 209 interventions, 70 correct an error and 22 introduce one. These results support task-specific correction while highlighting the remaining tradeoff between fixing errors and preserving reliable predictions.
Computational imaging can reduce the acquisition burden when the system is designed around task-relevant information rather than complete image recovery. We investigate this framework for sparse-view planar X-ray tomography (XCT) of layered objects, using measurement of the axial termination gap (ATG) of back-drilled printed-circuit-board (PCB) vias as a representative task. Instead of using 3D reconstruction, the proposed method estimates ATG directly from polychromatic projection data through a physics-based sensitivity model. A gated hierarchical estimator first fits ATG together with radiometric perturbations and selectively introduces global geometry and specimen-warp models only when a reduced chi-squared adequacy test and Bayesian information criterion justify the added degrees of freedom. The same low-dimensional perturbation model is incorporated into marginalized Fisher-information analyses, yielding task-specific conditional and calibration-aware Cram'er-Rao bounds. Numerical experiments show that, under a fixed total photon budget, task-driven metrology changes little over 12-64 views, while reconstruction-driven metrology is strongly view-dependent with much larger errors. The direct estimator also follows the expected photon-limited scaling. These results demonstrate a computational sensing strategy in which XCT acquisition, physical modeling, and inference are evaluated according to the quantitative information required by the task rather than for general-purpose volumetric image fidelity.
Yiming Yan, Hongyang Wang, Wenbo Chen, Xuan Wang, Jingguo Liu, Xian Li, Chen Qi, Leo Fan 等 15 位作者
Agentic reinforcement learning (RL) often requires repeated tool calls in task-specific environments. A training job may create thousands of these environments in bursts while sharing a cluster with other jobs. We present Sandman, a sandbox service built and operated at Amazon for agent tool execution and evaluation. It serves multiple production agentic RL workloads on shared Kubernetes. A training framework supplies a task image and uses commands and file operations in the returned sandbox. Files and processes persist across calls until the framework releases the environment. Sandman manages creation, readiness and cleanup. Each sandbox is a Kubernetes pod with a small daemon mounted into the unmodified image. In a scale test, Sandman held 11,858 concurrent tiny sandboxes. Four production SWE-bench Verified training campaigns created 3.1 million sandboxes in October 2026. We report the design trade-offs and operational lessons from handling creation bursts, preparing task environments, reclaiming abandoned sandboxes, and running evaluation under controlled conditions. We will open-source Sandman's daemon, provider and pool.
Generative recommendation (GR) represents items as semantic identifiers (SIDs) and autoregressively generates the next item's SID from interaction history. However, most methods assign the same SID length to all items despite highly unequal interaction frequencies. This uniform design repeatedly incurs the cost of fine-grained codes for popular items, while shortening codes to reduce this cost also compresses the semantic representations of long-tail items. Inspired by the compact forms of frequent expressions in natural language, we propose JevRec to encode items as variable-length words and generate the next item from interaction context. Linguistic encoding retains frequency-dependent semantic prefixes from shared codebooks, producing shorter SIDs for popular items and richer descriptions for long-tail items. Yet predicting a sequence-level length and generating token-level content impose different optimization requirements that may conflict under a shared output head. Our Jev-like decision modeling component therefore predicts a length distribution with a dedicated head and generates content conditioned on length through a separate head while sharing the interaction-history representation. Experiments on three public datasets show that JevRec outperforms all compared baselines across the evaluated recommendation metrics. Further analyses show lower training and inference time than TIGER and reduced gradient conflict between length and content prediction.
