论文检索

Arnav Goel, Pranjal A Chitale, Bhawna Paliwal, Bishal Santra, Amit Sharma

User behavior in the real world is diverse, cross-domain, and spans long time horizons. Existing user modeling benchmarks however remain narrow, focusing mainly on short sessions and next-item prediction within a single domain. Such limitations hinder progress toward robust and generalizable user models. We present HORIZON, a new benchmark that reformulates user modeling along three axes i.e. dataset, task, and evaluation. Built from a large-scale, cross-domain reformulation of Amazon Reviews, HORIZON covers 54M users and 35M items, enabling both pretraining and realistic evaluation of models in heterogeneous environments. Unlike prior benchmarks, it challenges models to generalize across domains, users, and time, moving beyond standard missing-positive prediction in the same domain. We propose new tasks and evaluation setups that better reflect real-world deployment scenarios. These include temporal generalization, sequence-length variation, and modeling unseen users, with metrics designed to assess general user behavior understanding rather than isolated next-item prediction. We benchmark popular sequential recommendation architectures alongside LLM-based baselines that leverage long-term interaction histories. Our results highlight the gap between current methods and the demands of real-world user modeling, while establishing HORIZON as a foundation for research on temporally robust, cross-domain, and general-purpose user models.

Yucan Guo, Miao Su, Saiping Guan, Zihao Sun, Xiaolong Jin, Jiafeng Guo, Xueqi Cheng

Retrieval-Augmented Generation (RAG) integrates non-parametric knowledge into Large Language Models (LLMs), typically from unstructured texts and structured graphs. While recent progress has advanced text-based RAG to multi-turn reasoning through Reinforcement Learning (RL), extending these advances to hybrid retrieval introduces additional challenges. Existing graph-based or hybrid systems typically depend on fixed or handcrafted retrieval pipelines, lacking the ability to integrate supplementary evidence as reasoning unfolds. Besides, while graph evidence provides relational structures crucial for multi-hop reasoning, it is substantially more expensive to retrieve. To address these limitations, we introduce RouteRAG, an RL-based framework that enables LLMs to perform multi-turn and adaptive graph-text hybrid RAG. RouteRAG jointly optimizes the entire generation process via RL, allowing the model to learn when to reason, what to retrieve from either texts or graphs, and when to produce final answers, all within a unified generation policy. To guide this learning process, we design a two-stage training framework that accounts for both task outcome and retrieval efficiency, enabling the model to exploit hybrid evidence while avoiding unnecessary retrieval overhead. Experimental results across five question answering benchmarks demonstrate that RouteRAG significantly outperforms existing RAG baselines, highlighting the benefits of end-to-end RL in supporting adaptive and efficient retrieval for complex reasoning.

Yucheng Zhou, Jianbing Shen

Autoregressive models have shown superior performance and efficiency in image generation, but remain constrained by high computational costs and prolonged training times in video generation. In this study, we explore methods to accelerate training for autoregressive video generation models through empirical analyses. Our results reveal that while training on fewer video frames significantly reduces training time, it also exacerbates error accumulation and introduces inconsistencies in the generated videos. To address these issues, we propose a Local Optimization (Local Opt.) method, which optimizes tokens within localized windows while leveraging contextual information to reduce error propagation. Inspired by Lipschitz continuity, we propose a Representation Continuity (ReCo) strategy to improve the consistency of generated videos. ReCo utilizes continuity loss to constrain representation changes, improving model robustness and reducing error accumulation. Extensive experiments on four class-to-video datasets demonstrate that our approach achieves superior performance to the baseline while halving the training cost without sacrificing quality.

