论文检索

Reinforcement Learning · Multi-agent

Yuan Sui, Bryan Hooi

Training large language models (LLMs) for non-verifiable tasks—such as creative writing, dialogue, and ethical reasoning—remains challenging due to the absence of ground-truth labels. While LLM-as-Judge approaches offer a scalable alternative to human feedback, they face a fundamental limitation: performance is constrained by the evaluator's own quality. If the judge cannot recognize good solutions, it cannot provide useful training signals, and evaluation biases (e.g., favoring verbosity over quality) remain unaddressed. This motivates meta-evaluation—the ability to evaluate and improve the evaluator itself. We introduce CoNL, a framework that unifies generation, evaluation, and meta-evaluation through multi-agent self-play. Our key insight: critique quality can be measured by whether it helps others improve their solutions. In CoNL, multiple agents sharing the same policy engage in structured conversations to propose, critique, and revise solutions. Critiques that enable other agents' solution improvements earn a diagnostic reward, creating explicit supervision for meta-evaluation and enabling joint optimization of generation and judging capabilities through self-play, without external judges or ground truth. Experiments on five benchmarks show that CoNL achieves consistent improvements over self-rewarding baselines while maintaining stable training.

Applications · Language, Speech and Dialog

Chang Liu, boyu shi, Xu Yang, Xin Geng

Mixture-of-Experts (MoE) language models organize knowledge into explicitly routed expert modules, making expert-level representations traceable and analyzable. By analyzing expert activation patterns in MoE language large models (LLMs), we find that a subset of experts is consistently activated across diverse knowledge domains. These common experts encode cross-domain, generalizable knowledge that is closely related to model generalization, naturally raising the question of how such identifiable expert knowledge can be practically reused. Motivated by this observation, we propose XPERT, a training-free framework that extracts, consolidates, and reuses expert knowledge from pre-trained MoE LLMs to support effective training of language models across different model scales. XPERT identifies cross-domain experts via inference-only analysis, refines their representations through tensor decomposition, and adapts the extracted knowledge to be reused in downstream models. Experiments on language understanding and dialogue generation benchmarks show that models benefiting from reused expert knowledge achieve consistently stronger performance and faster convergence compared to strong baselines. These results highlight MoE LLMs as structured and reusable knowledge sources, and demonstrate the value of expert-level knowledge reuse for improving model training.

Deep Learning · Large Language Models

Yupei Yang, Lin Yang, Wanxi Deng, Lin Qu, Fan Feng, Biwei Huang, Shikui Tu, Lei Xu

A reliable reward model is essential for aligning large language models (LLMs) with human preferences through reinforcement learning from human feedback (RLHF). However, standard reward models are susceptible to spurious features that are not causally related to human labels. This can lead to *reward hacking*, where high predicted reward does not translate into better behavior. In this work, we address this problem from a causal perspective by proposing a factored representation learning framework that decomposes the model’s contextual embedding into (1) causal factors that are sufficient for reward prediction and (2) non-causal factors that capture reward-irrelevant attributes such as length or sycophantic bias. The reward head is then constrained to depend only on the causal component. In addition, we introduce an adversarial head trained to predict reward from the non-causal factors, while applying gradient reversal to discourage them from encoding reward-relevant information. Experiments on both mathematical and dialogue tasks demonstrate that our method learns more robust reward models and consistently improves downstream RLHF performance over state-of-the-art baselines. Analyses on length and sycophantic bias further validate the effectiveness of our method in mitigating reward hacking behaviors.

Deep Learning · Large Language Models

Luca Moschella, Laura Manduchi, Ozan Sener

The growing size of Large Language Models (LLMs) makes efficient inference challenging, primarily due to the memory demands of the autoregressive Key-Value (KV) cache. Existing eviction or compression methods reduce cost but rely on heuristics, such as recency or past attention scores, which serve only as indirect proxies for a token’s future utility and introduce computational overhead. We reframe KV cache eviction as a reinforcement learning (RL) problem: learning to rank tokens by their predicted usefulness for future decoding. To this end, we introduce KV Policy (KVP), a framework of lightweight per-head RL agents trained on pre-computed generation traces using only key and value vectors. Each agent learns a specialized eviction policy guided by a holistic reward, derived from future utility, that evaluates the quality of the ranking across all cache budgets, requiring no modifications to the underlying LLM or additional inference. Evaluated across two different model families on the long-context benchmark RULER and the multi-turn dialogue benchmark OASST2-4k, KVP significantly outperforms baselines. Furthermore, zero-shot tests on standard downstream tasks (e.g., LongBench, BOOLQ, ARC) indicate that KVP generalizes well beyond its training distribution and to longer sequence lengths. These results demonstrate that learning to predict future token utility is a powerful and scalable paradigm for adaptive KV cache management.

