LLM-as-a-judge scales evaluation, but reasoning judges are slow and costly. We study JEV-as-a-Judge: evaluation with JEV, a decision-only judge that returns label probabilities instead of text, and whose confidence decides whether to accept its verdict or escalate to a reasoning judge. Against sixteen generative and reward-model judges, with blinded human adjudication, JEV comes within three points of GPT-6 wherever a verdict can be read off the text, at 0.36% of its fee and a 0.15-second median latency, and falls behind where the verdict must be derived, as in math, code, and logic. Its confidence marks this boundary. With a threshold frozen in advance, accepting confident verdicts and escalating the rest is 0.9 points more accurate than GPT-6 on 1,610 held-out pairs at 41% of its fee, and in a pre-specified live test on two new workloads the cascade matches GPT-6's accuracy exactly. Confidence routing weakens on style-adversarial pairs and reference-free prose; we close with a simple recipe for validating thresholds locally.
Understanding how humans interact with the world requires a representation of physical interactions at the hand surface. Tactile sensing captures these interactions, but collecting tactile data at scale requires specialized hardware. We introduce TactoHand, a model that predicts three dense hand-surface interaction fields---contact, proximity, and 3D force---from monocular video. We derive geometry-based contact and proximity labels and motion-derived force labels from eligible training data within a curated collection spanning approximately 50 million source frames. TactoHand uses a temporal transformer to aggregate visual evidence across frames, while three learned queries decode the corresponding hand-surface fields. A differentiable physics layer further constrains force predictions using observed object motion. Across in-domain and out-of-domain benchmarks, TactoHand improves contact and proximity estimation over prior methods. In addition, physics supervision reduces position and rotation errors in motion rollouts, and the predicted contact and proximity fields improve downstream hand-object reconstruction. Together, these results show a scalable, low-cost path to learning hand-surface interaction fields without collecting tactile measurements. Code will be released at https://github.com/zhenqis123/Tactohand.
Dexterous grasping is the foundational primitive in embodied AI, demanding massive data to train robust models. As real-world data collection is expensive, simulation has become the mainstream paradigm. Yet, while cluttered scenes best reflect real-world applications, learning to grasp within them is bottlenecked by a critical scarcity of large-scale data. To resolve this, we curate high-quality 3D objects and supporting bases, proposing a scalable seed-and-filter strategy that bypasses sluggish scene-level optimization. This yields an unprecedented benchmark comprising over 2.6 million scenes and 0.4B scene-specific grasp ground truths, featuring diverse realistic layouts paired with rich semantic and geometric observations. Furthermore, we introduce the OmniDex model to overcome the grasp multimodality and last-millimeter precision errors plaguing current generative models. By coupling Soft Winner-Takes-All learning with human-inspired physical constraints during training, and utilizing physics-driven ranking, our approach achieves robust dexterous grasping without the latency of post-optimization. Experimental results show that OmniDex model achieves state-of-the-art performance and strong generalization across diverse scenes, views, and unseen objects.
Chengquan Guo, Xiaodong Liu, Weiwei Yang, Mingqian Feng, Jialin Song, Qiyao Ma, Carolyn Buractaon, Christopher White 等 10 位作者
LLM agents are increasingly adopted in long-horizon tasks and interact with diverse external environments. Despite their growing capabilities, these agents remain vulnerable to real-world attacks such as indirect prompt injections, which manipulate agents through malicious instructions embedded in external environments. While existing risk assessment work has primarily focused on the injection content, the effect of where an injection appears has not been systematically characterized. Studying injection placement requires agents to operate over longer-horizon tasks and interact with diverse environmental surfaces. Thus, we build three agentic testbeds spanning tool use, coding, and daily workflow domains with diverse injection surfaces. Holding the injection content fixed, we find that agent susceptibility varies across injection placements. We refer to this phenomenon as the trust region effect: agents implicitly assign different levels of trust to different regions of their environments. For example, an agent powered by GPT-5.4 follows an unsafe injection instruction placed in a repository documentation file (i.e., a README or TODO file) 2.6 times more often on average than the same instruction placed in a code or data file. Building upon this observation, we develop learning-based attackers that generate injection content and select better injection placements. Our approach demonstrates that both where an injection is placed and what content it contains can affect the strength of indirect prompt injection attacks. To the best of our knowledge, this work provides the first systematic characterization for the effectiveness of different injection placements, spanning both coarse-grained environmental surfaces and fine-grained positions. The uneven trust that agents implicitly place across their environments provides important insights for future attack and defense design.
