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.
Zhenghao Chen, Lebin Zhou, Zicong Chen, Naifu Xue, Jiahao Li, Bin Li, Dong Xu, Yan Lu
Despite decades of progress, image codecs remain difficult to adapt to diverse real-world compression scenarios for two main reasons. First, most image codecs are optimized for a specific objective and require substantial redevelopment when the compression goal changes. Second, they typically assume clean inputs, whereas sensor-captured images often contain degradations that alter rate-quality behavior. To address these challenges, we propose Practical Image Coding Agents (PICA), a multi-agent framework that adapts existing codecs to diverse user requirements and potentially degraded inputs. PICA consists of four collaborative agents: a Perception agent that interprets user intent and analyzes image content, an Improvement agent that jointly considers bitrate and reconstruction quality to guide pre-coding restoration, a Coding agent that refines quality-control settings from global to regional levels, and an After-coding improvement agent that addresses residual degradations and compression-induced artifacts through restoration guided solely by image quality. To enable codec-agnostic regional adaptation, we introduce Agentic ROI Coding (AROIC), which treats the codec as a black box and allows the agent to automatically schedule, execute, evaluate, and reflect ROI coding plans without architectural modifications or codec-specific retraining. We integrate these four agents with a Scenario-Guided Quality-Rate (SGQR) evaluation strategy to enable more effective scenario-adaptive compression across heterogeneous codecs. Comprehensive experiments on clean, noisy, and blurred images across standard, perceptual, and machine-vision compression scenarios demonstrate that PICA consistently enhances the compression performance of both the conventional H.266/VVC codec and the neural Cheng2020 codec, outperforming competing compression methods on the Kodak, CLIC2020, and HEVC datasets.
Zhicheng Du, Pinyan Lu, Keyu Wang, Zihe Wang, Xiang Yan, Chunxue Yang
Online large language model (LLM) services face a pricing cold start: providers initially lack the historical data and user information needed for precise pricing, while users' willingness to pay evolves with repeated use. We formulate this challenge as a \emph{two-sided learning} problem between a provider and a user. The user learns from private usage experience, whereas the provider incurs a serving cost and dynamically adapts prices using only purchase decisions to maximize cumulative profit. The provider's central challenge is to learn to price against an evolving willingness to pay without observing the private experiences driving that evolution. We establish matching upper and lower bounds on the minimax regret over horizon $T$, identifying two cost-dependent regimes. When serving costs are sufficiently low, our Calibration-and-Tracking (CAT) algorithm achieves the optimal rate $Θ(T^{1/3})$. In the higher-cost regime, where the provider must learn whether continued service is profitable, Calibration-and-Tracking with Retirement (CAT-R) attains the optimal rate $Θ(\sqrt{T})$. This phase transition shows that serving costs affect not only profit margins but also the fundamental difficulty of learning to price LLM services. Simulations support the predicted regret scaling and show that our algorithms outperform standard online learning baselines in the evaluated settings.
Understanding how objects and their spatial relationships evolve in 3D space over time remains challenging for vision-language models (VLMs), particularly under changing viewpoints. Recent approaches incorporate geometric information into reasoning, but explicit supervision for temporally ordered 4D state evolution and its use as intermediate evidence remains limited. We introduce \textbf{Trace4D}, which trains VLMs to predict 4D state sequences and reason over their inter-timestep changes alongside temporal visual observations. Its 4D-grounded chain-of-thought (CoT) explicitly represents 4D states, analyzes temporal visual context, and connects both as evidence for the answer. We construct Trace4D-Train from in-the-wild videos with 600K 4D perception examples and approximately 70K grounded QA examples, and train with joint supervised fine-tuning followed by geometry-aware GRPO. At inference, Trace4D predicts 4D states and reasoning directly from video without reference annotations or an external geometric model. Experiments on Trace4D-Bench and VLM4D show improved QA accuracy over the base VLMs and benefits from reasoning over 4D state changes and temporal visual observations.
