Recently, training-free methods for improving large language models (LLMs) have attracted growing interest, with token-level attention tuning emerging as a promising and interpretable direction. However, existing methods typically rely on auxiliary mechanisms to identify important or irrelevant task-specific tokens, introducing potential bias and limiting applicability. In this paper, we uncover a surprising and elegant alternative: the semantically empty initial token is a powerful and underexplored control point for optimizing model behavior. Through theoretical analysis, we show that tuning the initial token's attention sharpens or flattens the attention distribution over subsequent tokens, and its role as an attention sink amplifies this effect. Empirically, we find that: (1) tuning its attention improves LLM performance more effectively than tuning other task-specific tokens; (2) the effect follows a consistent trend across layers, with earlier layers having greater impact, but varies across attention heads, with different heads showing distinct preferences in how they attend to this token. Based on these findings, we propose ZeroTuning, a training-free approach that improves LLM performance by applying head-specific attention adjustments to this special token. Despite tuning only one token, ZeroTuning achieves higher performance on text classification, multiple-choice, and multi-turn conversation tasks across models such as Llama, Qwen, and DeepSeek. For example, ZeroTuning improves Llama-3.1-8B by 11.71% on classification, 2.64% on QA tasks, and raises its multi-turn score from 7.804 to 7.966. The method is also robust to limited resources, few-shot settings, long contexts, quantization, decoding strategies, and prompt variations. Our work sheds light on a previously overlooked control point in LLMs, offering new insights into both inference-time tuning and model interpretability.
The paradigm of using Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) as evaluative judges has emerged as an effective approach in RLHF and inference-time scaling. In this work, we propose Multimodal Reasoner as a Judge (MR. Judge), a paradigm for empowering general-purpose MLLMs judges with strong reasoning capabilities. Instead of directly assigning scores for each response, we formulate the judgement process as a reasoning-inspired multiple-choice problem. Specifically, the judge model first conducts deliberate reasoning covering different aspects of the responses and eventually selects the best response from them. This reasoning process not only improves the interpretibility of the judgement, but also greatly enhances the performance of MLLM judges. To cope with the lack of questions with scored responses, we propose the following strategy to achieve automatic annotation: 1) Reverse Response Candidates Synthesis: starting from a supervised fine-tuning (SFT) dataset, we treat the original response as the best candidate and prompt the MLLM to generate plausible but flawed negative candidates. 2) Text-based reasoning extraction: we carefully design a data synthesis pipeline for distilling the reasoning capability from a text-based reasoning model, which is adopted to enable the MLLM judges to regain complex reasoning ability via warm up supervised fine-tuning. Experiments demonstrate that our MR. Judge is effective across a wide range of tasks. Specifically, our MR. Judge-7B surpasses GPT-4o by 9.9% on VL-RewardBench, and improves performance on MM-Vet during inference-time scaling by up to 7.7%.
Recent advances in Large Multimodal Models (LMMs) have significantly improved their reasoning and Optical Character Recognition (OCR) capabilities. However, their performance on complex logical reasoning tasks involving text-rich images remains underexplored. To bridge this gap, we introduce LogicOCR, a benchmark comprising 1,100 multiple-choice questions designed to evaluate LMMs' logical reasoning abilities on text-rich images, while minimizing reliance on domain-specific knowledge (e.g., mathematics). We construct LogicOCR by curating a text corpus from the Chinese National Civil Servant Examination and develop a scalable, automated pipeline to convert it into multimodal samples. First, we design prompt templates to steer GPT-Image-1 to generate images with diverse backgrounds, interleaved text-illustration layouts, and varied fonts, ensuring contextual relevance and visual realism. Then, the generated images are manually verified, with low-quality examples discarded. We evaluate a range of representative open-source and proprietary LMMs under both Chain-of-Thought (CoT) and direct-answer settings. Our multi-dimensional analysis reveals key insights, such as the impact of test-time scaling, input modality differences, and sensitivity to visual-text orientation. Notably, LMMs still lag in multimodal reasoning compared to text-only inputs, indicating that they have not fully bridged visual reading with reasoning. We hope LogicOCR will serve as a valuable resource for advancing multimodal reasoning research. The dataset is available at https://github.com/MiliLab/LogicOCR.
