Large Language Model (LLM)-based multi-agent systems are increasingly used to simulate human interactions and solve collaborative tasks. A common practice is to assign agents with personas to encourage behavioral diversity. However, this raises a critical yet underexplored question: do personas introduce biases into multi-agent interactions? This paper presents a systematic investigation into persona-induced biases in multi-agent interactions, with a focus on social traits like trustworthiness (how an agent's opinion is received by others) and insistence (how strongly an agent advocates for its opinion). Through a series of controlled experiments in collaborative problem-solving and persuasion tasks, we reveal that (1) LLM-based agents exhibit biases in both trustworthiness and insistence, with personas from historically advantaged groups (e.g., men and White individuals) perceived as less trustworthy and demonstrating less insistence; and (2) agents exhibit significant in-group favoritism, showing a higher tendency to conform to others who share the same persona. These biases persist across various LLMs, group sizes, and numbers of interaction rounds, highlighting an urgent need for awareness and mitigation to ensure the fairness and reliability of multi-agent systems.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Dataset diversity plays a pivotal role for the successful training of many machine learning models, particularly in the supervised fine-tuning (SFT) stage of large language model (LLM) development. Despite increasing recognition of its importance, systematic analyses of dataset diversity still remain underexplored. To address this gap, this work presents a systematic taxonomy of existing diversity-control strategies, which primarily focus on the instruction component, operating at either macroscopic (entire instruction semantics) or mesoscopic levels (instruction units), and furthermore introduces a novel analysis of microscopic diversity within the response component, specifically analyzing the statistical distribution of tokens in SFT training samples. In the experimental evaluation, we construct fixed-size datasets (e.g., 10,000 samples each) from a corpus of 117,000 open-source SFT samples, incorporating six distinct diversity-control strategies spanning macro-, meso-, and microscopic levels applied to both instructions and responses. We then fine-tune LLMs on these datasets to assess the six diversity-control strategies. Results reveal that while macroscopic and mesoscopic strategies lead to higher performance with increasing diversity, the microscopic strategy in responses exhibits not only a stronger correlation between model performance and the degree of diversity, but also superior performance with maximum diversity across all strategies. These findings offer actionable insights for constructing high-performance SFT datasets.
Semi-Supervised Synthetic Data Generation with Fine-Grained Relevance Control for Short Video Search Relevance Modeling
PDF ↗Synthetic data is widely adopted in embedding models to ensure diversity in training data distributions across dimensions such as difficulty, length, and language. However, existing prompt-based synthesis methods struggle to capture domain-specific data distributions, particularly in data-scarce domains, and often overlook fine-grained relevance diversity. In this paper, we present a Chinese short video dataset with 4-level relevance annotations, filling a critical resource void. Further, we propose a semi-supervised synthetic data pipeline where two collaboratively trained models generate domain-adaptive short video data with controllable relevance labels. Our method enhances relevance-level diversity by synthesizing samples for underrepresented intermediate relevance labels, resulting in a more balanced and semantically rich training data set. Extensive offline experiments show that the embedding model trained on our synthesized data outperforms those using data generated based on prompting or vanilla supervised fine-tuning(SFT). Moreover, we demonstrate that incorporating more diverse fine-grained relevance levels in training data enhances the model's sensitivity to subtle semantic distinctions, highlighting the value of fine-grained relevance supervision in embedding learning. In the search enhanced recommendation pipeline of Douyin's dual-column scenario, through online A/B testing, the proposed model increased click-through rate(CTR) by 1.45%, raised the proportion of Strong Relevance Ratio (SRR) by 4.9%, and improved the Image User Penetration Rate (IUPR) by 0.1054%.
