论文检索

Zhen Yang, Xinyue Zhang, Ping Jian, Chengzhi Li, Zhongbin Guo, Jiaping Feng, Wenpeng Lu

Despite the remarkable performance across numerous tasks, Large Language Models (LLMs) still exhibit notable deficiencies in temporal reasoning, even in simple event ordering tasks. For instance, a slight alteration in the temporal phrasing of the question (e.g., changing "Is event A before B?” to "Is event A after B?") can lead LLMs to hallucinate and produce inconsistent answers, reflecting a lack of robust temporal reasoning. Although many prior studies have focused on benchmarking and improving the temporal reasoning ability of LLMs, little is known about the intrinsic mechanisms within LLMs when performing temporal reasoning. In this work, we investigate the mechanistic interpretability of temporal ordering within event temporal reasoning through a structured "Identify-Interpret-Verify” pipeline. We first employ path patching to identify a sparse subset of attention heads that are causally responsible for reasoning outcomes. Detailed pattern analysis reveals that these key heads specialize in attending to either temporal keywords (semantic cues) or structural delimiters (syntactic cues). Furthermore, we rigorously validate the observed mechanism through comprehensive intervention-based experiments, ranging from head ablation to targeted attention modulation. We demonstrate that dynamically modulating the attention of these specific heads can robustly enhance model performance, which serves as strong empirical evidence that our identified mechanism faithfully captures the internal logic of temporal ordering in LLMs.

Hao An, Yibin Lou, Jiayi Guo, Yang Xu

Large language models (LLMs) often exhibit hallucinations due to their inability to accurately perceive their own knowledge boundaries. Existing abstention fine-tuning methods typically partition datasets directly based on response accuracy, causing models to suffer from severe label noise near the decision boundaries and consequently exhibit high rates of abstentions or hallucinations. This paper adopts a latent space representation perspective, revealing a "gray zone" near the decision hyperplane where internal belief ambiguity constitutes the core performance bottleneck. Based on this insight, we propose the **GeoDe** (**Geo**metric **De**noising) framework for abstention fine-tuning. This method constructs a truth hyperplane using linear probes and performs "geometric denoising" by employing geometric distance as a confidence signal for abstention decisions. This approach filters out ambiguous boundary samples while retaining high-fidelity signals for fine-tuning. Experiments across multiple models (Llama3, Qwen3) and benchmark datasets (TriviaQA, NQ, SciQ, SimpleQA) demonstrate that GeoDe significantly enhances model truthfulness and demonstrates strong generalization in out-of-distribution (OOD) scenarios. Code is available at https://github.com/Notbesidemoon/GeoDe.

Hwanjun Song

Much of the alignment tuning literature is organized around optimization objectives, while the construction of alignment data is often treated implicitly. In this survey, we adopt a data centric perspective and reframe alignment tuning as a pipeline design problem. We decompose alignment data construction into three interacting stages, response synthesis, preference evaluation, and preference instantiation, and use this framework to organize existing alignment methods into a unified taxonomy. Through this lens, we identify recurring design trade-offs and failure modes observed across prior alignment methods, and distill a set of high level principles that clarify how pipeline design choices influence the resulting optimization signal. Finally, we outline open challenges for alignment data pipelines, including prompt-level alignment, agentic settings, and alignment under evolving objectives.

Hongjin Qian, Siqi Bao, Zhao Cao, Zheng Liu

Information seeking bridges the knowledge gap between a query and its answer. Although LLMs perform well broadly, their ability to close this gap is limited by pretraining and degrades on specialized or up-to-date queries. A common remedy augments LLMs with external knowledge, either by injecting retrieved evidence into context or interleaving retrieval with reasoning. The former limits exploration of layered dependencies, while the latter is bounded by context length, constraining efficiency and scalability. For complex tasks with intricate dependencies and large text volumes, both approaches become inadequate.To tackle this bottleneck, we present AWARE (Agentic Knowledge Warehouse), an agentic knowledge warehousing framework that transforms heterogeneous, unstructured data into minimal, task-conditioned knowledge representations consumable by LLMs. Rather than exposing raw text, AWARE constructs knowledge through intent planning, online multi-threaded exploration, and map-reduce evidence integration, producing compact, LLM-ready context under finite budgets. Specifically, it applies offline document structuring to generate document headers that support controlled access, performs exploration with targeted refinement to recover layered information dependencies, and integrates distributed evidence into task-aware representations for downstream answer generation. Experiments on GAIA, WebWalker, and BrowseComp-Plus show improvements over all baselines