Long-form legal reasoning remains a key challenge for large language models (LLMs) in spite of recent advances in test-time scaling. We introduce LEXam, a novel benchmark derived from 340 law exams spanning 116 law school courses across a range of subjects and degree levels. The dataset comprises 4,886 law exam questions in English and German, including 2,841 long-form, open-ended questions and 2,045 multiple-choice questions. Besides reference answers, the open questions are also accompanied by explicit guidance outlining the expected legal reasoning approach such as issue spotting, rule recall, or rule application. Our evaluation on both open-ended and multiple-choice questions present significant challenges for current LLMs; in particular, they notably struggle with open questions that require structured, multi-step legal reasoning. Moreover, our results underscore the effectiveness of the dataset in differentiating between models with varying capabilities. Adopting an LLM-as-a-Judge paradigm with rigorous human expert validation, we demonstrate how model-generated reasoning steps can be evaluated consistently and accurately. Our evaluation setup provides a scalable method to assess legal reasoning quality beyond simple accuracy metrics. Project page: https://lexam-benchmark.github.io/
Large Language Models (LLMs) have achieved impressive performance in diverse natural language processing tasks, but specialized domains such as Web3 present new challenges and require more tailored evaluation. Despite the significant user base and capital flows in Web3, encompassing smart contracts, decentralized finance (DeFi), non-fungible tokens (NFTs), decentralized autonomous organizations (DAOs), on-chain governance, and novel token-economics, no comprehensive benchmark has systematically assessed LLM performance in this domain. To address this gap, we introduce the DMind Benchmark, a holistic Web3-oriented evaluation suite covering nine critical subfields: fundamental blockchain concepts, blockchain infrastructure, smart contract, DeFi mechanisms, DAOs, NFTs, token economics, meme concept, and security vulnerabilities. Beyond multiple-choice questions, DMind Benchmark features domain-specific tasks such as contract debugging and on-chain numeric reasoning, mirroring real-world scenarios. We evaluated 26 models, including ChatGPT, Claude, DeepSeek, Gemini, Grok, and Qwen, uncovering notable performance gaps in specialized areas like token economics and security-critical contract analysis. While some models excel in blockchain infrastructure tasks, advanced subfields remain challenging. Our benchmark dataset and evaluation pipeline are open-sourced on https://huggingface.co/datasets/DMindAI/DMind_Benchmark, reaching number one in Hugging Face's trending dataset charts within a week of release.
Majority voting is considered an effective method to enhance chain-of-thought reasoning, as it selects the answer with the highest "self-consistency" among different reasoning paths (Wang et al., 2023). However, previous chain-of-thought reasoning methods typically generate only a single answer in each trial, thereby ignoring the possibility of other potential answers. As a result, these alternative answers are often overlooked in subsequent voting processes. In this work, we propose to generate ranked answers in each reasoning process and conduct ranked voting among multiple ranked answers from different responses, thereby making the overall self-consistency more reliable. Specifically, we use three ranked voting methods: Instant-runoff voting, Borda count voting, and mean reciprocal rank voting. We validate our methods on six datasets, including three multiple-choice and three open-ended question-answering tasks, using both advanced open-source and closed-source large language models. Extensive experimental results indicate that our proposed method outperforms the baselines, showcasing the potential of leveraging the information of ranked answers and using ranked voting to improve reasoning performance. The code is available at https://github.com/szu-tera/RankedVotingSC.