Renzhe Xu, Haotian Wang, Xingxuan Zhang, Bo Li, Peng Cui
In this paper, we present the Proportional Payoff Allocation Game (PPA-Game), which characterizes situations where agents compete for divisible resources. In the PPA-game, agents select from available resources, and their payoffs are proportionately determined based on heterogeneous weights attributed to them. Such dynamics simulate content creators on online recommender systems like YouTube and TikTok, who compete for finite consumer attention, with content exposure reliant on inherent and distinct quality. We first conduct a game-theoretical analysis of the PPA-Game. While the PPA-Game does not always guarantee the existence of a pure Nash equilibrium (PNE), we identify prevalent scenarios ensuring its existence. Simulated experiments further prove that the cases where PNE does not exist rarely happen. Beyond analyzing static payoffs, we further discuss the agents' online learning about resource payoffs by integrating a multi-player multi-armed bandit framework. We propose an online algorithm facilitating each agent's maximization of cumulative payoffs over $T$ rounds. Theoretically, we establish that the regret of any agent is bounded by $O(\log^{1 + η} T)$ for any $η> 0$. Empirical results further validate the effectiveness of our online learning approach.
Liang Yuan, Xiaohu Du, Ruikun Luo, Jinze Wang, Mashud Rana, Kaibin Wang, Jiangwen Zhou, Bo Li 等 9 位作者
Web assistants increasingly answer questions about a user's private documents by pairing a small on-device model with a cloud LLM behind a Web API. In collaborative decoding, the document stays on the client, and the cloud receives only a public view that the user approves, such as the question alone, but the two models exchange messages while the answer is generated. We show that the message transcript can disclose private content that neither the public view nor the answer contains. In a planted-secret audit, speculative decoding with private drafts did not include the secret code in its answers, yet sent the code to the cloud in rejected drafts. We present ADLER, a decoding protocol for streaming LLM APIs in which every message to the cloud depends only on the public view and the answer the user sees, regardless of how the client combines the two models or how the cloud behaves. A client-side leakage ledger then bounds how much the answer itself reveals relative to a public reference process built from the cloud's replies. In the audit, no ADLER configuration exposed the secret beyond the answer. On GSM8K math problems with numerical values masked in the cloud's input, ADLER with an unlimited leakage budget raises the exact-match accuracy of a 0.5B client from 0.41 to 0.83 and of a 1.5B client from 0.67 to 0.85 under lenient answer parsing, close to the 0.92 achieved by uploading the full problem. It also exceeds a zero-leakage program baseline (0.69) by 0.14 and 0.16, respectively. In this baseline, the cloud writes a program that the client runs offline using the private numbers. The ledger records the disclosure charge associated with this gain; budgets of 16 nats or less leave accuracy near the cloud-only baseline on GSM8K.
Benjamin Pikus, Daniel Saunders, Mircea Mironenco, Salman Mohammadi, Mathew Mathew, Sirou Zhu, Divya Makkar, Xuhua Huang 等 14 位作者
A robot policy needs to understand what a task is asking for in a scene, and how to move the robot to accomplish that task. To achieve these two skills, we introduce Zero-real Agent Policy (ZAP), a recipe that teaches a small off-the-shelf vision-language model to control a robot via a simple tool harness. We use a pretrained 27B parameter model to bootstrap world understanding and reasoning, followed by supervised fine-tuning and GRPO on simple simulated tasks to teach it how to move the robot. Without observing real-world robot data during training, ZAP reaches 96.7% success on real-world tasks, matching GPT-6 Astra in our harness and outperforming the MolmoAct2 VLA (33.3%), while running 3x faster and 17x cheaper than GPT-6 Astra. Furthermore, ZAP also generalizes to an unseen embodiment as well as new objects, goals and strategies never seen during training. On a single-arm Franka, ZAP reaches 55.9% on RoboLab-120, 4.9 percentage points ahead of the leading VLA/WAM baseline.
Xinying Qian, Ying Zhang, Yu Zhao, Xubo Liu, Baohang Zhou, Jeff Z. Pan
Long-term memory enables Large Language Models (LLMs) to maintain coherent and personalized interactions across extended conversations. However, existing memory systems face two key challenges: (1) Information Entanglement, where heterogeneous information is mixed within coarse-grained memory units; and (2) Organizational Mismatch, where memory structures based primarily on semantic similarity may overlook other associations. To address these challenges, we propose ETAMem, a multi-view memory framework that decouples heterogeneous conversational information and aligns memory representations and organization with its distinct characteristics. Specifically, heterogeneous information parsing extracts atomic memory units and identifies affective information. View-specific memory structuring then organizes these units into three complementary hierarchical views: Event, Affect, and Time. ETAMem incrementally maintains memory structures with adaptively calibrated thresholds. For retrieval, entropy-based retrieval allocates query-specific budgets across views and performs hierarchical search within each view. Experiments on LoCoMo and LongMemEval-S demonstrate that ETAMem outperforms state-of-the-art baselines, achieving relative accuracy improvements of 12.46% and 7.96%, respectively, while maintaining competitive token efficiency. Further analyses demonstrate the complementary benefits of different memory views.