Mingchuan Ma, Dongqi Cai, Daqian Yang, Zhenyan Lu, Chen Chen, Teresa Lv, Yan Gao, Shangguang Wang 等 11 位作者
LLM pretraining is increasingly compute-intensive, while substantial inference serving compute remains idle during off-peak periods. However, such serving hardware is difficult to use for conventional first-order training due to its memory, networking and other system requirements. Zeroth-order (ZO) optimization offers a natural interface to this forward-only compute resource, but its gradient estimates break down in pretraining. Unlike fine-tuning, which can confine updates to a small parameter subspace, pretraining must update every parameter, and the alignment of a ZO gradient estimate with the true gradient shrinks as the number of updated parameters grows. We make a different observation: noisy ZO estimates can still provide useful curvature information, even when they are too inaccurate to serve as gradients. Based on this insight, we propose ZO-Curv, which leverages idle forward-only devices to estimate curvature and uses it to adapt and refine the update used during first-order training, thereby accelerating the process. We provide convergence guarantees that characterize the effect of finite-query curvature estimation. Experiments on LLMs from 125M to 1.5B parameters show that ZO-Curv consistently accelerates pretraining. At 770M parameters, ZO-Curv reaches AdamW's 20K-step validation loss in 3.9 hours instead of 7.5 hours, while achieving a lower final loss of 2.70 versus 2.84 under the same step budget. These results show that idle model serving compute can accelerate LLM pretraining when zeroth-order queries are used to estimate curvature rather than gradients directly.
Minyan Luo, Yuxin Zhang, Meng Zhao, Xincan Wang, Fuzhang Wu, Tong-Yee Lee, Oliver Deussen, Fan Tang 等 9 位作者
An outstanding creative advertising image should both attract attention and communicate a product's positioning. How to choose a persuasive conceptual bridge before rendering begins is challenging for advertising generation. We present AdMuse, a skill-guided multi-agent framework that distills high-quality advertisements into reusable creative skills for cross-domain advertising strategy. Rather than retrieving isolated exemplars, AdMuse abstracts recurring bridge patterns into a skill library that can guide new product briefs. At generation time, AdMuse coordinates skill routing, multi-candidate ideation and selection, and VLM-guided visual refinement: diverse skills propose alternative concepts, relative judgment selects the most promising direction, and vision-language feedback helps realize the selected intent in the final image. This design treats advertising creativity as a structured decision process rather than a single prompt-writing step. Experiments across diverse product and brand positionings show that AdMuse produces visually compelling advertisements that faithfully communicate product positioning.
Jie Huang, Cuiling Lan, Qize Yu, Xiulian Peng, Yan Lu
An effective agent memory system should adapt to a user's history and task requirements. Human-annotated supervision can guide this adaptation by providing feedback for evaluating and improving memory designs. Yet suitable labels are rarely available for new user histories, and manually constructing QA pairs and supporting evidence is costly. We introduce MemSteer, which turns unlabeled user histories into question-answer (QA) supervision with supporting evidence for memory-design optimization. MemSteer constructs an evidence-grounded answer plan that specifies how facts across related records jointly support a reference answer. The plan guides question generation, while source-grounded auditing and blind answer recovery check the resulting annotations. To study the utility of this supervision, we develop MemSteer-Opt, which combines QA-based design evaluation with evidence-guided failure diagnosis to revise memory store, retrieval, and context construction without updating model parameters. Experiments on ATM-Bench and RHELM show that this supervision is competitive with official-QA-guided optimization. Relative to Base-BM25, MemSteer-Opt improves ATM-Bench Question Score by 13.4 points and RHELM accuracy by 4.1 and 6.6 points without and with external data, respectively. The supervision also benefits two existing optimizers on ATM-Bench, while new-user experiments demonstrate continued design optimization using supervision generated from the target histories.
