Online continual learning aims to continuously train neural networks on a non-stationary data stream in a single pass. Rehearsal-based methods alleviate catastrophic forgetting by replaying stored samples, but the imbalance between current and previous data often leads to predictions biased toward current classes. Many approaches have been proposed to overcome the forgetting problem by correcting the bias; however, they still need to be improved in an online fashion. By decomposing the dot-product logits into an angle factor and a norm factor, our analysis reveals that the angle factor efficiently captures novel knowledge but is particularly susceptible to prediction bias, whereas the norm factor helps preserve historical knowledge. Based on these observations, we propose Dual-Logit Experience Replay (DLER), which exploits dual-logits for both model optimization and memory management. Specifically, DLER learns current samples using cosine logits and incorporates additional dot-product logits when replaying previous samples to mitigate prediction bias. To strengthen dual-logit learning under limited replay memory, an auxiliary supervised contrastive objective promotes angular compactness within each class, extending corrections learned from limited replay samples to other samples of the same class. Meanwhile, we introduce a memory update strategy that preferentially retains samples for which dot-product prediction assigns a higher probability to the correct class than cosine prediction, aligning sample selection with dual-logit training. Extensive experiments on three datasets demonstrate the effectiveness of DLER, particularly under limited memory budgets.
Decisions about the direction of research in power engineering determine which technologies mature and, through them, shape infrastructure that lasts decades, yet their consequences cannot be observed before they are taken, and existing forecasts of research fields extrapolate a single trajectory. Here we introduce research world models, which treat a research field as a dynamical system that can be acted upon, and establish one for power engineering from 1.1 million papers. A distilled language model reads each paper into role-typed concepts, two sophisticated, high-fidelity graph-temporal neural network cores learn how the resulting concept network evolves, and operators learned from natural experiments in the publication record estimate the effects of interventions. In temporally held-out tests, the model forecast years ahead which concepts would grow and which method-target pairs would be joined for the first time, outperforming tuned machine-learning baselines in most settings, and its estimated intervention effects tracked those realized in natural experiments it had not seen. Its training scheme and three of its operators replicated in two fields withheld from development. Questions posed in natural language are compiled by a domain-specialized 27-billion-parameter language model into executable actions and answered as paired simulations with and without the intervention. The answers were rated 4.1 of 5 for faithfulness by an independent language-model judge. By revealing the latent dynamics of a research field and its response to interventions, research world models offer a new way to simulate and test research decisions before they are made.
Zhaoyang Wang, Qianhui Wu, Boyu Gou, Demi Wang, Jiacheng Zhang, Peichun Hua, Qize Li, Shi Qiu 等 57 位作者
GUI web agents require realistic and interactive environments that remain stable during both evaluation and training. Live website evaluation suffers from changing content, access restrictions, and limited coverage of authenticated workflows, while existing synthetic environments for training often lack functional depth and visual fidelity. We propose WebHarbor, a suite of 153 executable and dockerized website environments that faithfully reconstruct the visual interfaces, content, and functional workflows of real-world web services under deterministic control. We build WebHarbor through task-driven human-agent collaboration, where environments evolve as new tasks expand their functional scope, coding agents implement the required changes, and human contributors audit visual fidelity, functional correctness, and task validity. The resulting suite comprises 3,518 tasks spanning multi-step and post-authentication workflows, where existing works often fail to cover. Beyond reliable evaluation, WebHarbor provides high-fidelity synthetic web environments for cost-efficient reinforcement learning (RL). We demonstrate that a Qwen 4B-level GUI agent trained with RL in WebHarbor generalizes to unseen tasks and performs competitively with leading open-source GUI web agents across four benchmarks. The code and data are publicly available at https://github.com/aiming-lab/WebHarbor.