Wenjun Wang, Yuchen Fang, Fengrui Liu, Zibo Liang, Kai Zheng
Large Language Model (LLM) agents have shown remarkable capabilities in autonomous decision-making by generating sequential trajectories of states, actions, and observations. However, in complex, long-horizon tasks, these agents frequently suffer from compounding errors and struggle to recover from failures. Despite advances in experience memory, extracting reliable and transferable correction knowledge from historical trajectories remains challenging. Existing methods, whether based on test-time reflection or offline memory construction, largely rely on LLMs to diagnose failures and formulate corrective strategies, which can produce brittle or task-specific insights that are difficult to generalize. To address this challenge, we propose Experience Memory Graph (EMG), a framework that reformulates agent failure recovery as a graph matching problem. During offline construction on the training set, we convert both failed exploration trajectories and successful expert trajectories into directed action decision graphs. By matching these graphs, we extract common subgraphs and graph edit paths that explicitly indicate how to correct failures, and store them in a memory graph with intra-task nodes and cross-task edges. At test time, EMG retrieves relevant insights and guides the agent in a single test-time attempt without additional reflection-and-replay rounds. Experiments on ALFWorld and ScienceWorld demonstrate consistent improvements over the evaluated memory construction and competitive baselines across seen and unseen splits.
Sahel Sharifymoghaddam, Yauhen Babakhin, Reza Esfandiarpoor, Oliver Holworthy, Jimmy Lin, Bo Liu
In agentic search, a document can match an agent-generated sub-query without helping answer the user's original question. Conventional rerankers score documents against the sub-query alone, overlooking the broader task and the agent's reasoning. We introduce the Agent-Context-Aware Reranker (ACARe), a compact 4B LLM trained to rerank documents using this context when available. Across multiple benchmarks and agent families, ACARe improves retrieval quality over its base reranker while reducing search effort. These benefits extend to answer generation: ACARe improves accuracy and reduces searches over no reranking, while matching or exceeding the base reranker's mean accuracy with more precise evidence. It also helps smaller agents narrow the retrieval-quality gap to larger agents, showing how better retrieval tools can make smaller agents more effective. The model, training data, and code will be released soon.
Bo Liu, Tong Li, Jiapeng Zhang, Zhu Xiao, Zhuo Tang, Kenli Li
Human mobility data are fundamental to Web-based location services and broader urban applications, and mobility generation provides synthetic trajectories when real data are costly or inaccessible. Although recent mobility generators have incorporated collective mobility patterns, they primarily focus on individual trajectories and aggregate population dynamics, leaving evolving community structures formed by user-location interactions insufficiently modeled. In this paper, we study community-consistent human mobility generation. To this end, we formulate mobility generation as the generation of dynamic user-location graphs to model community structures and propose MOTION, a Multi-agent framework for cOmmuniTy-consIstent human mObility geNeration. MOTION consists of user agents, location agents, and a global manager agent. MOTION first detects coherent communities from historical mobility using the Leiden algorithm and extracts community-conditioned interaction priors to guide the generation process. Guided by these priors, the user agents plan mobility, search destination candidates, and infer user movements from the individual view. In parallel, selectively activated location agents infer likely user arrivals from the community view. The manager coordinates both views, diagnoses interactions causing deviations, and attributes targeted reflection signals to revise user decisions, location attraction, and source-user selection. Experiments on two real-world datasets show that MOTION outperforms baselines by at least 4.05% across reported metrics of individual trajectories, population and flow distributions, and community and graph structures. Further analyses confirm its robustness, privacy preservation, and downstream utility. The code is available at https://anonymous.4open.science/r/MOTION-04C4.
