论文检索

Wenxuan Xie, Xuhong Wang

The integration of extensive, dynamic knowledge into Large Language Models (LLMs) remains a significant challenge due to the inherent entanglement of factual data and reasoning patterns. Existing solutions, ranging from non-parametric Retrieval-Augmented Generation (RAG) to parametric knowledge editing, are often constrained in practice by finite context windows, retriever noise, or the risk of catastrophic forgetting. In this paper, we propose DRIFT, a novel dual-model architecture designed to explicitly decouple knowledge extraction from the reasoning process. Unlike static prompt compression, DRIFT employs a lightweight knowledge model to dynamically compress document chunks into implicit fact tokens conditioned on the query. These dense representations are projected into the reasoning model’s embedding space, replacing raw, redundant text while maintaining inference accuracy. Extensive experiments show that DRIFT significantly improves performance on long-context tasks, outperforming strong baselines among comparably sized models. Our approach provides a scalable and efficient paradigm for extending the effective context window and reasoning capabilities of LLMs. Our code and data will be made public upon publication.

Qingfei Zhao, Ruobing Wang, Dingling Xu, Daren Zha, Ma Bowen, Zhichun Wang, Shijie Jia, Limin Liu, Xin Wang

Large language models (LLMs) have notably progressed in multi-step and long-chain reasoning. However, extending their reasoning capabilities to encompass deep interactions with search remains a non-trivial challenge, as models often fail to identify optimal reasoning–search interaction trajectories, resulting in suboptimal responses. We propose R-Search, a novel reinforcement learning framework for Reasoning–Search integration, designed to enable LLMs to autonomously execute multi-step reasoning with deep search interaction, and learn optimal reasoning–search interaction trajectories via multi-reward signals, improving response quality in complex logic- and knowledge-intensive tasks. R-Search guides the LLM to dynamically decide when to search or reason, while globally integrating key evidence to enhance deep knowledge interaction between reasoning and search. During RL training, R-Search provides multi-type rewards to jointly optimize the reasoning–search trajectory. Experiments on seven datasets show that R-Search significantly outperforms mainstream RAG baselines.

Xiao Li, Zhuo Chen, Jun Xia, Hongxin Xiang, Chao Wang, Wenjie Du, Yang Wang

While Retrieval-Augmented Generation (RAG) has become a standard paradigm for mitigating hallucinations in Large Language Models (LLMs), its effectiveness in complex medical reasoning remains limited. Existing RAG methods suffer from two main challenges: First, **Semantic Drift**: without explicit domain constraints, LLM-driven query decomposition often deviates from the original clinical intent, introducing substantial noise that degrades retrieval relevance. Second, **Concatenation Fallacy**: retrieved evidence from different semantic aspects is aggregated in a naive, unstructured manner, without modeling their inter-dependencies and potential conflicts, which ultimately undermines downstream reasoning. To address these challenges, we propose **Med-SRAF**, a multi-agent retrieval augmentation framework guided by medical domain knowledge. This framework reconstructs the traditional RAG process through two core mechanisms: (1) Intent-driven Semantic Routing, where a UMLS-based NavigationAgent dynamically maps queries to medical dimensions for strategic search space pruning; and (2) Evidence-based Agentic Fusion, where a FusionAgent resolves conflicts among dimension-specific evidence to build logically consistent reasoning chains. Extensive experiments on five widely used medical benchmarks show that Med-SRAF consistently outperforms existing general RAG baselines, achieving an average accuracy improvement of over **4.9%**, highlighting its effectiveness in robust and interpretable medical reasoning. Our code is at https://anonymous.4open.science/r/MultiAgent_RAG-F6DC.

