LLMs have shown strong performance on human-centric reasoning tasks. While previous evaluations have explored whether LLMs can infer intentions or detect deception, they often overlook the individualized reasoning styles that influence how people interpret and act in social contexts. Social deduction games (SDGs) provide a natural testbed for evaluating individualized reasoning styles, where different players may adopt diverse but contextually valid reasoning strategies under identical conditions. To address this, we introduce InMind, a cognitively grounded evaluation framework designed to assess whether LLMs can capture and apply personalized reasoning styles in SDGs. InMind enhances structured gameplay data with round-level strategy traces and post-game reflections, collected under both Observer and Participant modes. It supports four cognitively motivated tasks that jointly evaluate both static alignment and dynamic adaptation. As a case study, we apply InMind to the game Avalon, evaluating 11 state-of-the-art LLMs. General-purpose LLMs, even GPT-4o frequently rely on lexical cues, struggling to anchor reflections in temporal gameplay or adapt to evolving strategies. In contrast, reasoning-enhanced LLMs like DeepSeek-R1 exhibit early signs of style-sensitive reasoning. These findings reveal key limitations in current LLMs’ capacity for individualized, adaptive reasoning, and position InMind as a step toward cognitively aligned human–AI interaction.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Large Language Models (LLMs) excel at various natural language processing tasks but remain vulnerable to jailbreaking attacks that induce harmful content generation. In this paper, we reveal a critical safety inconsistency: LLMs can more effectively identify harmful requests as discriminators than defend against them as generators. This insight inspires us to explore aligning the model’s inherent discrimination and generation capabilities. To this end, we propose SDGO (Self-Discrimination-Guided Optimization), a reinforcement learning framework that leverages the model’s own discrimination capabilities as a reward signal to enhance generation safety through iterative self-improvement. Our method does not require any additional annotated data or external models during the training phase. Extensive experiments demonstrate that SDGO significantly improves model safety compared to both prompt-based and training-based baselines while maintaining helpfulness on general benchmarks. By aligning LLMs’ discrimination and generation capabilities, SDGO brings robust performance against out-of-distribution (OOD) jailbreaking attacks. This alignment achieves tighter coupling between these two capabilities, enabling the model’s generation capability to be further enhanced with only a small amount of discriminative samples. Our code and datasets are available at https://github.com/NJUNLP/SDGO.
Reinforcement learning from human feedback (RLHF) offers a promising approach to aligning large language models (LLMs) with human preferences. Typically, a reward model is trained or supplied to act as a proxy for humans in evaluating generated responses during the reinforcement training phase. However, current reward models operate as sequence-to-one models, allocating a single, sparse, and delayed reward to an entire output sequence. This approach may overlook the significant contributions of individual tokens toward the desired outcome. To this end, we propose a more fine-grained, token-level guidance approach for RL training. Specifically, we introduce RED, a novel REward reDistribition method that evaluates and assigns specific credit to each token using an off-the-shelf reward model. Utilizing these fine-grained rewards enhances the model’s understanding of language nuances, leading to more precise performance improvements. Notably, our method does not require modifying the reward model or introducing additional training steps, thereby incurring minimal computational costs. Experimental results across diverse datasets and tasks demonstrate the superiority of our approach.
Long chain-of-thought (CoT) supervision has become a common strategy to enhance reasoning in language models. While effective for large models, we identify a phenomenon we call Long CoT Degradation, in which small language models (SLMs; \leq3B parameters) trained on limited long CoT data experience significant performance deterioration. Through extensive experiments on the Qwen2.5, LLaMA3 and Gemma3 families, we demonstrate that this degradation is widespread across SLMs. In some settings, models trained on only 8k long CoT examples lose up to 75% of their original performance before fine-tuning. Strikingly, we further observe that for some particularly small models, even training on 220k long CoT examples fails to recover or surpass their original performance prior to fine-tuning. Our analysis attributes this effect to error accumulation: while longer responses increase the capacity for multi-step reasoning, they also amplify the risk of compounding mistakes. Furthermore, we find that Long CoT Degradation may negatively impacts downstream reinforcement learning (RL), although this can be alleviated by sufficiently scaled supervised fine-tuning (SFT). Our findings challenge common assumptions about the benefits of long CoT training for SLMs and offer practical guidance for building more effective small-scale reasoning models.
