Chemical Chain-of-Thought Functions as a Hallucination-Prone Molecular Scratchpad
Published in arXiv preprint, 2026
Recommended citation: Jiatong Li, Yuxuan Ren, Weida Wang, Xiaoyong Wei, and Yatao Bian. (2026). Chemical Chain-of-Thought Functions as a Hallucination-Prone Molecular Scratchpad. arXiv:2607.20935. https://arxiv.org/abs/2607.20935
Chemical reasoning language models are expected to derive molecular answers through faithful chain-of-thought (CoT). Across four reasoning model families and twelve chemistry tasks, however, hallucination is widespread and largely decoupled from answer correctness: correct answers often coexist with fabricated structural claims.
Attribution and perturbation analyses show that these traces can still act as causally useful molecular scratchpads. The results caution against treating CoT as direct evidence of faithful reasoning and motivate process-level supervision beyond answer-only evaluation.