Zezhou Wang, Ziyun Zhang, Xiaoyi Zhang Microsoft, Zhuzhong Qian, Yan Lu Microsoft Research Asia
Vision-language models are increasingly deployed as computer-use agents (CUAs) that operate desktops and browsers. Top-performing CUAs are framework-based systems that decompose planning and execution, while end-to-end screenshot-to-action policies are easier to deploy but lag behind on benchmarks such as OSWorld-Verified. GUI datasets like OSWorld pose two bottlenecks: they expose only a few hundred interactive, verifiable tasks and environments, and expert trajectories must be gathered by interacting with these environments, making such data hard to scale. We therefore ask how reinforcement learning from verifiable rewards (RLVR) can best exploit a small pool of exist expert trajectories to train end-to-end policies. Na"ively mixing these off-policy traces into on-policy RLVR is brittle: even after format conversion, expert trajectories exhibit structural mismatch and distribution shift from the learner. We propose BEPA (Bi-Level Expert-to-Policy Assimilation), which turns static expert traces into policy-aligned guidance via self-rolled reachable trajectories under the base policy (LEVEL-1) and a per-task, dynamically updated cache used in RLVR (LEVEL-2). On OSWorld-Verified, BEPA improves UITARS1.5-7B success from 22.87% to 32.13% and raises a held-out split from 5.74% to 10.30%, with consistent gains on MMBench-GUI and Online-Mind2Web. Our code and data are available at an anonymous repository: https://anonymous.4open.science/r/ACL_BEPA.
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
Zhengyang Ai Huawei Technologies Ltd., Zikang Shan Peking University, Xiaodong Ai Huawei Technologies Ltd., Jingxian Tang, Hangkai Hu Huawei Technologies Ltd., Pinyan Lu Huawei Technologies Ltd. and Shanghai University of Finance and Economics
Process supervision has emerged as a promising approach for enhancing LLM reasoning, yet existing methods fail to distinguish meaningful progress from mere verbosity, leading to limited reasoning capabilities and unresolved token inefficiency. To address this, we propose Stage-aware Hierarchical Advantage via Potential Estimation (SHAPE), a framework that formalizes reasoning as a trajectory through a state space of empirical solvability. SHAPE introduces a hierarchical credit assignment mechanism: at the segment level, it employs a stage-aware advantage function to prioritize efficient breakthroughs in low-potential states; at the token level, it utilizes entropy-driven redistribution to sharpen execution signals. Extensive experiments in math reasoning across three base models and five benchmarks demonstrate that SHAPE achieves an average accuracy gain of 3% with 30% reduced token consumption.
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
Yunfan Yang, Cuiling Lan Microsoft, Jitao Sang Beijing Jiaotong University, Yan Lu Microsoft Research Asia
Tables contain rich structured information, yet when stored as images their contents remain “locked” within pixels. Converting table images into LaTeX code enables faithful digitization and reuse, but current multimodal large language models (MLLMs) often fail to preserve structural, style, or content fidelity. Conventional post-training with reinforcement learning (RL) typically relies on a single aggregated reward, leading to reward ambiguity that conflates multiple behavioral aspects and hinders effective optimization. We propose Component-Specific Policy Optimization (CSPO), an RL framework that disentangles optimization across LaTeX tables components—structure, style, and content. In particular, CSPO assigns component-specific rewards and backpropagates each signal only through the tokens relevant to its component, alleviating reward ambiguity and enabling targeted component-wise optimization. To comprehensively assess performance, we introduce a set of hierarchical evaluation metrics. Extensive experiments demonstrate the effectiveness of CSPO, underscoring the importance of component-specific optimization for reliable structured generation.