Improving Reasoning Capabilities in Small Models through Mixture-of-layers Distillation with Stepwise Attention on Key Information
PDF ↗The significant computational demands of large language models have increased interest in distilling reasoning abilities into smaller models via Chain-of-Thought (CoT) distillation. Current CoT distillation methods mainly focus on transferring teacher-generated rationales for complex reasoning to student models. However, they do not adequately explore teachers’ dynamic attention toward critical information during reasoning. We find that language models exhibit progressive attention shifts towards key information during reasoning, which implies essential clues for drawing conclusions. Building on this observation and analysis, we introduce a novel CoT distillation framework that transfers the teacher’s stepwise attention on key information to the student model. This establishes structured guidance for the student’s progressive concentration on key information during reasoning. More importantly, we develop a Mixture of Layers module enabling dynamic alignment that adapts to different layers between the teacher and student. Our method achieves consistent performance improvements across multiple mathematical and commonsense reasoning datasets. To our knowledge, it is the first method to leverage stepwise attention within CoT distillation to improve small model reasoning.
Large Language Models (LLMs) based agent systems have made great strides in real-world applications beyond traditional NLP tasks. This paper proposes a new LLM-based Multi-Agent System (LLM-MAS) benchmark, Collab-Overcooked, built on the popular Overcooked-AI game with more applicable and challenging tasks in interactive environments. Collab-Overcooked extends existing benchmarks in two novel ways. First, it provides a multi-agent framework supporting diverse tasks and objectives and encourages collaboration through natural language communication. Second, it introduces a spectrum of process-oriented evaluation metrics to assess the fine-grained collaboration capabilities of different LLM agents, a dimension often overlooked in prior work. We conduct extensive experiments with 13 popular LLMs and show that, while the LLMs exhibit a strong ability in goal interpretation, there are significant shortcomings in active collaboration and continuous adaptation, which are critical for efficiently fulfilling complex tasks. Notably, we highlight the strengths and weaknesses of LLM-MAS and provide insights for improving and evaluating LLM-MAS on a unified and open-source benchmark. The environments, 30 open-ended tasks, and the evaluation package are publicly available at https://github.com/YusaeMeow/Collab-Overcooked.
Prominent Large Language Model (LLM) services from providers like OpenAI and Google excel at general tasks but often underperform on domain-specific applications. Current customization services for these LLMs typically require users to upload data for fine-tuning, posing significant privacy risks. While differentially private (DP) data synthesis presents a potential alternative, its application commonly results in low effectiveness due to the introduction of excessive noise on data for DP. To overcome this, we introduce *Llamdex*, a novel framework that facilitates LLM customization as a service, where the client uploads pre-trained domain-specific *models* rather than data. This client-uploaded model, optionally protected by DP with much lower noise, is inserted into the base LLM via connection modules. Significantly, these connecting modules are trained without requiring sensitive domain data, enabling clients to customize LLM services while preserving data privacy. Experiments demonstrate that Llamdex improves domain-specific accuracy by up to 26% over state-of-the-art private data synthesis methods under identical privacy constraints and, by obviating the need for users to provide domain context within queries, maintains inference efficiency comparable to the original LLM service.
