WebGPU enables private, offline LLM inference inside web pages, but browser dispatch, inter-process communication, synchronous readback and single-token kernels slow decoding and make conventional speculative decoding inefficient. We present a WebGPU runtime that addresses these browser-specific costs. Deterministic small-batch integer kernels accelerate multi-token verification while preserving exact agreement with greedy decoding in the same quantized runtime. A GPU-resident loop performs drafting, verification, acceptance and KV-cache maintenance without a per-token CPU decision. A draft-free context-mixing drafter reuses page content, conversation history and session memory, while a device-aware controller adapts the speculation width and the submission window to measured costs. Across seven models from 0.5B to 8B parameters, 11 tasks and three client GPUs, our runtime decodes 3.5 to 5.4$\times$ faster than WebLLM on a laptop GPU, 4.8 to 6.2$\times$ faster on an integrated GPU and 4.8 to 8.6$\times$ faster on a workstation GPU. The GPU-resident loop and speculation contribute a further 1.8 to 3.2$\times$ over the optimized kernels on the workstation. Outputs match the quantized greedy decoding of the runtime exactly, and the mean GSM8K accuracy matches that of WebLLM. Optional draft heads raise the peak speedup to 10.2$\times$, while session memory provides additional gains without downloading extra weights.
Prateek Kumar Rajput, Yewei Song, Jacques Klein, Tegawendé F. Bissyandé
Code language models complete, review, and repair programs without exposing the program structure encoded in their hidden states. In this paper, we test whether the hidden states of a code model can be described and edited using the information a compiler front end computes. For each token of a Python program, standard analyzers provide its position in the abstract syntax tree, its scope, its definition-use links, and its syntactic category. We fit a decomposition that expresses every hidden state of a frozen model as a sum of one term per token, each combining the token with this structural label, and we replace the model's own states with the decomposition at every layer while it predicts the output of a small program. If the model still answers correctly, the decomposition preserves the information the model uses. We apply this test to 14 open models, seven of them trained for code, and compare against a baseline decomposition of the same size built from token positions. The static-analysis decomposition preserves the model's answers better than the positional one in 10 of the 14 models, including six of the seven code models, and deleting the definition-use links from it removes most of that advantage on the two models tested. Editing the decomposition, for instance rebinding the argument of a call, makes StarCoder2-15B produce the answer of the edited program in 90 of 100 cases, compared with 19 for the positional baseline. The Qwen2.5 family is the exception on several measures. For most of the models tested, standard static analysis therefore provides both a description of what the hidden states encode about a program and a way to change it without retraining.
Long-horizon robotic tasks are commonly decomposed into semantic subgoals to simplify planning. However, subgoal progress in LLM-based planning is typically inferred implicitly from interaction histories. Inaccurate estimates can misguide execution and allow local errors to propagate. To address this, we propose ProLoop, a Progress-Aware Closed-Loop framework that derives explicit subgoal progress from structured relations between desired and execution states. A learned State-Relation Monitor assesses these relations, which are then aggregated into subgoal progress as ongoing, deviated, or completed. The resulting progress guides execution adaptation, enabling reliable progress-aware control throughout execution. Experiments on AgentBoard ALFWorld and a real-world robotic platform show that ProLoop improves success rate by 12.94% overall and by 15.46% on Hard tasks, while reducing environment interactions by 32.3%.