Recent advancements in text-guided 3D editing, particularly with 3D Gaussian Splatting (3DGS), have shown promise but still lag behind the flexibility of 2D editing tools. Existing methods face a critical trade-off between preserving multi-view consistency and allowing structural modifications. Image-based editors can also struggle to apply the intended edit to objects viewed from unusual angles. To address these limitations, we introduce VidEdit3D, a framework that reformulates 3D scene editing as a trajectory-based video editing task. By drawing a parallel between camera movement in static 3D scenes and temporal dynamics in video, we leverage the robust spatiotemporal priors of pre-trained Video Diffusion Models (VDMs) to encourage semantic and structural coherence across views. Specifically, we propose Trajectory-Guided Rendering (TGR) to select informative keyframes and construct smooth video clips aligned with VDM input requirements. To propagate edits from an anchor view to these clips, we develop Flow-Aligned Video Editing (FAVE), a training-free framework incorporating an Editing Consistency Guidance (ECG) mechanism. ECG contrasts source and edited image conditions at a shared latent state, using scheduled source-image amplification to strengthen adherence to the edited reference while maintaining original scene context. Finally, we introduce Region-Selective 3DGS Refinement (RS-GS) to strictly confine geometric and appearance updates to target primitives, reducing background attribute leakage. Experiments demonstrate stronger instruction alignment and better multi-view consistency than the baselines.
Andreas Kontogiannis, Ioannis Panageas, Vasilis Pollatos, Jingming Yan
Every nondegenerate bimatrix game has a Nash equilibrium of Shapley index +1, since all equilibria are isolated, have index +1 or -1, and their indices sum to +1. We prove that the following promise search problem is PPADS-complete: given a rational bimatrix game promised to be nondegenerate, find an exact Nash equilibrium of index +1. The nondegeneracy promise can be removed if nonregular equilibria are also accepted: finding either a regular Nash equilibrium of index +1 or a nonregular Nash equilibrium in an arbitrary rational bimatrix game is also PPADS-complete. To our knowledge, this is the first PPADS-complete equilibrium search problem whose instances are explicit rational normal form payoff matrices, rather than succinct circuits or Turing machines, and thereby addresses an open question posed by Daskalakis [2019].
Traditional industrial agents rely on modular pipelines, including Router, Retriever, Planner, Executor, Responder, Reviewer and so on, which inevitably fracture into a labyrinth of ad-hoc patches, leading to cascading errors and high latency. We propose OneModel, an applicable paradigm shift from external workflows to internalized knowledge representation. Unlike modular systems that slice fluid user intents into static steps, OneModel consolidates complex business logic and SOPs directly into the model’s parameters.Through Continual Pre-training (CPT) and logic-compilation SFT, we transform fragmented business rules into the model’s intuitive reasoning within a unified attention space. Deployed in our global financial service system, OneModel effectively breaks the impossible triangle of latency, accuracy, and complexity. Online A/B testing demonstrates end-to-end latency reduction of more than 50% (18.7s \rightarrow 8s) while the Intelligent Resolution Rate (IRR) jumps from 64.3% to 83.3%. The results demonstrate our paradigm OneModel effectively replaces brittle engineering logic with internalized cognitive intuition, offering a scalable and future-proof blueprint for transitioning industrial agents from complex, error-prone workflow to unified model architectures.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track) · 2026 · 正式发表
Yang Xiang, Yixin Ji Soochow University, Ruotao Xu, Dan Qiao Suzhou University, Zheming Yang ByteDance Inc., Juntao Li, Min Zhang Harbin Institute of Technology, Shenzhen and Harbin Institute of Technology
Large reasoning models (LRMs) have achieved remarkable performance in complex reasoning tasks, driven by their powerful inference-time scaling capability.However, LRMs often suffer from overthinking, which results in substantial computational redundancy and significantly reduces efficiency.Early-exit methods aim to mitigate this issue by terminating reasoning once sufficient evidence has been generated, yet existing approaches mostly rely on handcrafted or empirical indicators that are unreliable and impractical.In this work, we introduce Dynamic Thought Sufficiency in Reasoning (DTSR), a novel framework for efficient reasoning that enables the model to dynamically assess the sufficiency of its chain-of-thought (CoT) and determine the optimal point for early exit.Inspired by human metacognition, DTSR operates in two stages: (1) Reflection Signal Monitoring, which identifies reflection signals as potential cues for early exit, and (2) Thought Sufficiency Check, which evaluates whether the current CoT is sufficient to derive the final answer.Experimental results on the Qwen3 models show that DTSR reduces reasoning length by 28.9%–34.9% with minimal performance loss, effectively mitigating overthinking.We further discuss overconfidence in LRMs and self-evaluation paradigms, providing valuable insights for early-exit reasoning.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Fuwen Luo, Shengfeng Lou, Chi Chen, Ziyue Wang Tsinghua University, Tsinghua University, Chenliang Li Alibaba Group, Weizhou Shen, Jiyue Guo, Peng Li Tsinghua University 等 12 位作者