Jingyuan Yan, Qingchen Liu, Qichao Ma, Jiahu Qin

Complex flight tasks demand both intricate, long-horizon decision-making and precise operations, which imposes immense cognitive, knowledge and experience demands for pilots and highlights the need for advanced copilot systems. While Large Language Models (LLMs) bring powerful potential to this area, a comprehensive LLM-based copilot system—one that addresses deficiencies in task-level adaptability and fine-grained decision support while integrating with a high-fidelity environment—is critically lacking. To address this gap, we present FalconCopilot, pioneering the first such comprehensive system, composed of two parts: 1) Textual DCS, an interface built upon Digital Combat Simulator (DCS) World that unifies multi-modal cockpit data and piloting knowledge into a stable semantic interface for LLMs. Building on this interface, we introduce 2) FalconAgent, an LLM-powered copilot agent that performs optimized task planning, incorporating capabilities for multi-crew task allocation and procedural pruning. Our built-in human-AI interaction is grounded by a bidirectional feedback loop of runtime verification and human correction. In human-in-the-loop experiment, FalconCopilot shortens task completion time while attaining a level of performance approaching that of a human instructor.

Linfeng Gao, Qinggang Zhang, Baolong Bi, Bo Zeng, Zheng Yuan, Zerui Chen, Zhimin Wei, Shenghua Liu, Linlong Xu, Longyue Wang 等

Retrieval-Augmented Generation (RAG) systems often fail to maintain contextual faithfulness, generating responses that conflict with the provided context. Existing methods attempt to improve faithfulness through external interventions, such as specialized prompting, decoding-based calibration, or preference optimization. However, since these approaches treat the LLM as a black box, they lack a reliable mechanism to assess how these conflicts occur. Consequently, they tend to be brittle, data-intensive, and agnostic to the model’s internal reasoning process. In this paper, we move beyond black-box interventions to analyze the model’s internal reasoning process. We discover that conflicting and aligned knowledge states are linearly separable in the model’s latent space, and contextual noise systematically increases the entropy of these representations. Based on these findings, we propose ProbeRAG, a novel framework for faithful RAG that operates in three stages: (i) fine-grained knowledge pruning to filter irrelevant context, (ii) latent conflict probing to identify hard conflicts in the model’s latent space, and (iii) conflict-aware attention to modulate attention heads toward faithful context integration. Extensive experiments demonstrate that ProbeRAG substantially improves both accuracy and contextual faithfulness. The related resources are available at https://github.com/XMUDeepLIT/ProbeRAG.

Qirui Liu, Hao Chen, Weijie Shi, Jiajie Xu, Jia Zhu

Accurately identifying student misconceptions is crucial for personalized education but faces three challenges: (1) data scarcity with long-tail distribution, where authentic student reasoning is difficult to synthesize; (2) fuzzy boundaries between error categories with high annotation noise; (3) deployment paradox—large models overlook unconventional approaches due to pretraining bias and cannot be deployed on edge, while small models overfit to noise. Unlike traditional methods that increase diversity through large-scale data synthesis, we propose a two-stage knowledge distillation framework that mines high-value samples from existing data. The first stage performs standard distillation to transfer task capabilities. The second stage introduces a dual-layer marginal selection mechanism based on cognitive uncertainty, identifying four types of critical samples based on teacher model uncertainty and confidence differences. For different data subsets, we design difficulty-adaptive mechanism to balance hard/soft label contributions, enabling student models to inherit inter-class relationships from teacher soft labels while distinguishing ambiguous error types. Experiments show that with augmented training on only 10.30% of filtered samples, we achieve MAP@3 of 0.9585 (+17.8%) on the MAP-Charting dataset, and using only a 4B parameter model, we attain 84.38% accuracy on cross-topic tests of middle school algebra misconception benchmarks, significantly outperforming sota LLM (67.73%) and standard fine-tuned 72B models (81.25%). Our code is available at https://anonymous.4open.science/r/acl2026_map-5847/.

Yukang Feng, Jianwen Sun, Zelai Yang, Jiaxin Ai, Chuanhao Li, Zizhen Li, Fanrui Zhang, Kang He, Rui Ma, Jifan Lin 等