Xiaojie Gu, Ziying Huang, Weicong Hong, Jian Xie, Renze Lou, Kai Zhang

Large Language Models (LLMs) internalize vast world knowledge as parametric memory, yet inevitably inherit the staleness and errors of their source corpora. Consequently, ensuring the reliability and malleability of these internal representations is imperative for trustworthy real-world deployment. Knowledge editing offers a pivotal paradigm for surgically modifying memory without retraining. However, while recent editors demonstrate high success rates on standard benchmarks, it remains questionable whether current evaluation frameworks that rely on assessing output under specific prompting conditions can reliably authenticate genuine memory modification. In this work, we introduce a rigorous diagnostic framework that subjects models to discriminative self-assessment under diverse contextual pressures, specifically designed to scrutinize the subtle behavioral nuances induced by memory modifications. This probing reveals a pervasive phenomenon of Surface Compliance, where editors achieve high benchmark scores by merely mimicking target outputs without structurally overwriting internal beliefs. Moreover, we find that recursive modifications accumulate representational residues, triggering cognitive instability and permanently diminishing the reversibility of the model’s memory state. These insights underscore the risks of current editing paradigms and highlight the pivotal role of robust memory modification in building trustworthy, long-term sustainable LLM systems.

Xuchen Li, Jing Chen, Xuzhao Li, Hao Liang, Xiaohuan Zhou, Taifeng Wang, Wentao Zhang

In mathematical reasoning tasks, the advancement of Large Language Models (LLMs) relies heavily on high-quality training data with clearly defined and well-graded difficulty levels. However, existing data synthesis methods often suffer from limited diversity and lack precise control over problem difficulty, making them insufficient for supporting efficient training paradigms such as curriculum learning. To address these challenges, we propose MathMixup, a novel data synthesis paradigm that systematically generates high-quality, difficulty-controllable mathematical reasoning problems through hybrid and decomposed strategies. Automated self-checking and manual screening are incorporated to ensure semantic clarity and a well-structured difficulty gradient in the synthesized data. Building on this, we construct the MathMixupQA dataset and design a curriculum learning strategy that leverages these graded problems, supporting flexible integration with other datasets. Experimental results show that MathMixup and its curriculum learning strategy significantly enhance the mathematical reasoning performance of LLMs. Fine-tuned Qwen2.5-7B achieves an average score of 52.6% across seven mathematical benchmarks, surpassing previous state-of-the-art methods. These results fully validate the effectiveness and broad applicability of MathMixup in improving the mathematical reasoning abilities of LLMs and advancing data-centric curriculum learning.

Ziyan Wang, Yingpeng Du, Tianjun Wei, Haoyan Chua, Jieyi Bi, Jie Zhang, Zhu Sun

Large language models (LLMs) show potential for multi-interest analysis of users in recommender systems, going beyond heuristic assumptions in existing methods, e.g., co-occurring items indicate the same interest. Despite the effectiveness, two key challenges remain. First, the granularity of raw generation of LLMs for multi-interests is agnostic, possibly leading to overly fine or coarse interest grouping. Second, adopting LLM to analyze individual user behaviors lacks a global perspective on how items relate across users. In this paper, we propose an LLM-driven adaptive and representative multi-interest modeling framework to address these challenges. At the user-individual level, we exploit LLM analysis and alleviate the agnostic granularity by adaptively aggregating semantic clusters to collaborative multi-interests. At the user-crowd level, to mitigate the limited insights in individual behaviors, we formulate a max covering problem to expand the scope of LLM analysis with compactness and representativeness, disentangling interest representations from global perspectives. Experiments on real-world datasets show that our approach outperforms various baselines.

