The capabilities of Large Language Models (LLMs) are limited to some extent by pre-training, so some researchers optimize LLMs through post-training. Existing post-training strategies, such as memory-based retrieval or preference optimization, improve user alignment yet fail to enhance the model’s domain cognition. To bridge this gap, we propose a novel Dual-Phase Self-Evolution (DPSE) framework that jointly optimizes user preference adaptation and domain-specific competence. DPSE introduces a Censor module to extract multi-dimensional interaction signals and estimate satisfaction scores, which guide structured data expansion via topic-aware and preference-driven strategies. These expanded datasets support a two-stage fine-tuning pipeline: supervised domain grounding followed by frequency-aware preference optimization. Experiments across general NLP benchmarks and long-term dialogue tasks demonstrate that DPSE consistently outperforms Supervised Fine-Tuning, Preference Optimization, and Memory-Augmented baselines. Ablation studies validate the contribution of each module. In this way, our framework provides an autonomous path toward continual self-evolution of LLMs.
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
MTO: A Multi-turn Conversational Text-to-OverpassQL Dataset for Enhanced OpenStreetMap Query Generation
PDF ↗We propose a comprehensive framework for constructing multi-turn Text-to-OverpassQL dialogue datasets. Under this framework, we introduce the first multi-turn Text-to-OverpassQL dataset built upon the OverpassNL corpus. Our dataset comprises over 7,800 dialogues, each containing 2 to 4 user utterances, resulting in more than 20,000 individual utterances aligned with executable Overpass queries. To generate high-quality multi-turn dialogues, we design a four-stage pipeline. First, we convert Overpass queries into syntax trees using a custom parser developed based on the official OverpassQL grammar. This enables structural manipulation while preserving syntactic and executable validity. Second, we apply a diverse set of tree-editing templates, including both simple keyword-level changes and complex structural decompositions, to produce multiple valid and diverse Overpass queries. Third, we leverage a prompt-based approach to guide large language models in generating context-aware natural language questions, ensuring increasing inter-turn dependency across the dialogue. Finally, we implement a hybrid filtering strategy that combines manual annotation with model-assisted selection to validate alignment and correctness at scale. In addition to presenting the dataset, we evaluate the performance of several mainstream large language models and demonstrate that our end-to-end baseline model achieves competitive results. This work offers a new benchmark for studying executable semantic parsing and contextual understanding in map-based query tasks.
Large Language Models (LLMs) are widely used to generate plausible text on online platforms, without revealing the generation process.As users increasingly encounter such black-box outputs, detecting hallucinations has become a critical challenge.To address this challenge, we focus on developing a hallucination detection framework for black-box generators.Motivated by the observation that hallucinations, once introduced, tend to persist, we sample future contexts.The sampled future contexts provide valuable clues for hallucination detection and can be effectively integrated with various sampling-based methods.We extensively demonstrate performance improvements across multiple methods using our proposed sampling approach.
Do LLMs Really Know What They Don’t Know? Internal States Mainly Reflect Knowledge Recall Rather Than Truthfulness
PDF ↗Recent work suggests that LLMs "know what they don’t know", positing that hallucinated and factually correct outputs arise from distinct internal processes and can therefore be distinguished using internal signals.However, hallucinations have multifaceted causes: beyond simple knowledge gaps, they can emerge from training incentives that encourage models to exploit statistical shortcuts or spurious associations learned during pretraining.In this paper, we argue that when LLMs rely on such learned associations to produce hallucinations, their internal processes are mechanistically similar to those of factual recall, as both stem from strong statistical correlations encoded in the model’s parameters.To verify this, we propose a novel taxonomy categorizing hallucinations into Unassociated Hallucinations (UHs), where outputs lack parametric grounding, and Associated Hallucinations (AHs), which are driven by spurious associations. Through mechanistic analysis, we compare their computational processes and hidden-state geometries with factually correct outputs.Our results show that hidden states primarily reflect whether the model is recalling parametric knowledge rather than the truthfulness of the output itself. Consequently, AHs exhibit hidden-state geometries that largely overlap with factual outputs, rendering standard detection methods ineffective. In contrast, UHs exhibit distinctive, clustered representations that facilitate reliable detection.