RESF: Regularized-Entropy-Sensitive Fingerprinting for Black-Box Tamper Detection of Large Language Models
PDF ↗The proliferation of Machine Learning as a Service (MLaaS) has enabled widespread deployment of large language models (LLMs) via cloud APIs, but also raises critical concerns about model integrity and security. Existing black-box tamper detection methods, such as watermarking and fingerprinting, rely on the stability of model outputs—a property that does not hold for inherently stochastic LLMs. We address this challenge by formulating black-box tamper detection for LLMs as a hypothesis-testing problem. To enable efficient and sensitive fingerprinting, we derive a first-order surrogate for KL divergence—the entropy-gradient norm—to identify prompts most responsive to parameter perturbations. Building on this, we propose Regularized Entropy-Sensitive Fingerprinting (RESF), which enhances sensitivity while regularizing entropy to improve output stability and control false positives. To further distinguish tampering from benign randomness, such as temperature shifts, RESF employs a lightweight two-tier sequential test combining support-based and distributional checks with rigorous false-alarm control.Comprehensive analysis and experiments across multiple LLMs show that RESF achieves up to 98.80% detection accuracy under challenging conditions, such as minimal LoRA fine-tuning with five optimized fingerprints. RESF consistently demonstrates strong sensitivity and robustness, providing an effective and scalable solution for black-box tamper detection in cloud-deployed LLMs.
Large Audio-Language Models (LALMs) are augmented with the ability to perceive audio, demonstrating impressive capabilities in processing combined audio and text signals. However, their reliability when faced with conflicting inputs across modalities remains largely unexplored. This study examines how LALMs prioritize information when presented with inconsistent audio-text pairs. Through extensive evaluation across diverse audio understanding tasks, we reveal a concerning phenomenon: when inconsistencies exist between modalities, LALMs display a significant bias toward textual input, often disregarding audio evidence. This tendency leads to substantial performance degradation in audio-centric tasks and raises important reliability concerns for real-world applications. We further investigate the influencing factors of text bias, explore mitigation strategies through supervised fine-tuning, and analyze model confidence patterns that reveal persistent overconfidence even with contradictory inputs. These findings underscore the need for improved modality balancing during training and more sophisticated fusion mechanisms to enhance robustness when handling conflicting multi-modal inputs. The project is available at https://github.com/WangCheng0116/MCR-BENCH.
With thousand years of history, calligraphy serve as one of the representative symbols of Chinese culture. Increasing works try to digitize calligraphy by recognizing the context of calligraphy for better preservation and propagation. However, previous works stick to isolated single character recognition, not only requires unpractical manual splitting into characters, but also abandon the enriched context information that could be supplementary. To this end, we construct the pioneering end-to-end calligraphy recognition benchmark dataset, this dataset is challenging due to both the visual variations such as different writing styles and the textual understanding such as the domain shift in semantics. We further propose CalligraphicOCR (COCR) equipped with calligraphic image augmentation and action-based corrector targeted at the challenging root of this setting. Experiments demonstrate the advantage of our proposed model over cutting-edge baselines, underscoring the necessity of introducing this new setting, thereby facilitating a solid precondition for protecting and propagating the already scarce resources.
In this paper, we propose TrInk, a Transformer-based model for ink generation, which effectively captures global dependencies. To better facilitate the alignment between the input text and generated stroke points, we introduce scaled positional embeddings and a Gaussian memory mask in the cross-attention module. Additionally, we design both subjective and objective evaluation pipelines to comprehensively assess the legibility and style consistency of the generated handwriting. Experiments demonstrate that our Transformer-based model achieves a 35.56% reduction in character error rate (CER) and an 29.66% reduction in word error rate (WER) on the IAM-OnDB dataset compared to previous methods. We provide an demo page with handwriting samples from TrInk and baseline models at: https://akahello-a11y.github.io/trink-demo/
Single LLM, Multiple Roles: A Unified Retrieval-Augmented Generation Framework Using Role-Specific Token Optimization
PDF ↗Existing studies have optimized retrieval-augmented generation (RAG) across various sub-tasks, such as query understanding and retrieval refinement, but integrating these optimizations into a unified framework remains challenging. To tackle this problem, this work proposes RoleRAG, a unified RAG framework that achieves efficient multi-task processing through role-specific token optimization. RoleRAG comprises six modules, each handling a specific sub-task within the RAG process. Additionally, we introduce a query graph to represent the decomposition of the query, which can be dynamically resolved according to the decomposing state. All modules are driven by the same underlying LLM, distinguished by task-specific role tokens that are individually optimized. This design allows RoleRAG to dynamically activate different modules within a single LLM instance, thereby streamlining deployment and reducing resource consumption. Experimental results on five open-domain question-answering datasets demonstrate the effectiveness, generalizability, and flexibility of our framework.