Large language model (LLM)-based multi-agent systems (MAS) rely on role-specific prompts and execution models to solve complex tasks collaboratively. However, prompt optimization typically keeps model assignments fixed, leaving differences in role requirements and their associated inference costs outside the search. We present UniMASO, a Unified framework for heterogeneous Multi-Agent System Optimization through the joint adaptation of prompts and model assignments. UniMASO evaluates candidate updates through local, downstream, and global feedback, using cost-aware Pareto search to retain useful quality-cost trade-offs. Topology-aware co-evolution then coordinates these updates according to agent roles and dependencies. Cross-level trajectory memory further supplies historical successes and failures to guide subsequent optimization. Experiments on five benchmarks and five execution topologies show that UniMASO achieves competitive task performance at substantially lower deployment cost than the strongest homogeneous configurations. Under centralized execution, it improves the average score from 69.79 to 70.40 while reducing token-weighted deployment cost by 85.2% relative to DeepSeek-V4-Pro with MASPO. These results support joint prompt-model adaptation as an effective approach to efficient multi-agent collaboration. The code is available at https://github.com/Achilles008/UniMASO_main.
Mintong Kang, Stephen Rawls, Melissa Kazemi Rad, Jian Ni, Bo Li
Automatic red teaming increasingly relies on large language models (LLMs) to iteratively diagnose failed attacks and propose improved adversarial prompts or injections. Although prior work strengthens this loop through prompt engineering and search-harness design, the mechanisms underlying its two core roles, criticism and attack proposal, has not been explored. To further investigate the core roles and therefore improve red teaming effectiveness, we disentangle two design choices: whether criticism is generated by the target session itself (self-criticism) or by a separately instantiated critic (external criticism), and whether that criticism is translated into the next attack by the target session (self-proposal) or by a separate attacker (external proposal). Through our analysis, we find that self-criticism is more effective because it is grounded in the target's live execution context, including observations, constraints, and available reasoning; removing reasoning reduces downstream attack success by 6%-17%. In contrast, self-proposal is substantially less effective because awareness of the adversarial objective induces explicit refusal and non-refusal sandbagging. These findings motivate a role-separated red-teaming approach: the target session diagnoses why an attack failed, while a separately instantiated attacker generates the next proposal. On DTAP-BENCH, spanning Telecom, Workflow, CRM, and Code domains, under both direct misuse and indirect threat models, as well as multiple agent frameworks, this configuration achieves 63.2% overall attack success rate (ASR), much higher than baselines. Our results show that effective automatic red teaming depends not only on model capability and search strategy, but also on how execution context and adversarial responsibility are distributed across model sessions.
Despite the remarkable capabilities of large language models (LLMs) across diverse applications, they remain vulnerable to generating content that violates safety regulations and policies. To mitigate these risks, LLMs undergo safety alignment; however, they can still be effectively jailbroken. Off-the-shelf guardrail models are commonly deployed to monitor generations, but these models primarily focus on detection rather than ensuring safe decoding of LLM outputs. Moreover, existing efforts lack rigorous safety guarantees, which are crucial for the universal deployment of LLMs and certifiable compliance with regulatory standards. In this paper, we propose a Claim-based Stream Decoding (CSD) algorithm coupled with a statistical risk guarantee framework using conformal analysis. Specifically, our CSD algorithm integrates a stream guardrail model to safeguard sequential claims generated by LLMs and incorporates a backtracking mechanism to revise claims flagged with high safety risks. We provide theoretical guarantees demonstrating that the CSD algorithm achieves the desired generation distribution subject to safety constraints. Furthermore, we introduce a generation risk certification framework and derive a high-probability upper bound on the safety risk of the proposed CSD algorithm. We prove that our method can asymptotically control safety risk to any desired level. Empirical evaluations demonstrate the effectiveness and efficiency of the CSD algorithm compared to state-of-the-art safety decoding approaches. Additionally, we validate the soundness and tightness of the derived safety risk upper bound using realistic data.
We introduce OvisOCR3, a 0.8B model for a broad range of OCR-related tasks. OvisOCR3 unifies document parsing, general OCR, key information extraction (KIE), document figure captioning, and cross-page table reconstruction in a single model. We train the model on real and synthetic data through continued pretraining and use domain-balanced mid-training to balance supervision across tasks with different data volumes and response lengths. Multi-teacher on-policy distillation transfers expertise from five domain-specific 4B teachers into the shared model, followed by mixed-task reinforcement learning with domain-specific rewards. We compare OvisOCR3 with general-purpose vision-language models and specialized OCR models on public and in-house benchmarks covering all five task families. In these comparisons, it achieves an overall score of 96.74 on OmniDocBench v1.6 and the highest overall scores on PureDocBench, Real5-OmniDocBench, and MDPBench. On OCRBench v2, it ranks first, scoring 65.9 and 61.3 on the English and Chinese splits, respectively. It also outperforms the baselines on OCRBench and all four KIE benchmarks. On DocFigCapBench, it achieves the highest localization F1 and a caption score close to that of Qwen3.8-Max. For cross-page tables, it leads in Relation F1 on CrossTabBench and truncated-table TEDS on MPDocBench-Parse. These results demonstrate that a compact 0.8B model can achieve strong performance across diverse OCR-related tasks. OvisOCR3 is available at https://huggingface.co/ATH-MaaS/OvisOCR3.