Large Language Models (LLMs) are widely adopted for reflective error correction in task planning; however, existing approaches primarily rely on autonomous exploration, rendering them prone to accumulating erroneous experiences during early interactions with new environments. To address this limitation, we propose the Flexible Instructional Scaffolding Reflection Framework (FISRF), an agent--mentor architecture that estimates the robot's Zone of Proximal Development (ZPD) in task planning and adaptively selects reflection strategies and mentor guidance, thereby facilitating the accumulation of effective error-correction experience. Experiments in ALFWorld and on real-world robotic platforms demonstrate that FISRF reduces the average number of error-correction rounds by 37% and redundant reasoning by 25.1%, while improving the overall task success rate by 8%. Project webpage : https://fisrfplanning.github.io/
Long-horizon agents performing multi-step Web tasks must reuse earlier page observations and interactions to support subsequent actions. However, preserving relevant history does not ensure that it exposes the relations needed for the current decision. We call this gap representation mismatch. To this end, we propose MemPrism, a task-conditioned relational memory framework that separates persistent experience storage from decision-time working memory. MemPrism records interactions as the event stream and dynamically constructs relational views according to the current task context. A lightweight view policy selects the relation structure, evidence range, outcome condition, and granularity, while a deterministic composer and renderer transform historical facts into a temporary optical working-memory view for a frozen task policy. Experiments on long-horizon embodied and web-agent benchmarks show that MemPrism consistently improves the task performance, especially as trajectories become longer, while reducing memory token consumption. Furthermore, the learned view policy transfers across different VLMs without additional adaptation, demonstrating the effectiveness of task-conditioned relational views as a general memory interface for agents.
Tingxu Han, Yi Zhang, Wei Song, Chunrong Fang, Zhenyu Chen, Youcheng Sun, Lijie Hu
Agent skills, structured procedural knowledge packages injected at inference time, are increasingly used to augment LLM agents on software engineering tasks. However, their real utility in end-to-end development settings remains unclear, and no standard procedure tests an individual skill before installation. We present swe-skills-bench, a skill-centric, requirement-driven benchmark that measures how much individual agent skills improve performance on real-world software engineering (SWE) tasks. We start from 98 public SWE skills, each paired with a real GitHub repository pinned at a fixed commit. For each skill, we derive repository-specific requirements, create executable verifiers, and build a reproducible environment. Under each of four agent-model configurations, we run every task once without the skill and once with it, yielding 3,920 paired evaluations. For each requirement, we compute a paired delta, i.e., the change in outcome when the skill is added. We then aggregate these deltas across a skill's ten requirements into a per-skill utility profile, and use a statistical test to classify each skill as Golden (significantly helpful), Ineffective (no significant effect), or Harmful (significantly detrimental). Our results show that skills provide limited benefits: 92 of 98 skills have no statistically significant effect, and adding skills raises the average pass ratio only from 76.6% to 76.9%. The few helpful skills supply procedures that the agent would otherwise omit (gains of up to $+17.7$ points). In contrast, harmful skills anchor the agent on their own templates and override the task's explicit instructions (losses of up to -17.0 points). Our code is available at https://github.com/ControllableModel/SWE-Skills-Bench.
Jiawei Lu, Chen Zhang, Jinwei Lu, Yuanwei Song, Yuanfeng Song, Raymond Chi-Wing Wong
Users often request charts without knowing which database or table holds the relevant data. Such requests may be ambiguous, and users may ask follow-up questions to deepen their analysis or iteratively refine existing charts. A useful assistant must therefore discover relevant data sources among heterogeneous databases, ask clarifying questions when needed, and track evolving user intent while preserving constraints established in earlier turns. Yet recent text-to-visualization agents typically rely on static agent harnesses that are predefined and remain unchanged despite recurring failures across tasks. To address this limitation, we introduce RSIV-Harness, which keeps the language model weights fixed and applies Recursive Self-Improvement to revise the surrounding harness from the agent's own failure traces. To balance flexible self-improvement with stability, we introduce a four-layer framework that evaluates proposed harness revisions through anchor and memory checks, refining those that pass and restoring the best validated harness otherwise. We further introduce ODAC-Vis, a benchmark with 640 tasks for Open-Domain, Ambiguous, and Conversational text-to-visualization that jointly evaluates open-domain data discovery, ambiguity resolution, multi-turn visualization refinement, and executable chart generation. Experiments show that a single round of RSIV-Harness improves all six evaluation metrics relative to the seed harness, with particularly large gains in executability (+28.02 points) and chart structure (+34.01 points), demonstrating the ability to learn from task-solving experience and turn recurring failures across tasks into substantial performance improvements.
