We study whether transformers can learn to reason over parametric knowledge, a skill that even the most capable language models struggle with. Focusing on two representative reasoning types, composition and comparison, we consistently find that transformers learn implicit reasoning, but only through, i.e., extended training far beyond overfitting. The levels of generalization also vary across reasoning types: when faced with out-of-distribution examples, transformers fail to systematically generalize for composition but succeed for comparison. We delve into the model's internals throughout training, conducting analytical experiments that reveal: 1) the mechanism behind grokking, such as the formation of the generalizing circuit and its relation to the relative efficiency of generalizing and memorizing circuits, and 2) the connection between systematicity and the configuration of the generalizing circuit. Our findings guide data and training setup to better induce implicit reasoning and suggest potential improvements to the transformer architecture, such as encouraging cross-layer knowledge sharing. Furthermore, we demonstrate that for a challenging reasoning task with a large search space, GPT-4-Turbo and Gemini-1.5-Pro

Recent advances in Large Multimodal Models (LMMs) have revolutionized their reasoning and Optical Character Recognition (OCR) capabilities. However, their complex logical reasoning performance on text-rich images remains underexplored. To bridge this gap, we introduce LogicOCR, a benchmark comprising 2780 questions with two subsets, i.e., LogicOCR-Gen with 1100 multi-choice questions on generated images, and LogicOCR-Real with 1680 meticulously designed free-form questions on real-world images. For constructing LogicOCR-Gen, we first curate a text corpus from the Chinese National Civil Servant Examination, and customize an automatic pipeline to steer GPT-Image-1 to generate images with varied layouts and fonts, ensuring contextual relevance and visual realism. Then, the generated images are manually verified. We evaluate a range of representative LMMs under 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. Moreover, we propose TextCue, a training-free method that enhances LMMs' perception of image regions containing important text cues for solving questions. We leverage LMMs' attention maps and an off-the-shelf text segmentation specialist to determine the region, which is then cropped and enlarged to augment the original image. Experiments show its effectiveness, e.g., a 1.8% accuracy gain over LLaVA-OV-1.5-8B under the CoT setting. Our benchmark is available at https://github.com/MiliLab/LogicOCR.

The rapid rise of Large Language Models (LLMs) and Large Reasoning Models (LRMs) has been accompanied by an equally rapid increase of benchmarks used to assess them. However, due to both improved model competence resulting from scaling and novel training advances as well as likely many of these datasets being included in pre or post training data, results become saturated, driving a continuous need for new and more challenging replacements. In this paper, we discuss whether surpassing a benchmark truly demonstrates reasoning ability or are we simply tracking numbers divorced from the capabilities we claim to measure? We present an investigation focused on three model families, OpenAI, Anthropic, and Google, and how their reasoning capabilities across different benchmarks evolve over the years. We also analyze performance trends over the years across different reasoning tasks and discuss the current situation of benchmarking and remaining challenges. By offering a comprehensive overview of benchmarks and reasoning tasks, our work aims to serve as a first reference to ground future research in reasoning evaluation and model development.

Rapid advances in multimodal models demand benchmarks that rigorously evaluate understanding and reasoning in safety-critical, dynamic real-world settings. We present AccidentBench, a large-scale benchmark that combines vehicle accident scenarios with Beyond domains, safety-critical settings in air and water that emphasize spatial and temporal reasoning (e.g., navigation, orientation, multi-vehicle motion). The benchmark contains approximately 2000 videos and over 19000 human-annotated question--answer pairs spanning multiple video lengths (short/medium/long) and difficulty levels (easy/medium/hard). Tasks systematically probe core capabilities: temporal, spatial, and intent understanding and reasoning. By unifying accident-centric traffic scenes with broader safety-critical scenarios in air and water, AccidentBench offers a comprehensive, physically grounded testbed for evaluating models under real-world variability. Evaluations of state-of-the-art models (e.g., Gemini-2.5 Pro and GPT-5) show that even the strongest models achieve only about 18% accuracy on the hardest tasks and longest videos, revealing substantial gaps in real-world temporal, spatial, and intent reasoning. AccidentBench is designed to expose these critical gaps and drive the development of multimodal models that are safer, more robust, and better aligned with real-world safety-critical challenges. The code and dataset are available at: https://github.com/SafeRL-Lab/AccidentBench

This paper introduces TurnaboutLLM, a novel framework and dataset for evaluating the deductive reasoning abilities of Large Language Models (LLMs) by leveraging the interactive gameplay of detective games Ace Attorney and Danganronpa. The framework tasks LLMs with identifying contradictions between testimonies and evidences within long narrative contexts, a challenging task due to the large answer space and diverse reasoning types presented by its questions. We evaluate twelve state-of-the-art LLMs on the dataset, hinting at limitations of popular strategies for enhancing deductive reasoning such as extensive thinking and Chain-of-Thought prompting. The results also suggest varying effects of context size, the number of reasoning step and answer space size on model performance. Overall, TurnaboutLLM presents a substantial challenge for LLMs' deductive reasoning abilities in complex, narrative-rich environments.

Empirical evidence shows that human behaviour often deviates from game-theoretical rationality. For instance, humans may hold unrealistic expectations about future outcomes. As the evolutionary roots of such biases remain unclear, we investigate here how reasoning abilities and cognitive biases co-evolve using Evolutionary Game Theory. In our model, individuals in a population deploy a variety of unbiased and biased level-k reasoning strategies to anticipate others' behaviour in sequential interactions, represented by the Incremental Centipede Game. Positively biased reasoning strategies have a systematic inference bias towards higher but uncertain rewards, while negatively biased strategies reflect the opposite tendency. We find that selection consistently favours positively biased reasoning, with rational behaviour even going extinct. This bias co-evolves with bounded rationality, as the reasoning depth remains limited in the population. Interestingly, positively biased agents may co-exist with non-reasoning agents, thus pointing to a novel equilibrium. Longer games further promote positively biased reasoning, as they can lead to higher future rewards. The biased reasoning strategies proposed in this model may reflect cognitive phenomena like wishful thinking and defensive pessimism. This work therefore supports the claim that certain cognitive biases, despite deviating from rational judgment, constitute an adaptive feature to better cope with social dilemmas.