Recent advances in AI-assisted programming have empowered agents to execute complex workflows via command-line interfaces, however, existing benchmarks are limited by short task horizons, data contamination from GitHub scraping, and a lack of fine-grained evaluation metrics, fail to rigorously evaluate the long-horizon planning and execution capabilities essential for realistic software engineering. To address these gaps, we introduce LongCLI-Bench, a comprehensive benchmark designed to evaluate agentic capabilities across long-horizon, realistic, sequential engineering tasks. We curated 20 high-quality, long-horizon tasks from over 1,000 computer science assignments and real-world workflows, covering four engineering categories: from scratch, feature addition, bug fixing, and refactoring. LongCLI-Bench employs a dual-set testing protocol, which measures requirement fulfillment (fail(→)pass) and regression avoidance (pass(→)pass), and incorporates step-level scoring to pinpoint execution failures. Extensive experiments reveal that even state-of-the-art agents achieve pass rates below 20% in LongCLI-Bench. Step-level analysis further indicates that the majority of tasks stall at less than 30% completion, highlighting that critical failures often occur in the early stages. Although self-correction offers marginal gains, human-agent collaboration through plan injection and interactive guidance yields significantly higher improvements. These results highlight that future research must emphasize the development of synergistic human-agent workflows alongside advances in agents’ planning and execution capabilities to overcome key challenges in long-horizon task performance.

Miao Su, Yucan Guo, Zhongni Hou, Long Bai, Zixuan Li, Yufei Zhang, Guojun Yin, Wei Lin, Xiaolong Jin, Jiafeng Guo 等

Memory enables Large Language Model (LLM) agents to perceive, store, and use information from past dialogues, which is essential for personalization. However, existing methods fail to properly model the temporal dimension of memory in two aspects: 1) Temporal inaccuracy: memories are organized by dialogue time rather than their actual occurrence time; 2) Temporal fragmentation: existing methods focus on point-wise memory, losing durative information that captures persistent states and evolving patterns. To address these limitations, we propose Temporal Semantic Memory (TSM), a memory framework that models semantic time for point-wise memory and supports the construction and utilization of durative memory. During memory construction, it first builds a semantic timeline rather than a dialogue one. Then, it consolidates temporally continuous and semantically related information into a durative memory. During memory utilization, it incorporates the query’s temporal intent on the semantic timeline, enabling the retrieval of temporally appropriate durative memories and providing time-valid, duration-consistent context to support response generation. Experiments on LongMemEval and LoCoMo show that TSM consistently outperforms existing methods and achieves up to 12.2% absolute improvement in accuracy, demonstrating the effectiveness of the proposed method.

Shlok Shelat, Jay Raval, Souvik Roy, Manas Gaur

Large language models (LLMs) have demonstrated strong performance on formal language tasks, yet whether this reflects genuine symbolic reasoning or pattern matching on familiar constructions remains unclear. We introduce a benchmark for deterministic finite automata (DFA) construction from regular languages, comprising factual knowledge questions, seen construction problems from public sources, and two types of unseen problems: hand-crafted instances with multiple interacting constraints and systematically generated problems via Arden’s theorem. Models achieve perfect accuracy on factual questions and 84-90% on seen tasks. However, accuracy drops sharply on unseen problems (by 30-64%), with failures stemming from systematic misinterpretation of language constraints, incorrect handling of Kleene-star semantics, and a failure to preserve global consistency. We evaluate a three-stage hint protocol that enables correction of shallow errors but does not reliably resolve globally inconsistent or structurally flawed automata. Our analysis across multiple prompting strategies (direct, Chain-of-Thought, Tree-of-Thought) reveals that errors persist regardless of prompting approach, exposing a fundamental gap between LLMs’ ability to generate syntactically plausible DFAs and their capacity for semantically correct formal reasoning.

Peijie Wang, Ming-Liang Zhang, Jun Cao, Chao Deng, Dekang Ran, Pi Bu, Hongda Sun, Xuan Zhang, Yingyao Wang, Jun Song 等

Multimodal Large Language Models (MLLMs) have achieved remarkable progress but continue to struggle with geometric reasoning, primarily due to the perception bottleneck regarding fine-grained visual elements. While formal languages have aided plane geometry understanding, solid geometry which requires spatial understanding remains largely unexplored. In this paper, we address this challenge by designing a unified formal language that integrates plane and solid geometry, comprehensively covering geometric structures and semantic relations. We construct GDP-29K, a large-scale dataset comprising 20k plane and 9k solid geometry samples collected from diverse real-world sources, each paired with its ground-truth formal description. We propose a training paradigm combining Supervised Fine-Tuning with Reinforcement Learning via Verifiable Rewards, which effectively enforces syntactic correctness and geometric consistency. Experiments show that our approach achieves state-of-the-art parsing performance. Furthermore, we demonstrate that our parsed formal descriptions serve as a critical cognitive scaffold, significantly boosting MLLMs’ capabilities for downstream geometry reasoning tasks.