Large language models (LLMs) are prone to distraction by contextual information during reasoning. Previous work primarily focuses on improving the generation of the next token while overlooking the potential bias introduced by existing premises. We propose a novel decoding method to mitigate such biases. Our framework uses predicted logits to estimate the model’s confidence. By decomposing the full context into multiple premises, we gain a clearer understanding of the relevance of each premise to the question. During next-token prediction, we refine the output by contrasting the logits with the highest and lowest confidence. Our method effectively reveals how the model dynamically activates and adjusts its consideration of each premise as reasoning progresses.
Op-Fed: Opinion, Stance, and Monetary Policy Annotations on FOMC Transcripts Using Active Learning
PDF ↗The U.S. Federal Open Market Committee (FOMC) regularly discusses and sets monetary policy, affecting the borrowing and spending decisions of hundreds of millions of people. In this work, we release Op-Fed, a dataset of 1044 human-annotated sentences and their contexts from FOMC transcripts that captures monetary policy stance—specifically, whether an individual FOMC member expresses support for tightening or loosening policy. We faced two major technical challenges in dataset creation: imbalanced classes—we estimate fewer than 8% of sentences express a non-neutral stance toward monetary policy—and inter-sentence dependence—65% of instances require context beyond the sentence-level to annotate. To address these challenges, we developed a five-stage hierarchical schema to isolate aspects of opinion, monetary policy, and stance toward monetary policy, as well as the level of context needed. Second, we selected instances to annotate using active learning, approximately doubling the number of positive instances across all schema aspects. Using Op-Fed, we found a top-performing, closed-weight LLM achieves 0.80 zero-shot accuracy in opinion classification but only 0.61 zero-shot accuracy classifying stance toward monetary policy—below our human baseline of 0.89. We expect Op-Fed to be useful for future model training, confidence calibration, and as a seed dataset for future annotation efforts.
With the rise in capabilities of large language models (LLMs) and their deployment in real-world tasks, evaluating LLM alignment with human preferences has become an important challenge. Current benchmarks average preferences across all users to compute aggregate ratings, overlooking individual user preferences when establishing model rankings. Since users have varying preferences in different contexts, we call for personalized LLM benchmarks that rank models according to individual needs. We compute personalized model rankings using ELO ratings and Bradley-Terry coefficients for 115 active Chatbot Arena users and analyze how user query characteristics (topics and writing style) relate to LLM ranking variations. We demonstrate that individual rankings of LLM models diverge dramatically from aggregate LLM rankings, with Bradley-Terry correlations averaging only \rho = 0.04 (57% of users show near-zero or negative correlation) and ELO ratings showing moderate correlation (\rho = 0.43). Through topic modeling and style analysis, we find users exhibit substantial heterogeneity in topical interests and communication styles, influencing their model preferences. We further show that a compact combination of topic and style features provides a useful feature space for predicting user-specific model rankings. Our results provide strong quantitative evidence that aggregate benchmarks fail to capture individual preferences for most users, and highlight the importance of developing personalized benchmarks that rank LLM models according to individual user preferences.
Text-to-image (T2I) models have achieved remarkable progress in high-quality image synthesis, yet most benchmarks rely on simple, self-contained prompts, failing to capture the complexity of real-world captions. Human-written captions often involve multiple interacting subjects, rich contextual references, and abstractive phrasing, conditions under which current image-text encoders like CLIP struggle. To systematically study these deficiencies, we introduce ANCHOR, a large-scale dataset of 70K+ abstractive captions sourced from five major news media organizations. Analysis with ANCHOR reveals persistent failures in multi-subject understanding, context reasoning, and nuanced grounding. Motivated by these challenges, we propose Subject-Aware Fine-tuning (SAFE), which uses Large Language Models (LLMs) to extract key subjects and enhance their representation at the embedding-level. Experiments with contemporary models show that SAFE significantly improves image-caption consistency and human preference alignment, serving as a practical and scalable solution. The dataset and code will be released upon publication.