Guangwei Zhang, Jianing Zhu University of Texas at Austin, Cheng Qian University of Illinois at Urbana-Champaign, Neil Zhenqiang Gong Duke University, Rada Mihalcea University of Michigan, Zhaozhuo Xu Workato and Stevens Institute of Technology, Jingrui He University of Illinois at Urbana-Champaign, Jiaqi W. Ma University of Illinois Urbana-Champaign 等 14 位作者
We present Copyright Detective, the first interactive forensic system for detecting, analyzing, and visualizing potential copyright risks in LLM outputs. The system treats copyright infringement versus compliance as an evidence discovery process rather than a static classification task due to the complex nature of copyright law. It integrates multiple detection paradigms, including content recall testing, paraphrase-level similarity analysis, persuasive jailbreak probing, and unlearning verification, within a unified and extensible framework. Through interactive prompting, response collection, and iterative workflows, our system enables systematic auditing of verbatim memorization and paraphrase-level leakage, supporting responsible deployment and transparent evaluation of LLM copyright risks even with black-box access. In our experiments with GPT-4o-mini, we demonstrate that the specific persuasive strategy “Pathos” shifts the leakage distribution from about 0.1 (ROUGE-L) to 0.7. Our live system is hosted on Streamlit server, with a demonstration video included as supplementary material.
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 3: System Demonstrations) · 2026 · 正式发表
Guoliang Zhao, Zixin Cui, Chao Ye, Dengwu He, Fei Huang, Yubo Liu, Shuanglong Li, Tzungren Kuo 等 13 位作者
The matching paradigm is fundamental to large-scale information retrieval and is widely used in industrial search and advertising systems. Existing approaches employ Large Language Models (LLMs) primarily as feature extractors, underutilizing their full modeling capabilities. To address this limitation, we propose a novel matching paradigm, termed the Unified Generative and Discriminative LLM (UGD). It integrates two-tower, single-tower, and generative tasks within a unified LLM framework via attention-mask partitioning, enabling generative tasks to serve as auxiliary supervision for discriminative learning and facilitating distillation from single-tower to two-tower architectures through a multi-task fine-tuning mechanism. To satisfy online latency constraints, we further introduce a self-distillation variant of UGD with a KMeans-enhanced linearized RQVAE for prompt compression and quantization. This design compresses and quantizes landing-page documents during inference, improving serving efficiency and reducing storage overhead. Extensive experiments show that UGD achieves superior performance and strong practical value. The framework has been deployed in an industrial search engine serving hundreds of millions of users and hundreds of thousands of advertisers, significantly enhancing search experience. Open access upon publication.
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
Bo Li Baidu Inc, Chuan Wu, Shaolin Zhu Tianjin University
Mixture-of-Experts Multimodal Large Language Models (MoE MLLMs) suffer from a significant efficiency bottleneck during Expert Parallelism (EP) inference due to the straggler effect. This issue is worsened in the multimodal context, as existing token-count-based load balancing methods fail to address two unique challenges: (1) Information Heterogeneity, where numerous redundant visual tokens are treated equally to semantically critical ones, and (2) Modality Dynamics, where varying visual to text ratios across tasks lead to resource misallocation. To address these challenges, we propose MACS (Modality-Aware Capacity Scaling), a training-free inference framework. Specifically, MACS introduces an Entropy-Weighted Load mechanism to quantify the semantic value of visual tokens, addressing information heterogeneity. Additionally, the Dynamic Modality-Adaptive Capacity mechanism allocates expert resources based on the real-time modal composition of the input. Extensive experiments demonstrate that MACS significantly outperforms existing methods on various multimodal benchmarks, providing a novel and robust solution for the efficient deployment of MoE MLLMs in EP inference.