Video temporal understanding is crucial for multimodal large language models (MLLMs) to reason over events in videos. Despite recent advances in general video understanding, current MLLMs still struggle with fine-grained temporal reasoning. While reinforcement learning (RL) has been explored to address this issue recently, existing RL approaches remain limited in performance on time-sensitive tasks. In this work, we propose MUSEG, a novel RL-based method that enhances temporal understanding by introducing timestamp-aware multi-segment grounding. MUSEG enables MLLMs to align queries with multiple relevant video segments, promoting more comprehensive temporal reasoning. To facilitate effective learning, we design a customized RL training recipe with phased rewards that progressively guides the model toward temporally grounded reasoning. Extensive experiments on temporal grounding and time-sensitive video question answering (QA) tasks demonstrate that MUSEG significantly outperforms existing methods and generalizes well across diverse temporal understanding scenarios.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Guibin Zhang National University of Singapore, Haiyang Yu, Kaiming Yang, Bingli Wu, Fei Huang Alibaba Group US, Yongbin Li Alibaba Group, Shuicheng Yan National University of Singapore
Complex agentic AI systems, powered by a coordinated ensemble of Large Language Models (LLMs), tool and memory modules, have demonstrated remarkable capabilities on intricate, multi-turn tasks. However, this success is shadowed by prohibitive economic costs and severe latency, exposing a critical, yet underexplored, trade-off. We formalize this challenge as the Agent System Trilemma: the inherent tension among achieving state-of-the-art performance, minimizing monetary cost, and ensuring rapid task completion. To dismantle this trilemma, we introduce EvoRoute, a self-evolving model routing paradigm that transcends static, pre-defined model assignments. Leveraging an ever-expanding knowledge base of prior experience, EvoRoute dynamically selects Pareto-optimal LLM backbones at each step, balancing accuracy, efficiency, and resource use, while continually refining its own selection policy through environment feedback. Experiments on challenging agentic benchmarks such as GAIA and BrowseComp+ demonstrate that EvoRoute, when integrated into off-the-shelf agentic systems, not only sustains or enhances system performance but also reduces execution cost by up to 80% and latency by over 70%.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Li Zheng, Xin Zhang Ant Group and Microsoft, Shuyi He, Fei Li Wuhan University, Chong Teng Wuhan University, Jiang-Ming Yang, Donghong Ji Wuhan University, Zhuang Li Royal Melbourne Institute of Technology
Accurate comprehension and controllable generation of emotion and rhetoric are pivotal for enhancing the reasoning capabilities of large language models (LLMs). Existing studies mostly rely on external optimizations, lacking in-depth exploration of internal representation mechanisms, thus failing to achieve fine-grained steering at the neuron level. A handful of works on neurons are confined to emotions, neglecting rhetoric neurons and their intrinsic connections. Traditional neuron masking also exhibits counterintuitive phenomena, making reliable verification of neuron functionality infeasible. To address these issues, we systematically investigate the neurons representation mechanisms and inherent associations of 6 emotion categories and 4 core rhetorical devices. We propose a neuron identification framework that integrates multi-dimensional screening, and design an adaptive masking method incorporating dynamic filtering, attenuation masking, and feedback optimization, enabling reliable functional validation of neuron functionality. Through neuron regulation, we achieve directed induction of non-target sentences and enhancement of emotion tasks via rhetoric neurons. Experiments on 5 commonly used datasets validate the effectiveness of our method, providing a novel paradigm for the fine-grained steering of emotion and rhetoric expressions in LLMs.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yuming Yang Fudan University, Mingyoung Lai University of Toronto, Wanxu Zhao, Xiaoran Fan, Zhiheng Xi, Mingqi Wu, Chiyue Huang, Jun Zhao 等 16 位作者
