Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their further evolution is often hampered by the scarcity of high-quality training data and the heavy reliance of traditional methods on expert-labeled data. This reliance sets a ceiling on LLM performance and is particularly challenging in low data resource scenarios where extensive supervision is unavailable. To address this issue, we propose a novel paradigm named LANCE (**LAN**guage models as **C**ontinuous self-**E**volving data engineers) that enables LLMs to train themselves by autonomously generating, cleaning, reviewing, and annotating data with preference information. Our approach demonstrates that LLMs can serve as continuous self-evolving data engineers, significantly reducing the time and cost of post-training data construction. Through iterative fine-tuning on Qwen2 series models, we validate the effectiveness of LANCE across various tasks, showing that it can maintain high-quality data generation and continuously improve model performance. Across multiple benchmark dimensions, LANCE results in an average score enhancement of **3.64** for Qwen2-7B and **1.75** for Qwen2-7B-Instruct. This autonomous data construction paradigm not only lessens reliance on human experts or external models but also ensures data aligns with human preferences, offering a scalable path for LLM self-improvement, especially in contexts with limited supervisory data. Code is available at: https://github.com/Control-derek/LANCE.
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
Text-to-Image (T2I) models have made remarkable progress in generating images from text prompts, but their output quality and safety still depend heavily on how prompts are phrased. Existing safety methods typically refine prompts using large language models (LLMs), but they overlook the images produced, which can result in unsafe outputs or unnecessary changes to already safe prompts. To address this, we propose an iterative prompt refinement algorithm that uses Vision Language Models (VLMs) to analyze both the input prompts and the generated images. By leveraging visual feedback, our method refines prompts more effectively, improving safety while maintaining user intent and reliability comparable to existing LLM-based approaches. Additionally, we introduce a new dataset labeled with both textual and visual safety signals using off-the-shelf multi-modal LLM, enabling supervised fine-tuning. Experimental results demonstrate that our approach produces safer outputs without compromising alignment with user intent, offering a practical solution for generating safer T2I content. \textbf{\textcolor{red}{WARNING: This paper contains examples of harmful or inappropriate images generated by models.}}
Adapting cultural values in Large Language Models (LLMs) presents significant challenges, particularly due to biases and data limitations. Previous work aligns LLMs with different cultures using survey data, primarily from the World Values Survey (WVS). However, it remains unclear whether this approach effectively captures cultural nuances or produces distinct cultural representations for tasks like offensiveness classification. In this paper, we systematically investigate WVS-based training for cultural value adaptation and find that relying solely on survey data can homogenize cultural norms and interfere with factual knowledge. To address these issues, we propose augmenting WVS with encyclopedic and scenario-based cultural narratives from Wikipedia and NormAd. Our experiments across multiple cultures show that this approach captures more enhances differentiated cultural values and improves downstream classification performances.
As a crucial method in prompt engineering, In-Context Learning (ICL) enhances the generalization and knowledge utilization capabilities of Large Language Models (LLMs) (Dong et al., 2024). However, the lengthy retrieved contexts and limited token throughput in autoregressive models significantly constrain reasoning speed. To address this challenge, we propose N-Gram Trie Speculative Decoding, a novel approach that leverages the overlap between context and model output. This method constructs an n-gram trie from the context to generate drafts, accelerating token generation for LLMs. We evaluate our approach on summarization, Retrieval-Augmented Generation (RAG), and context-based Question Answering (QA) tasks. Experimental results on Vicuna-7B, Llama2-7B-Chat, and Llama3-8B-Instruct demonstrate substantial speed improvements without compromising accuracy. Compared with various strong baselines, our method achieves the highest mean speedup, showcasing its effectiveness and efficiency.
Visual-Language Models (VLMs) have shown remarkable performance across various tasks, particularly in recognizing geographic information from images. However, VLMs still show regional biases in this task. To systematically evaluate these issues, we introduce a benchmark consisting of 1,200 images paired with detailed geographic metadata. Evaluating four VLMs, we find that while these models demonstrate the ability to recognize geographic information from images, achieving up to 53.8% accuracy in city prediction, they exhibit significant biases. Specifically, performance is substantially higher for economically developed and densely populated regions compared to less developed (-12.5%) and sparsely populated (-17.0%) areas. Moreover, regional biases of frequently over-predicting certain locations remain. For instance, they consistently predict Sydney for images taken in Australia, shown by the low entropy scores for these countries. The strong performance of VLMs also raises privacy concerns, particularly for users who share images online without the intent of being identified. Our code and dataset are publicly available at https://github.com/uscnlp-lime/FairLocator.