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
Bobo Li National University of Singapore, Rui Wu Sichuan University, Zibo Ji, Meishan Zhang Harbin Institute of Technology (Shenzhen), China, Hao Fei National University of Singapore and University of Oxford, Min Zhang Harbin Institute of Technology, Shenzhen and Harbin Institute of Technology, Mong-Li Lee National University of Singapore, Wynne Hsu National University of Singapore
Large Language Model agents have rapidly evolved from static text generators into dynamic systems capable of executing complex autonomous workflows. To enhance reliability, multi-agent frameworks assigning specialized roles are increasingly adopted to enable self-reflection and mutual auditing. While such role-playing effectively leverages domain expert knowledge, we find it simultaneously induces a human-like cognitive bias known as Actor-Observer Asymmetry (AOA). Specifically, an agent acting as an actor (during self-reflection) tends to attribute failures to external factors, whereas an observer (during mutual auditing) attributes the same errors to internal faults. We quantify this using our new Ambiguous Failure Benchmark, which reveals that simply swapping perspectives triggers the AOA effect in over 20% of cases for most models. To tame this bias, we introduce ReTAS (Reasoning via Thesis-Antithesis-Synthesis), a model trained through dialectical alignment to enforce perspective-invariant reasoning. By integrating dialectical chain-of-thought with Group Relative Policy Optimization, ReTAS guides agents to synthesize conflicting viewpoints into an objective consensus. Experiments demonstrate that ReTAS effectively mitigates attribution inconsistency and significantly improves fault resolution rates in ambiguous scenarios.
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
Zhaoxin Feng, Zheng Chen Hong Kong University of Science and Technology, Jianfei Ma, Yip Tin Po, Emmanuele Chersoni The Hong Kong Polytechnic University, Bo Li
Alignment techniques often inadvertently induce sycophancy in LLMs. While prior studies studied this behaviour in direct-answer settings, the role of Chain-of-Thought (CoT) reasoning remains under-explored: does it serve as a logical constraint that mitigates sycophancy, or a tool for post-hoc rationalization that masks it? We evaluate a range of models across objective and subjective tasks to investigate the issue.Results show that reasoning generally reduces sycophancy in final decisions but also masks sycophancy in some samples, where models construct deceptive justifications through logical inconsistencies, calculation errors, and one-sided arguments etc. Furthermore, LLMs are more prone to sycophancy in subjective tasks and under authority-bias. Our mechanistic analysis reveals that the tendency of sycophancy in LLMs is dynamic during the reasoning process rather than being pre-determined at the input.
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
Jinming Wu ByteDance Inc., Zihao Deng, Wei Li Bytedance, Yiding Liu Tiktok, Bo You ByteDance Inc., Bo Li, Zejun MA TikTok, Ziwei Liu Nanyang Technological University
Robust deployment of large multimodal models (LMMs) in real-world scenarios requires access to external knowledge sources, given the complexity and dynamic nature of real-world information. Existing approaches such as retrieval-augmented generation (RAG) and prompt engineered search agents rely on rigid pipelines, often leading to inefficient or excessive search behaviors. We present MMSearch-R1, the first end-to-end reinforcement learning framework that enables LMMs to perform on-demand, multi-turn search in real-world Internet environments. Our framework integrates both image and text search tools, allowing the model to reason about when and how to invoke them guided by an outcome-based reward with a search penalty. To support training, We collect a multimodal search VQA dataset through a semi-automated pipeline that covers diverse visual and textual knowledge needs and curate a search-balanced subset with both search-required and search-free samples, which proves essential for shaping efficient and on-demand search behavior. Extensive experiments on knowledge-intensive and info-seeking VQA tasks show that our model not only outperforms traditional RAG-based baselines of the same model size, but also matches the performance of a larger RAG-based model while reducing search calls by over 30%. We further analyze key empirical findings to offer actionable insights for advancing research in multimodal search.
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
Lei Wei, Xiao Peng Alibaba Group, Tt, Guannan Zhang, Chenhao Jiang, Hongyu Li Alibaba Group, Lanbo Lin, Yuanwu Xu Alibaba Group 等 11 位作者
Large Language Model (LLM) agents struggle with ultra-long-horizon tasks requiring hundreds or thousands of interaction steps. Traditional context management approaches face a fundamental dilemma: preserving complete histories rapidly exhausts context windows and forces crude truncation, while aggressive summarization discards critical information prematurely. We propose Predictive Adaptive Context Extraction (PACE), a novel framework that reconceptualizes context management as a Next Step Prediction problem. Inspired by neural attention, PACE dynamically constructs context by adjusting historical memory granularity based on its predicted relevance for the next action. Comprehensive evaluation across diverse benchmarks and models demonstrates that PACE consistently improves task success rates, with larger gains on complex tasks and robust cross-lingual performance. Crucially, PACE enables agents to sustain effective reasoning for 4,897 interaction steps in ultra-long-horizon scenarios, achieving a 66.2 improvement over the full-context ReAct baseline and 5.1 over advanced folding baselines. This fundamentally advances the capability of LLM-based agents in previously intractable long-horizon scenarios. Our code and data are available at https://anonymous.4open.science/r/PACE-B000/.