Long chain-of-thought (CoT) trajectories provide rich supervision signals for distilling reasoning from teacher to student LLMs. However, both prior work and our experiments show that trajectories from stronger teachers do not necessarily yield better students, highlighting the importance of data-student suitability in distillation. Existing methods assess suitability primarily through student likelihood, favoring trajectories that align closely with the student model’s current behavior but overlooking more informative ones. Addressing this, we propose Rank–Surprisal Ratio (RSR), a simple metric that captures both alignment and informativeness to assess the suitability of a reasoning trajectory. RSR is motivated by the observation that effective trajectories typically balance learning signal strength and behavioral alignment by combining low absolute probability with relatively high-ranked tokens under the student model.Concretely, RSR is defined as the ratio of a trajectory’s average token-wise rank to its average negative log-likelihood, and is straightforward to compute and interpret. Across five student models and reasoning trajectories from 11 diverse teachers, RSR strongly correlates with post-training reasoning performance (average Spearman 0.86), consistently outperforming existing metrics. We further demonstrate its practical utility in both trajectory selection and teacher selection.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Language agents, i.e., LLM agents, progress rapidly and are increasingly deployed in production environments. This trend underscores the urgent need for rigorous and realistic evaluations. However, most existing benchmarks evaluate agents in simplified, idealized settings. They typically rely on pre-packaged tool interfaces, overlook critical steps, and assume inputs are clean and fully specified. Consequently, they understate the difficulty of real deployments, where uncertainty and noise are ubiquitous and agents must proactively explore the environment to uncover new tools. To bridge this gap, we present AgentGym2, a new evaluation framework with task instances grounded in real-world end-to-end working demands. Beyond reasoning and planning, it measures agents’ ability to execute end-to-end procedures, discover tools via exploration, compose tools for unseen tasks, and remain robust to noisy and underspecified information. Experiments on 15 proprietary and open-source models show that even SOTA systems like Gemini and GPT-5 struggle on AgentGym2, revealing a substantial gap between the capability of current agents and the demands of real-world applications.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Lang Feng Nanyang Technological University, Fuchao Yang Nanyang Technological University, Feng Chen, Xin Cheng Chongqing University, Haiyang Xu, Zhenglin Wan, Ming Yan Alibaba Group, Bo An Nanyang Technological University
Recent advances in large language models (LLMs) enable agentic systems trained with reinforcement learning (RL) over multi-turn interaction, but practical deployment is bottlenecked by rapidly growing textual histories that inflate token and memory costs. We introduce AgentOCR, a framework that exploits visual tokens’ superior information density by representing the accumulated observation-action history as a compact rendered image. To make multi-turn rollouts scalable, AgentOCR proposes segment optical caching. By decomposing history into hashable segments and maintaining a visual cache, this mechanism eliminates redundant re-rendering. Beyond fixed rendering, AgentOCR introduces agentic self-compression, where the agent actively emits a compression rate and is trained with compression-aware reward to adaptively balance task success and token efficiency. We conduct extensive experiments on challenging agentic benchmarks, ALFWorld and search-based QA. Remarkably, AgentOCR preserves over 95% of text-based agent performance while substantially reducing token consumption (>50%), yielding consistent token and memory efficiency. Further analysis validates a 20\times rendering speedup from optical caching and effective self-compression balancing. Our code is available at https://github.com/langfengQ/AgentOCR.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Xuanyu Lei, Chenliang Li Alibaba Group, Yuning Wu Alibaba Group, Kaiming Liu Tsinghua University, Tsinghua University, Weizhou Shen, Peng Li Tsinghua University, Ming Yan Alibaba Group, Fei Huang Alibaba Group US 等 10 位作者