Xingzhi Qian, Sungho Park, Rongyuan Tan, Jianbo Yu, Jue Zhang, Lorenzo Cavallaro, Menglin Xia, Yanzhi Chen 等 16 位作者
Harness optimization can substantially improve LLM agents without modifying model weights, but existing methods primarily optimize response quality. Extending them to joint quality--cost optimization introduces a challenge: applying the deployment preference during intermediate updates can prematurely reject candidates that are valuable for later improvements. We introduce ParetoSaddler, which addresses this mismatch by separating intermediate candidate promotion from final deployment selection. At each iteration, an optimizer updates a parent harness to produce a candidate, and then compares the two over repeated executions and uses Bayesian bootstrap Pareto support to promote candidates showing credible progress in either quality or cost. Promoted candidates receive broader development evaluation and remain available as parents for later optimization steps. We evaluate ParetoSaddler on ClawMark and GAIA2 using domain-disjoint splits. On ClawMark, ParetoSaddler improves quality from 70.86\% to 81.35\% while reducing cost from \$1.17 to \$1.01 per task; on GAIA2, it improves quality from 56.22\% to 62.89\% while reducing cost from \$0.78 to \$0.45 per task. The Pareto-selected harnesses also achieve higher quality at lower cost than quality-only harness optimization. Analysis shows that Pareto-aware optimization retains intermediate candidates that would be discarded by deployment preference but can support later improvements, while Bayesian bootstrap screening filters out candidates whose gains are weakly supported across runs.
Robust Instruction Compliance in Cooperative Multi-Agent Reinforcement Learning
Wo Wei Lin, Ethan Rathbun, Enrico Marchesini, Xiang Zhi Tan
AI 速览：提出 MAVIC，通过在指令边界修正 Bellman 备份，缓解多智能体强化学习中外部语言指令打断宏动作时的价值估计不一致问题。
Multi-agent reinforcement learning (MARL) in real-world use cases may need to adapt to external natural language instructions that interrupt ongoing behavior and conflict with long-horizon objectives. However, conditioning rewards on instructions introduces a fundamental failure mode as Bellman updates couple value estimates across instruction contexts, leading to inconsistent values when instructions interrupt macro-actions. We propose Macro-Action Value Correction for Instruction Compliance (MAVIC), which corrects Bellman backups at instruction boundaries by correcting the incoming instruction objective and restoring the continuation value under the current objective. Unlike reward shaping, MAVIC modifies the bootstrapping target itself, enabling consistent value estimation under stochastic instruction switching within a unified policy. We provide theoretical analysis and an actor-critic implementation, and show that MAVIC achieves high instruction compliance while preserving base task performance in increasingly complex cooperative multi-agent environments.
LLM-based assistants on web-scale service platforms increasingly address local-service needs through multi-step planning and tool use. Under dynamic business conditions and evolving user needs, outcome-level evaluation offers limited evidence of where execution fails or how service drifts across continued interaction. We present ATLAS, a dual-horizon diagnostic evaluation framework for industrial tool-use agents. Within-turn evaluation uses trajectory-wise signals across thinking and reflection, tool and skill execution, and response generation to identify execution failures and unmet capability requirements. Across-turn evaluation uses user-wise signals spanning multiple lifecycles to diagnose failures in maintaining or updating user intent, constraints, and corrections. Fine-grained signals use rule-based checks and LLM judgments with explicit evidence scopes and decision boundaries. To support reliable diagnosis at production scale, LLM judges are calibrated against dynamically maintained references from business logs, and selected judges are distilled into compact diagnostic models. The resulting diagnostic profiles help developers attribute failures, prioritize targeted interventions, and reassess improvements. Applied in Meituan Xiaotuan, a million-scale DAU local-services assistant, ATLAS achieves high diagnostic fidelity across calibrated judges and distilled models. Blinded human auditing and a 14-day online A/B test with nearly two million users per arm confirm that the ATLAS-guided agent improves recommendation relevance, factual reliability, user engagement (+2.84% answer read-through rate), and order volume (+6.13% paid orders), leading to full production deployment.
Danny Y. C. Wang, Lars Arne Jordanger, Jerry Chun-Wei Lin
In rapidly evolving e-commerce industry, the capability of selecting high-quality data for model training is essential. This study introduces the High-Utility Sequential Pattern Mining using SHAP values (HUSPM-SHAP) model, a utility mining-based active learning strategy to tackle this challenge. We found that the parameter settings for positive and negative SHAP values impact the model's mining outcomes, introducing a key consideration into the active learning framework. Through extensive experiments aimed at predicting behaviors that do lead to purchases or not, the designed HUSPM-SHAP model demonstrates its superiority across diverse scenarios. The model's ability to mitigate labeling needs while maintaining high predictive performance is highlighted. Our findings demonstrate the model's capability to refine e-commerce data processing, steering towards more streamlined, cost-effective prediction modeling.