UrBLiMP: A Benchmark for Evaluating the Linguistic Competence of Large Language Models in Urdu
PDF ↗Evaluating how large language models (LLMs) capture the grammatical structure of low-resource languages remains underexplored. This paper presents the Urdu Benchmark of Linguistic Minimal Pairs (UrBLiMP)—a diagnostic suite of 5,696 minimal pairs that contrast grammatical acceptability across ten core syntactic and morpho-syntactic phenomena in Urdu. The dataset is constructed from the Urdu Treebank and diverse text corpora, and human validation achieves a 96.1% inter-annotator agreement, confirming its reliability. We evaluate twenty one multilingual LLMs, including LLaMA-3-70B and Gemma-3-27B-PT, and additionally assess the proprietary GPT-4o model using grammar-prompting techniques. GPT-4o (grammar-prompted) attains the highest average accuracy (97.4%), reaching near-human performance on regular phenomena such as aspect agreement and ergativity. However, all models continue to struggle with flexible syntactic patterns like word-order variation and long-distance subject–verb agreement. UrBLiMP provides the first controlled evaluation framework for probing morpho-syntactic competence in Urdu and highlights both the progress and remaining challenges of multilingual and proprietary LLMs in low-resource settings.
Scientific reasoning is a key aspect of human intelligence, requiring the integration of multimodal inputs, domain expertise, and multi-step inference across various subjects. Existing benchmarks for multimodal large language models (MLLMs) often fail to capture the complexity and traceability of reasoning processes necessary for rigorous evaluation. To fill this gap, we introduce SciVQR, a multimodal benchmark covering 54 subfields in mathematics, physics, chemistry, geography, astronomy, and biology. SciVQR includes domain-specific visuals, such as equations, charts, and diagrams, and challenges models to combine visual comprehension with reasoning. The tasks range from basic factual recall to complex, multi-step inferences, with 46% including expert-authored solutions. SciVQR not only evaluates final answers but also examines the reasoning process, providing insights into how models reach their conclusions. Our evaluation of leading MLLMs, including both proprietary and open-source models, reveals significant limitations in handling complex multimodal reasoning tasks, underscoring the need for improved multi-step reasoning and better integration of interdisciplinary knowledge in advancing MLLMs toward true scientific intelligence. The dataset and evaluation code are publicly available at https://github.com/CASIA-IVA-Lab/SciVQR.
Large Language Models (LLMs) have demonstrated strong mathematical reasoning abilities through supervised fine-tuning and reinforcement learning. However, existing Process Reward Models (PRMs) are vulnerable to reward hacking and require expensive, large-scale annotation of reasoning steps, limiting their reliability and scalability. To address the first problem, we propose a novel reward model approach, Hierarchical Reward Model (HRM), which evaluates both individual and consecutive reasoning steps from fine-grained and coarse-grained level. HRM excels at assessing multi-step mathematical reasoning coherence, particularly in cases where a flawed step is later corrected through self-reflection. Furthermore, to address the inefficiency of autonomously annotating PRM training data via Monte Carlo Tree Search (MCTS), we propose a lightweight data augmentation strategy, Hierarchical Node Compression (HNC), which merges consecutive reasoning steps within the tree structure. Applying HNC to MCTS-generated reasoning trajectories increases the diversity and robustness of HRM training data, while introducing controlled noise with minimal computational overhead. Empirical results on the PRM800K dataset demonstrate that HRM, in conjunction with HNC, achieves superior stability and reliability in evaluation compared to PRM. Furthermore, cross-domain evaluations on MATH500 and GSM8K dataset confirm HRM’s superior generalization and robustness across diverse mathematical reasoning tasks.
