LLM role-playing seeks to portray arbitrary characters in interactive narratives, yet existing systems often lack immersion and adapt ability. They typically under-model dynamic environment information and assume a largely static scene/cast, offering limited support for multi-character orchestration, scene transitions, and on-the-fly character introduction. We propose an adaptive multi-agent interaction framework dubbed AdaMARP, which featuring an immersive message format that interleaves [Thought], (Action), Environment, and Speech, and an explicit Scene Manager that controls role-playing via discrete actions (init_scene, pick_speaker, switch_scene, add_role, end) with rationales. To train these abilities, we construct AdaRPSet for the Actor Model and AdaSMSet for supervising or chestration decisions, and introduce AdaptiveBench for trajectory-level evaluation. Experiments across multiple backbones and scales show consistent gains: AdaRPSet improves character consistency, environment grounding, and narrative coherence—an 8B actor outperforming several commercial LLMs, while AdaSMSet enables smoother scene transitions and more natural role introductions, surpassing Claude Sonnet 4.5 with only 14B LLMs.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
WoW-Bench: Evaluating Fine-Grained Acoustic Perception in Audio-Language Models via Marine Mammal Vocalizations
PDF ↗Large audio-language models (LALMs) extend language understanding into the auditory domain, yet their ability to perform low-level listening, such as pitch and duration detection, remains underexplored. However, low-level listening is critical for real-world, out-of-distribution tasks where models must reason about unfamiliar sounds based on fine-grained acoustic cues. To address this gap, we introduce the World-of-Whale benchmark (WoW-Bench) to evaluate low-level auditory perception and cognition using marine mammal vocalizations. We use marine mammal vocalizations as out-of-distribution sound events to better assess models’ low-level listening and so that the models do not rely on prior knowledge of the sound events. WoW-bench is composed of a Perception benchmark for categorizing novel sounds and a Cognition benchmark, inspired by Bloom’s taxonomy, to assess the abilities to remember, understand, apply, and analyze sound events. For the Cognition benchmark, we additionally introduce distractor questions to evaluate whether models are truly solving problems through listening rather than relying on other heuristics. Experiments with state-of-the-art LALMs show performance far below human levels, indicating a need for stronger auditory grounding in LALMs.
Large language models (LLMs) have achieved good performance in multiple reasoning tasks. However, they are limited to adapt the rapid knowledge updates in the real-world scenario without retraining the entire LLM or modifying the model weights. Excluding these consuming methods, knowledge graphs (KGs) are used as external memory under knowledge updating because of their structural knowledge and efficient updating ability, which is yet limited by the gap between structural KG and LLM, and the deficient entity-independent semantics. To this end, we propose an LLM reasoning framework with hierarchical relational retrieval for large-scale knowledge updating, named G-HiRel. To integrate the structural edited KG into continuous LLMs, G-HiRel generates hierarchical instructions based on natural language questions. In order to handle the knowledge inconsistency between the KG and LLM and obtain the entity independence, G-HiRel utilizes a designed hierarchical relational retrieval for relational path candidates, which are selected by a designed semantics-based strategy. Finally, top entity-independent relational paths are instantiated and integrated into LLMs to generate the answer, in order to verify the reasoning performance under knowledge edits. Extensive experiments of G-HiRel on three benchmarks show that G-HiRel achieves superiority in terms of accuracy and interpretability. The code of G-HiRel is available at the link: https://github.com/HJJ-designed/G-HiRel.
Existing information extraction (IE) tasks increasingly adopt in-context learning (ICL) with large language models. However, current approaches either show inconsistent performance across model scales or lack systematic optimization and generalizability. Building on this, we propose BCL-IE (Bayesian In-Context Learning Framework for Information Extraction), the first optimization framework that uses particle filtering with Bayesian updates to systematically refine label representations across IE tasks. Through four steps—initialization, observation, weight update, and resampling, BCL-IE generalizes to both sequence labeling and relation classification paradigms. Extensive experiments demonstrate substantial improvements over existing approaches (up to 30%), achieving prior performance while other methods either fail to generalize or show limited effectiveness.