Generative recommendation (GR) models for ranking typically adopt a fast-thinking direct prediction paradigm with strong practical efficiency. However, ranking requires fine-grained scoring across multiple candidate items, while existing GR models still mainly rely on one-shot prediction. This raises the question whether additional inference-time computation can further improve ranking quality. Recently, inspired by Chain-of-Thought (CoT) reasoning in large language models (LLMs), some work has begun to introduce reasoning into recommendation systems. Some of these methods rely on LLM reasoning frameworks to perform explicit reasoning in textual space, but such designs are not suitable for more widely deployed non-textual recommendation models. A more relevant direction is to introduce latent reasoning into ID-based recommendation models. Although recent work has extended latent reasoning to industrial ranking, how successive reasoning steps should improve fine-grained candidate scores remains less explored. To address this problem, we propose Refiner, a latent residual reasoning framework with parameter-efficient fine-tuning for generative ranking. Our approach introduces three key components: (1) a residual reasoning mechanism that progressively refines ranking scores through cumulative additive corrections, (2) history queries that provide condensed user behavior representations to enhance candidate-specific reasoning, and (3) a parameter-efficient fine-tuning strategy that leverages LoRA modules to adapt pre-trained models for latent reasoning. Extensive experiments on two public benchmarks, an offline industrial dataset, and an online A/B test demonstrate that Refiner consistently improves over direct prediction, validating its effectiveness and practical utility.
Scaling coding agents requires a continuing supply of executable data for training, benchmarking, and continuous evaluation. Each task must couple a realistic software state with a specification, development tools, and reliable verification. To expand this supply, we present Change2Task, a system grounded in repository history that converts merged pull requests (PRs) into verified tasks on healthy modern revisions of the same repository. It aligns historical evidence with evolved code, reconstructs task states through Patch Reversal, Code Mapping, or Agent Reconstruction, and validates the lifecycle from a healthy base to a task state and a restored state. By deriving multiple tasks grounded in developer evidence from maintained environments, Change2Task provides executable data for coding agent training and evaluation while reducing repeated environment setup, storage, and task construction effort. We evaluate five coding agent task families: Bug Fix, Feature Addition, Test Generation, Application Programming Interface (API) Migration, and Security Repair. Starting from 1,130 source changes eligible for construction, Change2Task achieves 79.6% verified task construction success across these task families. On a matched candidate set, it recovers 29.2% more verified tasks than a construction baseline based on PRs. Historical and reconstructed cases achieve 89.7% aggregate matched outcome agreement under agent evaluation, while reuse of modern bases reduces measured expenditure across the complete pipeline by 10.8%. Four matched studies examine construction stages, semantic review, branch consistency, and resource trade-offs. An industrial deployment further constructs 1,000 tasks from private PR histories and evaluates four models on a fixed 500-task cohort.
Enterprise LLM agents can dramatically improve workplace productivity, but their core capability, retrieving and using internal context to act on a user’s behalf, also creates new risks for sensitive information leakage. We introduce CI-Work, a Contextual Integrity (CI)-grounded benchmark that simulates enterprise workflows across five information-flow directions and evaluates whether agents can convey essential content while withholding sensitive context in dense retrieval settings.Our evaluation of frontier models reveals that privacy failures are prevalent (violation rates range from 15.8%-50.9%, with leakage reaching up to 26.7%) and uncovers a counterintuitive trade-off critical for industrial deployment: higher task utility often correlates with increased privacy violations.Moreover, the massive scale of enterprise data and potential user behavior further amplify this vulnerability. Simply increasing model size or reasoning depth fails to address the problem. We conclude that safeguarding enterprise workflows requires a paradigm shift, moving beyond model-centric scaling toward context-centric 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
ChenZhuo Zhao, Xinda Wang Peking University, Pu Zhao, Yue Huang, Junting Lu, Ziqian Liu, Qingwei Lin Microsoft Research, Saravan Rajmohan Microsoft 等 9 位作者
Automatic prompt optimization is a practical alternative to fine-tuning for adapting large language models (LLMs), yet existing approaches often trade off signal quality against computational cost. Methods that rely on generative feedback can be informative but expensive to scale, while sampling-based optimization typically requires many evaluations and exhibits high variance. Even loss-driven prompt optimization remains limited by costly segment attribution that scales with prompt length and by overfitting to a single evaluator, which weakens transfer across model families and domains. We propose Gradient-guided Multi-judge Prompt Optimization (GMPO), a scalable framework that improves both efficiency and robustness. GMPO uses a first-order gradient approximation to score segment importance in a continuous masking direction, requiring only one forward and one backward pass. GMPO further employs a generate multi-judge design in which candidate prompt edits are proposed by a generator and selected using cross-entropy losses aggregated from multiple lightweight judge models, reducing evaluator bias and improving generalization. Experiments across math, reasoning, instruction-following evaluation, and safety robustness benchmarks demonstrate consistent gains with substantially lower optimization overhead.