Model compression through post-training pruning offers a way to reduce model size and computational requirements without significantly impacting model performance. However, the effect of pruning on the fairness of LLM-generated summaries remains unexplored, particularly for opinion summarisation where biased outputs could influence public views. In this paper, we present a comprehensive empirical analysis of opinion summarisation, examining three state-of-the-art pruning methods and various calibration sets across three open-source LLMs using four fairness metrics. Our systematic analysis reveals that pruning methods have larger impact on fairness than calibration sets. Building on these insights, we propose High Gradient Low Activation (HGLA) pruning, which identifies and removes parameters that are redundant for input processing but influential in output generation. Our experiments demonstrate that HGLA can better maintain or even improve fairness compared to existing methods, showing promise across models and tasks where traditional methods have limitations. Our human evaluation shows HGLA-generated outputs are fairer than existing state-of-the-art pruning methods.
This research investigates both explicit and implicit social biases exhibited by Vision-Language Models (VLMs). The key distinction between these bias types lies in the level of awareness: explicit bias refers to conscious, intentional biases, while implicit bias operates subconsciously. To analyze explicit bias, we directly pose questions to VLMs related to gender and racial differences: (1) Multiple-choice questions based on a given image (e.g., “What is the education level of the person in the image?”) (2) Yes-No comparisons using two images (e.g., “Is the person in the first image more educated than the person in the second image?”) For implicit bias, we design tasks where VLMs assist users but reveal biases through their responses: (1) Image description tasks: Models are asked to describe individuals in images, and we analyze disparities in textual cues across demographic groups. (2) Form completion tasks: Models draft a personal information collection form with 20 attributes, and we examine correlations among selected attributes for potential biases. We evaluate Gemini-1.5, GPT-4V, GPT-4o, LLaMA-3.2-Vision and LLaVA-v1.6. Our code and data are publicly available at https://github.com/uscnlp-lime/VisBias.
Evaluating statement autoformalization, translating natural language mathematics into formal languages like Lean 4, remains a significant challenge, with few metrics, datasets, and standards to robustly measure progress. In this work, we present a comprehensive approach combining improved metrics, robust benchmarks, and systematic evaluation, to fill this gap. First, we introduce BEq+, an automated metric that correlates strongly with human judgment, along with ProofNetVerif, a new dataset for assessing the quality of evaluation metrics, containing 3,752 annotated examples. Second, we develop two new autoformalization benchmarks: ProofNet#, a corrected version of ProofNet, and RLM25, with 619 new pairs of research-level mathematics from six formalization projects. Through systematic experimentation across these benchmarks, we find that current techniques can achieve up to 45.1% accuracy on undergraduate mathematics but struggle with research-level content without proper context. Our work establishes a reliable foundation for evaluating and advancing autoformalization systems.
Large Language Models (LLMs) have demonstrated robust performance in Semantic Parsing (SP) for well-defined queries with unambiguous intent and answerable responses. However, practical user questions frequently deviate from these ideal conditions, challenging the applicability of existing benchmarks. To address this issue, we introduce SQUAB, an automatic dataset generator of Ambiguous and Unanswerable questions. SQUAB generates complex, annotated SP tests using a blend of SQL and LLM capabilities. Results show that SQUAB reduces test generation costs by up to 99% compared to human-based solutions while aligning with real-world question patterns. Furthermore, these tests challenge LLM performance while revealing disparities between public and proprietary datasets. This highlights the need for a dynamic, automatic dataset generator as SQUAB. The code is designed for user extension to accommodate new ambiguous and unanswerable patterns and is available at https://anonymous.4open.science/r/squab-8716/.
VeriFact: Enhancing Long-Form Factuality Evaluation with Refined Fact Extraction and Reference Facts
PDF ↗Large language models (LLMs) excel at generating long-form responses, but evaluating their factuality remains challenging due to complex inter-sentence dependencies within the generated facts. Prior solutions predominantly follow a decompose-decontextualize-verify pipeline but often fail to capture essential context and miss key relational facts. In this paper, we introduce VeriFact, a factuality evaluation framework designed to enhance fact extraction by identifying and resolving incomplete and missing facts to support more accurate verification results. Moreover, we introduce FactRBench , a benchmark that evaluates both precision and recall in long-form model responses, whereas prior work primarily focuses on precision. FactRBench provides reference fact sets from advanced LLMs and human-written answers, enabling recall assessment. Empirical evaluations show that VeriFact significantly enhances fact completeness and preserves complex facts with critical relational information, resulting in more accurate factuality evaluation. Benchmarking various open- and close-weight LLMs on FactRBench indicate that larger models within same model family improve precision and recall, but high precision does not always correlate with high recall, underscoring the importance of comprehensive factuality assessment.