Editing images using natural language instructions has become a natural and expressive way to modify visual content; yet, evaluating the performance of such models remains challenging. Existing evaluation approaches often rely on image-text similarity metrics like CLIP, which lack precision. In this work, we introduce a new benchmark designed to evaluate text-guided image editing models in a more grounded manner, along two critical dimensions: (i) functional correctness, assessed via automatically generated multiple-choice questions that verify whether the intended change was successfully applied; and (ii) image content preservation, which ensures that non-targeted regions of the image remain visually consistent using an object-aware masking technique and preservation scoring. The benchmark includes over 1000 high-quality editing examples across 20 diverse content categories, each annotated with detailed editing instructions, evaluation questions, and spatial object masks. We conduct a large-scale study comparing GPT-Image-1, the latest flagship in the text-guided image editing space, against several state-of-the-art editing models, and validate our automatic metrics against human ratings. Results show that GPT-Image-1 leads in instruction-following accuracy, but often over-modifies irrelevant image regions, highlighting a key trade-off in the current model behavior. GIE-Bench provides a scalable, reproducible framework for advancing more accurate evaluation of text-guided image editing.
Recent advances in vision language models (VLMs) have enabled broad progress in the general medical field. However, pathology still remains a more challenging subdomain, with current pathology specific VLMs exhibiting limitations in both diagnostic accuracy and reasoning plausibility. Such shortcomings are largely attributable to the nature of current pathology datasets, which are primarily composed of image description pairs that lack the depth and structured diagnostic paradigms employed by real world pathologists. In this study, we leverage pathology textbooks and real world pathology experts to construct high-quality, reasoning-oriented datasets. Building on this, we introduce Patho-R1, a multimodal RL-based pathology Reasoner, trained through a three-stage pipeline: (1) continued pretraining on 3.5 million image-text pairs for knowledge infusion; (2) supervised fine-tuning on 500k high-quality Chain-of-Thought samples for reasoning incentivizing; (3) reinforcement learning using Group Relative Policy Optimization and Decoupled Clip and Dynamic sAmpling Policy Optimization strategies for multimodal reasoning quality refinement. To further assess the alignment quality of our dataset, we propose PathoCLIP, trained on the same figure-caption corpus used for continued pretraining. Comprehensive experimental results demonstrate that both PathoCLIP and Patho-R1 achieve robust performance across a wide range of pathology-related tasks, including zero-shot classification, cross-modal retrieval, Visual Question Answering, and Multiple Choice Question. Our project is available at the Patho-R1 repository: https://github.com/Wenchuan-Zhang/Patho-R1.
Plain language summaries (PLSs) are essential for facilitating effective communication between clinicians and patients by making complex medical information easier for laypeople to understand and act upon. Large language models (LLMs) have recently shown promise in automating PLS generation, but their effectiveness in supporting health information comprehension remains unclear. Prior evaluations have generally relied on automated scores that do not measure understandability directly, or subjective Likert-scale ratings from convenience samples with limited generalizability. To address these gaps, we conducted a large-scale crowdsourced evaluation of LLM-generated PLSs using Amazon Mechanical Turk with 150 participants. We assessed PLS quality through subjective Likert-scale ratings focusing on simplicity, informativeness, coherence, and faithfulness; and objective multiple-choice comprehension and recall measures of reader understanding. Additionally, we examined the alignment between 10 automated evaluation metrics and human judgments. Our findings indicate that while LLMs can generate PLSs that appear indistinguishable from human-written ones in subjective evaluations, human-written PLSs lead to significantly better comprehension. Furthermore, automated evaluation metrics fail to reflect human judgment, calling into question their suitability for evaluating PLSs. This is the first study to systematically evaluate LLM-generated PLSs based on both reader preferences and comprehension outcomes. Our findings highlight the need for evaluation frameworks that move beyond surface-level quality and for generation methods that explicitly optimize for layperson comprehension.