Shu Yang, Jingyu Hu, Tong Li, Hanqi Yan, Wenxuan Wang, Di Wang

We introduce AutoMonitor-Bench, the first benchmark designed to systematically evaluate the reliability of LLM-based misbehavior monitors across diverse tasks and failure modes. AutoMonitor-Bench consists of 3,010 carefully annotated test samples spanning question answering, code generation, and reasoning, with paired misbehavior and benign instances. We evaluate monitors using two complementary metrics: Miss Rate (MR) and False Alarm Rate (FAR), capturing failures to detect misbehavior and oversensitivity to benign behavior respectively. Evaluating 12 proprietary and 10 open-source LLMs, we observe substantial variability in monitoring performance and a consistent trade-off between MR and FAR, revealing an inherent safety–utility tension. To further explore the limits of monitor reliability, we construct a large-scale training corpus of 153,581 samples and fine-tune Qwen3-4B-Instruction, to investigate whether training on known, relatively easy-to-construct misbehavior datasets improves monitoring performance on unseen and more implicit misbehaviors. Our results highlight the challenges of reliable, scalable misbehavior monitoring and motivate future work on task-aware designing and training strategies for LLM-based monitors.

Wenbin Hu, Huihao Jing, Haochen Shi, Changxuan Fan, Haoran Li, Yangqiu Song

Ensuring the safety and compliance of large language models (LLMs) is of paramount importance. However, existing LLM safety datasets often rely on ad-hoc taxonomies for data generation and suffer from a significant shortage of rule-grounded, real-world cases that are essential for robustly protecting LLMs. In this work, we address this critical gap by constructing a comprehensive safety dataset from a compliance perspective. Using a powerful web-searching agent, we collect a rule-grounded, real-world case dataset OmniCompliance-100K, sourced from multi-domain authoritative references. The dataset spans 74 regulations and policies across a wide range of domains, including security and privacy regulations, content safety and user data privacy policies from leading AI companies and social media platforms, financial security requirements, medical device risk management standards, educational integrity guidelines, and protections of fundamental human rights. In total, our dataset contains 12,985 distinct rules and 106,009 associated real-world compliance cases. Our analysis confirms a strong alignment between the rules and their corresponding cases. We further conduct extensive benchmarking experiments to evaluate the safety and compliance capabilities of advanced LLMs across different model scales. Our experiments reveal several interesting findings that have great potential to offer valuable insights for future LLM safety research.

Yifu Chen, Shengpeng Ji, Qian Chen, Tianle Liang, Yangzhuo Li, Ziqing Wang, Wen Wang, Jingyu Lu, Haoxiao Wang, Xueyi Pu 等

End-to-end spoken dialogue models have garnered significant attention because they offer a higher potential ceiling in expressiveness and perceptual ability than cascaded systems. However, the intelligence and expressiveness of current open-source spoken dialogue models often remain below expectations. Motivated by the success of online reinforcement learning(RL) in other domains, one might attempt to directly apply preference optimization to spoken dialogue models, yet this transfer is non-trivial. We analyze these obstacles from the perspectives of reward modeling and rollout sampling, focusing on how sparse preference supervision interacts with dense speech generation under shared-parameter updates. Based on the analysis, we propose a modality-aware adaptive post-training recipe that makes RL practical for spoken dialogue: it constrains preference updates to the semantic channel and improves acoustic behavior via explicit anchoring, while dynamically regulating their mixture from rollout statistics to avoid unreliable preference gradients. We evaluate the method across multiple spoken dialogue benchmarks and representative architectures, and observe consistent improvements in semantic quality and speech expressiveness.