This paper introduces a new paradigm for generative error correction (GER) framework in audio-visual speech recognition (AVSR) that reasons over modality-specific evidences directly in the language space. Our framework, **DualHyp**, empowers a large language model (LLM) to compose independent N-best hypotheses from separate automatic speech recognition (ASR) and visual speech recognition (VSR) models. To maximize the effectiveness of DualHyp, we further introduce **RelPrompt**, a noise-aware guidance mechanism that provides modality-grounded prompts to the LLM. RelPrompt offers the temporal reliability of each modality stream, guiding the model to dynamically switch its focus between ASR and VSR hypotheses for an accurate correction. Under various corruption scenarios, our framework attains up to 57.7% error rate gain on the LRS2 benchmark over standard ASR baseline, contrary to single-stream GER approaches that achieve only 10% gain. To facilitate research within our DualHyp framework, we release the code and the dataset comprising ASR and VSR hypotheses at https://github.com/sungnyun/dualhyp.
Logic Jailbreak: Efficiently Unlocking LLM Safety Restrictions Through Formal Logical Expression
PDF ↗Despite substantial advancements in aligning LLMs with human values, current safety mechanisms remain susceptible to jailbreak attacks. We attribute this vulnerability to the distributional discrepancies between alignment-oriented prompts and malicious prompts. To investigate this, and drawing inspiration from logic-driven NLP tasks, we introduce LogiBreak, a universal black-box jailbreak method that utilizes logical expression translation to bypass LLM safety mechanisms. By converting harmful natural language prompts into formal logical expressions, LogiBreak exploits the distributional gap between alignment data and logic-expressed inputs, preserving the underlying semantic intent and readability while evading safety constraints. Furthermore, to fill the gap of existing benchmarks that lack systematic resources specifically targeting logical expression-based attacks against LLM robustness, we construct a novel multilingual logical expression jailbreak dataset for evaluation. Our evaluations of LogiBreak in five languages demonstrate its effectiveness and generalizability in various linguistic contexts. The code is available at https://github.com/Applied-Machine-Learning-Lab/ACL2026_Logibreak.
The tasks of idiom understanding and dialect understanding are both well-established benchmarks in natural language processing. In this paper, we propose combining them, and using regional idioms as a test of dialect understanding. Towards this end, we propose three new benchmark datasets for the Quebec dialect of French: QFrCoRE, which contains 4,633 instances of idiomatic phrases, and QFrCoRT, which comprises 171 regional instances of idiomatic words, and a new benchmark for French Metropolitan expressions, MFrCoE, which comprises 4,938 phrases.We explain how to construct these corpora, so that our methodology can be replicated for other dialects. Our experiments with 111 LLMs reveal a critical disparity in dialectal competence: while models perform well on French Metropolitan, 65.77% of them perform significantly worse on Quebec idioms, with only 9.0% favoring the regional dialect. These results confirm that our benchmarks are a reliable tool for quantifying the dialect gap and that prestige-language proficiency does not guarantee regional dialect understanding.
Designing effective task-level prompts is crucial for improving the performance of Large Language Models (LLMs). While prior work on instruction induction demonstrates that LLMs can infer better instructions with limited examples, existing approaches often rely on input-output pairs, where obtaining labeled answers can be difficult or costly. To address this limitation, we propose Strategy-Induct, a framework that derives task-level instructions solely from a small set of example questions without requiring labeled answers. Our approach first prompts the model to generate explicit reasoning strategies for each question, forming (strategy, question) pairs. These pairs are then used to induce a task instruction that guides reasoning. Experiments across multiple tasks and model scales demonstrate that Strategy-Induct outperforms state-of-the-art methods in question-only settings. Furthermore, we observe that jointly utilizing LLMs and Large Reasoning Models for both task instruction generation and inference can lead to further performance improvements.
Large Language Models (LLMs) have fundamentally transformed natural language processing (NLP), demonstrating remarkable capabilities across a wide spectrum of tasks. However, when applied to instruction-based text editing, LLMs continue to exhibit some limitations. Different from free-form generation, instruction-based editing requires precise, targeted modifications that respect two essential properties: faithfully implementing the specific instruction and local fidelity. Existing approaches often overlook these properties, treating editing as a generic text generation problem. As a result, they either over-edit or fail to apply modifications consistently. To address this gap, we propose HyperEdit, a framework that adaptively processes each editing request to best align with it. To achieve this, HyperEdit generates request-specific dynamic weights that guide the editing process. The computational overhead of producing these weights is minimized through a carefully designed hypernetwork. With this design, HyperEdit achieves a relatively 9% improvement over the state-of-the-art editing model.