Large audio-language models (LALMs) have achieved near-human performance in sentence-level transcription and emotion recognition. However, existing evaluations focus mainly on surface-level perception, leaving the capacity of models for contextual and inference-driven reasoning in speech-based scenarios insufficiently examined. To address this gap, we introduce SpeechR, a unified benchmark for evaluating reasoning over speech in large audio-language models. SpeechR evaluates models along three key dimensions: factual retrieval, procedural inference, and normative judgment. It includes three distinct evaluation formats. The multiple-choice version measures answer selection accuracy. The generative version assesses the coherence and logical consistency of reasoning chains. The acoustic-feature version investigates whether variations in stress and emotion affect reasoning performance. Evaluations on eleven state-of-the-art LALMs reveal that high transcription accuracy does not translate into strong reasoning capabilities. SpeechR establishes a structured benchmark for evaluating reasoning in spoken language, enabling more targeted analysis of model capabilities across diverse dialogue-based tasks.

Recent works on large language models (LLMs) have demonstrated the impact of prompting strategies and fine-tuning techniques on their reasoning capabilities. Yet, their effectiveness on clinical natural language inference (NLI) remains underexplored. This study presents the first controlled evaluation of how prompt structure and efficient fine-tuning jointly shape model performance in clinical NLI. We inspect four classes of prompting strategies to elicit reasoning in LLMs at different levels of abstraction, and evaluate their impact on a range of clinically motivated reasoning types. For each prompting strategy, we construct high-quality demonstrations using a frontier model to distil multi-step reasoning capabilities into smaller models (4B parameters) via Low-Rank Adaptation (LoRA). Across different language models fine-tuned on the NLI4CT benchmark, we found that prompt type alone accounts for up to 44% of the variance in macro-F1. Moreover, LoRA fine-tuning yields consistent gains of +8 to 12 F1, raises output alignment above 97%, and narrows the performance gap to GPT-4o-mini to within 7.1%. Additional experiments on reasoning generalisation reveal that LoRA improves performance in 75% of the models on MedNLI and TREC Clinical Trials Track. Overall, these findings demonstrate that (i) prompt structure is a primary driver of clinical reasoning performance, (ii) compact models equipped with strong prompts and LoRA can rival frontier-scale systems, and (iii) reasoning-type-aware evaluation is essential to uncover prompt-induced trade-offs. Our results highlight the promise of combining prompt design and lightweight adaptation for more efficient and trustworthy clinical NLP systems, providing insights on the strengths and limitations of widely adopted prompting and parameter-efficient techniques in highly specialised domains.

Existing benchmarks fail to capture a crucial aspect of intelligence: physical reasoning, the integrated ability to combine domain knowledge, symbolic reasoning, and understanding of real-world constraints. To address this gap, we introduce PhyX: the first large-scale benchmark designed to assess models capacity for physics-grounded reasoning in visual scenarios. PhyX includes 3K meticulously curated multimodal questions spanning 6 reasoning types across 25 sub-domains and 6 core physics domains: thermodynamics, electromagnetism, mechanics, modern physics, optics, and wave\&acoustics. In our comprehensive evaluation, even state-of-the-art models struggle significantly with physical reasoning. GPT-4o, Claude3.7-Sonnet, and GPT-o4-mini achieve only 32.5%, 42.2%, and 45.8% accuracy respectively-performance gaps exceeding 29% compared to human experts. Our analysis exposes critical limitations in current models: over-reliance on memorized disciplinary knowledge, excessive dependence on mathematical formulations, and surface-level visual pattern matching rather than genuine physical understanding. We provide in-depth analysis through fine-grained statistics, detailed case studies, and multiple evaluation paradigms to thoroughly examine physical reasoning capabilities. To ensure reproducibility, we implement a compatible evaluation protocol based on widely-used toolkits such as VLMEvalKit, enabling one-click evaluation. More details are available on our project page: https://phyx-bench.github.io/.

Large language models (LLMs) have shown impressive capabilities in handling complex tasks through long-chain reasoning. However, the extensive reasoning steps involved can significantly increase computational costs, posing challenges for real-world deployment. Recent efforts have focused on optimizing reasoning efficiency by shortening the Chain-of-Thought (CoT) reasoning processes through various approaches, such as length-aware prompt engineering, supervised fine-tuning on CoT data with variable lengths, and reinforcement learning with length penalties. Although these methods effectively reduce reasoning length, they still necessitate an initial reasoning phase. More recent approaches have attempted to integrate long-chain and short-chain reasoning abilities into a single model, yet they still rely on manual control to toggle between short and long CoT. In this work, we propose a novel approach that autonomously switches between short and long reasoning chains based on problem complexity. Our method begins with supervised fine-tuning of the base model to equip both long-chain and short-chain reasoning abilities. We then employ reinforcement learning to further balance short and long CoT generation while maintaining accuracy through two key strategies: first, integrating reinforcement learning with a long-short adaptive group-wise reward strategy to assess prompt complexity and provide corresponding rewards; second, implementing a logit-based reasoning mode switching loss to optimize the model's initial token choice, thereby guiding the selection of the reasoning type. Evaluations on mathematical datasets demonstrate that our model can dynamically switch between long-chain and short-chain reasoning modes without substantially sacrificing performance. This advancement enhances the practicality of reasoning in large language models for real-world applications.