LLMs’ Logical Reasoning: A New Framework for Detailed Evaluation

FineLogic assesses logical reasoning in large language models across accuracy, stepwise soundness and representation alignment. Research demonstrates supervised fine-tuning enhances generalisation, with symbolic reasoning promoting structurally sound inference. Improvements stem from step-by-step generation, not shortcut learning or internalised correctness.

The capacity for logical reasoning remains a critical, yet incompletely understood, attribute of large language models (LLMs). Current evaluation metrics frequently focus on the correctness of a final answer, overlooking the integrity of the process by which it is reached. A collaborative study, detailed in ‘Dissecting Logical Reasoning in LLMs: A Fine-Grained Evaluation and Supervision Study’, introduces a novel framework – FineLogic – to assess reasoning across accuracy, stepwise validity, and internal representation alignment. Researchers Yujun Zhou, Jiayi Ye, Zipeng Ling, Yufei Han, Yue Huang, Haomin Zhuang, Zhenwen Liang, Kehan Guo, Taicheng Guo, Xiangqi Wang, and Xiangliang Zhang, representing institutions including the University of Notre Dame, MBZUAI, the University of Pennsylvania, and INRIA, present both the evaluation framework and an analysis of how different training supervision methods impact the development of logical reasoning capabilities within LLMs. Their work demonstrates that focused supervision enhances generalisation and promotes more structurally sound inference.

FineLogic: A Multifaceted Evaluation of Logical Reasoning in Large Language Models

Current evaluation of large language models (LLMs) frequently relies on overall accuracy, offering limited insight into how these models arrive at their conclusions. A new framework, FineLogic, provides a detailed assessment of logical reasoning, moving beyond simple performance metrics to analyse the reasoning process itself. This approach assesses reasoning across three key dimensions: overall benchmark performance, the logical consistency of each step (stepwise soundness), and the alignment of the model’s internal representations with logical principles (representation-level alignment).

Experiments demonstrate that supervised fine-tuning consistently improves generalisation performance in LLMs, even when applied to tasks differing from the training data and those requiring processing of extended contexts. Crucially, the study highlights the benefits of employing symbolic supervision styles during fine-tuning. These styles encourage the development of more structurally sound and atomic inference chains – prompting models to break down problems into smaller, logically connected steps. This fosters a more transparent and interpretable reasoning process.

Analysis of the model’s internal representations reveals that fine-tuning primarily enhances reasoning by improving the step-by-step generation process. This suggests a shift from pattern recognition towards genuine reasoning, with models demonstrably improving their ability to construct logically coherent chains. This indicates active engagement in deductive reasoning, rather than reliance on memorised patterns or statistical correlations.

Traditional benchmarks often fail to capture the nuances of logical reasoning, making it difficult to accurately assess LLM capabilities. FineLogic addresses these limitations with its three-dimensional assessment. Benchmark performance measures overall accuracy, while stepwise soundness evaluates the logical validity of each individual step. Representation alignment assesses how well the model’s internal numerical representations reflect the underlying logical structure of the problem.

Experiments involved training LLMs using four distinct supervision styles, varying the presentation of symbolic information to investigate its impact on reasoning ability. The results confirm that training with symbolic reasoning styles encourages the development of more structurally sound and atomic inference chains.

The analysis extends beyond behavioural performance to examine the internal mechanisms driving improved reasoning. Fine-tuning primarily enhances reasoning by improving the model’s ability to generate step-by-step solutions, suggesting a fundamental shift in cognitive processes. Researchers employed representation-level probing – analysing the model’s internal numerical representations – to determine how well they correspond to the logical structure of the problem. This technique identified specific patterns suggesting the model is indeed engaging in deductive thought.

This work establishes a rigorous methodology for evaluating logical reasoning in LLMs and provides insights into the impact of different training strategies. By focusing on the quality and structure of reasoning, rather than solely on final answers, FineLogic facilitates a more interpretable assessment of model capabilities. The findings underscore the importance of step-by-step generation as a key mechanism for improving reasoning performance and suggest that symbolic supervision can promote the development of more robust and logically sound inference chains.

Future work will investigate the transferability of these findings to different model architectures and reasoning tasks. Exploring the impact of varying levels of symbolic supervision could further refine the training process and optimise reasoning performance. Additionally, investigating the potential for combining symbolic supervision with other training techniques, such as reinforcement learning, could lead to even more powerful reasoning systems.

Researchers also plan to use FineLogic to diagnose and address specific weaknesses in model reasoning abilities, developing targeted interventions to improve performance on challenging tasks. Ultimately, the goal is to create AI systems that can not only solve complex problems but also explain their reasoning in a clear and understandable way, fostering trust and transparency in AI decision-making.