Recent advances in Large Language Models (LLMs) have enabled strong performance in long-form writing, but current training paradigms remain limited: Supervised Fine-Tuning (SFT) remains constrained by data saturation and performance ceilings, while Reinforcement Learning with Verifiable Reward (RLVR), though successful in verifiable domains like math and code, cannot be directly migrated to open-ended long-form writing due to a lack of ground-truths. To further advance long-form writing, we present Writing-RL: an Adaptive Curriculum Reinforcement Learning framework to advance long-form writing capabilities beyond SFT. The framework consists of three key components: Margin-aware Data Selection strategy that prioritizes samples with high learning potential, Pairwise Comparison Reward mechanism that provides discriminative learning signals in the absence of verifiable rewards, and Dynamic Reference Scheduling approach, which plays a critical role by adaptively adjusting task difficulty based on evolving model performance. Experiments on 7B-scale writer models show that Writing-RL effectively improves long-form writing performance over strong SFT baselines. Furthermore, we observe that models trained with long-output RL generalize surprisingly well to long-input reasoning tasks, potentially offering a promising perspective for rethinking long-context training.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Taiqiang Wu, Runming Yang, Tao Liu Tsinghua University, Jiahao Wang, Ngai Wong The University of Hong Kong
Model merging, typically on Instruct and Thinking models, has shown remarkable performance for efficient reasoning. In this paper, we systematically revisit the simplest merging method that interpolates two weights directly. Particularly, we observe that model interpolation follows a three-stage evolutionary paradigm with distinct behaviors on the reasoning trajectory. These dynamics provide a principled guide for navigating the performance-cost trade-off. Empirical results demonstrate that a strategically interpolated model surprisingly surpasses sophisticated model merging baselines on both efficiency and effectiveness. We further validate our findings with extensive ablation studies on model layers, modules, and decoding strategies. Ultimately, this work demystifies model interpolation and offers a practical framework for crafting models with precisely targeted reasoning capabilities.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zhengxi Lu, Jiabo Ye, Fei Tang Nankai University, Yongliang Shen Zhejiang University, Haiyang Xu, Ziwei Zheng Xi'an Jiaotong University, Weiming Lu Zhejiang University, Ming Yan Alibaba Group 等 11 位作者
GUI agents have demonstrated remarkable progress in automating complex user interface interactions. However, training such agents for long-horizon tasks remains challenging. Single-turn reinforcement learning conditions on expert histories during training but self-generated histories during deployment, causing distribution mismatch. Online multi-turn methods eliminate this gap via environment interaction but suffer from sparse rewards and prohibitive costs. We propose \underline{\textbf{E}}xperience-driven \underline{\textbf{M}}ulti-turn \underline{\textbf{P}}olicy \underline{\textbf{O}}ptimization (\textbf{EMPO}), which leverages expert trajectories as environment experiences for on-policy multi-turn training. The agent constructs self-generated history throughout rollouts; when actions match expert experiences, the trajectory provides valid state transitions, and a Patch Module recovers mismatched steps to maintain on-policy rollouts. EMPO further incorporates discounted future rewards and dual-level advantage estimation to capture long-horizon dependencies. We also propose \textbf{AndroidControl-Real}, an evaluation metric strongly correlated with real-world performance (R^2=0.934). With only 1K public trajectories as RL experiences, our method achieves substantial gains over the base model (e.g., +12.0% on AndroidWorld and +23.8% on AITW) and achieves competitive performance against strong baselines such as UI-TARS-7B and GPT-4o, demonstrating better generalization than prior single-turn RL approaches. Code available: https://anonymous.4open.science/r/UI-S1-0DAF.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zijun Liu, Zhennan Wan, Peng Li Tsinghua University, Ming Yan Alibaba Group, Fei Huang Alibaba Group US, Yang Liu
With the rapid advancement of post-training techniques for reasoning and information seeking, large language models (LLMs) can incorporate a large quantity of retrieved knowledge to solve complex tasks. However, the limited context window of LLMs obstructs scaling the amount of external knowledge input, prohibiting further improvement. Existing context window extension methods inevitably cause information loss. LLM-based multi-agent methods emerge as a new paradigm to handle massive input in a distributional manner, where we identify two core bottlenecks in existing agent orchestration designs. In this work, we develop a multi-agent framework, ExtAgents, to overcome the bottlenecks and enable better scalability in inference-time knowledge integration without longer-context training. Benchmarked with our enhanced multi-hop question answering test, \boldsymbol\inftyBench+, and other public test sets including long survey generation, ExtAgents significantly enhances the performance over existing non-training methods with the same amount of external knowledge input, regardless of whether it falls within or exceeds the context window. Moreover, the method maintains efficiency due to high parallelism. We believe further study in the coordination of LLM agents on increasing external knowledge input could benefit real-world applications.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表