Utility of Text-Only Chain-of-Thought for Human Verification

Determine whether text-only chain-of-thought explanations generated by large language models improve end users’ (students, teachers, and non-expert verifiers) ability to understand and verify mathematical reasoning.

Background

LLMs frequently produce chain-of-thought (CoT) explanations to show intermediate reasoning steps on math and reasoning tasks. Although CoT improves model performance on benchmarks, its presentation is typically long and text-heavy, which may impose cognitive load on end users and impede error detection.

The paper motivates interactive alternatives to CoT by noting that it is not established whether standard text-only CoT outputs actually help human users comprehend and verify mathematical reasoning, particularly in educational contexts where clarity and error checking are essential.