ReasonOps: Operator Segmentation for LLM Reasoning Traces

In three linesReasonOps is an unsupervised method for analyzing chain-of-thought traces from LLMs. It identifies 7 recurring reasoning operators (backtracking, inferring, hypothesizing) from 44,662 traces across 12 models on 8 benchmarks. These operators enable source model identification at 70-76% accuracy and predict answer correctness before trace completion.

Summary generated by Claude — human-verified