Large language model post-training relies on reinforcement learning to improve model capability and alignment quality. However, the off-policy training paradigm introduces distribution shift, which often pushes the policy beyond the trust region, leading to training instabilities manifested as fluctuations in policy entropy and unstable gradients. Although PPO-Clip mitigates this issue through importance clipping, it still overlooks the global distributional shift of actions. To address these challenges, we propose using the entropy ratio between the current and previous policies as a new global metric that effectively quantifies the relative change in policy exploration throughout updates. Building on this metric, we introduce an Entropy Ratio Clipping (ERC) mechanism that imposes bidirectional constraints on the entropy ratio. This stabilizes policy updates at the global distribution level and compensates for the inability of PPO-clip to regulate probability shifts of un-sampled actions. We integrate ERC into both DAPO and GPPO reinforcement learning algorithms. Experiments across multiple benchmarks show that ERC consistently improves performance.
Large Language Models exhibit advanced reasoning capabilities that enable them to address complex tasks, but these capabilities also increase the risk of generating harmful content, particularly in multi-turn dialogues. Existing inference-phase safety alignment methods face three major challenges. First, they lack the relationship consideration between question and response, making the model easy to provide harmful content toward complex scenarios. Second, they are difficult to adapt to defense instruction. Third, these methods fail to effectively leverage historical information for safe response generation. To address these challenges, we propose CogGSE, an inference-time safety alignment framework that explicitly models the cognitive process of problem solving through a structured cognitive analysis graph. We retrieve a question-specific graph to ensure the safety information is tailored to the query. To fully exploit historical information in multi-turn settings, we retrieve relevant graphs from previous turns and selectively retain safety-related nodes, which are jointly used with the current-turn graph to guide safe response generation. This design enables transparent, controllable reasoning while maintaining strong safety guarantees. Extensive experiments demonstrate the effectiveness of our approach in multiple safety scenarios.
As the parameter size of language models continues to grow, effective model compression is required to reduce their computational and memory overhead. Existing compression methods suffer from bottleneck issues: when the compression ratio is increased, performance degrades significantly. Low-rank decomposition and quantization are two prominent compression methods that have been proven to significantly reduce the computational and memory requirements of Large Language Models (LLMs) while maintaining model accuracy. Evidently, combining these two methods will break through the existing compression bottleneck. However, how these two methods interact when combined remains a critical question for developers, as many assume they are orthogonal, meaning their combination would not introduce additional errors beyond those independently introduced by each method. This paper provides the first mathematical proof that low-rank decomposition and quantization are non-orthogonal. We validate these findings through a series of experiments on large language models. Our results demonstrate that these methods are non-orthogonal, and their combination leads to significant performance degradation. Importantly, we propose a novel approach Diagonal Adhesive Method (DAM), which can effectively combine the two methods and mitigate the performance loss. Our research provides deep insights into model compression and lays a solid theoretical and experimental foundation for future related studies.
Read the Room, Read the Image: Understanding Indirect Speech Acts in Multimodal Visual Contexts
PDF ↗Indirect speech acts (ISAs) require pragmatic reasoning over context, as directive intent cannot be inferred from surface form alone. Prior text-based studies and existing multimodal benchmarks largely overlook this requirement, focusing instead on explicitly encoded context or perceptual recognition, and thus underexplore context-dependent pragmatic understanding—particularly in high-context languages such as Korean. We introduce READI, a multimodal benchmark for evaluating ISA understanding through integrated reasoning over visual context and dialogue. READI models graded indirectness grounded in pragmatic theory and formulates the task as vision-based pragmatic question answering (V-PQA), supporting cross-lingual evaluation in English and Korean. Experiments show that even state-of-the-art multimodal models struggle with visually grounded indirect speech acts, with performance declining as indirectness increases, underscoring the need for benchmarks that explicitly target contextual pragmatic reasoning.
