Correct Answers, Invalid Traces: What Verifiable Grade-School Math Reveals About Chain-of-Thought Traces
On hard iGSM problems, 31.6% of correct answers have invalid chain-of-thought traces, undermining CoT monitoring.
Using the synthetic iGSM grade-school math benchmark, the authors check whether correct answers come with valid chain-of-thought traces. Out of distribution, answer correctness and trace validity diverge: on the hardest items, 31.6% of correct answers have invalid traces, and more than half of those pass syntactic and arithmetic checks but fail semantic dependency checks. Corrupted or swapped training traces can still yield high accuracy, which weakens CoT monitoring as evidence of genuine reasoning.
- iGSM exposes quantities and dependencies so traces can be checked programmatically.
- On the hardest instances, 31.6% of correct answers have invalid traces.
- Over half pass syntax and arithmetic but fail semantic dependency checks.
- Shuffling 10% of training-trace sentences still preserves near-clean accuracy.
- Results weaken using chain-of-thought traces for monitoring and auditing.
From the original article
1 min read at arxiv.orgChain-of-thought traces are widely read as records of how models reach their answers, informing debugging, agent auditing, and claims about reasoning. Testing this interpretation is difficult because natural-language thinking traces are rarely mechanically verifiable.
By Ratish Puduppully · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.