Cui Encheng, Shaowen Peng, Kazuhiro Ito, XU Jinsha, Hisada Shohei, Shoko Wakamiya, Eiji Aramaki

Multi-Agent Systems (MAS) are commonly used to improve reasoning diversity and robustness by simulating interactions among agents with distinct roles. However, prior work often entangles the contribution of the multi-agent architecture with that of prompt conditioning, making the source of observed diversity gains unclear. We address this confound with a controlled study on divergent thinking tasks, using identical prompt conditioning for MAS and single agent baseline. Under these matched conditions, single agent setups consistently outperform multi-agent systems in semantic diversity. We attribute this gap to information visibility: parallel agents often converge on overlapping ideas, whereas a single agent model can condition on its own generation to avoid redundancy. We further find that a Multi-Output strategy, which prompts a single agent to produce multiple responses within a single inference pass, achieves the highest diversity without degrading logical validity. Together, these results point to a more efficient and effective way to expand diversity, with implications for the design of more efficient agentic frameworks.

Tingxu Han, Wei Song, Ziqi Ding, Ziming Li, Chunrong Fang, Yuekang Li, Dongfang Liu, Zhenyu Chen, Zhenting Wang

Large language models (LLMs) increasingly mediate decisions in domains where unfair treatment of demographic groups is unacceptable. Existing work probes when biased outputs appear, but gives little insight into the mechanisms that generate them, leaving existing mitigations largely fragile. In this paper, we conduct a systematic investigation of LLM unfairness and propose DiffHeads—a lightweight debiasing framework for LLMs. We first compare Direct-Answer (DA) prompting to Chain-of-Thought (CoT) prompting across eight representative open- and closed-source LLMs. DA will trigger the nature-bias component of the LLM and reduce measured unfairness by 391.9%- 534.5% in both one- and two-turn dialogues. Next, we define a token-to-head contribution score that traces each token’s influence back to individual attention heads. This reveals a small cluster of bias heads that activate under DA but stay largely dormant with CoT, providing the first causal link between prompting strategy and bias emergence. Finally, building on this insight, we propose DiffHeads, which identify bias heads through differential activation analysis between DA and CoT and selectively mask only those heads. DiffHeads reduces unfairness by 49.4%, and 40.3% under DA and CoT, respectively, without harming model utility.

Wenhao Yu, Zhicong Lu, Bo Lv, Fangyin Ma, Kaiwen Wei, Shihao Yang, Nayu Liu

Knowledge editing aims to efficiently update factual information in Large Language Models (LLMs) without full retraining. However, existing methods still suffer from performance degradation in batch knowledge editing. We identify that semantic representation entanglement, such as overlapping concepts and shared syntactic patterns, accumulates interference in the representation space and reduces editing precision. To bridge this gap, in this paper, we propose Orthogonal Representation Editing (ORE), which performs edits in the hidden representation space of LLMs by constructing a general semantic subspace and enforcing orthogonal constraints on edit vectors, effectively decoupling semantic entanglement. Furthermore, we introduce a gated non-linear representation head to enable adaptive learning of editing locations and precise control over knowledge injection. Extensive experiments show that ORE outperforms existing methods and achieves superior performance in cross-lingual knowledge editing scenarios. We release our code at https://github.com/YVVH/ORE.

Rongjie Huang, Dongchao Yang, Wenxiang Guo, Huadai Liu, Xize Cheng, Zehan Wang, Zhou Zhao, Xixin Wu, Helen M. Meng

Flow-matching generative models have created significant milestones in text-to-audio generation, powered by scalable training with increased data, computational resources, and model size, while their scalable inference remains less explored. In this work, we propose MaskAudioFlow, a continuous flow-matching transformer with masked generative modeling designed for scaling text-to-audio inference-time prediction. Specifically, MaskAudioFlow 1) masks spans of audio frames in training and approximates the continuous velocity vector field with flow-matching objective, and 2) performs inference via masked prediction, where we mask out generation and re-predict them through iterative decoding. To reduce the gap between generation and human preferences, we fine-tune MaskAudioFlow using reward signals from text-audio correspondence and perceptual aesthetics. Experimental results demonstrate that MaskAudioFlow achieves state-of-the-art performance in text-to-audio generation, effectively scaling inference-time computation through iterative masked prediction. Moreover, the preference-tuned model demonstrates superior text-audio alignment faithfulness and enhanced perceptual aesthetics. Audio samples are available at https://MaskAudio.github.io