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
Kairui Hu Nanyang Technological University, Penghao Wu Nanyang Technological University, Fanyi Pu, Wang Xiao, Xiang Yue Carnegie Mellon University, Bo Li, Yuanhan Zhang, Ziwei Liu Nanyang Technological University
Humans acquire knowledge through three cognitive stages: perceiving information, comprehending knowledge, and adapting knowledge to solve novel problems. Videos serve as an effective medium for knowledge acquisition, facilitating a natural progression through these learning stages. However, existing video benchmarks fail to evaluate the knowledge acquisition capabilities of Large Multimodal Models (LMMs). To address this gap, we introduce Video-MMMU, a multi-modal, multi-discipline, multi-track benchmark that evaluates LMMs’ ability to acquire knowledge from college-level, educational videos. Video-MMMU features a collection of 300 videos and 900 human-annotated questions across six disciplines, evaluating knowledge acquisition through stage-aligned question-answer pairs: Perception, Comprehension, and Adaptation. Beyond measuring final accuracy, Video-MMMU proposes the performance gain metric that quantifies an LMM’s learning gain from video, shifting the focus of evaluation from absolute performance to learning efficiency. Our evaluation reveals a substantial gap between human learners and current LMMs, highlighting the need to improve models’ ability to learn and adapt knowledge from video content.
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
Linxuan Du, Guangquan Xue, Xiaobo Liang, Qipeng Huang, Yuyang Ding, Xinyu Shi, Zhang Yijun China Mobile Communications Company Limited Research Institute, Ji Qi 等 11 位作者
Despite the potential of multi-turn self-reflection to improve LLM reasoning, its effectiveness in practice is severely constrained by a failure mode we term the Echo Trap.Specifically, this phenomenon gives rise to two coupled problems: (1) the model becomes limited by its inherent capabilities and tends to repeat earlier reflections to preserve reward signals; (2) once such “copy” behavior is reinforced, the model ceases to try new strategies, leading to exploration collapse.We attribute this issue to imprecise credit assignment during training, as standard GRPO assigns rewards at the trajectory level, making it difficult to distinguish which reflection steps contribute to improved outcomes.To address this limitation, we propose a tree-structured extension of GRPO for multi-turn self-reflection, which enables more accurate advantage estimation.Through extensive experiments, we analyze the Echo Trap and demonstrate that our method effectively mitigates behavior collapse and improves performance across multiple benchmarks.
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
Yi Jiang, Sendong Zhao, Jianbo Li, Haochun Wang Harbin Institute of Technology, Lizhe Zhang, Yan Liu, Bing Qin Harbin Institute of Technology
Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs), especially for knowledge-intensive tasks. Despite its advantages, current RAG methods often struggle to fully exploit knowledge during generation. In particular, the synergy between the model’s internal parametric knowledge and external retrieved knowledge remains limited. Retrieved contents may sometimes mislead generation, while certain generated content can guide the model toward more accurate outputs. In this work, we propose Collaborative Chain-of-Agents, a framework designed to enhance explicitly synergy over both parametric and retrieved knowledge. Specifically, we first introduce CoCoA-zero, a multi-agent RAG framework that first performs conditional knowledge induction and then reasons answers. Building on this, we develop CoCoA, a long-chain training strategy that synthesizes extended multi-agent reasoning trajectories from CoCoA-zero to fine-tune the LLM. This strategy enhances the model’s capability to explicitly integrate and jointly leverage parametric and retrieved knowledge. Experimental results demonstrate the superiority of CoCoA in open-domain QA and multi-hop QA.
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
The ability to model sparse and underspecified rewards, characteristic of human preferences, is fundamental to scaling Reinforcement Learning (RL). Current preference-based reward modeling largely relies on verifiable rewards, where human-annotated labels define rule-based signals. However, these methods face a fundamental bottleneck we term the Matryoshka Doll Problem: a recursive dependency where each reward verifier requires a meta-verifier, leading to continuous and costly dependence on human annotation. In this work, we propose Dual RM, which couples discriminative and generative reward models (DisRMs and GenRMs) under a non-parametric meta-reward. Rather than verifying the correctness of GenRM’s reasoning, the meta-reward evaluates its practical impact on response quality. Specifically, GenRM identifies multi-dimensional evaluation rubrics and iteratively refines the response, while DisRM quantifies the quality shifts induced by each rubric. Furthermore, we implement rubric-based test-time scaling to improve sample efficiency and preference alignment under both DPO and GRPO. Our experiments demonstrate that Dual RM achieves strong performance across major preference benchmarks. Notably, even when trained exclusively on language modality, it exhibits robust cross-modal transfer on Omni-RewardBench.