From Sampling to Cognition: Modeling Internal Cognitive Confidence in Language Models for Robust Uncertainty Calibration
PDF ↗Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of tasks, yet they generally lack self-awareness, often displaying overconfidence when confronted with questions beyond their knowledge boundaries. This limitation severely hinders their trustworthiness in high-stakes scenarios. Existing calibration methods typically rely on sampling accuracy, derived from multiple outputs, as a proxy for model confidence. However, this coarse-grained metric fails to capture the model’s internal cognitive states, such as confusion, hallucination, or persistent belief in false knowledge. To address this, we propose CogConf (Cognitive Confidence), a cognitively grounded uncertainty signal that extends sampling accuracy by incorporating the semantic diversity of incorrect answers and the model’s abstention behaviors. By shifting the focus from sampling-based to cognition-oriented uncertainty modeling, CogConf offers a more faithful reflection of the model's internal beliefs. Building on this signal, we introduce CogAlign, a simple yet effective alignment framework that explicitly aligns the model’s verbalized confidence with CogConf, thereby producing uncertainty estimates that better reflect the model’s internal cognition. Experimental results on six knowledge-intensive in-domain and out-of-domain QA datasets demonstrate that CogConf robustly characterizes the model's internal uncertainty. Building on this foundation, CogAlign guides the model's expression to significantly enhance the trustworthiness and utility of its uncertainty calibration without compromising its underlying QA capabilities, while also demonstrating strong cross-task generalization and output stability. Offering a new pathway toward building more trustworthy LLMs.
Self-training large language models (LLMs) with generated reasoning paths has emerged as a promising approach to improve performance on complex reasoning tasks. However, most existing methods rely on correctness-based supervision, treating samples that reach the correct answer as high-quality despite potentially flawed intermediate steps, leading to noisy training signals. In this work, we propose K-STaR (Knowledge-aware Self-Taught Reasoner), a self-training framework that verifies reasoning paths through knowledge elicitation and integration as a proxy, without requiring any external reward models or dense step-by-step annotations. K-STaR models reasoning as a structured composition of knowledge units and automatically assigns process rewards to intermediate steps via consistency and frequency analysis, ensuring that only knowledge-grounded reasoning paths are retained. Experiments on mathematical and commonsense reasoning tasks show that K-STaR consistently discovers higher-quality reasoning paths and achieves superior self-training performance compared to prior methods. Our results highlight the importance of moving beyond correctness-centric supervision toward knowledge-grounded self-improvement.
The high memory demands of the Key-Value (KV) Cache during the inference of Large Language Models (LLMs) severely restrict their deployment in resource-constrained platforms. Quantization can effectively alleviate the memory pressure caused by KV Cache. However, existing methods either rely on static one-size-fits-all precision allocation or fail to dynamically prioritize critical KV in long-context tasks, forcing memory-accuracy-throughput tradeoffs. In this work, we propose a novel mixed-precision quantization method for KV Cache named KVmix. KVmix leverages gradient-based importance analysis to evaluate how individual Key and Value projection matrices affect the model loss, enabling layer-specific bit-width allocation for mix-precision quantization. It dynamically prioritizes higher precision for important layers while aggressively quantizing less influential ones, achieving a tunable balance between accuracy and efficiency. KVmix introduces a dynamic long-context optimization strategy that adaptively keeps full-precision KV pairs for recent pivotal tokens and compresses older ones, achieving high-quality sequence generation with low memory usage. Additionally, KVmix provides efficient low-bit quantization and CUDA kernels to optimize computational overhead. On LLMs such as Llama and Mistral, KVmix achieves near-lossless inference performance with extremely low quantization configuration (Key 2.19bit Value 2.38bit), while delivering a remarkable 4.9× memory compression and a 5.3× speedup in inference throughput.