Kawin Mayilvaghanan, Siddhant Gupta, Ayush Kumar

Large Language Models (LLMs) are increasingly deployed in contact-center Quality Assurance (QA) to automate agent performance evaluation and coaching feedback. While LLMs offer unprecedented scalability and speed, their reliance on web-scale training data raises concerns regarding demographic and behavioral biases that may distort workforce assessment. We present a counterfactual fairness evaluation of LLM-based QA systems across 13 dimensions spanning three categories: Identity, Context, and Behavioral Style. Fairness is quantified using the Counterfactual Flip Rate (CFR), the frequency of binary judgment reversals, and the Mean Absolute Score Difference (MASD), the average shift in coaching or confidence scores across counterfactual pairs. Evaluating 18 LLMs on 3,000 real-world contact center transcripts, we find systematic disparities, with CFR ranging from 5.4% to 13.0% and consistent MASD shifts across confidence, positive, and improvement scores. Larger, more strongly aligned models show lower unfairness, though fairness does not track accuracy. Contextual priming of historical performance induces the most severe degradations (CFR up to 16.4%), while implicit linguistic identity cues remain a persistent bias source. Finally, we analyze the efficacy of fairness-aware prompting, finding that explicit instructions yield only modest improvements in evaluative consistency. Our findings underscore the need for standardized fairness auditing pipelines prior to deploying LLMs in high-stakes workforce evaluation.

Youyuan Lin, Yuan Li, Yahan Yu, Fei Cheng, Shin’ya Nishida, Chenhui Chu

Generative vision-language models (VLMs) can edit and synthesize images, yet their ability to adapt visual assets across markets remains under-evaluated.We study cross-market image transcreation via movie posters, where localization must preserve a movie’s identity while matching market-specific design preferences and multilingual typography.We introduce the Movie Poster Transcreation Benchmark (MPTc-Bench), a cross-market benchmark of 582 aligned poster examples spanning 34 target markets, and define two task variants: \textbf{Surface} (text-centric localization) and \textbf{Deep} (preference-level style adaptation).We propose a two-stage planner-editor pipeline in which an VLM planner specifies executable edits and an image editor renders them.We evaluate in a triplet setup (source, human target-market poster, model output) using information-preservation checks, LLM-as-a-judge ratings for aesthetics and target-market fit, and objective similarity signals.Across multiple planners and editors, experiments reveal substantial gaps between model outputs and human target-market posters, highlighting open challenges for market-aware generation.MPTc-Bench enables controlled, quantitative progress on cross-market image editing beyond understanding-centric benchmarks.

Zhipeng Liu, Yifan Zheng, Fanqi Kong, Ziming Zhao

Large Language Model (LLM) serving systems increasingly face strict time-to-first-token (TTFT) service-level objectives (SLOs), yet TTFT remains highly sensitive to router-side queueing effects. Prefill costs scale with prompt length, decode lengths are uncertain, and prefix locality creates strong performance skew across requests. Despite major advances in continuous batching and KV-cache management, today’s routers are often agnostic to request cost, which makes them vulnerable to head-of-line blocking and tail-latency amplification under mixed workloads. We propose QUARTZ, a quantile-aware routing and queueing layer for LLM serving that predicts conservative quantile-based request-cost proxies, rather than point estimates, using lightweight router-visible signals. QUARTZ uses these quantiles together with backlog-aware router signals to guide worker selection and admission decisions that better align with TTFT tail SLOs while preserving fairness. We implement QUARTZ as a router upgrade for SGLang and evaluate it on representative interactive and retrieval-augmented workloads. The results show reductions in TTFT tail latency and SLO violations across heterogeneous workloads.