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
Ziyun Zhang, Zezhou Wang, Xiaoyi Zhang Microsoft, Zongyu Guo Microsoft Research, Jiahao Li Microsoft Research Asia, Bin Li Microsoft, Yan Lu Microsoft Research Asia
GUI agents that interact with graphical interfaces on behalf of users are a promising direction for practical AI assistants, yet training them is hindered by scarce suitable environments. We present InfiniteWeb, a system that automatically generates functional web environments at scale for GUI agent training. While LLMs perform well on generating a single webpage, building a realistic and functional website with many interconnected pages faces challenges. We address these challenges through unified specification, task-centric test-driven development, and combining website seed variation with reference design images. Our system also generates verifiable task evaluators enabling dense reward signals for reinforcement learning. Experiments show that our system surpasses commercial coding agents at realistic website construction, and GUI agents trained on our generated environments achieve significant performance improvements on OSWorld and Online-Mind2Web, demonstrating the effectiveness of the proposed system.
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
Generative speech technologies are progressing rapidly, but evaluating the perceptual quality of synthetic speech remains a core challenge. Existing methods typically rely on scalar scores or binary decisions, which lack interpretability and generalization across tasks and languages. We present SpeechLLM-as-Judges, a new paradigm for enabling large language models (LLMs) to conduct structured and explanation-based speech quality evaluation. To support this direction, we introduce SpeechEval, a large-scale dataset containing 32,207 multilingual speech clips and 128,754 annotations spanning four tasks: quality assessment, pairwise comparison, improvement suggestion, and deepfake detection. Based on this resource, we develop SQ-LLM, a speech-quality-aware LLM trained with chain-of-thought reasoning and reward optimization to improve capability. Experimental results show that SQ-LLM delivers strong performance across tasks and languages, revealing the potential of this paradigm for advancing speech quality evaluation. The relevant code, models, and data are publicly available at https://github.com/NKU-HLT/SpeechLLM-as-Judges.
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
Sikuan Yan Ludwig-Maximilians-Universität München and Huawei Technologies Ltd., Xiufeng Yang, Zuchao Huang, Ercong Nie Shanghai Jiaotong University, Zifeng Ding, Zonggen Li University of Hong Kong, Xiaowen Ma, Jinhe Bi 等 13 位作者
Large Language Models (LLMs) have demonstrated impressive capabilities across a wide range of NLP tasks, but they remain fundamentally stateless, constrained by limited context windows that hinder long-horizon reasoning. Recent efforts to address this limitation often augment LLMs with an external memory bank, yet most existing pipelines are static and heuristic-driven, lacking a learned mechanism for deciding what to store, update, or retrieve. We present Memory-R1, a reinforcement learning (RL) framework that equips LLMs with the ability to actively manage and utilize external memory through two specialized agents: a Memory Manager that learns structured operations, including ADD, UPDATE, DELETE, and NOOP; and an Answer Agent that pre-selects and reasons over relevant entries. Both agents are fine-tuned with outcome-driven RL (PPO and GRPO), enabling adaptive memory management with minimal supervision. With only 152 training QA pairs, Memory-R1 outperforms strong baselines and generalizes across diverse question types, three benchmarks (LoCoMo, MSC, LongMemEval), and multiple model scales (3B–14B).
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
Chaoren Wang, Heng Lu, Xueyao Zhang The Chinese University of Hong Kong, Shenzhen, Shujie Liu, Yan Lu Microsoft Research Asia, Jinyu Li Microsoft, Zhizheng Wu Chinese University of Hong Kong, Shenzhen
Although Speech Large Language Models have achieved notable progress, a substantial modality reasoning gap remains: their reasoning performance on speech inputs is markedly weaker than on text. This gap could be associated with representational drift across Transformer layers and behavior deviations in long-chain reasoning. To address this issue, we introduce TARS, a reinforcement-learning framework that aligns text-conditioned and speech-conditioned trajectories through an asymmetric reward design. The framework employs two dense and complementary signals: representation alignment, which measures layer-wise hidden-state similarity between speech- and text-conditioned trajectories, and behavior alignment, which evaluates semantic consistency between generated outputs and reference text completions. Experiments on challenging reasoning benchmarks, including MMSU and OBQA, show that our approach significantly narrows the modality reasoning gap and achieves state-of-the-art performance among 7B-scale Speech 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