Chain-of-thought (CoT) prompting enhances reasoning in large language models (LLMs) but often leads to verbose and redundant outputs, thus increasing inference cost. We hypothesize that many reasoning steps are unnecessary for producing correct answers. To investigate this, we start with a systematic study to investigate what is the minimum reasoning required for a model to reach a stable decision. Based on the insights, we propose three inference-time strategies to improve efficiency: (1) early stopping via answer consistency, (2) boosting the probability of generating end-of-reasoning signals, and (3) a supervised method that learns when to stop based on internal activations. Experiments across five benchmarks and five open-weights LLMs show that our methods largely reduce token usage with little or no accuracy drop. In particular, on NaturalQuestions, Answer Consistency reduces tokens by over 40% while further improving accuracy. Our work underscores the importance of cost-effective reasoning methods that operate at inference time, offering practical benefits for real-world applications.
Spoken language from older adults often deviates from written norms due to omission, disordered syntax, constituent errors, and redundancy, limiting the usefulness of automatic transcripts in downstream tasks. We present COAS2W, a Chinese spoken-to-written corpus of 10,004 utterances from older adults, each paired with a written version, fine-grained error labels, and four-sentence context. Fine-tuned lightweight open-source models on COAS2W outperform larger closed-source models. Context ablation shows the value of multi-sentence input, and normalization improves performance on downstream translation tasks. COAS2W supports the development of inclusive, context-aware language technologies for older speakers. Our annotation convention, data, and code are publicly available at https://github.com/Springrx/COAS2W.
GUI agents powered by vision-language models (VLMs) show promise in automating complex digital tasks. However, their effectiveness in real-world applications is often limited by scarce training data and the inherent complexity of these tasks, which frequently require long-tailed knowledge covering rare, unseen scenarios. We propose RAG-GUI , a lightweight VLM that leverages web tutorials at inferencetime. RAG-GUI is first warm-started via supervised finetuning (SFT) and further refined through self-guided rejection sampling fine-tuning (RSF). Designed to be model-agnostic, RAG-GUI functions as a generic plug-in that enhances any VLM-based agent. Evaluatedacross three distinct tasks, it consistently outperforms baseline agents and surpasses other inference baselines by 2.6% to 13.3% acrosstwo model sizes, demonstrating strong generalization and practical plug-and-play capabilities in real-world scenarios.
Recently, Multimodal Large Language Models (MLLMs) have gained significant attention across various domains. However, their widespread adoption has also raised serious safety concerns.In this paper, we uncover a new safety risk of MLLMs: the output preference of MLLMs can be arbitrarily manipulated by carefully optimized images. Such attacks often generate contextually relevant yet biased responses that are neither overtly harmful nor unethical, making them difficult to detect. Specifically, we introduce a novel method, **P**reference **Hi**jacking (**Phi**), for manipulating the MLLM response preferences using a preference hijacked image. Our method works at inference time and requires no model modifications. Additionally, we introduce a universal hijacking perturbation – a transferable component that can be embedded into different images to hijack MLLM responses toward any attacker-specified preferences. Experimental results across various tasks demonstrate the effectiveness of our approach. The code for Phi is accessible at https://github.com/Yifan-Lan/Phi.
Detecting misogyny in multimodal content remains a notable challenge, particularly in culturally conservative and low-resource contexts like Bangladesh. While existing research has explored hate speech and general meme classification, the nuanced identification of misogyny in Bangla memes, rich in metaphor, humor, and visual-textual interplay, remains severely underexplored. To address this gap, we introduce BanMiMe, the first comprehensive Bangla misogynistic meme dataset comprising 2,000 culturally grounded samples where each meme includes misogyny labels, humor categories, metaphor localization, and detailed human-written explanations. We benchmark the various performance of open and closed-source vision-language models (VLMs) under zero-shot and prompt-based settings and evaluate their capacity for both classification and explanation generation. Furthermore, we systematically explore multiple fine-tuning strategies, including standard, data-augmented, and Chain-of-Thought (CoT) supervision. Our results demonstrate that CoT-based fine-tuning consistently enhances model performance, both in terms of accuracy and in generating meaningful explanations. We envision BanMiMe as a foundational resource for advancing explainable multimodal moderation systems in low-resource and culturally sensitive settings.