Zee Hen Tang, Mi-Yen Yeh

Existing evaluations of large language models primarily rely on single-agent dilemmas or static binary-choice tasks, offering limited insight into how cooperation contexts influence LLM behavior. We introduce CoopValue, a multi-agent evaluation framework that assesses LLMs’ value preferences through cooperative scenarios. CoopValue includes 1,778 scenarios spanning all pairwise conflicts among the 10 Schwartz values and three cooperation types: reciprocal, coopetitive, and altruistic. We evaluate 24 LLMs across 8 model families and examine how their value preferences vary across different cooperative contexts, showing the importance of assessing LLM value preferences in interactive, context-sensitive settings to guide the selection and deployment of LLMs aligned with desired cooperative behavior.

Zhe Yang, Yi Huang, Yaqin Chen, Mengfei Guo, Xiaoting Wu, Junlan Feng

In the realm of domain-specific natural language understanding (NLU) tasks, acquiring high-quality labeled data is often arduous, thereby posing significant challenges for effective model training. Multi-task learning (MTL) addresses these limitations by jointly optimizing multiple tasks within a unified framework. In this paper, we introduce a novel sparse NLU multi-task learning framework that decomposes the language model into modular skill components and employs a dynamic, learnable skill-combination mechanism to adaptively handle diverse tasks. Extensive experiments on benchmark NLU datasets demonstrate that our proposed method surpasses conventional multi-task learning approaches in performance.

Henan Sun, Kaichi Yu, Yuyao Wang, Bowen Liu, Xunkai Li, Rong-Hua Li, Nuo Chen, Jia Li

Reasoning ability has become a central focus in the advancement of Large Reasoning Models (LRMs). Although notable progress has been achieved on several reasoning benchmarks such as MATH500 and LiveCodeBench, existing benchmarks for algorithmic reasoning remain limited, failing to answer a critical question: Do LRMs truly master algorithmic reasoning? To answer this question, we propose AlgBench, an expert-curated benchmark that evaluates LRMs under an algorithm-centric paradigm.AlgBench consists of over 3,000 original problems spanning 27 algorithms, constructed by ACM algorithmic experts and organized under a comprehensive taxonomy, including Euclidean-structured, non-Euclidean-structured, non-optimized, local-optimized, global-optimized, and heuristic-optimized categories. Empirical evaluations on leading LRMs (e.g., Gemini-3-Pro, DeepSeek-v3.2-Speciale and GPT-o3) reveal substantial performance heterogeneity: while models perform well on non-optimized tasks (up to 92%), accuracy drops sharply to around 49% on globally optimized algorithms such as dynamic programming. Further analysis uncovers strategic over-shifts, wherein models prematurely abandon correct algorithmic designs due to necessary low-entropy tokens. These findings expose fundamental limitations of problem-centric reinforcement learning and highlight the necessity of an algorithm-centric training paradigm for robust algorithmic reasoning.

Hanlin Wang, Chak Tou Leong, Jian Wang, Wenjie Li

Recent advancements in large language models (LLMs) have enabled agents to tackle complex embodied tasks through environmental interaction. However, these agents still make suboptimal decisions and perform ineffective actions, as they often overlook critical environmental feedback that differs from their internal beliefs. Through a formal probing analysis, we characterize this as \textit{belief inertia}, a phenomenon where agents stubbornly adhere to prior beliefs despite explicit observations. To address this, we advocate active belief intervention, moving from passive understanding to active management. We introduce the Estimate-Verify-Update (EVU) mechanism, which empowers agents to predict expected outcomes, verify them against observations through explicit reasoning, and actively update prior beliefs based on the verification evidence. EVU is designed as a unified intervention mechanism that generates textual belief states explicitly, and can be integrated into both prompting-based and training-based agent reasoning methods. Extensive experiments across three embodied benchmarks demonstrate that EVU consistently yields substantial gains in task success rates. Further analyses validate that our approach effectively mitigates belief inertia, advancing the development of more robust embodied agents. Our code is available at https://github.com/WangHanLinHenry/EVU.