Each year, tens of millions of essays are written and graded in college-level English courses. Students are asked to analyze literary and cultural texts through a process known as close reading, in which they gather textual details to formulate evidence-based arguments. Despite being viewed as a basis for critical thinking and widely adopted as a required element of university coursework, close reading has never been evaluated on large language models (LLMs), and multi-discipline benchmarks like MMLU do not include literature as a subject. To fill this gap, we present KRISTEVA, the first close reading benchmark for evaluating interpretive reasoning, consisting of 1331 multiple-choice questions adapted from classroom data. With KRISTEVA, we propose three progressively more difficult sets of tasks to approximate different elements of the close reading process, which we use to test how well LLMs may seem to understand and reason about literary works: 1) extracting stylistic features, 2) retrieving relevant contextual information from parametric knowledge, and 3) multi-hop reasoning between style and external contexts. Our baseline results find that, while state-of-the-art LLMs possess some college-level close reading competency (accuracy 49.7% - 69.7%), their performances still trail those of experienced human evaluators on 10 out of our 11 tasks.
Long chain-of-thought (CoT) is an essential ingredient in effective usage of modern large language models, but our understanding of the reasoning strategies underlying these capabilities remains limited. While some prior works have attempted to categorize CoTs using predefined strategy types, such approaches are constrained by human intuition and fail to capture the full diversity of model behaviors. In this work, we introduce the CoT Encyclopedia, a bottom-up framework for analyzing and steering model reasoning. Our method automatically extracts diverse reasoning criteria from model-generated CoTs, embeds them into a semantic space, clusters them into representative categories, and derives contrastive rubrics to interpret reasoning behavior. Human evaluations show that this framework produces more interpretable and comprehensive analyses than existing methods. Moreover, we demonstrate that this understanding enables performance gains: we can predict which strategy a model is likely to use and guide it toward more effective alternatives. Finally, we provide practical insights, such as that training data format (e.g., free-form vs. multiple-choice) has a far greater impact on reasoning behavior than data domain, underscoring the importance of format-aware model design.
Large Language Models (LLMs) show growing promise in autonomous driving by reasoning over complex traffic scenarios to generate path plans. However, their tendencies toward overconfidence, and hallucinations raise critical safety concerns. We introduce SafePath, a modular framework that augments LLM-based path planning with formal safety guarantees using conformal prediction. SafePath operates in three stages. In the first stage, we use an LLM that generates a set of diverse candidate paths, exploring possible trajectories based on agent behaviors and environmental cues. In the second stage, SafePath filters out high-risk trajectories while guaranteeing that at least one safe option is included with a user-defined probability, through a multiple-choice question-answering formulation that integrates conformal prediction. In the final stage, our approach selects the path with the lowest expected collision risk when uncertainty is low or delegates control to a human when uncertainty is high. We theoretically prove that SafePath guarantees a safe trajectory with a user-defined probability, and we show how its human delegation rate can be tuned to balance autonomy and safety. Extensive experiments on nuScenes and Highway-env show that SafePath reduces planning uncertainty by 77\% and collision rates by up to 70\%, demonstrating effectiveness in making LLM-driven path planning more safer.
Language Models (LMs) continue to advance, improving response quality and coherence. Given Internet-scale training datasets, LMs have likely encountered much of what users may ask them to generate in some form during their training. A plethora of evaluation benchmarks have been constructed to assess model quality, response appropriateness, and reasoning capabilities. However, the human effort required for benchmark construction is rapidly being outpaced by the size and scope of the models under evaluation. Having humans build a benchmark for every possible domain of interest is impractical. Therefore, we propose a methodology for automating the construction of fact-based synthetic data model evaluations grounded in document populations. This work leverages the same LMs to evaluate domain-specific knowledge automatically, using only grounding documents (e.g., a textbook) as input. This synthetic data benchmarking approach corresponds well with human curated questions producing a Spearman ranking correlation of 0.97 and a benchmark evaluation Pearson accuracy correlation of 0.75. This novel approach supports generating both multiple choice and open-ended synthetic data questions to gain diagnostic insight of LM capability. We apply this methodology to evaluate model performance on two recent arXiv preprints, discovering a surprisingly strong performance from Gemma-3 models on open-ended questions. Code is available at https://github.com/mmajurski/grounded-synth-lm-benchmark
We present HealthBench, an open-source benchmark measuring the performance and safety of large language models in healthcare. HealthBench consists of 5,000 multi-turn conversations between a model and an individual user or healthcare professional. Responses are evaluated using conversation-specific rubrics created by 262 physicians. Unlike previous multiple-choice or short-answer benchmarks, HealthBench enables realistic, open-ended evaluation through 48,562 unique rubric criteria spanning several health contexts (e.g., emergencies, transforming clinical data, global health) and behavioral dimensions (e.g., accuracy, instruction following, communication). HealthBench performance over the last two years reflects steady initial progress (compare GPT-3.5 Turbo's 16% to GPT-4o's 32%) and more rapid recent improvements (o3 scores 60%). Smaller models have especially improved: GPT-4.1 nano outperforms GPT-4o and is 25 times cheaper. We additionally release two HealthBench variations: HealthBench Consensus, which includes 34 particularly important dimensions of model behavior validated via physician consensus, and HealthBench Hard, where the current top score is 32%. We hope that HealthBench grounds progress towards model development and applications that benefit human health.