CliCARE: Grounding Large Language Models in Clinical Guidelines for Decision Support over Longitudinal Cancer Electronic Health Records
PDF ↗Large Language Models (LLMs) hold significant promise for improving clinical decision support and reducing physician burnout by synthesizing complex, longitudinal cancer Electronic Health Records (EHRs). However, their implementation in this critical field faces three primary challenges: the inability to effectively process the extensive length and fragmented nature of patient records for accurate temporal analysis; a heightened risk of clinical hallucination, as conventional grounding techniques such as Retrieval-Augmented Generation (RAG) do not adequately incorporate process-oriented clinical guidelines; and unreliable evaluation metrics that hinder the validation of AI systems in oncology. To address these issues, we propose CliCARE, a framework for Grounding Large Language Models in Clinical Guidelines for Decision Support over Longitudinal Cancer Electronic Health Records. The framework operates by transforming unstructured, longitudinal EHRs into patient-specific Temporal Knowledge Graphs (TKGs) to capture long-range dependencies, and then grounding the decision support process by aligning these real-world patient trajectories with a normative guideline knowledge graph. This approach provides oncologists with evidence-grounded decision support by generating a high-fidelity clinical summary and an actionable recommendation. We validated our framework using large-scale, longitudinal data from a private Chinese cancer dataset and the public English MIMIC-IV dataset. In these settings, CliCARE significantly outperforms baselines, including leading long-context LLMs and Knowledge Graph-enhanced RAG methods. The clinical validity of our results is supported by a robust evaluation protocol, which demonstrates a high correlation with assessments made by oncologists.
Res-Bench: Benchmarking the Robustness of Multimodal Large Language Models to Dynamic Resolution Input
PDF ↗Multimodal Large Language Models (MLLMs) increasingly support dynamic image resolutions. However, current evaluation paradigms primarily assess semantic performance, overlooking the critical question of resolution robustness - whether performance remains stable across varying input resolutions. To address this gap, we introduce Res-Bench, a comprehensive benchmark comprising 14,400 samples across 12 resolution levels and six core capability dimensions. We designed a novel evaluation framework that goes beyond traditional accuracy metrics to capture performance stability. This framework introduces multiple robustness metrics: Spearman's correlation for assessing resolution-performance trends, and Absolute/Relative Continuous Error (ACE/RCE) for measuring performance volatility. Using these metrics, we conducted a large-scale evaluation of leading MLLMs. Our analysis encompasses: (1) model-centric and task-centric robustness examination, (2) investigation of preprocessing strategies including padding and super-resolution, and (3) exploration of fine-tuning for stability enhancement.
Recent advances in Speech Large Language Models (Speech LLMs) have led to great progress in speech understanding tasks such as Automatic Speech Recognition (ASR) and Speech Emotion Recognition (SER). However, whether these models can achieve human-level auditory perception, particularly in terms of their ability to comprehend latent intentions and implicit emotions in real-world spoken language, remains underexplored. To this end, we introduce the Human-level Perception in Spoken Speech Understanding (HPSU), a new benchmark for fully evaluating the human-level perceptual and understanding capabilities of Speech LLMs. HPSU comprises over 20,000 expert-validated spoken language understanding samples in English and Chinese. It establishes a comprehensive evaluation framework by encompassing a spectrum of tasks, ranging from basic speaker attribute recognition to complex inference of latent intentions and implicit emotions. To address the issues of data scarcity and high cost of manual annotation in real-world scenarios, we developed a semi-automatic annotation process. This process fuses audio, textual, and visual information to enable precise speech understanding and labeling, thus enhancing both annotation efficiency and quality. We systematically evaluate various open-source and proprietary Speech LLMs. The results demonstrate that even top-performing models still fall considerably short of human capabilities in understanding genuine spoken interactions. Consequently, HPSU will be useful for guiding the development of Speech LLMs toward human-level perception and cognition.