Large Language Models (LLMs), constrained by limited context windows, often face significant performance degradation when reasoning over long contexts. To address this, Retrieval-Augmented Generation (RAG) retrieves and reasons over chunks but frequently sacrifices logical coherence due to its reliance on similarity-based rankings. Similarly, divide-and-conquer frameworks (DCF) split documents into small chunks for independent reasoning and aggregation. While effective for local reasoning, DCF struggles to capture long-range dependencies and risks inducing conflicts by processing chunks in isolation. To overcome these limitations, we propose ToM, a novel Tree-oriented MapReduce framework for long-context reasoning. ToM leverages the inherent hierarchical structure of long documents (e.g., main headings and subheadings) by constructing a DocTree through hierarchical semantic parsing and performing bottom-up aggregation. Using a Tree MapReduce approach, ToM enables recursive reasoning: in the Map step, rationales are generated at child nodes; in the Reduce step, these rationales are aggregated across sibling nodes to resolve conflicts or reach consensus at parent nodes. Experimental results on 70B+ LLMs show that ToM significantly outperforms existing divide-and-conquer frameworks and retrieval-augmented generation methods, achieving better logical coherence and long-context reasoning.
Recent advancements in tool learning have enabled large language models (LLMs) to integrate external tools, enhancing their task performance by expanding their knowledge boundaries. However, relying on tools often introduces trade-offs between performance, speed, and cost, with LLMs sometimes exhibiting overreliance and overconfidence in tool usage. This paper addresses the challenge of aligning LLMs with their knowledge boundaries to make more intelligent decisions about tool invocation. We propose a multi-objective alignment framework that combines probabilistic knowledge boundary estimation with dynamic decision-making, allowing LLMs to better assess when to invoke tools based on their confidence. Our framework includes two methods for knowledge boundary estimation—consistency-based and absolute estimation—and two training strategies for integrating these estimates into the model’s decision-making process. Experimental results on various tool invocation scenarios demonstrate the effectiveness of our framework, showing significant improvements in tool efficiency by reducing unnecessary tool usage.
Existing video-language models (Video-LLMs) typically rely on concatenating visual tokens with textual inputs for joint modeling. However, this token-level alignment leads to significant inefficiency, especially when scaling to long videos with dense visual inputs. In this work, we propose a video-to-parameter efficiency paradigm named ViPE that eliminates redundant visual tokens by transforming video content into visual perceptual weights, which are directly injected into the LLM’s parameters. ViPE consists of a visual injection module that compresses video features into a small set of perceptual queries using a hierarchical merge strategy, and a visual perception module that integrates the resulting representations into the LLM through a lightweight LoRA-like mechanism. ViPE achieves performance comparable to token-based baselines such as LLaVA, while reducing FLOPs by 85% and inference time by up to 65%, demonstrating a highly efficient and scalable solution for video understanding.
Recent advances in test-time scaling have enabled Large Language Models (LLMs) to display sophisticated reasoning abilities via extended Chain-of-Thought (CoT) generation. Despite their impressive reasoning abilities, Large Reasoning Models (LRMs) frequently display unstable behaviors, e.g., hallucinating unsupported premises, overthinking simple tasks, and displaying higher sensitivity to prompt variations. This raises a deeper research question: \textit{How can we represent the reasoning process of LRMs to map their minds?} To address this, we propose a unified graph-based analytical framework for fine-grained modeling and quantitative analysis of LRM reasoning dynamics. Our method first clusters long, verbose CoT outputs into semantically coherent reasoning steps, then constructs directed reasoning graphs to capture contextual and logical dependencies among these steps. Through a comprehensive analysis of derived reasoning graphs, we also reveal that key structural properties, such as exploration density, branching, and convergence ratios, strongly correlate with models’ performance. The proposed framework enables quantitative evaluation of internal reasoning structure and quality beyond conventional metrics and also provides practical insights for prompt engineering and cognitive analysis of LLMs. Code and resources will be released to facilitate future research in this direction.
Large Language Models (LLMs) are catalyzing a paradigm shift in scientific discovery, evolving from task-specific automation tools into increasingly autonomous agents and fundamentally redefining research processes and human-AI collaboration. This survey systematically charts this burgeoning field, placing a central focus on the changing roles and escalating capabilities of LLMs in science. Through the lens of the scientific method, we introduce a foundational three-level taxonomy—Tool, Analyst, and Scientist—to delineate their escalating autonomy and evolving responsibilities within the research lifecycle. We further identify pivotal challenges and future research trajectories such as robotic automation, self-improvement, and ethical governance. Overall, this survey provides a conceptual architecture and strategic foresight to navigate and shape the future of AI-driven scientific discovery, fostering both rapid innovation and responsible advancement.