Dissecting Logical Reasoning in LLMs: A Fine-Grained Evaluation and Supervision Study

TL;DR

This paper introduces FineLogic, a detailed evaluation framework for logical reasoning in LLMs, revealing how different supervision styles influence reasoning quality and process, and providing insights for improving LLM reasoning capabilities.

Contribution

The paper proposes a novel fine-grained evaluation framework and analyzes the impact of various supervision formats on LLM reasoning abilities.

Findings

Natural language supervision improves out-of-distribution generalization.

Symbolic supervision enhances structural soundness of reasoning steps.

Fine-tuning mainly refines step-by-step reasoning rather than answer convergence.

Abstract

Logical reasoning is a core capability for large language models (LLMs), yet existing benchmarks that rely solely on final-answer accuracy fail to capture the quality of the reasoning process. To address this, we introduce FineLogic, a fine-grained evaluation framework that assesses logical reasoning across three dimensions: overall accuracy, stepwise soundness, and representation-level probing. Leveraging this framework, we conduct a comprehensive study on how different supervision formats in fine-tuning shape reasoning abilities. We fine-tune LLMs on four supervision styles: one in natural language and three symbolic variants. We find a key trade-off: natural language supervision excels at generalization to out-of-distribution and long-chain problems, whereas symbolic supervision is superior at instilling structurally sound, atomic reasoning steps. Furthermore, our probing analysis indicates that fine-tuning primarily refines the model's step-by-step generation process, rather than improving its ability to converge on an answer early. Together, our framework and analysis provide a more rigorous lens for evaluating and improving logical reasoning in LLMs. The code is available at https://github.com/YujunZhou/FineLogic.