[Preprint] Reliable Control-Point Selection for Steering Reasoning in Large Language Models

Steering vectors can control reasoning behaviors in LLMs without retraining, but current methods detect behavioral boundaries through keyword matching—treating every match as genuine. We find that 93.3% of keyword-detected boundaries are behaviorally unstable. Our method, stability filtering, retains only boundaries where the model consistently reproduces the behavior, achieving 0.784 accuracy on MATH-500 (+5.0 over SEAL) with cross-model transfer.

Code: [GitHub].