前瞻式蒸馏：验证教师建议对未来轨迹的影响

A smaller AI (student) learns better when it checks whether a bigger AI’s (Teacher) advice actually helps later.

The system does not blindly copy the teacher; it watches what the student does next.

A correction counts only when it improves the smaller AI’s next several decisions in practice.

On-policy distillation trains a smaller agent from teacher feedback on situations the student actually reaches, rather than only from clean examples.

The problem is that early mistakes change later situations, so a teacher action that looks better now can still send the student somewhere worse.

FutureBridge-OPD finds the student turn with the biggest teacher disagreement, replaces that action with a short teacher bridge, and keeps the original path for comparison.

The same frozen student continues from both branches for a few turns, creating a direct test of what the correction causes next.

The method keeps the teacher action only when its branch contains a larger share of future choices that the teacher considers better than the student's baseline.

– arxiv. org/abs/2608.01953

Title: "Look Ahead Before You Distill: Future Trajectory Validation of Teacher Guidance for Agentic On-Policy Distillation"