From Proof to Program: Characterizing Tool-Induced Reasoning Hallucinations in Large Language Models

TL;DR

This paper investigates how tool use in large language models can lead to reasoning failures, introducing the concept of Tool-Induced Myopia (TIM) and proposing a framework to improve reasoning with tools.

Contribution

It characterizes TIM as a new failure mode in tool-augmented language models and develops a framework to mitigate reasoning degradation caused by tool use.

Findings

Tool use increases answer accuracy but worsens reasoning coherence.

Models shift from arithmetic errors to reasoning failures with more tool use.

The proposed framework improves both accuracy and reasoning depth.

Abstract

Tool-augmented Language Models (TaLMs) can invoke external tools to solve problems beyond their parametric capacity. However, it remains unclear whether these tool-enabled gains reflect trustworthy reasoning. Focusing on the Code Interpreter tool, we show that even when tools are selected and executed correctly, TaLMs treat tool outputs as substitutes for reasoning, producing solutions that appear correct but lack coherent justification. We term this failure mode Tool-Induced Myopia (TIM), and study it using PYMATH, a benchmark of 1,679 competition-level mathematical problems for which Python code is helpful but not sufficient. We further develop a multi-dimensional evaluation suite to quantify reasoning degradation in TaLMs relative to their non-tool counterparts. Our findings reveal that while TaLMs achieve up to a 19.3 percentage point gain in final-answer accuracy, their reasoning behavior consistently deteriorates (e.g., non-tool LLMs win up to 41.5% more often in pairwise comparisons of the reasoning process). This degradation intensifies with tool use; the more frequently a model invokes tools, the less coherent its reasoning becomes. Moreover, tool use shifts errors from arithmetic mistakes toward global reasoning failures (logic, assumption, creativity); with TIM present in ~55% of high-risk cases. Finally, we propose a preference-optimization-based framework that realigns TaLMs to use tools as assistive evidence, improving both final-answer accuracy and reasoning depth under tool use. Codes and data are available at: https://github.com/megagonlabs/TIM.