Predicting Future Behaviors in Reasoning Models Enables Better Steering

arXiv·medium signal

Kortukov, Komorowski, and Klein show that test-time steering of large reasoning models (LRMs) works substantially better when the intervention is informed by first predicting the model's own future behavior from its hidden representations. For agent builders, this is a direct lever on controlling unpredictable reasoning-model behavior at inference time without retraining. It reframes steering as a predict-then-intervene loop rather than a blind activation edit.