Applications · Language, Speech and Dialog

Jingwang Huang, Jie Zhang, Haoyang Zeng, Changzai Pan, Xianjie Wu, Guanting Dong, Jiaheng Liu, Wei Zhang, Mingyu Zheng, Chunxiao Liu 等

Recent advances in large language models (LLMs) have substantially expanded the scope of Table Question Answering (TableQA). However, existing benchmarks primarily treat TableQA as a passive, single-turn natural language understanding task, lacking the capacity to evaluate autonomous reasoning and tool-call trajectories in realistic, multi-turn scenarios. To bridge this gap, we introduce TableAgent-Bench, a large-scale bilingual benchmark that reformulates TableQA as proactive, agentic interactions over structurally complex, multi-table environments. With a topology-aware construction strategy, TableAgent-Bench captures dynamic intent evolution through 1,310 multi-turn dialogues grounded in 2,275 industrial tables. Furthermore, we propose the Table-centric Agent Evaluation Framework (TAEF) to assess agent interactions with complex table structures. Specifically, TAEF integrates a specialized agent toolset and 4 metric categories to systematically diagnose intermediate failure modes, assessing performance across table localization, tool-invocation rationality, and trajectory-level pass rate. Extensive experiments with 25 state-of-the-art LLM agents reveal a substantial capability gap, with even the strongest model Gemini-3-Pro-Preview achieving only 53.4% information coverage. We expect TableAgent-Bench to serve as a rigorous testbed for developing and evaluating agents capable of robust table-centric reasoning.

Applications · Computer Vision

Rongjiang Zhu, Wei Kang, Zeqi Liu, Chen junyu, Shuo Yang, Xinxiao Wu

Traditional 3D visual grounding typically assumes that natural language expressions unambiguously refer to target objects in a 3D scene. However, in practical applications, human instructions are often ambiguous or insufficient, which may lead existing models to associate the query with multiple possible objects, resulting in incorrect results. In this paper, we propose a new task, 3D visual grounding with referential ambiguity, which allows for referential ambiguity in language descriptions, making it more broadly applicable to real-world scenarios. To tackle this task, we propose an interactive grounding framework that performs multi-round question-answer interactions, in which the model actively generates clarifying questions and receives human-provided answers to acquire additional object attributes, spatial relationships, and other contextual information, thereby resolving referential ambiguity and achieving accurate grounding. To support the learning of interactive grounding, we construct a large-scale dataset named AmbiRefer3D, which contains 47,085 samples with 141,255 annotations of question-answer dialogues that capture interactive disambiguation processes, covering 7,316 indoor 3D scenes. Furthermore, we establish multi-round evaluation metrics to measure both disambiguation efficiency and grounding accuracy.

Applications · Language, Speech and Dialog

Siddhant Arora, Haidar Khan, Kai Sun, Xin Dong, Sajal Choudhary, Seungwhan Moon, Xinyuan Zhang, Adithya Sagar, Surya Appini, Kaushik Patnaik 等

End-to-end speech-in, speech-out dialogue systems are emerging as a powerful alternative to traditional ASR–LLM–TTS pipelines but remain prone to hallucinations due to limited factual grounding. While text-based dialogue models have effectively mitigated this issue through tools such as web search APIs, extending such capabilities to speech-in, speech-out systems remains underexplored. A key challenge is that tool integration increases latency, disrupting conversational flow. To mitigate this, we propose Streaming Retrieval-Augmented Generation (Stream RAG), a novel framework that reduces latency by predicting tool queries in parallel with user speech, even before the user finishes speaking. Specifically, we develop a post-training pipeline that teaches the model when to issue tool calls and how to generate spoken summaries using retrieved text results, thereby improving both accuracy and responsiveness. To evaluate our approach, we construct AudioCRAG, a benchmark created by converting queries from the publicly available CRAG dataset into speech form. Experimental results show that Stream RAG improves QA accuracy by over 20.0% absolute on AudioCRAG and achieves state-of-the-art performance, including outperforming cascaded systems, on the SLUE-SQA benchmark, while reducing latency by up to 57%. Stream RAG is modality-agnostic and can be applied equally to typed input, paving the way for more agentic, real-time AI assistants.