Dynamic retrieval-augmented generation (RAG) allows large language models (LLMs) to fetch external knowledge on demand, offering greater adaptability than static RAG. A central challenge in this setting lies in determining the optimal timing for retrieval. Existing methods often trigger retrieval based on low token-level confidence, which may lead to delayed intervention after errors have already propagated. We introduce Entropy-Trend Constraint (ETC), a training-free method that determines optimal retrieval timing by modeling the dynamics of token-level uncertainty. Specifically, ETC utilizes first- and second-order differences of the entropy sequence to detect emerging uncertainty trends, enabling earlier and more precise retrieval. Experiments on six QA benchmarks with three LLM backbones demonstrate that ETC consistently outperforms strong baselines while reducing retrieval frequency. ETC is particularly effective in domain-specific scenarios, exhibiting robust generalization capabilities. Ablation studies and qualitative analyses further confirm that trend-aware uncertainty modeling yields more effective retrieval timing. The method is plug-and-play, model-agnostic, and readily integrable into existing decoding pipelines. Implementation code is included in the supplementary materials.
Language Drift in Multilingual Retrieval-Augmented Generation: Characterization and Decoding-Time Mitigation
PDF ↗Multilingual Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to perform knowledge-intensive tasks in multilingual settings by leveraging retrieved documents as external evidence. However, when the retrieved evidence differs in language from the user query and in-context exemplars, the model often exhibits language drift by generating responses in an unintended language. This phenomenon is especially pronounced during reasoning-intensive decoding, such as Chain-of-Thought (CoT) generation, where intermediate steps introduce further language instability. In this paper, we systematically study output language drift in multilingual RAG across multiple datasets, languages, and LLM backbones. Our controlled experiments reveal that the drift results not from comprehension failure but from decoder-level collapse, where dominant token distributions and high-frequency English patterns dominate the intended generation language. We further observe that English serves as a semantic attractor under cross-lingual conditions, emerging as both the strongest interference source and the most frequent fallback language. To mitigate this, we propose Soft Constrained Decoding (SCD), a lightweight, training-free decoding strategy that gently steers generation toward the target language by penalizing non-target-language tokens. SCD is model-agnostic and can be applied to any generation algorithm without modifying the architecture or requiring additional data. Experiments across three multilingual datasets and multiple typologically diverse languages show that SCD consistently improves language alignment and task performance, providing an effective and generalizable solution in multilingual RAG.
Existing neural vocoders have demonstrated promising performance by leveraging Mel-spectrum as an acoustic feature for conditional audio generation. Nonetheless, they remain constrained by an inherent ``performance-cost'' dilemma that significantly hinders the development of this field. This paper revisits this foundational task from a novel degradation perspective, where Mel-spectrum is regarded as a special signal degradation process from the target spectrum. Drawing inspiration from traditional sparse signal recovery problems, we propose DegVoC, a GAN-based neural vocoder with a two-step solution procedure. First, by exploiting degradation priors, we attempt to retrieve the initial spectral structure from Mel-domain representations as an initial solution via a simple linear transformation. Based on that, we introduce a deep prior solver that accounts for the heterogeneous distribution of sub-bands in the time-frequency domain. A convolution-style attention module with a large kernel size is specially devised for efficient inter-frame and inter-band contextual modeling. With 3.89 M parameters and substantially reduced inference complexity, DegVoC achieves state-of-the-art performance across objective and subjective evaluations, outperforming existing GAN-, DDPM- and flow-matching-based baselines.
Automatically solving table reasoning tasks remains challenging due to three main factors: (1) diverse and hierarchical table structures that hinder comprehension, (2) the heavy reliance on complex logical and numerical reasoning—which makes purely text-based methods prone to hallucinations—and (3) the necessity of multi-step processing to handle intricate tasks involving multiple and lengthy tables. To address these challenges, we introduce TaREx, a novel framework that unifies table representation, integrates code-driven execution, and supports interactive multi-step reasoning. TaREx employs a reinforcement learning-based training pipeline to optimize its reasoning policy for complex tasks. Experimental results show that TaREx achieves state-of-the-art performance across a wide range of table reasoning benchmarks, both in-domain and out-of-domain. These include fundamental tasks such as table question answering (TQA) and table fact verification (TFV), as well as advanced tabular data analysis tasks. The results highlight TaREx’s effectiveness and scalability in advancing automated table reasoning.