Analyze Like a Venture Capitalist: Information-Gain and Knowledge Enhanced Graph Reasoning for Startup Success Prediction
PDF ↗Most venture capital (VC) investments fail, while a few deliver outsized returns. Predicting startup success requires synthesizing relational evidence across company fundamentals, investor track records, and investment networks through explicit reasoning, which traditional machine learning and graph neural networks lack. Large language models excel at reasoning, but applying them to VC prediction must address: selecting compact evidence subgraphs from large investment networks, one-sided label noise where failures may be latent successes, and grounding decisions in structured VC domain knowledge. We present MIRAGE-VC, an evidence-grounded reasoning framework with three innovations. First, an information-gain-driven retriever distills networks into compact evidence subgraphs. Second, a dual-layer knowledge base grounds reasoning in VC principles. Third, a noise-aware mechanism down-weights mislabeled negatives via improved Positive-Unlabeled (PU) estimation. MIRAGE-VC achieves +5.9% F1 and +22.1% Precision@5 over state-of-the-art baselines. Expert evaluation confirms professional-quality rationales. We further validate our approach on public data with consistent improvements. Code and reasoning results are available at: https://github.com/ZhangDataLab/MIRAGE-VC.git
As high-quality data becomes harder to obtain, reward models are increasingly important. Beyond the costly RLHF stage, they are now used at inference time to guide LLM generation and in data selection for post-training. These methods bring efficiency and performance gains, but current reward models often fail to prevent untrustworthy behaviors such as privacy leaks and stereotypes. Re-training reward models to address these issues is expensive, since it requires large-scale human preference data. We propose SelfRW, a lightweight intrinsic reward that needs no extra fine-tuning or auxiliary models. By pruning current LLMs to approximate an “trust” and an “untrust” token distribution, we compute the log-probability difference as an auxiliary reward. When integrated into reward-guided sampling, SelfRW significantly reduces untrustworthy outputs while preserving task performance. It also improves reward-guided data selection, yielding better post-trained models. Experiments with two reward models and four LLMs on privacy, bias, and stereotype benchmarks show that combining SelfRW consistently improves trustworthiness (over 10% in privacy tasks and 20% in bias tasks) with minimal impact on general utility benchmarks.
This work investigates the ability of large language models (LLMs) to generate mathematical equations from scientific texts. Prior work faces challenges in unstructured grounding, multi-equation dependency, and human-aligned evaluation. To address this, we construct a dataset of AI research papers, pairing contextual passages with ground-truth equations and variable descriptions. We develop an explainable equation generation workflow and evaluate it across diverse open- and closed-source LLMs. Our evaluation protocol combines automatic metrics, LLM-based rubrics, and human judgments to assess accuracy, explainability, and human-LLM alignment. Results show that LLMs achieve moderate performance on lexical and syntactic similarity, but struggle with semantic accuracy. LLM-based evaluations show limited alignment with human judgments, highlighting challenges in assessing equation quality. These findings provide insights for improving equation generation models and developing more reliable evaluation methods for scientific creativity. We provide code and data for reproducibility.
We are entering an era in which individuals and organizations increasingly deploy dedicated AI agents that interact and collaborate with other agents.However, the dynamics of multi-agent collaboration under privacy constraints remain poorly understood.In this work, we present PAC\text{-}Bench, a benchmark for systematic evaluation of multi-agent collaboration under privacy constraints.Experiments on PAC\text{-}Bench show that privacy constraints substantially degrade collaboration performance and make outcomes depend more on the initiating agent than the partner.Further analysis reveals that this degradation is driven by recurring coordination breakdowns, including early-stage privacy violations, overly conservative abstraction, and privacy-induced hallucinations.Together, our findings identify privacy-aware multi-agent collaboration as a distinct and unresolved challenge that requires new coordination mechanisms beyond existing agent capabilities.