Recent advancements in large language models (LLMs) showcase varied multilingual capabilities across tasks like translation, code generation, and reasoning. Previous assessments often limited their scope to fundamental natural language processing (NLP) or isolated capability-specific tasks. To alleviate this drawback, we aim to present a comprehensive multilingual multitask benchmark. First, we introduce P-MMEval, a large-scale benchmark covering fundamental and capability-specialized datasets. Furthermore, P-MMEval delivers consistent language coverage across various datasets and provides parallel samples. Finally, we conduct extensive experiments on representative multilingual model series to compare performances across models and tasks, explore the relationship between multilingual performances and factors such as tasks, model sizes, languages, and prompts, and examine the effectiveness of knowledge transfer from English to other languages. The resulting insights are intended to offer valuable guidance for future research.
Co-Evolving LLMs and Embedding Models via Density-Guided Preference Optimization for Text Clustering
PDF ↗Large language models (LLMs) have shown strong potential in enhancing text clustering when combined with traditional embedding models. However, existing methods predominantly treat LLMs as static pseudo-oracles, i.e., unidirectionally querying them for similarity assessment or data augmentation, while never seeking feedback from embedding models to improve them. In this work, we propose a training framework that enables bidirectional refinement between LLMs and embedding models. We first design task-aware prompts to guide the LLM in generating interpretations for the input texts. These interpretations are projected into the embedding space, in which interpretations that are preferred by the embedding model are selected based on their distribution densities. The selected interpretations are then used to fine-tune the LLM via preference optimization to prioritize the generation of helpful interpretations. Meanwhile, we enhance the embedding model via contrastive learning on the generated interpretations and perform clustering on the output embeddings, leading to iterative co-training between the LLM and the embedding model. Experiments on 14 benchmark datasets across 5 tasks demonstrate the effectiveness of our method.
Quantization is widely adopted to accelerate inference and reduce memory consumption in large language models (LLMs). While activation-weight joint quantization enables efficient low-precision decoding, it suffers substantial performance degradation on multi-step reasoning tasks. We propose QSPEC, a novel quantization paradigm that decouples efficiency from quality by integrating two complementary schemes via speculative decoding: low-precision joint quantization for fast drafting and high-precision weight-only quantization for accurate verification. QSPEC reuses both weights and KV cache across stages, enabling near-zero-cost switching without retraining or auxiliary models. Compared to high-precision baselines, QSPEC achieves up to 1.64x speedup without quality degradation, and outperforms state-of-the-art speculative decoding methods by up to 1.55x in batched settings. Furthermore, QSPEC supports plug-and-play deployment and generalizes well across model scales, quantization methods, and workloads. These properties make QSPEC a practical and scalable solution for high-fidelity quantized LLM serving under memory-constrained scenarios.
pFedGPT: Hierarchically Optimizing LoRA Aggregation Weights for Personalized Federated GPT Models
PDF ↗Federated finetuning of Large Language Models (LLMs) using Low-Rank Adaptation (LoRA) offers computational efficiency and preserves data privacy. However, applying LoRA in federated settings faces significant challenges: standard approaches struggle with data heterogeneity, and existing personalization techniques fail to precisely adapt shared global knowledge to individual client needs. To address these issues, we propose pFedGPT, a framework that leverages Hierarchical Bayesian Optimization (HBO) for fine-grained, personalized LoRA aggregation. pFedGPT intelligently partitions LoRA parameters based on model structure and client information, then employs HBO to hierarchically search for optimal, module-specific weights. This enables a nuanced integration of the downloaded global LoRA state with each client’s local model, precisely capturing client-specific requirements. To manage the optimization cost inherent in HBO, pFedGPT incorporates efficient multi-fidelity evaluations and a curriculum learning strategy. Extensive experiments demonstrate that pFedGPT achieves state-of-the-art (SOTA) performance on personalized FL benchmarks, showcasing robustness and scalability while introducing only minimal (approx. 4%) additional optimization overhead. Our results also underscore the limitations of traditional FL methods for LoRA-based LLM personalization, highlighting the need for tailored approaches like pFedGPT.