Applications · Computer Vision

Yue Jiang, Xue JIANG, Lihua Zhang, Zhiqiang Wang, Yuhang Lu, Peng Wang, Bo Han, Feng Zheng, Dingkang Yang

Multimodal Large Language Models (MLLMs) demonstrate remarkable visual understanding, yet their reliability in interactive settings is severely undermined by {hallucination snowballing}: a phenomenon where initial errors amplify across conversational turns, leading to a collapse in coherence. This failure reveals a fundamental vulnerability where models progressively neglect visual grounding in favor of over-relying on polluted textual history. Existing benchmarks are predominantly confined to single-turn VQA or simplistic dialogues, which fail to capture the complex dynamics of error propagation in realistic, long-horizon interactions. To address this, we introduce MM-Snowball, the first benchmark for fine-grained diagnosis of hallucination snowballing within 6-turn dialogues. Extensive evaluation shows that our benchmark poses a significant challenge even to advanced MLLMs and reveals the inefficacy of existing mitigation methods designed for single-turn tasks. To counteract this degradation, we propose Conflict-Aware Visual Rectification (CAVR). This training-free framework mitigates snowballing through a synergistic dual-mechanism that refreshes visual grounding at the representation level and rectifies output distributions at the logit level, effectively re-anchoring the model to visual facts. Experiments demonstrate that our proposed CAVR achieves state-of-the-art performance, offering a promising path toward more reliable interactive AI.

Applications · Computer Vision

Yang Bai, Tingfeng Wang, Bin Yang, Min Cao, Jinqiao Wang, Mang Ye

Traditional text-based person retrieval approaches typically rely on single-shot textual queries, which are generally incomplete or vague in real-world scenarios. Recently, chat-based person retrieval methods enable iterative query refinement via question-answering interactions between the system and users. However, these methods fall short of direct user interaction with retrieved candidates during conversation, making it challenging to effectively refine the retrieval results. To address these limitations, we propose multimodal interactive person retrieval (MInterPR), a new retrieval paradigm that allows users to iteratively refine retrieved candidates by providing feedback on visual differences from the target person. To support this task, we establish MInterPEDES, a multimodal conversational dataset constructed by augmenting existing question-answering dialogues with synthesized visual feedbacks. Furthermore, to tackle the challenge of accurate and efficient semantics understanding in multimodal dialogues, we propose a multimodal conversational memory-enhanced framework MNEMO, which incorporates an atomic turn encoding (ATE) module to model each dialogue turn independently, and a dialogue memory aggregation (DMA) module to capture the fine-grained information and cross-turn dependencies. Extensive experiments demonstrate that MNEMO achieves substantial improvements in both retrieval accuracy and generalization ability, highlighting its promising potential in real-world scenarios. The code and dataset will be released to facilitate future research.

Applications · Language, Speech and Dialog

Donghang Wu, Tianyu Zhang, Yuxin Li, Hexin Liu, Chen Chen, EngSiong Chng, Yoshua Bengio

During conversational interactions, humans subconsciously engage in concurrent thinking while listening to a speaker. Although this internal cognitive processing may not always manifest as explicit linguistic structures, it is instrumental in formulating high-quality responses. Inspired by this cognitive phenomenon, we propose a novel **F**ull-duplex **LA**tent and **I**nternal **R**easoning method named FLAIR that conducts *latent* thinking simultaneously with speech perception. Unlike conventional "thinking" mechanisms in NLP, which require post-hoc generation, our approach aligns seamlessly with spoken dialogue systems: during the user’s speaking phase, it recursively feeds the latent embedding output from the previous step into the next step, enabling continuous reasoning that strictly adheres to causality without introducing additional latency. To enable this latent reasoning, we design an Evidence Lower Bound-based objective that supports efficient supervised finetuning via teacher forcing, circumventing the need for explicit reasoning annotations. Experiments demonstrate the effectiveness of this think-while-listening design, which achieves competitive results on a range of speech benchmarks. Furthermore, FLAIR robustly handles conversational dynamics and attains competitive performance on full-duplex interaction metrics.