Analogical reasoning is at the core of human cognition, serving as an important foundation for a variety of intellectual activities. While prior work has shown that LLMs can represent task patterns and surface-level concepts, it remains unclear whether these models can encode high-level relational concepts and apply them to novel situations through structured comparisons. In this work, we explore this fundamental aspect using proportional and story analogies, and identify three key findings. First, LLMs effectively encode the underlying relationships between analogous entities; both attributive and relational information propagate through mid-upper layers in correct cases, whereas reasoning failures reflect missing relational information within these layers. Second, unlike humans, LLMs often struggle not only when relational information is missing, but also when attempting to apply it to new entities. In such cases, strategically patching hidden representations at critical token positions can facilitate information transfer to a certain extent. Lastly, successful analogical reasoning in LLMs is marked by strong structural alignment between analogous situations, whereas failures often reflect degraded or misplaced alignment. Overall, our findings reveal that LLMs exhibit emerging but limited capabilities in encoding and applying high-level relational concepts, highlighting both parallels and gaps with human cognition.
Recent advances in Large Language Models (LLMs) - particularly model scaling and test-time techniques - have greatly enhanced the reasoning capabilities of language models at the expense of higher inference costs. To lower inference costs, prior works train router models or deferral mechanisms that allocate easy queries to a small, efficient model, while forwarding harder queries to larger, more expensive models. However, these trained router models often lack robustness under domain shifts and require expensive data synthesis techniques such as Monte Carlo rollouts to obtain sufficient ground-truth routing labels for training. In this work, we propose Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning (STEER), a domain-agnostic, framework that performs fine-grained, step-level routing between smaller and larger LLMs without utilizing external models. STEER leverages confidence scores from the smaller model’s logits prior to generating a reasoning step, so that the large model is invoked only when necessary. Extensive evaluations using different LLMs on a diverse set of challenging benchmarks across multiple domains such as Mathematical Reasoning, Multi-Hop QA, and Planning tasks indicate that STEER achieves competitive or enhanced accuracy while reducing inference costs (up to +20% accuracy with 48% less FLOPs compared to solely using the larger model on AIME), outperforming baselines that rely on trained external modules. Our results establish model-internal confidence as a robust, domain-agnostic signal for model routing, offering a scalable pathway for efficient LLM deployment.
In today’s world, emotional support is increasingly essential, yet it remains challenging for both those seeking help and those offering it. Multimodal approaches to emotional support show great promise by integrating diverse data sources to provide empathetic, contextually relevant responses, fostering more effective interactions. However, current methods have notable limitations, often relying solely on text or converting other data types into text, or providing emotion recognition only, thus overlooking the full potential of multimodal inputs. Moreover, many studies prioritize response generation without accurately identifying critical emotional support elements or ensuring the reliability of outputs. To overcome these issues, we introduce MULTIMOOD, a new framework that (i) leverages multimodal embeddings from video, audio, and text to predict emotional components and to produce responses responses aligned with professional therapeutic standards. To improve trustworthiness, we (ii) incorporate novel psychological criteria and apply Reinforcement Learning (RL) to optimize large language models (LLMs) for consistent adherence to these standards. We also (iii) analyze several advanced LLMs to assess their multimodal emotional support capabilities. Experimental results show that MultiMood achieves state-of-the-art on MESC and DFEW datasets while RL-driven trustworthiness improvements are validated through human and LLM evaluations, demonstrating its superior capability in applying a multimodal framework in this domain.