Zijun Min, Bingshuai Liu, Ante Wang, Long Zhang, Anxiang Zeng, Haibo Zhang, Jinsong Su

Reinforcement Learning with Verifiable Rewards (RLVR) offers a promising framework for optimizing large language models in reasoning tasks. However, existing RLVR algorithms focus on different granularities, and each has complementary strengths and limitations. Group Relative Policy Optimization (GRPO) updates the policy with token-level importance ratios, which preserves fine-grained credit assignment but often suffers from high variance and instability. In contrast, Group Sequence Policy Optimization (GSPO) applies single sequence-level importance ratios across all tokens in a response that better matches sequence-level rewards, but sacrifices token-wise credit assignment. In this paper, we propose Dynamic Hybrid Policy Optimization (DHPO) to bridge GRPO and GSPO within a single clipped surrogate objective. DHPO combines token-level and sequence-level importance ratios using weighting mechanisms. We explore two variants of the mixing mechanism, including an averaged mixing and an entropy-guided mixing. To further stabilize training, we employ a branch-specific clipping strategy that constrains token-level and sequence-level ratios within separate trust regions before mixing, preventing outliers in either branch from dominating the update. Across seven challenging mathematical reasoning benchmarks, experiments on both dense and MoE models from the Qwen3 series show that DHPO consistently outperforms GRPO and GSPO. Our code is publicly available at https://github.com/XMUDeepLIT/DHPO.

Lingling Shi, Haoyu Jin, Ruiyu Fang, Shuangyong Song, Jinsong Su, Yongxiang Li, Xuelong Li

Large Language Models have shown strong performance in Machine Translation, yet they often suffer from paraphrasing errors, omissions, or hallucinations when the input contains translation-specific elements (e.g., URLs, slang, and idioms) that require strict preservation or controlled transformation, undermining the reliability of critical details.We propose CEMT, a Controllable Element-Oriented Machine Translation framework inspired by the analysis–strategy–generation paradigm in human translation. CEMT first employs an Element Detection Module to identify translation-specific elements, and then introduces a Translation Module that decomposes the translation process into linguistically grounded analysis, strategy formulation, and final generation, thereby guiding the reliable translation of these elements. We further introduce a CoT Judge model during training that provides step-wise supervision over the accuracy and consistency of the translation process.On the WMT23/24 Chinese–English benchmarks, CEMT improves performance over existing Machine Translation models while significantly reducing element-level constraint violations.

Hao Li, Yiqun Zhang, Zhaoyan Guo, Chenxu Wang, Shengji Tang, Qiaosheng Zhang, Yang Chen, Biqing Qi, Peng Ye, Lei Bai 等

Large language model (LLM) routing assigns each query to the most suitable model from an ensemble. We introduce LLMRouterBench, a large-scale benchmark and unified framework for LLM routing. It comprises over 400K instances from 21 datasets and 33 models. Moreover, it provides comprehensive metrics for both performance-oriented and performance-cost trade-off routing, and integrates 10 representative routing baselines. Using LLMRouterBench, we systematically re-evaluate the field. While confirming strong model complementarity—the central premise of LLM routing—we find that many routing methods exhibit similar performance under unified evaluation, and several recent approaches, including commercial routers, fail to reliably outperform a simple baseline. Meanwhile, a substantial gap remains to the Oracle, driven primarily by persistent model-recall failures. We further show that backbone embedding models have limited impact, that larger ensembles exhibit diminishing returns compared to careful model curation, and that the benchmark also enables latency-aware analysis. All code and data are available at https://github.com/ynulihao/LLMRouterBench.