Jinxiang Xie, Zihao Li, Wei He, Rui Ding, Shi Han, Dongmei Zhang

Text analysis of tabular data relies on two core operations: summarization for corpus-level theme extraction and tagging for row-level labeling. A critical limitation of employing large language models (LLMs) for these tasks is their inability to meet the high standards of output stability demanded by data analytics. To address this challenge, we introduce CAST (Consistency via Algorithmic Prompting and Stable Thinking), a framework that enhances output stability by constraining the model’s latent reasoning trajectory. CAST combines (i) Algorithmic Prompting to impose a procedural scaffold over valid reasoning transitions and (ii) Thinking-before-Speaking to enforce explicit intermediate commitments before final generation. To measure progress, we introduce CAST-S and CAST-T, stability metrics for bulleted summarization and tagging, and validate their alignment with human judgments. Experiments across publicly available benchmarks on multiple LLM backbones show that CAST consistently achieves the best stability among all baselines, improving Stability Score by up to 16.2%, while maintaining or improving output quality.

Changhao Pan, Rui Yang, Han Wang, Zhuan Zhou, Xuming He, Wenxiang Guo, Ziyue Jiang, Ruiqi Li, Yu Zhang, Chenyuhao Wen 等

Recent advances in speech generation have enabled high-fidelity synthesis, yet systematic evaluation of models under long-context conditions remains largely underexplored. A comprehensive evaluation benchmark for long-form speech is indispensable for two reasons: 1) existing test scenarios are often confined to limited domains, creating a significant gap with the diverse downstream applications; 2) existing metrics overlook critical long-text factors such as consistency and coherence, failing to generalize reliably. To this end, we propose LFSBench, a comprehensive benchmark that decomposes “long-form speech quality” into specific, disentangled dimensions. LFSBench has three key properties. 1) Rich speech scenarios: Focusing on long-form speech generation and multi-speaker dialog generation, LFSBench covers acoustics, semantics, and expressiveness challenges, and consists of 1,101 samples spanning 17 common speech scenarios; 2) Comprehensive evaluation dimensions: Along the acoustics, semantics, and expressiveness axes, LFSBench defines an automated evaluation protocol with seven metrics to provide a comprehensive, accurate, and standardized assessment; 3) Valuable Insights: Through extensive experiments, we reveal that current models still struggle in highly expressive scenarios and exhibit a notable gap in consistency and hierarchy compared to real recordings.

Lei Yin, Wentao Cheng, Zhida Qin, Tianyu Huang, Yidong Li, Gangyi Ding

Automatically generating 3D games in commercial game engines remains a non-trivial challenge, as it involves complex engine-related workflows for generating assets such as scenes, blueprints, and code. To address this challenge, we propose a novel multi-agent system, AutoUE, which coordinates multiple agents to end-to-end generate 3D games, covering model retrieval, scene generation, gameplay and interaction code synthesis, and automated game testing for evaluation. In order to mitigate tool-use hallucinations in LLMs, we introduce a retrieval-augmented generation mechanism that grounds agents with relevant UE tool documentation. Additionally, we incorporate game design patterns and engine constraints into the code generation process to ensure the generation of correct and robust code. Furthermore, we design an automated play-testing pipeline that generates and executes runtime test commands, enabling systematic evaluation of dynamic behaviors. Finally, we construct a game generation dataset and conduct a series of experiments that demonstrate AutoUE’s ability to generate 3D games end-to-end, and validate the effectiveness of these designs.

Giang Do, Hung Le, Truyen Tran

In the era of Large Language Models (LLMs), the Mixture of Experts (MoE) architecture has emerged as an effective approach for training extremely large models with improved computational efficiency. This success builds upon extensive prior research aimed at enhancing expert specialization in MoE-based LLMs. However, the nature of such specializations and how they can be systematically interpreted remain open research challenges. In this work, we investigate this gap by posing a fundamental question: *Do domain-specific experts exist in MoE-based LLMs?* To answer the question, we evaluate ten advanced MoE-based LLMs ranging from 3.8B to 120B parameters and provide empirical evidence for the existence of domain-specific experts. Building on this finding, we propose **Domain Steering Mixture of Experts (DSMoE)**, a training-free framework that introduces zero additional inference cost and outperforms both well-trained MoE-based LLMs and strong baselines, including Supervised Fine-Tuning (SFT). Experiments on four advanced open-source MoE-based LLMs across both target and non-target domains demonstrate that our method achieves strong performance and robust generalization without increasing inference cost or requiring additional retraining.