Deep Learning · Large Language Models

Haotong Sun, Yinghui Jiang, Bocheng Xu, Jianye Xie

Long-horizon dialogue agents suffer from *latent state drift*: what an agent says, what it internally represents, and what it stores in memory can diverge silently across turns. This creates *asymmetric rupture risk*—many locally coherent exchanges undone by a single high-cost contradiction. We propose **BRIDGE** (**B**ehavioral **R**easoning through **I**ntegrated **D**ynamic **G**ated **E**volution), which performs *triangular fixed-point refinement* to explicitly couple Observable ($\mathcal{O}$), Latent ($\mathcal{L}$), and Memory ($\mathcal{M}$) before decoding each response. We prove that under mild conditions, the refinement operator converges to a unique fixed point, providing a theoretical guarantee that the agent’s internal state remains self-consistent before each response. Empirically, BRIDGE achieves the highest scores on both PersonaGym (4.59 avg., surpassing Claude-3.7-Sonnet) and CoSER (59.5\% avg., +3.1 over Claude-3.7-Sonnet), with gains concentrated on persona-specific metrics (+8.0 Character Fidelity over Qwen2.5-32B-Instruct)—while updating only 0.85\% trainable parameters of the frozen backbone. We also provide a Lyapunov-style uniform drift bound for tiered memory updates, grounding bounded persona evolution in long-horizon interaction.

Applications · Language, Speech and Dialog

Doyoung Kim, Youngjun Lee, Joeun Kim, Jihwan Bang, Hwanjun Song, Susik Yoon, Jae-Gil Lee

Conversational query reformulation (CQR) has become indispensable for improving retrieval in dialogue-based applications. However, existing approaches typically rely on reference passages for optimization, which are **impractical** to acquire in real-world scenarios. To address this limitation, we introduce a novel **reference-free** preference optimization framework ***DualReform*** that generates **pseudo reference passages** from **commonly-encountered** conversational datasets containing only queries and responses. DualReform attains this goal through two key innovations: (1) **response-based inference**, where responses serve as proxies to infer pseudo reference passages, and (2) **response refinement via the dual-role of CQR**, where a CQR model refines responses based on the shared objectives between response refinement and CQR. Despite not relying on reference passages, ***DualReform*** achieves 96.9--99.1% of the retrieval accuracy attainable only with reference passages and surpasses the state-of-the-art method by up to 31.6%.

Deep Learning · Large Language Models

Minsoo Kim, Arnav Kundu, Han-Byul Kim, Richa Dixit, Minsik Cho

Modern large language models (LLMs) extend context lengths to millions of tokens, enabling coherent, personalized responses grounded in long conversational history. However, the Key-Value (KV) cache grows linearly with the extended dialogue history, causing the model’s memory footprint to quickly exceed device limits. While recent KV cache compression methods attempt to reduce memory usage, most apply cache eviction after processing the entire context, incurring unbounded peak memory usage. Additionally, query-dependent eviction narrows the cache semantics to a single query, leading to failure cases in multi-turn conversations. In this paper, we introduce EpiCache, a training-free KV cache management framework for long conversational question answering (LongConvQA) under fixed memory budgets. EpiCache bounds cache growth through block-wise prefill and preserves topic-relevant context via episodic KV compression, which clusters conversation history into coherent episodes and performs episode-specific KV cache eviction. Across three LongConvQA benchmarks (LongMemEval, Realtalk, and LoCoMo), EpiCache improves accuracy by up to 30\%, achieves near-full-cache accuracy under $4$–$6\times$ compression, and reduces latency and peak memory by up to $2.4\times$ and $3.7\times$, respectively.