Zhenqin Li, ShengYong Ding, Shuangyin Li

Question-Answer Generation (QAG) is essential for alleviating the cold-start problem in domain-specific large language model (LLM) post-training, where high-quality data is severely scarce.Effective training samples include rich semantic diversity and rigorous factual consistency.Thus, it is necessary to consider the inherent tension between semantic breadth and factual fidelity.However, it is extremely challenging to trade off semantic diversity against factual consistency, in that generalization across the semantic space must be achieved effectively and reliably, and factual integrity must be ensured as well.To address this issue, we propose an effective framework, namely DiFRa, that integrates continuous concept diffusion with discrete knowledge graph constraints to balance semantic diversity and factual consistency.Specifically, the proposed DiFRa models discrete concepts as a continuous latent distribution to sample embeddings that capture rich semantic variations, and constructs a refined knowledge graph as explicit factual constraints.Then, a diversity and consistency aware mechanism is designed to dynamically integrate both embeddings and the knowledge graph for QA pairs generation.Furthermore, we introduce SeFa, which harmonizes semantic entropy and consistency scores to quantify the trade-off between diversity and correctness.Extensive experiments demonstrate that DiFRa consistently outperforms the baseline models, validating its efficacy in reconciling the tension to generate semantically diverse and factually consistent QA pairs. The source code is publicly available.

Kailun Lyu, Fu Zhang, Zehan Li, Jingwei Cheng

Zero-Shot Relation Extraction (ZSRE) aims to predict unseen relations for given entity pairs in sentences. Existing methods typically operate from a local perspective, predicting the relation for each entity pair (given its corresponding sentence) in isolation. Consequently, they often fail to distinguish between unseen, semantically similar relations, particularly when the sentence phrasing is ambiguous.To address this limitation, we propose **G-NSR**, a novel ZSRE framework built upon a **G**lobal **N**euro-**S**ymbolic **R**easoner architecture, specifically designed to enable global reasoning across a set of predictions. The key idea is to model the logical relationships among multiple predictions, and perform neuro-symbolic reasoning to ensure logically consistent and more accurate predictions. Specifically, we first introduce Duality Type-Constrained Relation Schemas, which formulate each candidate relation as a pair of complementary positive-negative propositions. These propositions are then synthesized by our designed Neuro-Symbolic Reasoner, which explicitly models their logical interdependencies. By approximating logical rules, the reasoner allows high-confidence predictions to serve as evidence for refining incorrect results, ensuring the final predictions are logically consistent and more accurate. Extensive experiments on widely used datasets demonstrate that our method significantly outperforms existing approaches and establishes new state-of-the-art results across all evaluation settings. Our code is available at https://anonymous.4open.science/r/G-NSR

Guangyi Liu, Pengxiang Zhao, Liang Liu, Zhiming Chen, Yuxiang Chai, Yaozhen Liang, WenHao Wang, Siheng Chen, Zhengxi Lu, Shuai Ren 等

Mobile GUI agents show promise in automating tasks but face significant generalization challenges in long-tail scenarios. While learning from few-shot demonstrations is an emerging solution, its progress is hindered by two critical gaps: the lack of a comprehensive benchmark for systematic evaluation on mobile devices, and the absence of a systematic framework designed to learn from demonstrations in this domain. To address these gaps, we introduce LearnGUI, the first comprehensive benchmark designed for studying demonstration-based learning in mobile agents, comprising 2,252 offline and 101 online tasks. We further develop LearnAct, a modular agent framework engineered to systematically extract, retrieve, and leverage knowledge from visual demonstrations. Extensive evaluations across six backbone models validate our approach: LearnAct achieves dramatic improvements for general-purpose models (e.g., Gemini-2.5-Pro: 38.5%→58.9%) and specialized models alike (e.g., UI-TARS-7B-SFT’s online success rate: 18.1%→32.8%), demonstrating consistent gains across model architectures. Our work provides a robust benchmark and a systematic framework, paving the way for more adaptable and practical mobile agents. Our code and data are publicly available at https://lgy0404.github.io/LearnAct/.

