Large language models (LLMs) frequently produce source code that seems correct and well-formed, yet includes hallucinated elements that cause downstream test failures. In this study, we benchmark state-of-the-art uncertainty quantification methods and existing baselines for the task of hallucination detection in source code and introduce a diff-based pipeline to construct a code dataset annotated with line-level hallucinations. Building on this, we train a lightweight Transformer-based detector that uses LLM internal representations to identify hallucinations, substantially outperforming existing methods across several code generation domains. The detector also shows particular promise for enabling self-correction in LLM-based coding agents. We release the first publicly available dataset of line-level code hallucinations, along with the corresponding source code and trained hallucination detectors https://github.com/datapaf/CodeHallucinationDetection
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
The Confidence Paradox: Unveiling the Latent Discriminative Power of Diffusion Large Language Models in Mathematical Reasoning
PDF ↗Diffusion large language models (DLLMs) have emerged as a promising alternative to autoregressive (AR) generation, uniquely offering token-level probabilities under bidirectional context. However, the semantics of their native uncertainty estimates remain underexplored. In this work, we uncover a calibration paradox inherent to the bidirectional generation mechanism of state-of-the-art DLLMs. Concretely, we demonstrate that diffusion confidence is structurally distinct from AR likelihood. Notably, LLaDA-8B is highly miscalibrated (31.2% ECE) on mathematical reasoning benchmarks, yet possesses superior discriminative power (0.826 AUROC), significantly outperforming comparable AR baselines in single-pass settings (0.611 AUROC). We diagnose that this paradox arises because diffusion confidence functions less like a probability of correctness and more like a proxy for structural consistency enabled by the model’s bidirectional access to the entire solution path. We further show that lightweight post-hoc calibration can reconcile this gap, reducing ECE by over 60% while preserving the strong ranking signal. Our findings suggest that DLLMs offer a unique, cost-efficient uncertainty signal for reasoning tasks that complements expensive AR approaches.
Large language models (LLMs), like human language learners, show patterns influenced by their dominant training language. Just as humans display language patterns influenced by their native tongue (semantic accents) when learning new languages, LLMs often default to English-centric responses even when generating in other languages. However, we observe that explicitly providing cultural context in prompts significantly improves the models’ ability to generate culturally localized responses. We term this phenomenon the explicit-implicit localization gap, indicating that while cultural knowledge exists within LLMs, it may not naturally surface in multilingual interaction without explicitly including cultural context. In this paper, we (1) quantify this gap in multiple LLMs using a new cultural localization benchmark and find large (>10%) gaps in the majority of investigated models. (2) Demonstrate a fundamental trade-off between localization accuracy and output diversity. (3) Through mechanistic interpretability, we identify the underlying localization mechanisms within LLMs and show that these mechanisms are both language and task agnostic, with individual steering vectors effectively generalizing across different languages and culturally-relevant tasks.
In fine-grained sparse Mixture-of-Experts (MoE) models, a large pool of specialized experts replaces a small homogeneous set, shifting performance and throughput to be governed by inference-time expert activation. Yet most existing optimization recipes implicitly assume a fixed activation budget (e.g., a constant Top-k per layer), whose behavior in fine-grained MoEs is poorly understood. We first characterize runtime skipping strategies, quantifying the accuracy–efficiency trade-off of (i) uniform fixed activation and (ii) static layer-wise Top-k allocation found by search. Our analysis reveals that static skipping can already provide substantial throughput gains, but optimal static schedules vary significantly across models and routing mechanisms. We therefore introduce Adaptive Skipping with Entropy-Penalized Thresholding (ASET), a training-free policy that adapts token-level activation using router confidence and entropy while remaining within the model’s original budget. Across the fine-grained MoEs we study, static skipping policies yield 10–78% throughput gains with minimal performance degradation, including \ge10% improvement on DeepSeek-V3 without measurable loss. On the OLMoE testbed, ASET yields a Pareto frontier between average activation and task quality. Overall, these results identify expert skipping as a practical lever for faster fine-grained MoE inference, with adaptive activation helping when fixed budgets are too rigid.