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
Zhenguang Wang, Bo Li, Wenhui Tan, Peng Cao Northeastern University, Yang Wang, Jia Duan Nanjing Medical University, Fei Wang, Osmar Zaiane University of Alberta
Standard depression assessment relies on instruments such as the clinician-rated Hamilton Depression Rating Scale (HAMD) and the patient-reported Patient Health Questionnaire (PHQ-8), but manual scoring is time-consuming and subject to inter-rater variability. Prior automated approaches typically regress a single total score or coarse severity category, lacking the fine-grained subscore-level supervision needed for precise clinical diagnosis. To address this, we propose MTSP (Multi-Task Subscore Prediction), a fine-grained model for subscore prediction via multi-task learning. MTSP achieves state-of-the-art performance on the public E-DAIC dataset (MAE 3.48, RMSE 4.57) and generalizes well to the public PDCH and a large-scale private clinical dataset (CIDH), outperforming total-score regression baselines and Qwen3-14B direct scoring. We further show that multi-task learning is essential, subscore-level supervision improves assessment by better capturing symptom-cluster structure, and prior constraints plus task-level self-paced learning enhance robustness to subscore difficulty and annotation noise.
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
Bo Li Peking University, Mingda Wang Hebei University of Technology, Gexiang Fang, Shikun Zhang Peking University, Wei Ye Peking University
We revisit retrieval-augmented generation (RAG) by embedding retrieval control directly into generation. Instead of treating retrieval as an external intervention, we express retrieval decisions within token-level decoding, enabling end-to-end coordination without additional controllers or classifiers. Under the paradigm of Retrieval as Generation, we propose GRIP (Generation-guided Retrieval with Information Planning), a unified framework in which the model regulates retrieval behavior through control-token emission. Central to GRIP is Self-Triggered Information Planning, which allows the model to decide when to retrieve, how to reformulate queries, and when to terminate, all within a single autoregressive trajectory. This design tightly couples retrieval and reasoning and supports dynamic multi-step inference with on-the-fly evidence integration. To supervise these behaviors, we construct a structured training set covering answerable, partially answerable, and multi-hop queries, each aligned with specific token patterns. Experiments on five QA benchmarks show that GRIP surpasses strong RAG baselines and is competitive with GPT-4o while using substantially fewer parameters. Code and resources are provided in the supplementary materials.
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
Ziyi Wang, Yuxuan Lu, Wenbo Li, Amirali Amini Northeastern University and Sharif University of Technology, Bo Sun Northeastern University, Yakov Bart Northeastern University, Weimin Lyu Amazon, Jiri Gesi Amazon Science 等 16 位作者
Can Large Language models (LLMs) accurately simulate the next web action of a specific user? While LLMs have shown promising capabilities in generating believable human behaviors, evaluating their ability to mimic real user behaviors remains an open challenge, largely due to the lack of high-quality, publicly available datasets that capture both the observable actions and the internal reasoning of an actual human user. To address this gap, we introduce OPeRA, a novel dataset of Observation, Persona, Rationale, and Action collected from real human participants during online shopping sessions. OPeRA is the first public dataset that comprehensively captures: user personas, browser observations, fine-grained web actions, and self-reported just-in-time rationales. We developed both an online questionnaire and a custom browser plugin to gather this dataset with high fidelity. Using OPeRA, we establish the first benchmark to evaluate how well current LLMs can predict a specific user’s next action and rationale with a given persona and history. This dataset lays the groundwork for future research into LLM agents that aim to act as personalized digital twins for human.
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
Bo Li Peking University, Mingda Wang Hebei University of Technology, Shikun Zhang Peking University, Wei Ye Peking University
Instruction tuning relies on large instruction–response corpora whose quality and composition strongly affect downstream performance. We propose Answer Divergence-Guided Selection (ADG), which selects instruction data based on the geometric structure of multi-sample outputs. ADG draws several high-temperature generations per instruction, maps responses into an embedding space, and computes an output divergence score that jointly encodes dispersion magnitude and shape anisotropy. High scores correspond to instructions whose answers are both far apart and multi-modal, rather than clustered paraphrases along a single direction. Across two backbones and three public instruction pools, fine-tuning on only 10K ADG-selected examples consistently outperforms strong selectors on six benchmarks spanning reasoning, knowledge, and coding. Analyses further show that both dispersion magnitude and shape anisotropy are necessary, supporting answer divergence as a practical signal for instruction data selection. Code and appendix are included in the supplementary materials.
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