Junkai Li, Yunghwei Lai, Tianyi Zhu, Zheng Long Lee, Weizhi Ma, Yang Liu

Formulating a treatment plan is inherently a complex reasoning and refinement task rather than a simple generation problem. However, existing large language models (LLMs) mainly rely on one-shot output without explicit verification, which may result in rough, incomplete, and potentially unsafe treatment plans. To address these limitations, we propose TheraAgent, an agentic framework that replaces one-shot generation with an iterative generate-judge-refine pipeline. By mirroring the actual reasoning process of human experts who iteratively revise treatment plans, our framework progressively transforms coarse and incomplete drafts into precise, comprehensive, and safer therapeutic regimens. To facilitate the critical judge component, we introduce TheraJudge, a treatment-specific evaluation module integrated into the inference loop to enforce clinical standards. Experiments show TheraAgent achieves state-of-the-art results on HealthBench, leading in Accuracy and Completeness. In expert evaluations, it attains an 86% win rate against physicians, with superior Targeting and Harm Control. Moreover, the highly agreement between TheraJudge and HealthBench evaluations confirms the reliability of our framework.

Hanmeng Zhong, Wentao Wu, Linqing Chen, Peng Zhou

Navigating biopharmaceutical intellectual property necessitates precisely associating visual chemical structures with their textual referents across lengthy documents. Despite its critical role in drug discovery, this multimodal coreference task remains underexplored. It presents unique challenges, including handling Markush structures and distinguishing the atom-level differences between adjacent structures. To bridge this gap, we define the multimodal Chemical Structure-Text coreference and introduce CheST, the first dataset explicitly designed for the task. Furthermore, to satisfy the strict logical consistency in the task, we propose RULER, a RULE-guided multimodal Reinforcement learning framework built upon an SFT cold start. RULER utilizes rule-driven reward functions operationalizing multidimensional consistencies, acting as a domain-specific "verifier" to obtain the correct domain knowledge. Experimental results demonstrate that RULER achieves a 40% improvement over the strongest baseline–Gemini-2.5-Pro, demonstrating the superior efficacy.

Jiahang Lin, Kai Hu, Binghai Wang, Yuhao Zhou, Zhiheng Xi, Honglin Guo, Shichun Liu, Junzhe Wang, Shihan Dou, Enyu Zhou 等

Conventional Retrieval-Augmented Generation (RAG) systems often struggle with complex multi-hop queries over long documents due to their single-pass retrieval. We introduce **MM-Doc-R1**, a novel framework that employs an agentic, vision-aware workflow to address long document visual question answering through iterative information discovery and synthesis. To incentivize the information seeking capabilities of our agents, we propose **Similarity-based Policy Optimization (SPO)**, addressing baseline estimation bias in existing multi-turn reinforcement learning (RL) algorithms like GRPO. Our core insight is that in multi-turn RL, the more semantically similar two trajectories are, the more accurate their shared baseline estimation becomes. Leveraging this, SPO calculates a more precise baseline by similarity-weighted averaging of rewards across multiple trajectories, unlike GRPO which inappropriately applies the initial state’s baseline to all intermediate states. This provides a more stable and accurate learning signal for our agents, leading to superior training performance that surpasses GRPO. Our experiments on the MMLongbench-Doc benchmark show that **MM-Doc-R1** outperforms previous baselines by **10.4%**. Furthermore, **SPO** demonstrates superior performance over **GRPO**, boosting results by **5.0%** with Qwen3-8B and **6.1%** with Qwen3-4B. These results highlight the effectiveness of our integrated framework and novel training algorithm in advancing the state-of-the-art for complex, long-document visual question answering.