Junjie Ye, Changhao Jiang, Zhengyin Du, Yufei Xu, Xuesong Yao, Zhiheng Xi, Xiaoran Fan, Qi Zhang, Tao Gui, Xuanjing Huang 等

Effective tool use is essential for large language models (LLMs) to interact with their environment. However, progress is limited by the lack of efficient reinforcement learning (RL) frameworks specifically designed for tool use, due to challenges in constructing stable training environments and designing verifiable reward mechanisms. To address this, we propose an automated environment construction pipeline, incorporating scenario decomposition, document generation, function integration, complexity scaling, and localized deployment. This enables the creation of high-quality training environments that provide detailed and measurable feedback without relying on external tools. Additionally, we introduce a verifiable reward mechanism that evaluates both the precision of tool use and the completeness of task execution. When combined with trajectory data collected from the constructed environments, this mechanism integrates seamlessly with standard RL algorithms to facilitate feedback-driven model training. Experiments on LLMs of varying scales demonstrate that our approach significantly enhances the models’ tool-use performance without degrading their general capabilities. Our analysis suggests that these gains result from improved context understanding and reasoning, driven by updates to the lower-layer MLP parameters in models. Code and data are available at https://github.com/bytedance/FTRL.

Bo Gao, Chang Liu, Yuyang Miao, SiYuan Ma, Ser-Nam Lim

Recent advancements in Large Generative Models (LGMs) have revolutionized multi-modal generation. However, generating illustrated storybooks remains an open challenge, where prior works mainly decompose this task into separate stages, and thus, holistic multi-modal grounding remains limited. Besides, while safety alignment is studied for text- or image-only generation, existing works rarely integrate child-specific safety constraints into narrative planning and sequence-level multi-modal verification. To address these limitations, we propose BookAgent, a safety-aware multi-agent collaboration framework designed for high-quality, safety-aware visual narratives. Different from prior story visualization models that assume a fixed storyline sequence, BookAgent targets end-to-end storybook synthesis from a user draft by jointly planning, scripting, illustrating, and globally repairing inconsistencies. To ensure precise multi-modal grounding, BookAgent dynamically calibrates page-level alignment between textual scripts and visual layouts. Furthermore, BookAgent calibrates holistic consistency from the temporal dimension, by verifying-then-rectifying global inconsistencies in character identity and storytelling logic. Extensive experiments demonstrate that BookAgent significantly outperforms current methods in narrative coherence, visual consistency, and safety compliance, offering a robust paradigm for reliable agents in complex multi-modal creation. The implementation will be publicly released at https://github.com/bogao-code/BookAgent/tree/main.

Han Bao, Penghao Zhang, Yue Huang, Zhengqing Yuan, Yanchi Ru, SU Rui, Yujun Zhou, Xiangqi Wang, Kehan Guo, Nitesh V Chawla 等

Large Language Models (LLMs) are increasingly integrated into real-world decision-making, including in the domain of public policy. Yet, their ability to comprehend and reason about policy-related content remains underexplored. To fill this gap, we present PolicyBench, the first large-scale bilingual benchmark evaluating policy comprehension, comprising 21K cases across a broad spectrum of policy areas, capturing the diversity and complexity of real-world governance. Following Bloom’s taxonomy, the benchmark assesses three core capabilities: (1) Memorization: factual recall of policy knowledge, (2) Understanding: conceptual and contextual reasoning, and (3) Application: problem-solving in real-life policy scenarios. Building on this benchmark, we further propose PolicyMoE, a domain-specialized Mixture-of-Experts (MoE) model with expert modules aligned to each cognitive level. The proposed models demonstrate stronger performance on application-oriented policy tasks than on memorization or conceptual understanding, and yields the highest accuracy on structured reasoning tasks. Our results reveal key limitations of current LLMs in policy understanding and suggest paths toward more reliable, policy-focused models