Ru Peng, Tianyu Zhao, Xijun Gu, Zhiting Fan, Haokai Xu, Jinyang Zhang, Yawen Zeng, Yihong Zhuang, Kexin Yang, Junyang Lin 等

High-quality, diverse data are vital for large language models (LLMs) but remain scarce and costly. Data synthesis is a viable alternative and succeeds on closed tasks, yet the humanities and social sciences (HSS) are overlooked, and their open-ended nature makes synthesis challenging.Moving beyond prior capability-centric, fragmented attempts, we adopt a subject-centric paradigm, define the first HSS domain system covering 14 mainstream fields, and introduce HSS-Synth—the first data synthesis pipeline for HSS.HSS-Synth comprises: (1) constructing seed document from web corpora via multi-step filtering and text refinement evaluated by a judge; (2) specifying “requirements + persona” to backtranslate seed document into diverse yet faithful instructions with strict Q&A alignment check; and (3) breaking LLM response limits via teacher-forced Answering that fed seed documents during response to anchor semantics, reduce hallucinations, and preserve tone and integrity.HSS-Synth yields 237k high-quality, diverse instruction-tuning samples that outperform 14 leading baselines on 16 benchmarks. The fine-tuned Qwen3-8B-Base set new SOTA and approached official Qwen3-8B, improving both human preference and knowledge capability without performance seesaws. Extensive experiments demonstrate the HSS-Synth’s robustness and transferability.Our code is publicly available at https://github.com/pengr/HSS-Synth.

Renxuan Tan, Rongpeng Li, Zhifeng Zhao, Honggang Zhang

Transcending the single-preference paradigm, aligning LLMs with diverse human values is pivotal for robust deployment. Contemporary Multi-Objective Preference Alignment (MPA) approaches predominantly rely on static linear scalarization or rigid gradient projection to navigate these trade-offs. However, by enforcing strict conflict avoidance or simultaneous descent, these paradigms often prematurely converge to local stationary points. While mathematically stable, these points represent a conservative compromise where the model sacrifices potential global Pareto improvements to avoid transient local trade-offs. To break this deadlock, we propose Pareto-Lenient Consensus (PLC), a game-theoretic framework that reimagines alignment as a dynamic negotiation process. Unlike rigid approaches, PLC introduces consensus-driven lenient gradient rectification, which dynamically tolerates local degradation provided there is a sufficient dominant coalition surplus, thereby empowering the optimization trajectory to escape local suboptimal equilibrium and explore the distal Pareto-optimal frontier. Theoretical analysis validates PLC can facilitate stalemate escape and asymptotically converge to a Pareto consensus equilibrium. Moreover, extensive experiments show that PLC surpasses baselines in both fixed-preference alignment and global Pareto frontier quality. This work highlights the potential of negotiation-driven alignment as a promising avenue for MPA. Our codes are available at https://anonymous.4open.science/r/aaa-6BB8.

Jin Zhong, Jinglin Liang, Tongtong Yang, Zijian Xie, Shuangping Huang, Hanlin Gu

Adapting large language models (LLMs) to domain-specific tasks via fine-tuning is often infeasible: models are protected by intellectual property, while sensitive data cannot be shared due to privacy regulations. A promising paradigm, Offsite Tuning (OT), addresses this challenge by constructing an emulator of the original model. Data owners leverage the emulator to train an adapter on downstream data, which is then plugged back into the original model, enabling knowledge transfer without transmitting either the original model or the raw data. However, emulators constructed by existing OT-based methods often retain substantial inference capabilities, thereby exposing model capability privacy and posing risks of misuse. To address this, we propose Loss Landscape Elevation Offsite Tuning (LLEOT), a framework that secures data privacy as well as model parameter and capability privacy. At its core, Loss Landscape Elevation (LLE) enforces a fixed margin between the loss landscapes of the emulator and the original model. We theoretically demonstrate that LLE simultaneously (i) degrades emulator inference via perplexity amplification and (ii) preserves gradient alignment, ensuring consistent convergence for adapter training. Extensive experiments confirm that LLEOT achieves strong adaptation performance while effectively mitigating emulator misuse. Code is available at https://github.com/Z-eloto/LLEOT.