Best-of-n sampling improves the accuracy of large language models (LLMs) and large reasoning models (LRMs) by generating multiple candidate solutions and selecting the one with the highest reward. The key challenge for reasoning tasks is designing a scoring function that can identify correct reasoning chains without access to ground-truth answers. We propose Probabilistic Confidence Selection and Ranking for Reasoning Chains (PiCSAR): a simple, training-free method that scores each candidate generation using the joint log-likelihood of the reasoning and final answer. This method utilises both the scores of the reasoning path (*reasoning confidence*) and the final answer (*answer confidence*). PiCSAR achieves substantial gains across several benchmarks (+11.7 on AIME2024, +9.81 on AIME2025), outperforming baselines with at least 2x fewer samples in 20 out of 25 comparisons. Our analysis reveals that correct reasoning chains exhibit higher reasoning and answer confidence, justifying the effectiveness of PiCSAR.
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
XGUARD: A Graded Benchmark for Evaluating Safety Failures of Large Language Models on Extremist Content
PDF ↗Large Language Models (LLMs) can generate content spanning ideological rhetoric to explicit instructions for violence. However, existing safety evaluations often rely on simplistic binary labels (safe/unsafe), overlooking the nuanced spectrum of risk these outputs pose. To address this, we present XGUARD, a benchmark and evaluation framework designed to assess the severity of extremist content generated by LLMs on a multi-level grading. It includes 3,840 red-teaming prompts generated using templates informed by real-world extremist scenarios from social media, forums, and news. The framework categorizes model responses into five danger levels (0–4) defined by degree of extremist endorsement, enabling nuanced analysis of failure frequency and severity. We introduce the interpretable Attack Severity Curve (ASC) to visualize vulnerabilities and compare defense mechanisms across threat intensities. Using XGUARD, we evaluate five popular LLMs and two lightweight defense strategies, revealing key insights into current safety gaps and trade-offs between robustness and expressive freedom. Our work underscores the value of graded safety metrics for building trustworthy LLMs. The code and dataset is available at https://github.com/Abishethvarman/XGUARD
TRUST: Towards Robust Social Bot Detection via Uncertainty-Guided Pseudo-Labeling and Graph Structure Purification
PDF ↗Social bots threaten online platforms by mimicking human behavior and forming deceptive connections, enabling the dissemination of misinformation while evading detection. Existing graph-based detection models leverage graph neural networks (GNNs) to capture relational structures and multimodal user features. However, such models are vulnerable to deceptive message propagation, where bots deliberately interact with legitimate users. These interactions create heterophilous edges–connections between nodes with different labels (i.e. human and bot)–which undermine the homophily assumption that connected users typically share similar characteristics. In this work, we propose a novel framework to mitigate deceptive message propagation through node-level uncertainty estimation and graph structure purification. The framework comprises three key components: (1) Node uncertainty estimation employs evidential deep learning with an error-sensitive uncertainty loss to obtain calibrated node-wise uncertainty; (2) Uncertainty-guided pseudo-label generation assigns pseudo-labels to low-uncertainty nodes using a dynamic threshold; (3) Graph structure purification selectively disconnects heterophilous edges identified between differently labeled nodes. Extensive experiments on three benchmark datasets and six GNN backbones demonstrate that our framework consistently enhances detection performance and serves as an effective general-purpose enhancement module for social bot detection.
Generating long sequences with structural coherence remains a fundamental challenge for autoregressive models across sequential generation tasks. In symbolic music generation, this challenge is particularly pronounced, as existing methods are constrained by the severe error accumulation inherent in autoregressive models, leading to poor performance in music quality and structural integrity. In this paper, we propose the Anchored Cyclic Generation (ACG) paradigm, which relies on anchor features from previously generated musical content to guide subsequent generation during the autoregressive process, effectively mitigating error accumulation in autoregressive methods. Based on the ACG paradigm, we further propose the Hierarchical Anchored Cyclic Generation (Hi-ACG) framework, which employs a systematic global-to-local generation strategy and is highly compatible with our specifically designed piano token, an efficient musical representation. The experimental results demonstrate that compared to traditional autoregressive models, the ACG paradigm reduces cosine distance by an average of 34.7% between predicted feature vectors and ground-truth semantic vectors. In long-sequence symbolic music generation tasks, the Hi-ACG framework significantly outperforms existing mainstream methods in both subjective and objective evaluations. Furthermore, the framework exhibits excellent task generalization capabilities, achieving superior performance in related tasks such as music completion.