Yongtong Gu, Songze Li, Xia Hu

The increasing misuse of AI-generated texts (AIGT) has motivated the rapid development of AIGT detection methods. However, the reliability of these detectors remains fragile against adversarial evasions. Existing attack strategies often rely on white-box assumptions or demand prohibitively high computational and interaction costs, rendering them ineffective under practical black-box scenarios. In this paper, we propose Multi-stage Alignment for Style Humanization (MASH), a novel framework that evades black-box detectors based on style transfer. MASH sequentially employs style-injection supervised fine-tuning, direct preference optimization, and inference-time refinement to shape the distributions of AI-generated texts to resemble those of human-written texts. Experiments across 6 datasets and 5 detectors demonstrate the superior performance of MASH over 11 baseline evaders. Specifically, MASH achieves an average Attack Success Rate (ASR) of 92%, surpassing the strongest baselines by an average of 24%, while maintaining superior linguistic quality.

Gyunyeop Kim, Sangwoo Kang

Reinforcement learning with verifiable rewards has improved reasoning in language models, but it typically relies on a ground-truth answer or an external verifier, which limits applicability and increases cost. We propose an answer-free training objective that derives rewards solely from the model’s own probabilities by exploiting prompt paraphrases as multiple semantic views of the same intent. For each paraphrase set, we generate candidate responses, rescore each response under the other paraphrased prompts via teacher forcing, and define a cross-prompt consensus reward that serves as a practical internal training signal, favoring responses supported across views rather than those that fit only a single phrasing. We optimize this reward using a policy update with an all-pairs objective and advantage broadcasting across prompt–response pairs. The framework naturally supports prefix-level training, enabling a controllable cost–signal trade-off. Experiments on RobustAlpacaEval and out-of-domain reasoning benchmarks (OpenBookQA, AQuA, HumanEval) show strong in-domain gains and competitive or improved average out-of-domain performance over pre-trained and answer-free training baselines on LLaMA3.2-3B and Qwen3-4B, alongside analyses demonstrating reward–performance alignment and the importance of design choices such as excluding self-view scores and ensembling-based candidates. All experiment code is available at our GitHub.

Yongxuan Wu, Runyu Chen, Peiyu Liu, Hongjin Qian

Long-context understanding poses significant challenges in natural language processing, particularly for real-world dialogues characterized by high redundancy and uneven information density. Although large language models (LLMs) achieve impressive results on existing benchmarks, these datasets fail to reflect the complexities of such texts, limiting their applicability to practical scenarios. To bridge this gap, we construct the first spoken long-text dataset, derived from live streams, designed to reflect the redundancy-rich and conversational nature of real-world scenarios. We construct tasks in three categories: retrieval, reasoning, and hybrid tasks. We then evaluate both popular LLMs and specialized methods to assess their ability to understand long contexts in these tasks. Our results show that current methods exhibit strong task-specific preferences and perform poorly on highly redundant inputs, with no single method consistently outperforming others. We propose a new baseline that better handles redundancy in spoken text and achieves strong performance across tasks. Our findings highlight key limitations of current methods and suggest future directions for improving long-context understanding. Finally, our benchmark fills a gap in evaluating long-context spoken language understanding and provides a practical foundation for developing real-world e-commerce systems. The code and benchmark are available at https://github.com/Yarayx/livelongbench.

Yiguo Deng, Xia Lei, Yuan Zhang, Long Ye

Traditional psychological counseling struggles to meet public demand due to high costs, social stigma, and limited accessibility. Recently, large language models (LLMs) have shown great potential in healthcare, offering new opportunities to build accessible mental health dialogue systems. However, current LLMs often lack accurate modeling of cognitive empathy, especially the ability to understand users’ emotions and their underlying psychological causes. To address this, we propose CogEmp, a dialogue generation model tailored for the Chinese cultural context that integrates cognitive empathy. The model follows a three-stage decision pipeline: emotion and cause recognition, contextual understanding, and empathetic response generation. First, the model identifies the user’s fine-grained emotions and their underlying causes within the Chinese context, laying the foundation for personalized emotional comprehension. Then, it retrieves semantically similar counseling cases to extract topic and strategy information, thereby constructing a context-aware representation. Finally, guided by the extracted multi-dimensional cues, the model drives LLMs to generate empathetic responses that are both contextually appropriate and professionally grounded. Experiments conducted on Chinese mental health datasets show that CogEmp outperforms existing approaches in key evaluation metrics, particularly in empathy, comprehensibility, and professionalism.