Medical report generation from medical images is a vital AI task that helps doctors with diagnosis and marks a significant step toward creating general AI-powered medical systems. However, previous methods either fail to optimize factual accuracy or heavily depend on expert preference data. To overcome these challenges, we propose MedQPA, an automatic and generalizable report evaluation technique that uses question proposing and answering to enable controllable, structured reasoning grounded in medical domain knowledge and the factual correctness of the report. Additionally, we design MedQPA-Gen, a medical report generation pipeline that maximizes the MedQPA score through prompt engineering and reinforcement learning with MedQPA as a reward signal. We demonstrate that MedQPA is an accurate evaluation metric that closely correlates with human preferences. More importantly, MedQPA-Gen achieves higher human preference scores and better performance on downstream tasks. We open-source code at this repo https://github.com/MedQPA-gen/MedQPA-gen.
Chain-of-Relations: Faithful and Efficient LLM Reasoning over Knowledge Graphs via Relation-Centric Exploration
PDF ↗Knowledge graph question answering (KGQA) serves as an essential benchmark for KG-enhanced large language models. Among various approaches, agent-based methods have emerged as an effective solution.Existing methods adopt entity-centric exploration that incrementally constructs reasoning paths by selecting and connecting intermediate entities. However, they face two critical limitations. (1) Entity incompleteness vulnerability arises when some intermediate entities lack semantic information beyond opaque IDs, preventing relevance evaluation and leading to discarding valid reasoning paths.(2) Premature entity pruning occurs because beam search retains only top-ranked entities at each step, eliminating candidates before their relevance can be verified.To address these challenges, this paper proposes Chain-of-Relations (CoR) with relation-centric exploration and global entity filtering, reducing dependence on entity completeness and ensuring complete candidate retrieval before constraint validation.Experiments on three benchmark datasets show that CoR consistently outperforms strong baselines in both F1 score and KG-grounded Rate.
Although Large Language Models undergo rigorous safety alignment, they remain vulnerable to adversarial attacks. Existing methods, particularly gradient-based prompt optimization, suffer from high computational costs and produce uninterpretable, high-perplexity inputs. While recent logit-space attacks improve efficiency, they often rely on cumbersome auxiliary models or complex pipelines. In this work, we propose Sparse Index-Based Intervention (SIBI), a white-box, inference-time jailbreak that bypasses guardrails via lightweight, sparse logit editing. SIBI operates without gradients or auxiliary models, modifying pre-softmax logits using a compact, tokenizer-aligned dictionary of penalty and reward tokens. By incorporating temperature-consistent scaling and a mixed-norm trust region, the method ensures attack effectiveness while preserving generation fluency. On standard benchmarks, SIBI achieves high attack success rates while reducing computational overhead and space overhead compared to optimization baselines.
Despite the rapid progress of large vision-language models (LVLMs), fine-grained, state-conditioned GUI interaction remains challenging. Current evaluations offer limited coverage, imprecise target-state definitions, and an overreliance on final-task success, obscuring where and why agents fail.To address this gap, we introduce FineState-Bench, a benchmark that evaluates whether an agent can correctly ground an instruction to the intended UI control and reach the exact target state.FineState-Bench comprises 2,209 instances across desktop, web, and mobile platforms, spanning four interaction families and 23 UI component types, with each instance explicitly specifying an exact target state for fine-grained state setting.We further propose FineState-Metrics, a four-stage diagnostic pipeline with stage-wise success rates: Localization Success Rate (SR@Loc), Interaction Success Rate (SR@Int), Exact State Success Rate at Locate (ES-SR@Loc), and Exact State Success Rate at Interact (ES-SR@Int), and a plug-and-play Visual Diagnostic Assistant (VDA) that generates a Description and a bounding-box Localization Hint to diagnose visual grounding reason via controlled w/ vs. w/o comparisons.On FineState-Bench, exact goal-state success remains low: ES-SR@Int peaks at 32.8% on Web and 22.8% on average across platforms. With VDA localization hints, Gemini-2.5-Flash gains +14.9 ES-SR@Int points, suggesting substantial headroom from improved visual grounding, yet overall accuracy is still insufficient for reliable fine-grained state-conditioned interaction Github.