Synergizing Semantic Anchors and Ordinal Smoothed Cross-Entropy for Speech Fluency Classification
PDF ↗Speech fluency is a core indicator of second language proficiency and a critical component of Computer-Assisted Pronunciation Training (CAPT) systems. Accurate assessment requires models to perceive both macroscopic speech flow trends and microscopic local anomalies. However, existing methods struggle to bridge the semantic gap between static expert priors and dynamic temporal representations, while often overlooking the inherent ordinal nature of fluency scores. To address these challenges, we first construct a set of expert features targeting fluency disruptions and rhythmic regularity to provide explicit linguistic priors. Building on this, we propose the Multimodal Multi-Stream Fusion Classification (MMSFC) network. It employs a Mutual Cross-Attention (MCA) mechanism that leverages these expert features as “semantic anchors” to actively guide Whisper’s temporal representations and integrate decoder contexts, achieving deep interaction between global priors and local dynamics. Furthermore, we propose the Ordinal Smoothed Cross-Entropy (OSCE) loss. By constructing distance-aware soft target distributions coupled with confidence-adaptive smoothing and boundary enhancement, OSCE explicitly models ordinal relationships to resolve boundary ambiguity. Experiments on SpeechOcean762 show MMSFC achieves 83.40% accuracy, significantly outperforming strong baselines. Notably, OSCE also demonstrates superior generalization potential in cross-domain CV and NLP tasks. Our code is available at https://github.com/speech26ai/MMSFCCode.
Persuasive dialogue generation plays a vital role in decision-making, negotiation, counseling, and behavior change, yet it remains a challenging problem. In complex persuasion where the persuadee’s internal states are not expressed clearly, the persuader must interpret responses, infer the persuadee’s latent mental states (e.g., beliefs and desires), and translate them into targeted, strategy-consistent actions; however, current approaches often produce generic or weakly grounded responses even when such cues are identified. Moreover, although large language models (LLMs) can generate persuasive content, their performance varies substantially across domains due to uneven knowledge coverage and limited reasoning generalization. To address these challenges, we propose MA^{2}P, a meta-cognitive autonomous intelligent agent framework for complex persuasion. Specifically, we develop an autonomous multi-agent architecture that coordinates perception management, mental-state inference, strategy execution, memory maintenance, and performance evaluation. To mitigate cross-domain performance variation, we further design a meta-cognitive configurator that selects an appropriate meta-strategy from a structured knowledge base at the outset, thereby guiding subsequent reasoning and planning. Experimental results show that our approach achieves a higher persuasion success rate than baselines.
Large language models have recently advanced automated program repair, yet most existing approaches provide only post-hoc natural-language explanations that are neither executable nor verifiable. This limitation is especially critical for quantum programs, where correctness hinges on subtle semantic properties such as circuit equivalence and fidelity preservation. We propose Explainable Quantum Program Repair, a framework that couples repair generation with machine-checkable executable explanations. Given a buggy quantum circuit, a language model proposes candidate repairs together with structured transformation rationales, which are compiled into proof traces and validated using formal verification backends, including circuit equivalence checking, ZX-calculus reasoning, stabilizer analysis, and quantum simulation. Only repairs whose explanations are fully verified are accepted. Experiments on QASMBench with mutation-generated quantum program bugs demonstrate that our approach achieves competitive repair success while substantially improving semantic precision and explanation faithfulness over baselines that rely on unconstrained or purely natural-language explanations.
Large language models (LLMs) are widely used in decision-making across diverse domains. Ensuring the generation of safe and reliable responses is critical for the effective deployment of LLM-based applications, particularly in high-stakes domains such as healthcare and finance. Most of these applications typically use carefully crafted prompts to guide response generation; however, the relationship between prompts and the reliability of LLM-generated responses is not yet fully understood. To address this gap, we propose a novel prompt-response concept model that explains the relationship between the amount of task-relevant information (informativeness) provided in the prompt and the LLM-generated response uncertainty by identifying four sources of response uncertainty: prompt underspecification, model quality, task variability, and semantic redundancy. We prove that response uncertainty decreases as prompt informativeness or model quality increases, mirroring the behavior of epistemic uncertainty in probabilistic models. Our experimental results on real-world datasets further validate our proposed model and corroborate the theoretical results.
