Correct Answers, Invalid Traces: What Verifiable Grade-School Math Reveals About Chain-of-Thought Traces
Correct answers can arrive with traces that fail the benchmark’s own dependency checks.
The paper tests chain-of-thought traces on iGSM, where each needed quantity and dependency can be verified step by step. On the hardest out-of-distribution cases, 31.6% of correct answers came with invalid traces. More than half of those invalid traces still passed syntactic and arithmetic checks, failing only on semantic dependencies. The authors also found that corrupted or swapped trace supervision could preserve high answer accuracy while destroying trace verifiability. ArXiv · AI/CL/LG's note
The paper tests chain-of-thought traces on iGSM, where each needed quantity and dependency can be verified step by step. On the hardest out-of-distribution cases, 31.6% of correct answers came with invalid traces. More than half of those invalid traces still passed syntactic and arithmetic checks, failing only on semantic dependencies. The authors also found that corrupted or swapped trace supervision could preserve high answer accuracy while destroying trace verifiability. ArXiv · AI/CL/LG's note
score 5