CoT Path Correctness Gap

A correct final answer does not imply a correct reasoning path — treat CoT traces as debugging evidence, not as verified proofs of intermediate steps.

Chain-of-thought prompting yields readable intermediate steps, but Wei et al. emphasize there is no guarantee those paths are correct even when the final answer matches gold. Paths can support right or wrong answers; factual faithfulness of generations remains an open problem. Kojima et al.’s CommonsenseQA sample study makes this concrete: among correct predictions, ~22% still had incorrect CoT (often failing to narrow to one choice while the first listed option matched gold). Wei et al. note accidental correct finals are rarer in free-response math than in multiple-choice / binary settings — choose eval formats carefully when auditing faithfulness. Among incorrects, traces can be logically coherent yet miss common-sense priors — see CoT failure modes.

For harness design, do not equate answer accuracy with verified reasoning — especially when using CoT traces as audit artifacts or as feedback into reflection and error correction / self-critique prompting.

Prefer independent oracles (gold-standard matching, functional completion testing, unit tests) over trusting the monologue. When intermediate steps must be reliable (agent plans, tool arguments), add explicit validation — plan-validate-execute, action format enforcement — rather than assuming CoT self-explains correctly. Self-consistency decoding can reduce reliance on any single path without proving faithfulness.

---

Sources

• Chain-of-Thought Prompting (Wei et al.), §6
• Large Language Models are Zero-Shot Reasoners (Kojima et al.), Appendix C.1