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
Lihao Sun Microsoft, Hang Dong Microsoft Research, Bo Qiao Microsoft Research, Qingwei Lin Microsoft Research, Dongmei Zhang Microsoft, Saravan Rajmohan Microsoft
This work characterizes large language models’ chain-of-thought generation as a structured trajectory through representation space. We show that mathematical reasoning traverses functionally ordered, step-specific subspaces that become increasingly separable with layer depth. This structure already exists in base models, while reasoning training primarily accelerates convergence toward termination-related subspaces rather than introducing new representational organization. While early reasoning steps follow similar trajectories, correct and incorrect solutions diverge systematically at late stages. This late-stage divergence enables mid-reasoning prediction of final-answer correctness with ROC–AUC up to 0.87. Furthermore, we introduce trajectory-based steering, an inference-time intervention framework that enables reasoning correction and length control based on derived ideal trajectories. Together, these results establish reasoning trajectories as a geometric lens for interpreting, predicting, and controlling LLM reasoning behavior.
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
Song Jin, Juntian Zhang, Xun Zhang, Zeying Tian, Fei Jiang Meituan, Guojun Yin Meituan, Wei Lin, Yong Liu Renmin University of China 等 9 位作者
Recent advancements in Vision-Language Models (VLMs) have revolutionized general visual understanding. However, their application in the food domain remains constrained by benchmarks that rely on coarse-grained categories, single-view imagery, and inaccurate metadata. To bridge this gap, we introduce DiningBench, a hierarchical, multi-view benchmark designed to evaluate VLMs across three levels of cognitive complexity: Fine-Grained Classification, Nutrition Estimation, and Visual Question Answering. Unlike previous datasets, DiningBench comprises 3,021 distinct dishes with an average of 5.27 images per entry, incorporating fine-grained “hard” negatives from identical menus and rigorous, verification-based nutritional data. We conduct an extensive evaluation of 29 state-of-the-art open-source and proprietary models. Our experiments reveal that while current VLMs excel at general reasoning, they struggle significantly with fine-grained visual discrimination and precise nutritional reasoning. Furthermore, we systematically investigate the impact of multi-view inputs and Chain-of-Thought reasoning, identifying five primary failure modes. DiningBench serves as a challenging testbed to drive the next generation of food-centric VLM research. All codes are released in https://github.com/meituan/DiningBench.
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
Zhiyuan Peng, Xin Yin, Pu Zhao, Fangkai Yang, Lu Wang Microsoft, Ran Jia Microsoft, Xu Chen, Qingwei Lin Microsoft Research 等 10 位作者
Large language models and agents have achieved remarkable progress in code generation. However, existing benchmarks focus on isolated function/class-level generation (e.g., ClassEval) or modifications to existing codebases (e.g., SWE-Bench), neglecting complete microservice repository generation that reflects real-world 0-to-1 development workflows. To bridge this gap, we introduce RepoGenesis, the first multilingual benchmark for repository-level end-to-end web microservice generation, comprising 106 repositories (60 Python, 46 Java) across 18 domains and 11 frameworks, with 1,258 API endpoints and 2,335 test cases verified through a “review-rebuttal” quality assurance process. We evaluate open-source agents (e.g., DeepCode) and commercial IDEs (e.g., Cursor) using Pass@1, API Coverage (AC), and Deployment Success Rate (DSR). Results reveal that despite high AC (up to 73.91%) and DSR (up to 100%), the best-performing system achieves only 23.67% Pass@1 on Python and 21.45% on Java, exposing deficiencies in architectural coherence, dependency management, and cross-file consistency. Notably, RepoGenesis-8B, fine-tuned on RepoGenesis (train), achieves performance comparable to GPT-5 mini, demonstrating the quality of RepoGenesis for advancing microservice generation. We release our benchmark at https://github.com/pzy2000/RepoGenesis.