Machine Behavior in Relational Moral Dilemmas: Moral Rightness, Predicted Human Behavior, and Model Decisions
PDF ↗Human moral judgment is context-dependent and changes based on interpersonal relationships. As large language models (LLMs) increasingly serve as decision-support systems, it is critical to understand if they encode these social nuances. We characterize LLM behavior using the Whistleblower’s Dilemma, systematically varying two experimental factors: crime severity and relational closeness. Our study compares three evaluative perspectives: (1) moral rightness (general prescriptive norms), (2) predictive human behavior (how models expect people to navigate social situations), and (3) models’ own decision-making. By analyzing the reasoning processes, we find a clear cross-perspective divergence: moral rightness remains consistently fairness-oriented, while predicted human behavior shifts with relational context toward loyalty. Crucially, the model decisions mirror moral rightness judgments, rather than their behavioral predictions. This cross-perspective inconsistency suggests that LLM decision-making favors abstract rules over the social sensitivity found in their internal modeling, potentially producing conflicting expectations in real-world deployments.
The saturation of high-quality pre-training data has shifted research focus toward evolutionary systems capable of continuously generating novel artifacts, leading to the success of AlphaEvolve. However, the progress of such systems is hindered by the lack of rigorous, quantitative evaluation. To tackle this challenge, we introduce CreativeBench, a benchmark for evaluating machine creativity in code generation, grounded in a classical cognitive framework. Comprising two subsets – CreativeBench-Combo and CreativeBench-Explore – the benchmark targets combinatorial and exploratory creativity through an automated pipeline utilizing reverse engineering and self-play. By leveraging executable code, CreativeBench objectively distinguishes creativity from hallucination via a unified metric defined as the product of quality and novelty. Our analysis of state-of-the-art models reveals distinct behaviors: (1) scaling significantly improves combinatorial creativity but yields diminishing returns for exploration; (2) larger models exhibit “convergence-by-scaling,” becoming more correct but less divergent; and (3) reasoning capabilities primarily benefit constrained exploration rather than combination. Finally, we propose EvoRePE, a plug-and-play inference-time steering strategy that internalizes evolutionary search patterns to consistently enhance machine creativity.
Language Reasoning Models (LRMs) achieve strong performance by scaling test-time computation but often suffer from "overthinking", producing excessively long reasoning traces that increase latency and memory usage. Existing LRMs typically enforce conciseness with uniform length penalties, which over-compress crucial early deduction steps at the sequence level and indiscriminately penalize all queries at the group level. To solve these limitations, we propose PACE, a dual-level framework for prefix-protected and difficulty-aware compression under hierarchical supervision. At the sequence level, prefix-protected optimization employs decaying mixed rollouts to maintain valid reasoning paths while promoting conciseness. At the group level, difficulty-aware penalty dynamically scales length constraints based on query complexity, maintaining exploration for harder questions while curbing redundancy on easier ones. Extensive experiments on DeepSeek-R1-Distill-Qwen (1.5B/7B) demonstrate that PACE achieves a substantial reduction in token usage (up to 55.7%) while simultaneously improving accuracy (up to 4.1%) on math benchmarks, with generalization ability to code, science, and general domains.
Large language models (LLMs) have shown remarkable capabilities in text-to-SQL generation, yet existing approaches remain prone to hallucinations and lack verification mechanisms. Current methods such as Chain-of-Thought (CoT) and Program-of-Thought (PoT) typically rely on intermediate reasoning that is either purely textual or executed only as a final step, leaving the reasoning process opaque and prone to grounding and logical hallucinations. In this paper, we introduce Verifiable Execution Tracing (VET), a novel reasoning paradigm that transforms text-to-SQL from unverifiable textual rationales into step-wise executable semantics. VET addresses these limitations by constraining the reasoning process within a candidate schema space and formulating it as a sequence of executable Python steps. Crucially, each step is executed against the real database to produce observable intermediate results, which serve as immediate verification feedback and transform the traditionally opaque generation process into a transparent, debuggable interaction with database reality.Experiments show consistent gains under matched, training-free settings, achieving 70.93% execution accuracy on BIRD and 37.04% on Spider 2.0-lite, with particularly strong improvements on complex queries.