Advanced speech synthesis technologies have enabled highly realistic speech generation, posing security risks that motivate research into audio deepfake detection (ADD). While state space models (SSMs) offer linear complexity, pure causal SSMs architectures often struggle with the content-based retrieval required to capture global frequency-domain artifacts. To address this, we explore the scaling properties of hybrid architectures by proposing XLSR-MamBo, a modular framework integrating an XLSR front-end with synergistic Mamba-Attention backbones. We systematically evaluate four topological designs using advanced SSM variants, Mamba, Mamba2, Hydra, and Gated DeltaNet. Experimental results demonstrate that the MamBo-3-Hydra-N3 configuration achieves competitive performance compared to other state-of-the-art systems on the ASVspoof 2021 LA, DF, and In-the-Wild benchmarks. This performance benefits from Hydra’s native bidirectional modeling, which captures holistic temporal dependencies more efficiently than the heuristic dual-branch strategies employed in prior works. Furthermore, evaluations on the DFADD dataset demonstrate robust generalization to unseen diffusion- and flow-matching-based synthesis methods. Crucially, our analysis reveals that scaling backbone depth effectively mitigates the performance variance and instability observed in shallower models. These results demonstrate the hybrid framework’s ability to capture artifacts in spoofed speech signals, providing an effective method for ADD. Codes are publicly available at https://github.com/saki-ciallo/XLSR-MamBo.
Thought-Action Graph Reasoning: Faithful and Efficient Reasoning of Large Language Models via Reusing Past Experience
PDF ↗Large language models (LLMs) often hallucinate in question answering (QA) tasks due to a lack of factual knowledge. While integrating knowledge graphs (KGs) with LLMs has alleviated this issue, existing methods suffer from poor generalization or low reasoning efficiency, and critically, they overlook the learning and reuse of reasoning paths from past experiences. To address these challenges, we introduce Thought-Action Graph (TAG), a structured repository of reasoning experiences. TAG decomposes successful LLM-KG interaction trajectories into fine-grained semantic operators, which are stored in TAG constructed by the thought layer and action layer. Building upon TAG, we propose a novel KGQA paradigm \textemdash TAG-Reasoning (TAGR). TAGR first retrieves and assembles reasoning blueprints from TAG, and then guides LLM to efficiently execute on KG according to them. Through this approach, TAGR transforms the computationally expensive online exploration process of LLMs into an offline process of TAG retrieval and assembly. Experimental results on multiple KGQA benchmarks demonstrate that TAGR significantly outperforms state-of-the-art methods across key metrics, while drastically reducing the number of LLM calls and generated tokens. This work opens new avenues for building continual learning, efficient, and faithful KGQA systems.
Structured Dialogue Refinement: Building Retrieval-Augmented Question Answering on Goal-Oriented Dialogues
PDF ↗Retrieval-Augmented Generation (RAG) is widely used for question answering over well-structured document corpora. However, a large amount of real-world problem-solving knowledge is captured in goal-oriented dialogues, where common ground misalignment between users and helpers gives rise to sparse, diffuse, and dynamically refined evidence that challenges standard RAG pipelines. We propose Structured Dialogue Refinement (SDR), a unified framework that adapts dialogue corpora for RAG at both the retrieval and generation stages without altering the underlying pipeline. Specifically, SDR introduces Dual Dialogue Querying for intent-aligned retrieval via issue-centric and solution-centric pseudo-documents, and Graph-Structured Dialogues coupled with a relevance-driven subgraph selection strategy to enable effective utilization of conversational evidence. We further adopt a nugget-based evaluation setup for dialogue-grounded RAG, enabling fine-grained analysis of retrieval coverage and grounded answer generation. Experiments demonstrate that SDR substantially improves both retrieval quality and grounded QA performance under dialogue-specific structural challenges.
Prior work has shown that instruction-tuned large language models (LLMs) are less well calibrated than their base pre-trained counterparts. However, little is known about the frequently used chat template’s effect on the calibration of conversational LLMs. In this work, we investigate the mechanisms driving this miscalibration by decoupling the effects of the post-training algorithm and the chat format. We find that, while instruction tuning fundamentally harms calibration, the chat template aggravates the issue through an “ownership bias” – models are significantly more confident in their *own* answers than in identical answers provided by a user. Extensive experiments across six recent open-weight LLMs, three benchmarks, and three confidence elicitation methods show that models assign up to 26% higher confidence to their own responses. Leveraging this insight, we propose a simple inference-time strategy: framing the model’s answer as user input during confidence elicitation. This approach significantly reduces overconfidence and improves calibration by up to 26% without the need for retraining, narrowing the gap between base and instruction-tuned models.
The increasing adoption of large language models (LLMs) in software engineering necessitates rigorous security evaluation of their generated code. However, existing benchmarks often lack relevance to real-world AI-assisted programming scenarios, making them inadequate for assessing the practical security risks associated with AI-generated code in production environments. To address this gap, we introduce A.S.E (AI Code Generation Security Evaluation), a repository-level evaluation benchmark designed to closely mirror real-world AI programming tasks, offering a comprehensive and reliable framework for assessing the security of AI-generated code. Our evaluation of leading LLMs on A.S.E reveals several key findings. In particular, current LLMs still struggle with secure coding. The complexity in repository-level scenarios presents challenges for LLMs that typically perform well on snippet-level tasks. Moreover, a larger reasoning budget does not necessarily lead to better code generation. These observations offer valuable insights into the current state of AI code generation and help developers identify the most suitable models for practical tasks. They also lay the groundwork for refining LLMs to generate secure and efficient code in real-world applications.