Despite recent advances in video understanding, the capabilities of Large Video Language Models (LVLMs) to perform video-based causal reasoning remains underexplored, largely due to the absence of relevant and dedicated benchmarks for evaluating causal reasoning in visually grounded and goal-driven settings. To fill this gap, we introduce a novel benchmark named Video-based long-form Causal Reasoning (VCRBench). We create VCRBench using procedural videos of simple everyday activities, where the steps are deliberately shuffled with each clip capturing a key causal event, to test whether LVLMs can identify, reason about, and correctly sequence the events needed to accomplish a specific goal. Moreover, the benchmark is carefully designed to prevent LVLMs from exploiting linguistic shortcuts, as seen in multiple-choice or binary QA formats, while also avoiding the challenges associated with evaluating open-ended QA. Our evaluation of state-of-the-art LVLMs on VCRBench suggests that these models struggle with video-based long-form causal reasoning, primarily due to their difficulty in modeling long-range causal dependencies directly from visual observations. As a simple step toward enabling such capabilities, we propose Recognition-Reasoning Decomposition (RRD), a modular approach that breaks video-based causal reasoning into two sub-tasks of video recognition and causal reasoning. Our experiments on VCRBench show that RRD significantly boosts accuracy on VCRBench, with gains of up to 25.2%. Finally, our thorough analysis reveals interesting insights, for instance, that LVLMs primarily rely on language knowledge for complex video-based long-form causal reasoning tasks.
The rise of large language models and multi-agent systems has sparked growing interest in AI scientists capable of autonomous biological research. However, existing benchmarks either focus on reasoning without data or on data analysis with predefined statistical answers, lacking realistic, data-driven evaluation settings. Here, we introduce the Biological AI Scientist Benchmark (BaisBench), a benchmark designed to assess AI scientists' ability to generate biological discoveries through data analysis and reasoning with external knowledge. BaisBench comprises two tasks: cell type annotation on 31 expert-labeled single-cell datasets, and scientific discovery through answering 198 multiple-choice questions derived from the biological insights of 41 recent single-cell studies. Systematic experiments on state-of-the-art AI scientists and LLM agents showed that while promising, current models still substantially underperform human experts on both tasks. We hope BaisBench will fill this gap and serve as a foundation for advancing and evaluating AI models for scientific discovery. The benchmark can be found at: https://github.com/EperLuo/BaisBench.