Confidence in LLMs is a useful indicator of model uncertainty and answer reliability. Existing work mainly focused on single-turn scenarios, while research on confidence in complex multi-turn interactions is limited. In this paper, we investigate whether LLM-based search agents have the ability to communicate their own confidence through verbalized confidence scores after long sequences of actions, a significantly more challenging task compared to outputting confidence in a single interaction. Experimenting on open-source agentic models, we first find that models exhibit much higher task accuracy at high confidence while having near-zero accuracy when confidence is low. Based on this observation, we propose Test-Time Scaling (TTS) methods that use confidence scores to determine answer quality, encourage the model to try again until reaching a satisfactory confidence level. Results show that our proposed methods significantly reduce token consumption while demonstrating competitive performance compared to baseline fixed budget TTS methods.
As large language models (LLMs) are increasingly deployed, ensuring their safe use is paramount. Jailbreaking, adversarial prompts that bypass model alignment to trigger harmful outputs, present significant risks, with existing studies reporting high success rates in evading common LLMs. However, previous evaluations have focused solely on the models, neglecting the full deployment pipeline, which typically incorporates additional safety mechanisms like content moderation filters. To address this gap, we present a systematic evaluation of jailbreak attacks targeting LLM safety alignment, assessing their success across the full inference pipeline, including both input and output filtering stages. Our findings yield two key insights: first, nearly all evaluated jailbreak techniques can be detected by at least one safety filter, suggesting that prior assessments may have overestimated the practical success of these attacks; second, while safety filters are effective in detection, there remains room to better balance recall and precision to further optimize protection and user experience.We highlight critical gaps and call for further refinement of detection accuracy and usability in LLM safety systems.
Wikipedia is a critical resource for modern NLP, serving as a rich repository of up-to-date and citation-backed information on a wide variety of subjects. The reliability of Wikipedia—its groundedness in its cited sources—is vital to this purpose. This work analyzes both how grounded Wikipedia is and how readily fine-grained grounding evidence can be retrieved. To this end, we introduce PeopleProfiles—a large-scale, multi-level dataset of claim support annotations on biographical Wikipedia articles. We show that: ~22% of claims in Wikipedia *lead* sections are unsupported by the article body; ~30% of annotated claims in the article *body* are unsupported by their (publicly accessible) sources; and real-world Wikipedia citation practices often differ from documented standards. Finally, we show that complex evidence retrieval remains a challenge—even for recent reasoning rerankers.
State-of-the-art large multi-modal models (LMMs) face challenges when processing high-resolution images, as these inputs are converted into enormous visual tokens, many of which are irrelevant to the downstream task. In this paper, we propose Multi-turn Grounding-based Policy Optimization (MGPO), an end-to-end reinforcement learning (RL) framework that enables LMMs to iteratively focus on key visual regions by automatically cropping sub-images, based on model-predicted grounding coordinates within a multi-turn conversation framework. Compared to supervised fine-tuning (SFT), which requires costly additional grounding annotations, our approach highlights that LMMs can emerge robust grounding abilities during the RL training process, leveraging only a binary reward function derived from the correctness of the final answer. Additionally, we observe that LMMs struggle to autonomously trigger visual grounding during the rollout process. To address this cold start problem, we design a multi-turn conversational template and restrict policy loss computation to model outputs generated across multiple dialogue rounds, thereby promoting stable optimization. Extensive experiments demonstrate that, when trained on standard visual-question-short answering data without grounding annotations, MGPO effectively elicits stronger grounding capabilities compared to GRPO, leading to 5.4% improvement on in-distribution MME-Realworld and 5.2% improvement on the challenging out-of-distribution (OOD) V* Bench. Notably, MGPO post-training on Qwen2.5-VL-7B with 21K samples surpasses OpenAI’s o1 and GPT-4o models on the OOD V* Bench.