The Best of Both Worlds: Combining Parallel and Sequential Inference Scaling via Aggregation Fine-Tuning
PDF ↗Scaling data and model size has been proven effective for boosting the performance of large language models. In addition to training-time scaling, recent studies have revealed that increasing test-time computational resources can further improve performance. In this work, we introduce Aggregation Fine-Tuning (AFT), a supervised fine-tuning paradigm where the model learns to synthesize multiple draft responses, referred to as proposals, into a single, refined answer, termed aggregation. At inference time, we apply a propose-and-aggregate strategy that iteratively generates and aggregates proposals, effectively scaling inference-time computation without relying on external guidance such as a reward model. Empirical results across benchmark datasets demonstrate that AFT-trained models achieve substantial gains with test-time scaling, outperforming best-of-N baselines while eliminating the need for external reward signals. Notably, an AFT model, fine-tuned from Llama3.1-8B-Base with only 64k data, achieves a 41.3% LC win rate on AlpacaEval 2, surpassing significantly larger LLMs such as Llama3.1-405B-Instruct and GPT-4. By combining sequential refinement and parallel sampling, the propose-and-aggregate framework scales inference-time computation in a flexible manner.
Large language models (LLMs) are increasingly released as open-weight models with safeguards against harmful requests. Nevertheless, sentence completion remains vulnerable to incomplete harmful prompts. In this work, we formalize this phenomenon as incomplete prompt jailbreaks (IPJ) and provide a systematic empirical characterization of when and how incomplete prompts elicit harmful continuations. We analyze diverse attractor types associated with incomplete sentence continuation and show that LLMs systematically delay refusal until sentence termination. We further demonstrate that training models to refuse incomplete harmful prompts via parameter tuning is insufficient, failing to generalize across both content domains and attractor types. To enable fine-grained control, we identify two functional neurons: termination and continuation neurons. By clarifying their roles in sentence completion, we highlight the potential of neuron-level interventions for more precise and robust IPJ defenses.
pQuant: Towards Effective Low-Bit Language Models via Decoupled Linear Quantization-Aware Training
PDF ↗Quantization-Aware Training from scratch has emerged as a promising approach for building efficient large language models (LLMs) with extremely low-bit weights (sub 2-bit), which can offer substantial advantages for edge deployment. However, existing methods still fail to achieve satisfactory accuracy and scalability. In this work, we identify a parameter democratization effect as a key bottleneck: the sensitivity of all parameters becomes homogenized, severely limiting expressivity. To address this, we propose pQuant, a method that decouples parameters by splitting linear layers into two specialized branches: a dominant 1-bit branch for efficient computation and a compact high-precision branch dedicated to preserving the most sensitive parameters. Through tailored feature scaling, we explicitly guide the model to allocate sensitive parameters to the high-precision branch. Furthermore, we extend this branch into multiple, sparsely-activated experts, enabling efficient capacity scaling. Extensive experiments indicate our pQuant achieves state-of-the-art performance in extremely low-bit quantization.
The Uniform Information Density (UID) hypothesis proposes that effective communication is achieved by maintaining a stable flow of information. In this work, we revisit this principle in the context of Large Language Model (LLM) reasoning, asking whether step-level uniformity reflects reasoning quality. To this end, we introduce a novel framework to quantify uniformity of information flow at both local and global levels, using an entropy-based stepwise density metric. Across experiments on seven reasoning benchmarks, we see a counter-intuitive pattern: while high-quality reasoning exhibit smooth step-by-step transitions (local uniformity) and structured, non-uniform information flow at the trajectory level (global non-uniformity). The results demonstrate that these uniformities outperform alternative internal signals as predictors of reasoning quality, and such divergence with human communication is not a model deficiency, but a byproduct of distinct objectives between human communication and LLM reasoning.
Train in Vain: Functionality-Preserving Poisoning to Prevent Unauthorized Use of Code Datasets
PDF ↗The widespread availability of large-scale code datasets has accelerated the development of code large language models (CodeLLMs), raising concerns about unauthorized dataset usage. Dataset poisoning offers a proactive defense by reducing the utility of such unauthorized training. However, existing poisoning methods often require full-dataset poisoning and introduce transformations that break code compilability. In this paper, we introduce FunPoison, a functionality-preserving poisoning approach that injects short, compilable weak-use fragments into executed code paths. FunPoison leverages reusable statement-level templates with automatic repair and conservative safety checking to ensure side-effect freedom, while a type-aware synthesis module preserves type correctness, suppresses static-analysis warnings, and improves stealth. Extensive experiments across multiple CodeLLMs and code-generation benchmarks show that FunPoison achieves effective poisoning by contaminating only 10% of the dataset, while maintaining 100% compilability and functional correctness. FunPoison also remains robust against advanced code sanitization techniques, including detection, purification, rewriting, static-analysis, and formatting defenses.
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.
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.