Large Language Models are versatile, general-purpose tools with a wide range of applications. Recently, the advent of "reasoning models" has led to substantial improvements in their abilities in advanced problem-solving domains such as mathematics and software engineering. In this work, we assessed the ability of reasoning models to directly perform chemistry tasks, without any assistance from external tools. We created a novel benchmark, called ChemIQ, which consists of 796 questions assessing core concepts in organic chemistry, focused on molecular comprehension and chemical reasoning. Unlike previous benchmarks, which primarily use multiple choice formats, our approach requires models to construct short-answer responses, more closely reflecting real-world applications. The reasoning models, exemplified by OpenAI's o3-mini, correctly answered 28%-59% of questions depending on the reasoning level used, with higher reasoning levels significantly increasing performance on all tasks. These models substantially outperformed the non-reasoning model, GPT-4o, which achieved only 7% accuracy. We found that Large Language Models can now convert SMILES strings to IUPAC names, a task earlier models were unable to perform. Additionally, we show that the latest reasoning models can elucidate structures from 1H and 13C NMR data, correctly generating SMILES strings for 74% of molecules containing up to 10 heavy atoms, and in one case solving a structure comprising 21 heavy atoms. For each task, we found evidence that the reasoning process mirrors that of a human chemist. Our results demonstrate that the latest reasoning models have the ability to perform advanced chemical reasoning.
Large Language Models (LLMs) excel in natural language tasks, but less is known about their reasoning capabilities over tabular data. Prior analyses devise evaluation strategies that poorly reflect an LLM's realistic performance on tabular queries. Moreover, we have a limited understanding of the robustness of LLMs towards realistic variations in tabular inputs. Therefore, we ask: Can general-purpose LLMs reason over tabular data, really?, and focus on two questions 1) are tabular reasoning capabilities of general-purpose LLMs robust to real-world characteristics of tabular inputs, and 2) how can we realistically evaluate an LLM's performance on analytical tabular queries? Building on a recent tabular reasoning benchmark, we first surface shortcomings of its multiple-choice prompt evaluation strategy, as well as commonly used free-form text metrics such as SacreBleu and BERT-score. We show that an LLM-as-a-judge procedure yields more reliable performance insights and unveil a significant deficit in tabular reasoning performance of LLMs. We then extend the tabular inputs reflecting three common characteristics in practice: 1) missing values, 2) duplicate entities, and 3) structural variations. Experiments show that the tabular reasoning capabilities of general-purpose LLMs suffer from these variations, stressing the importance of improving their robustness for realistic tabular inputs.
The primary goal of this study is to develop and evaluate an innovative prompting technique, AnaQuest, for generating multiple-choice questions (MCQs) using a pre-trained large language model. In AnaQuest, the choice items are sentence-level assertions about complex concepts. The technique integrates formative and summative assessments. In the formative phase, students answer open-ended questions for target concepts in free text. For summative assessment, AnaQuest analyzes these responses to generate both correct and incorrect assertions. To evaluate the validity of the generated MCQs, Item Response Theory (IRT) was applied to compare item characteristics between MCQs generated by AnaQuest, a baseline ChatGPT prompt, and human-crafted items. An empirical study found that expert instructors rated MCQs generated by both AI models to be as valid as those created by human instructors. However, IRT-based analysis revealed that AnaQuest-generated questions - particularly those with incorrect assertions (foils) - more closely resembled human-crafted items in terms of difficulty and discrimination than those produced by ChatGPT.
As Large Language Models (LLMs) become widely accessible, a detailed understanding of their knowledge within specific domains becomes necessary for successful real world use. This is particularly critical in public health, where failure to retrieve relevant, accurate, and current information could significantly impact UK residents. However, currently little is known about LLM knowledge of UK Government public health information. To address this issue, this paper introduces a new benchmark, PubHealthBench, with over 8000 questions for evaluating LLMs' Multiple Choice Question Answering (MCQA) and free form responses to public health queries. To create PubHealthBench we extract free text from 687 current UK government guidance documents and implement an automated pipeline for generating MCQA samples. Assessing 24 LLMs on PubHealthBench we find the latest private LLMs (GPT-4.5, GPT-4.1 and o1) have a high degree of knowledge, achieving >90% accuracy in the MCQA setup, and outperform humans with cursory search engine use. However, in the free form setup we see lower performance with no model scoring >75%. Importantly we find in both setups LLMs have higher accuracy on guidance intended for the general public. Therefore, there are promising signs that state of the art (SOTA) LLMs are an increasingly accurate source of public health information, but additional safeguards or tools may still be needed when providing free form responses on public health topics.