Do Language Models Use Logophoric Cues? Evidence from Mandarin Chinese Long-Distance Reflexive
PDF ↗Resolving anaphora requires integrating syntactic, semantic, and discourse information. Mandarin Chinese offers a particularly revealing case through the reflexive ziji, whose interpretation permits long-distance binding licensed by logophoric cues (i.e., cues relevant to discourse perspective). While these cues have been extensively studied in linguistic theory and psycholinguistic experiments, it remains an open question to what extent such cues are captured by computational models.We investigate this question by probing large language models’ sensitivity to four logophoric cues known to license long-distance binding of ziji: predicate type, perspective marking, discourse topicality, and discourse relation. Using minimal pairs and surprisal-based measures, we assess whether models exhibit systematic biases toward non-local antecedents in logophoric contexts.Across two model families, we find that (i) models exhibit above-chance sensitivity to all four cues; (ii) lexically anchored cues are more robustly captured than discourse-level cues; and (iii) some cues generalize cross-lingually, whereas others appear to depend on language-specific training data. Taken together, these findings provide non-English evidence that large language models capture certain aspects of logophoricity, yet continue to struggle with discourse-level representations that are central to human anaphora resolution. Code and data are available at: https://github.com/yunfang-dong/mandarin-logophoricity-llm
Large Language Models (LLMs) are reshaping recommender systems by leveraging extensive world knowledge and semantic reasoning to interpret user intent. However, effectively integrating these capabilities with collaborative signals while avoiding prohibitive inference latency remains a critical bottleneck. To address this, we propose a trajectory-driven internalization framework to develop a Single-agent Trajectory-Aligned Recommender (STAR). Specifically, to internalize complex reasoning capabilities into a single efficient model, we first design a multi-agent teacher system capable of multi-turn tool usage and reflection. This teacher utilizes a Collaborative Signal Translation mechanism to explicitly convert latent behavioral patterns into descriptive natural language evidence to enhance reasoning accuracy. Subsequently, a trajectory-driven distillation pipeline transfers this agentic logic, including planning, tool usage, and self-reflection, into the compact STAR model. Extensive experiments demonstrate that STAR surpasses its teacher by 8.7% to 39.5% while eliminating iterative latency, paving the way for real-time, reasoning-enhanced recommendation.
We investigate the integration of human-like working memory constraints into the Transformer architecture and implement several cognitively inspired attention variants, including fixed-width windows based and temporal decay based attention mechanisms. Our modified GPT-2 models are trained from scratch on developmentally plausible datasets (10M and 100M words). Performance is evaluated on grammatical judgment tasks (BLiMP) and alignment with human reading time data. Our results indicate that these cognitively-inspired constraints, particularly fixed-width attention, can significantly improve grammatical accuracy especially when training data is scarce. These constrained models also tend to show a stronger alignment with human processing metrics. The findings suggest that such constraints may serve as a beneficial inductive bias, guiding models towards more robust linguistic representations, especially in data-limited settings.
Automated analysis of customer feedback on social media is hindered by three challenges: the high cost of annotated training data, the scarcity of evaluation sets, especially in multilingual settings, and privacy concerns that prevent data sharing and reproducibility. We address these issues by developing a generalizable synthetic data generation pipeline applied to a case study on customer distress detection in French public transportation. Our approach utilizes backtranslation with fine-tuned models to generate 1.7 million synthetic tweets from a small seed corpus, complemented by synthetic reasoning traces. We train 600M-parameter reasoners with English and French reasoning that achieve 77-79% accuracy on human-annotated evaluation data, matching or exceeding SOTA proprietary LLMs and specialized encoders. Beyond reducing annotation costs, our pipeline preserves privacy by eliminating the exposure of sensitive user data. Our methodology can be adopted for other use cases and languages.
Long Story Short: Disentangling Compositionality and Long-Caption Understanding in Contrastive VLMs
PDF ↗Contrastive vision-language models (VLMs) have made significant progress in binding visual and textual information, yet understanding long, compositional captions remains an open challenge. While these capabilities are often assumed to be closely related, the conditions under which they reinforce each other remain unclear. In this paper, we empirically analyze when compositional reasoning and long-caption understanding transfer across tasks, and when this relationship fails. Through controlled experiments across diverse training objectives, datasets, and architectural designs, we find a bidirectional but sensitive relationship between the two capabilities. Models trained on poorly grounded captions or with limited parameter updates fail to generalize, while high-quality long-caption data with strong visual grounding promotes both capabilities simultaneously. We further show that architectural choices aimed at preserving general alignment, such as frozen positional embeddings, can inadvertently limit compositional learning. Our analysis provides actionable guidelines for data selection and model design to improve VLM generalization.