Zelin Li Microsoft, Jia Leng, Dawei Song Beijing Institute of Technology and Open University, Yangen Hu
Direct Alignment Algorithms (DAAs) such as DPO simplify RLHF by optimizing policies directly from preference pairs. However, the Bradley–Terry probability-gap objective can induce likelihood displacement and, under weak KL constraints, may even reduce the probability of preferred responses, while implicit rewards can be limited in generalizaiton. We propose Reward Alignment Optimization (RAO), a point-wise direct alignment method that uses an explicit reward model to specify exact target generation probabilities and align the policy offline towards them. Our key insight is a theoretical principle we call “prefix consistency”, which links the normalization terms of prompts that share a prefix. Leveraging this property, RAO decouples target reward differentials from bias terms, prevents decreasing preferred-response probabilities, and better exploits reward information both within and across prompts. Extensive experiments on multiple base LLMs show that RAO consistently outperforms existing DAAs while enabling controllable target probability distributions.
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
Yanzhi Tian Beijing Institute of Technology, Cunxiang Wang, Zeming Liu, Heyan Huang Beijing Institute of Technology, Wenbo Yu, Dawei Song Beijing Institute of Technology and Open University, Jie Tang Tsinghua University, Tsinghua University, Yuhang Guo
Large Language Models (LLMs) have significantly advanced Machine Translation (MT), applying them to linguistically complex domains-such as Social Network Services, literature etc. In these scenarios, translations often require handling non-literal expressions, leading to the inaccuracy of MT metrics. To systematically investigate the reliability of MT metrics, we first curate a meta-evaluation dataset focused on non-literal translations, namely MENT. MENT encompasses four non-literal translation domains and features source sentences paired with translations from diverse MT systems, with 7,530 human-annotated scores on translation quality. Experimental results reveal the inaccuracies of traditional MT metrics and the limitations of LLM-as-a-Judge, particularly the knowledge cutoff and score inconsistency problem. To mitigate these limitations, we propose RATE, a novel agentic translation evaluation framework, centered by a reflective Core Agent that dynamically invokes specialized sub-agents. Experimental results indicate the efficacy of RATE, achieving an improvement of at least 3.2 points in combined system- and segment-level correlation with human judgments compared with current methods. Further experiments demonstrate the robustness of RATE to general-domain MT evaluation. Code and dataset are available at: https://github.com/BITHLP/RATE.
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
Yuhan Chen Capital Normal University, Zizhuo Shen Beijing Language and Culture University and IFLYTEK CO.LTD., Miaomiao Cheng, Xu Han, Jiefu Gong IFLYTEK CO.LTD., Shijin Wang State Key Laboratory of Cognitive Intelligence, Wei Song Capital Normal University
Simulating student writing behavior offers a promising pathway to scalable feedback evaluation and teacher training. However, existing LLM-based approaches tend to model overly capable learners who readily understand and over-apply feedback, resulting in pedagogically implausible behavior. In this work, we introduce pedagogical realism as a guiding principle for student writing simulation, emphasizing bounded cognition, selective feedback comprehension, and developmentally plausible learning processes. To operationalize this idea, we propose CPT-Agent, a cognitively grounded framework that decouples cognitive ability from writing proficiency and models their interaction during writing and revision. CPT-Agent combines probabilistic modeling of cognitive development, proficiency-controlled text generation, and structured memory for skill accumulation. Experiments show that it (1) produces clearly distinguishable proficiency levels, (2) generates cognitively plausible revisions consistent with instructional theories, and (3) achieves strong agreement with expert judgments in evaluating feedback quality. These results highlight the importance of modeling cognitive constraints in LLM-based student simulation and demonstrate the potential of pedagogically realistic agents for automated feedback assessment and teacher development.
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