Large language models (LLMs) have made significant strides in mathematical reasoning, particularly at the elementary level. However, they continue to face substantial challenges when confronted with complex, advanced mathematical problems. In contrast to humans—who can effectively draw upon prior experiences in solving similar problems and retrieve relevant knowledge and theorems from memory—LLMs often struggle to accurately identify analogous problems and to recall or apply appropriate theorems. To overcome these limitations, we introduce a novel framework for constructing a template-theorems knowledge base, leveraging the capabilities of large language models. Inspired by the associative mechanisms of human cognition, our approach abstracts real-world problems into generalized templates and establishes intricate linkages between these templates and pertinent theorems. This design enables the efficient expansion of a comprehensive knowledge base, even when starting from a limited set of seed examples. Moreover, we develop an efficient retrieval strategy that, given a new problem, systematically extracts and presents the most relevant knowledge from the knowledge base as contextual input to the LLM. Extensive experiments on multiple public mathematical datasets and models demonstrate that our approach consistently surpasses conventional methods. Comprehensive ablation studies further corroborate the effectiveness of both our knowledge base construction and retrieval modules.
LC3: Long Cross-Language Code Clone Detection Enhanced by Opcode Sequences and Affinity Aggregation
PDF ↗Cross-language code clone detection, which identifies functionally similar code across programming languages, is critical for ensuring synchronized evolution and reducing maintenance costs in multi-platform software development. While zero-shot approaches have emerged as a practical solution to data scarcity, state-of-the-art methods still face two major limitations: an insufficiency in learning language-agnostic representations and information loss during the processing of long code. To address these challenges, we propose LC3, a novel framework for robust zero-shot cross-language code clone detection. To overcome the language-agnostic representation insufficiency, LC3 fuses source code with its underlying opcode sequences, leveraging a bimodal architecture and adversarial training to learn a language-agnostic representation. To resolve long-code information loss, LC3 introduces a semantic affinity aggregation strategy. This strategy synthesizes a robust clone score from a complete pairwise similarity matrix computed between segmented code blocks, overcoming the limitations of both simple truncation and aggregation. Extensive experiments show that LC3 significantly outperforms state-of-the-art zero-shot baselines, especially in challenging long-code scenarios.
Rationalization model has recently garnered significant attention for enhancing the interpretability of natural language processing by first using a generator to select the most relevant pieces from the text with respect to the label, before passing the text input to the predictor. However, the robustness of the rationalization models is not sufficiently investigated. Specifically, this paper explores the robustness of rationalization models against backdoor attacks, which has been ignored by previous studies. Surprisingly, we find that conventional backdoor attack techniques fail to inject triggers into the rationalization model because its generator can filter out bad triggers. Considering this, we further propose a novel backdoor attack method named as BadRNL designed specially for the rationalization models. The core idea of BadRNL is first to search for the personalized trigger for each specific dataset and then manipulate the rationales and labels to conduct attacks. Besides, BadRNL controls the order of sample learning through poison-priority sampling strategies. Experimental results show that our method can successfully craft the predictions of samples containing triggers while maintaining the performance of the model on clean data.
Rethinking the Sampling Criteria in Reinforcement Learning for LLM Reasoning: A Competence-Difficulty Alignment Perspective
PDF ↗The low sampling efficiency during the rollout phase poses a significant challenge to scaling reinforcement learning for large language model reasoning. Existing methods attempt to improve efficiency by scheduling problems based on problem difficulties. However, these approaches suffer from unstable and biased estimations of problem difficulty and fail to capture the alignment between model competence and problem difficulty in RL training, leading to suboptimal results. To address these challenges, we introduce Competence-Difficulty Alignment Sampling (CDAS). This approach allows for accurate and stable estimation of problem difficulties by aggregating historical performance discrepancies across problems. Subsequently, model competence is quantified to adaptively select problems whose difficulties align with the model's current competence using a fixed-point system. Extensive experiments in mathematical RL training show that CDAS consistently outperforms strong baselines, achieving the highest average accuracy of 45.89%. Furthermore, CDAS reduces the training step time overhead by 57.06% compared to the widely-used Dynamic Sampling strategy, verifying the efficiency of CDAS. Additional experiments on different tasks, model architectures, and model sizes demonstrate the generalization capability of CDAS.