Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces

In three linesStudy on long chain-of-thought traces used for LLM supervised fine-tuning. Researchers identify "harmful continuation": when reasoning continues after the answer is sufficiently supported. Removing these continuations improves fine-tuning outcomes. They propose HCC (Harmful Continuation Cut), a lightweight proxy to detect these boundaries.

Summary generated by Claude — human-verified