As large language models (LLMs) often generate plausible but incorrect content, error detection has become increasingly critical to ensure truthfulness.However, existing detection methods often overlook a critical problem we term as **self-consistent error**, where LLMs repeatedly generate the same incorrect response across multiple stochastic samples.This work formally defines self-consistent errors and evaluates mainstream detection methods on them.Our investigation reveals two key findings: (1) Unlike inconsistent errors, whose frequency diminishes significantly as the LLM scale increases, the frequency of self-consistent errors remains stable or even increases.(2) All four types of detection methods significantly struggle to detect self-consistent errors.These findings reveal critical limitations in current detection methods and underscore the need for improvement.Motivated by the observation that self-consistent errors often differ across LLMs, we propose a simple but effective cross‐model probe method that fuses hidden state evidence from an external verifier LLM.Our method significantly enhances performance on self-consistent errors across three LLM families.
Traditional robot simulators focus on physical process modeling and realistic rendering, often suffering from high computational costs, inefficiencies, and limited adaptability. To handle this issue, we concentrate on behavior simulation in robotics to analyze and validate the logic behind robot behaviors, aiming to achieve preliminary evaluation before deploying resource-intensive simulators and thus enhance simulation efficiency. In this paper, we propose BeSimulator, a modular and novel LLM-powered framework, as an attempt towards behavior simulation in the context of text-based environments. By constructing text-based virtual environments and performing semantic-level simulation, BeSimulator can generalize across scenarios and achieve long-horizon complex simulation. Inspired by human cognition paradigm, it employs a “consider-decide-capture-transfer” four-phase simulation process, termed Chain of Behavior Simulation (CBS), which excels at analyzing action feasibility and state transition. Additionally, BeSimulator incorporates code-driven reasoning to enable arithmetic operations and enhance reliability, and reflective feedback to refine simulation. Based on our manually constructed behavior-tree-based simulation benchmark, BTSIMBENCH, our experiments show a significant performance improvement in behavior simulation compared to baselines, ranging from 13.60% to 24.80%. Code and data are available at https://github.com/Dawn888888/BeSimulator.
Large language models (LLMs) rely on massive amounts of training data, however, the quantity of empirically observed data is limited. To alleviate this issue, lots of LLMs leverage synthetic data to enhance the quantity of training data. Despite significant advancements in LLMs, the efficiency and scalability characteristics of data synthesis during pre-training phases remain insufficiently explored. In this work, we propose a novel data synthesis framework, Cognitive Combination Synthesis (CCS), designed to achieve highly efficient and scalable data synthesis. Specifically, our methodology mimics human cognitive behaviors by recombining and interconnecting heterogeneous data from diverse sources thereby enhancing advanced reasoning capabilities in LLMs. Extensive experiments demonstrate that: (1) effective data organization is essential, and our mapping-based combination learning approach significantly improves data utilization efficiency; (2) by enhancing data diversity, accuracy, and complexity, our synthetic data scales beyond 100B tokens, revealing CCS’s strong scalability. Our findings highlight the impact of data organization methods on LLM learning efficiency and the significant potential of scalable synthetic data to enhance model reasoning capabilities.
Recall with Reasoning: Chain-of-Thought Distillation for Mamba’s Long-Context Memory and Extrapolation
PDF ↗Mamba’s theoretical infinite-context potential is limited in practice when sequences far exceed training lengths. This work explores unlocking Mamba’s long-context memory ability by a simple-yet-effective method, Recall with Reasoning (RwR), by distilling chain-of-thought (CoT) summarization from a teacher model. Specifically, RwR prepends these summarization as CoT prompts during fine-tuning, teaching Mamba to actively recall and reason over long contexts. Experiments on LONGMEMEVAL and HELMET show that RwR outperforms existing long-term memory methods on the Mamba model. Furthermore, under similar pre-training conditions, RwR improves the long-context performance of Mamba relative to comparable Transformer/hybrid baselines while preserving short-context capabilities, all without changing the architecture.