Forecasting Future Behavior as a Learning Task

The paper introduces a novel approach to forecasting the behavior of large reasoning models (LRMs) by training Behavior Forecasters that utilize the reasoning trajectory directly, bypassing traditional explanation methods. The proposed method, evaluated on tasks predicting answer repetition and the impact of input removal, demonstrates superior accuracy compared to established models like GPT-5.4 and Claude Opus-4.6, while significantly reducing inference costs. This advancement highlights the potential for more efficient and interpretable AI systems by leveraging the inherent information within reasoning trajectories.