DR-HM: Distill-then-Reinforce Training with Cognition-Aware Data Synthesis for Harmful Meme Detection
PDF ↗Harmful memes convey offensive intent through implicit associations between visual symbols and text, requiring a broad understanding of cultural stereotypes and visual metaphors. Small-scale Multimodal Large Language Models (MLLMs) often lack the knowledge required to identify such implicit hate, whereas Large-scale MLLMs, despite their broader knowledge, exhibit systematic labeling bias. To address these challenges, we propose DR-HM, a Distill-then-Reinforce training framework with cognition-aware data synthesis for harmful meme detection, which aims to transfer knowledge from closed-source models while mitigating their biases. DR-HM introduces a six-step structured data synthesis scheme with self-refinement that decomposes meme analysis into a progressive, human-inspired reasoning process from entity recognition to harmfulness judgment. Based on the synthesized reasoning data, we further adopt a Distill-then-Reinforce training strategy. This approach combines a two-stage Supervised Fine-Tuning (SFT) with an Adaptive Group Relative Policy Optimization (A-GRPO) algorithm, which incorporates class-ratio-aware reward weighting and dynamic KL coefficients. Experiments on three benchmark datasets show that the proposed approach consistently outperforms existing methods and achieves an accuracy of 84.7% on the FHM dataset, approaching the reported performance of human annotators.
Large Language Model (LLM) Agents have demonstrated remarkable capabilities in task automation and intelligent decision-making, driving the widespread adoption of agent development frameworks such as LangChain and AutoGen. However, these frameworks predominantly serve developers with extensive technical expertise—a significant limitation considering that only 0.03% of the global population possesses the necessary programming skills. This stark accessibility gap raises a fundamental question: Can we enable everyone, regardless of technical background, to build their own LLM agents using natural language alone? To address this challenge, we introduce AutoAgent - a Fully-Automated and highly Self-Developing framework that enables users to create and deploy LLM agents through Natural Language Alone. Operating as an autonomous Agent Operating System, AutoAgent comprises four key components: i) Agentic System Utilities, ii) LLM-powered Actionable Engine, iii) Self-Managing File System, and iv) Self-Play Agent Customization module. This lightweight yet powerful system enables efficient and dynamic creation and modification of tools, agents, and workflows without coding requirements or manual intervention. Beyond its code-free agent development capabilities, AutoAgent also serves as a versatile multi-agent system for General AI Assistants. Comprehensive evaluations on the GAIA benchmark demonstrate AutoAgent’s effectiveness in generalist multi-agent tasks, surpassing existing state-of-the-art methods. Furthermore, AutoAgent’s Retrieval-Augmented Generation (RAG)-related capabilities have shown consistently superior performance compared to many alternative LLM-based solutions.
High-Throughput and Memory-Efficient Zeroth-Order Fine-tuning LLMs with Distributed Parallel Computing
PDF ↗Fine-tuning large language models (LLMs) remains resource-intensive due to their sheer scale. While zeroth-order (ZO) optimization provides a memory-efficient alternative by eliminating backward passes, its application to multi-hundred-billion-parameter models is constrained by GPU memory and compute throughput. The ZO2 framework addresses the memory bottleneck by offloading model parameters to CPU memory and overlapping transformer block transfer with dual forward computation on a single GPU. However, ZO2 remains limited by its single-device execution and achieves modest throughput. In this work, we present DistZO2 (Distributed Zeroth-Order Offloading), a high-throughput, memory-efficient framework for distributed zeroth-order fine-tuning of LLMs. DistZO2 introduces three parallel strategies: (1) Perturbation Parallelism (PertP), which parallelizes the two perturbed forward passes across devices; (2) Distributed Data Parallelism (DDP), adapted to the scalar-gradient nature of ZO training; and (3) a unified 2D Parallelism design that combines PertP and DDP. To further mitigate communication bottlenecks introduced by parameter offloading, we propose a hardware-aware communication strategy that slices parameter blocks and redistributes them across GPUs via high-speed interconnects such as NVLink. DistZO2 scales zeroth-order fine-tuning to modern multi-GPU systems, preserving ZO2’s memory efficiency while substantially improving training throughput. In our experiments on OPT-175B, DistZO2 achieves a 3x speedup over ZO2 with distributed computing.