Applications · Language, Speech and Dialog

Haohao Luo, Zexi Li, Yuexiang Xie, Wenhao Zhang, Yaliang Li, Ying Shen

Deep Research (DR) agents extend Large Language Models (LLMs) beyond parametric knowledge by autonomously retrieving and synthesizing evidence from large web corpora into long-form reports, enabling a long-horizon agentic paradigm. However, unlike real-time conversational assistants, DR is computationally expensive and time-consuming, creating an autonomy-interaction dilemma: high autonomy on ambiguous user queries often leads to prolonged execution with unsatisfactory outcomes. To address this, we propose IntentRL, a framework that trains proactive agents to clarify latent user intents before starting long-horizon research. To overcome the scarcity of open-ended research data, we introduce a scalable pipeline that expands a few seed samples into high-quality dialogue turns via a shallow-to-deep intent refinement graph. We further adopt a two-stage reinforcement learning (RL) strategy: Stage I applies RL on offline dialogues to efficiently learn general user-interaction behavior, while Stage II uses the trained agent and a user simulator for online rollouts to strengthen adaptation to diverse user feedback. Extensive experiments show that IntentRL significantly improves both intent hit rate and downstream task performance, outperforming the built-in clarify modules of closed-source DR agents and proactive LLM baselines.

Social Aspects · Safety

pengcheng li, Jie Zhang, Tianwei Zhang, Han Qiu, Zhang kejun, Weiming Zhang, Nenghai Yu, Wenbo Zhou

Safety alignment in large language models is typically evaluated under isolated queries, yet real-world use is inherently multi-turn. Although multi-turn jailbreaks are empirically effective, the structure of conversational safety failure remains insufficiently understood. In this work, we study safety failures from a state-space perspective and show that many multi-turn failures arise from structured contextual state evolution rather than isolated prompt vulnerabilities. We introduce STAR, a state-oriented diagnostic framework that treats dialogue history as a state transition operator and enables controlled analysis of safety behavior along interaction trajectories. Rather than optimizing attack strength, STAR provides a principled probe of how aligned models traverse the safety boundary under autoregressive conditioning. Across multiple frontier language models, we find that systems which appear robust under static evaluation can undergo rapid and reproducible safety collapse under structured multi-turn interaction. Mechanistic analysis reveals monotonic drift away from refusal-related representations and abrupt phase transitions induced by role-conditioned context. Together, these findings motivate viewing language model safety as a dynamic, state-dependent process defined over conversational trajectories.

Social Aspects · Safety

Yuyang Jiang, Chacha Chen, Teng Wu, Liwen Sun, Han Liu, Shi Feng, Chenhao Tan

*Debate*, where AI agents argue opposing positions, has emerged as a key approach to scalable oversight. However, debate faces a fundamental tension: models are incentivized to be persuasive to the judge, which may not always align with epistemic honesty. In this work, we propose an alternative paradigm: *disagreement resolution*, which reframes the interaction mechanism from adversarial debate to collaborative truth seeking. Drawing on principles from human mediation and conflict resolution, where mediators facilitate dialogue to help disputing parties reach consensus rather than adjudicating between them, we design an automated pipeline that adapts these strategies to AI oversight. Unlike standard debate where models argue for fixed positions, our pipeline directs models to collaboratively identify points of disagreement, examine the evidence for conflicting claims, and converge toward consensus or isolate the specific ``crux'' of their disagreement. We find that Disagreement Resolution consistently helps non-expert models identify the truth, achieving 62.1\% judging accuracy compared to 49.2\% for standard debate. Our results provide encouraging empirical evidence for rethinking the scalable oversight protocol from adversarial persuasion to collaborative truth-seeking.

Deep Learning · Attention Mechanisms

Siran Liu, Zheng Cao, Yongchao He

Efficient long-context understanding is increasingly vital for large language model (LLM) applications such as multi-turn dialogue and program analysis. However, the core self-attention scales quadratically with sequence length, creating a fundamental computational bottleneck. Existing sparse attention methods alleviate this issue but face trade-offs: training-based methods are costly and cannot be directly applied as acceleration plugins for other models, while inference-time methods often compromise efficiency or cross-modal generality. To address these limitations, we present UniSparse, a unified mechanism that introduces the notion of composite tokens—compact representations that aggregate multi-granularity contextual information. Building on this abstraction, UniSparse dynamically constructs sparse attention through multi-granularity compression and block-level selection, enabling efficient and hardware-friendly execution on GPU. Across multiple modalities and tasks ranging from synthetic benchmarks to real-world applications, UniSparse consistently surpasses state-of-the-art sparse attention methods (e.g., MInference, XAttention, FlexPrefill) in both accuracy and efficiency, achieving $\ge$ 99\% of full-attention accuracy and up to 2.61$\times$ faster attention computation than FlashAttention.