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
Zihao Cheng, Zeming Liu, Yingyu Shan, Xinyi Wang, Xiangrong Zhu, Yunpu Ma Ludwig-Maximilians-Universität München and Siemens Corporate Research, Hongru Wang, Yuhang Guo 等 10 位作者
While large language model–powered agents can self-evolve by accumulating experience or by dynamically creating new assets (i.e., tools or expert agents), existing frameworks typically treat these two evolutionary processes in isolation. This separation overlooks their intrinsic interdependence: the former is inherently bounded by a manually predefined static toolset, while the latter generates new assets from scratch without experiential guidance, leading to limited capability growth and unstable evolution. To address this limitation, we introduce a novel paradigm of co-evolutionary Capability Expansion and Experience Distillation. Guided by this paradigm, we propose the Mem^{\textbf{2}}Evolve, which integrates two core components: Experience Memory and Asset Memory. Specifically, Mem^{2}Evolve leverages accumulated experience to guide the dynamic creation of assets, thereby expanding the agent’s capability space while simultaneously acquiring new experience to achieve co-evolution. Extensive experiments across 6 task categories and 8 benchmarks demonstrate that Mem^{2}Evolve achieves improvement of 18.53% over standard LLMs, 11.80% over agents evolving solely through experience, and 6.46% over those evolving solely through asset creation, establishing it as a substantially more effective and stable self-evolving agent framework.
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
Hongbo Jin, Kuanwei Lin, Wenhao Zhang, Yichen Jin Peking University, Ge Li Peking University
Reinforcement Learning (RL) is crucial for empowering Video-LLMs with complex spatiotemporal reasoning. However, current RL paradigms predominantly rely on random data shuffling or naive curriculum strategies based on scalar difficulty metrics. We argue that scalar metrics fail to disentangle two orthogonal challenges in video understanding: Visual-Temporal Perception Load and Cognitive Reasoning Depth. To address this, we propose VideoCuRL, a novel framework that decomposes difficulty into these two axes. We employ efficient, training-free proxies—optical flow/keyframe entropy for visual complexity and Calibrated Surprisal for cognitive complexity—to map data onto a 2D curriculum grid. A competence-aware Diagonal Wavefront strategy then schedules training from base alignment to complex reasoning. Furthermore, we introduce Dynamic Sparse KL and Structured Revisiting to stabilize training against reward collapse and catastrophic forgetting. Extensive experiments show that VideoCuRL surpasses strong RL baselines on reasoning (+2.5% on VSI-Bench) and perception (+2.9% on VideoMME) tasks. Notably, VideoCuRL eliminates the prohibitive inference overhead of generation-based curricula, offering a scalable solution for robust video post-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
Jie Cao, Tianwei Lin, Bo Yuan, Rolan Yan, Hongyang He Manifolda.Ai, Wenqiao Zhang Zhejiang University, Juncheng Li Zhejiang University, Dongping Zhang China Jiliang University 等 10 位作者
Recent studies integrate Low-Rank Adaptation (LoRA) and Mixture-of-Experts (MoE) to further enhance the performance of parameter-efficient fine-tuning (PEFT) methods in Large Language Model (LLM) applications. Existing methods employ homogeneous MoE-LoRA architectures composed of LoRA experts with either similar or identical structures and capacities. However, these approaches often suffer from representation collapse and expert load imbalance, which negatively impact the potential of LLMs. To address these challenges, we propose a heterogeneous Mixture-of-Adapters (MoA) approach. This method dynamically integrates PEFT adapter experts with diverse structures, leveraging their complementary representational capabilities to foster expert specialization, thereby enhancing the effective transfer of pre-trained knowledge to downstream tasks. MoA supports two variants: (i) Soft MoA achieves fine-grained integration by performing a weighted fusion of all expert outputs; (ii) Sparse MoA activates adapter experts sparsely based on their contribution, achieving this with negligible performance degradation. Experimental results demonstrate that heterogeneous MoA outperforms homogeneous MoE-LoRA methods in both performance and parameter efficiency. Our project is available at https://github.com/DCDmllm/MoA.
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