Emotion Trajectory-aware Retrieval for Markov-driven Emotion Anticipation in LLM-based Emotional Support Conversation
PDF ↗Emotional support conversation (ESC) aims to alleviate users’ psychological stress. Selecting the appropriate strategy is crucial for effective emotional support. Current strategy planner-based methods prioritize immediate responses while neglecting users’ future reactions. Some studies retrieve historical examples with similar emotions to the current utterance, then anticipating future emotions based on next-turn emotions of historical examples. However, their retrievals focus on the current emotion (i.e. a single-turn emotion state), while they ignore the evolution of user’s emotion before the current state. We argue that retrievals considering the whole emotional trajectories enables models to capture the dynamic emotional needs, thereby enhancing the anticipation of future emotions. To this end, we propose Markov-driven emotion anticipation framework with emotion trajectory-aware retrieval for LLM-based ESC, which anticipates future emotion states to guide strategy planning and achieve sustained emotional support. First, we construct a dynamic emotion memory and perform hierarchical retrieval that combines semantic matching and emotion trajectory alignment. Then, we model emotional transitions as Markov chains, leveraging trajectory-aware retrieval to estimate future emotion. Finally, we use the anticipated emotion to steer LLMs in generating candidate strategies and introduce active online learning to optimize the planner, boosting its robustness on diverse users. Experiments on two datasets with two models shows that our method excels all baselines.
Event Causality Identification (ECI) requires models to determine whether a given pair of events in a context exhibits a causal relationship. While Large Language Models (LLMs) have demonstrated strong performance across various NLP tasks, their effectiveness in ECI remains limited due to biases in causal reasoning, often leading to overprediction of causal relationships (causal hallucination). To mitigate these issues and enhance LLM performance in ECI, we propose SERE, a structural example retrieval framework that leverages LLMs’ few-shot learning capabilities. SERE introduces an innovative retrieval mechanism based on three structural concepts: (i) Conceptual Path Metric, which measures the conceptual relationship between events using edit distance in ConceptNet; (ii) Syntactic Metric, which quantifies structural similarity through tree edit distance on syntactic trees; and (iii) Causal Pattern Filtering, which filters examples based on predefined causal structures using LLMs. By integrating these structural retrieval strategies, SERE selects more relevant examples to guide LLMs in causal reasoning, mitigating bias and improving accuracy in ECI tasks. Extensive experiments on multiple ECI datasets validate the effectiveness of SERE.
Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations
PDF ↗Improving the Theory of Mind (ToM) capability of Large Language Models (LLMs) is crucial for effective social interactions between these AI models and humans. However, the existing benchmarks often measure ToM capability improvement through story-reading, multiple-choice questions from a third-person perspective, while ignoring the first-person, dynamic, and open-ended nature of human-AI (HAI) interactions. To directly examine how ToM improvement techniques benefit HAI interactions, we first proposed the new paradigm of interactive ToM evaluation with both perspective and metric shifts. Next, following the paradigm, we conducted a systematic study of four representative ToM enhancement techniques using both four real-world datasets and a user study, covering both goal-oriented tasks (e.g., coding, math) and experience-oriented tasks (e.g., counseling). Our findings reveal that improvements on static benchmarks do not always translate to better performance in dynamic HAI interactions. This paper offers critical insights into ToM evaluation, showing the necessity of interaction-based assessments in developing next-generation, socially aware LLMs for HAI symbiosis.
Although large language models (LLMs) are increasingly used as annotators at scale, they are typically treated as a pragmatic fallback rather than a faithful estimator of human perspectives. This work challenges that presumption. By framing perspective-taking as the estimation of a latent group-level judgment, we characterize the conditions under which modern LLMs can outperform human annotators, including in-group humans, when predicting aggregate subgroup opinions on subjective tasks, and show that these conditions are common in practice. This advantage arises from structural properties of LLMs as estimators, including low variance and reduced coupling between representation and processing biases, rather than any claim of lived experience. Our analysis identifies clear regimes where LLMs act as statistically superior frontline estimators, as well as principled limits where human judgment remains essential. These findings reposition LLMs from a cost-saving compromise to a principled tool for estimating collective human perspectives.