Huayi Lai, Shichao Song, Simin Niu, Hanyu Wang, Jiawei Yang, Zhouxing Wang, Zhiqiang Yin, Xun Liang

Role-playing agents(RPAs) are widely used to steer large language models(LLMs) toward role-consistent behavior, yet existing benchmarks mainly evaluate surface-level fidelity and offer limited insight into decision making under role–alignment value conflicts. To address this gap, we introduce RoleCDE, the first benchmark designed to evaluate RPAs under structured conflicts between role-specific values and alignment-oriented constraints. RoleCDE formulates role-aware decision making as cognitive dilemma scenarios, jointly evaluating role–scenario grounding, value conflict resolution, and decision tendencies. The benchmark is constructed at scale, covering approximately 8k diverse role profiles and scenarios and nearly 240k dilemma instances across three difficulty levels and eight role categories. Evaluation of several mainstream LLMs reveals a "Role Value Decoupling" phenomenon, where agents systematically default to alignment- and morality-consistent decisions rather than role-specific values when the two conflict, even under explicit role conditioning. This behavior is largely invariant to dilemma difficulty but varies substantially across role categories. We further show that RoleCDE-based fine-tuning effectively mitigates this decoupling by improving value trade-off reasoning, while preserving general role-playing fidelity and general reasoning performance. Code is available at: https://github.com/rabbitrose/RoleCDE.

Mengwei Wang, Simin Niu, Xun Liang, Yuefeng Ma, Sensen Zhang, Jiawei Yang, Shichao Song, Hanyu Wang, Huayi Lai

Treating random masking as a performance plug-in for large language models (LLMs) offers three advantages: low coupling to the task, the model, and training resources. However, the critical drawback is that its gains are highly stochastic. Motivated by this, we propose play-it-by-ear masking performance plug-in (PibE-MPP), which enables LLMs to adaptively select masking target combinations for each task, retaining these advantages and mitigating the drawback. Specifically, we pose two core questions—what are the masking targets and what is the masking strategy under 7 constraints obtained from these advantages and a drawback. For the first question, we select all attention heads in the last layer as masking targets by constructing a first-order Markov process with alternating hidden state and information fusion. The feasibility of this target is validated by random masking experiments. For the second question, we first construct a small yet interpretable candidate set by proposing a three-axis mapping and a mean-based criterion for fusion features of masking targets. We then propose an axis-variance minimization to select a compact masking-target combination, reducing sensitivity to outlier targets. Experiments on 6 LLMs (Qwen and LLaMA) and 24 datasets demonstrate PibE-MPP’s effectiveness and generality, gain stability, and domain performance, and verify the necessity of its final module, providing empirical evidence of its transferability across tasks and models. The code is available at https://github.com/wtctcop/PibE-MPP.

SungHo Kim, Juhyeong Park, Eda Atalay, SangKeun Lee

Korean is a morphologically rich language with a featural writing system in which each character is systematically composed of subcharacter units known as Jamo. These subcharacters not only determine the visual structure of Korean but also encode frequent and linguistically meaningful morphophonological processes. However, most current Korean language models (LMs) are based on subword tokenization schemes, which are not explicitly designed to capture the internal compositional structure of characters. To address this limitation, we propose \textbf{SCRIPT}, a model-agnostic module that injects subcharacter compositional knowledge into Korean PLMs.\textbf{SCRIPT} allows to enhance subword embeddings with structural granularity, without requiring architectural changes or additional pre-training.As a result, \textbf{SCRIPT} consistently enhances all baselines across various Korean natural language understanding (NLU) and generation (NLG) tasks. Moreover, beyond performance gains, detailed linguistic analyses show that \textbf{SCRIPT} reshapes the embedding space in a way that better captures grammatical regularities and semantically cohesive variations. Our code is available at [https://github.com/SungHo3268/SCRIPT](https://github.com/SungHo3268/SCRIPT).