Deep Learning · Large Language Models

Wenhong Zhu, Ruobing Xie, Rui Wang, Pengfei Liu

Knowledge distillation (KD) is a powerful paradigm for compressing large language models (LLMs), whose effectiveness depends on intertwined choices of divergence direction, optimization strategy, and data regime. We break down the design of existing KD methods and present a unified view that establishes connections between them, reformulating KD as a reweighted log-likelihood objective at the token level. We further propose Hybrid Policy Distillation (HPD), which integrates the complementary advantages of forward and reverse KL to balance mode coverage and mode-seeking behaviors, and combines off-policy data with lightweight approximate on-policy sampling. We validate HPD on long-generation math reasoning as well as short-generation dialogue and code tasks, demonstrating improved optimization stability, computational efficiency, and final performance across diverse model families and scales.

Social Aspects · Safety

Yudong Yang, Xuezhen Zhang, Zhifeng Han, Siyin Wang, Jimin Zhuang, Zengrui Jin, Jing Shao, Guangzhi Sun, Chao Zhang

Recent progress in large language models (LLMs) has enabled understanding of both speech and non-speech audio, but has also exposed new safety risks arising from complex audio inputs that are inadequately handled by current safeguards. We introduce SACRED-Bench (Speech–Audio Composition for RED-teaming) to evaluate the robustness of LLMs under complex audio-based attacks. Unlike existing perturbation-based methods that rely on noise optimization or white-box access, SACRED-Bench exploits speech–audio composition to enable effective black-box attacks. SACRED-Bench adopts three composition mechanisms: (a) speech overlap, (b) multi-speaker dialogue, and (c) mixtures of speech and non-speech audio. These mechanisms focus on evaluating safety in settings where benign and harmful intents co-occur within a single auditory scene. Moreover, questions in SACRED-Bench are designed to implicitly refer to content in the audio, such that no explicit harmful information appears in the text prompt alone. Experiments demonstrate that even Gemini 2.5 Pro, a state-of-the-art proprietary LLM with safety guardrails fully enabled, still exhibits a 66% attack success rate. To bridge this gap, we propose SALMONN-Guard, the first guard model that jointly inspects speech, audio, and text for safety judgments, reducing the attack success rate to 20\%. Our results highlight the need for audio-aware defenses to ensure the safety of multimodal LLMs.

Applications · Language, Speech and Dialog

Yong Ren, Jingbei Li, Haiyang Sun, Yujie Chen, Cheng Yi, Yechang Huang, Hao Gu, Ye Bai, Xuerui Yang

Recent advances in Large Audio Language Models (LALMs) have extended Text-to-Speech (TTS) to interactive role-play scenarios, which demand high expressiveness and strict adherence to role-play instructions. However, existing models struggle to maintain stylistic consistency with character profiles and scene descriptions across multi-turn dialogues. A critical bottleneck is the lack of objective metrics for quantifying speaking style. To bridge this gap, we propose **Mean Continuation Log-Probability (MCLP)** as both an evaluation metric and a reward signal, validated on LALM-based Role-Play TTS (RP-TTS) tasks. Critically, we leverage the In-Context Learning capability of pre-trained LALMs to formulate MCLP via a continuation log-probability prediction. This metric quantifies stylistic consistency by measuring the likelihood of the ground-truth speech conditioned on the generated speech. Furthermore, we employ MCLP as a reinforcement learning reward to enhance the style alignment between generated speech and Role-Play instructions. To facilitate evaluation, we construct an RP-TTS dataset with rich scene and character annotations. Experimental results demonstrate that our method significantly outperforms strong LALM baselines on both objective and subjective metrics. Audio samples are available at https://mclp-tts.github.io/mclp/.