Forecasting Future Behavior as a Learning Task

In three linesNew approach to predict large reasoning model (LRM) behavior without explanation methods. Authors train Behavior Forecasters on reasoning trajectories to forecast answer stability and input modification impact. Evaluation on three datasets: forecasters outperform GPT-5.4 and Claude Opus-4.6 at fraction of inference cost.

Summary generated by Claude — human-verified