SPOT: Sparse Probing and Outcome Calibration for On-Policy Distillation
SPOT chooses only the most useful token positions to probe, then distills targets based on downstream outcomes.
The paper argues that standard on-policy distillation can underweight plausible continuations because teacher probabilities do not show whether uncertainty is useful or just spread across a long tail. SPOT scores positions using teacher entropy, top-k mass, and student-teacher mismatch, then spends its probing budget there. It tests teacher-proposed candidates through verifier-scored student continuations and builds a KL-regularized target that stays anchored to the teacher while favoring candidates that lead to better results. The authors report gains across multiple student models and reasoning benchmarks. HF Daily Papers' note
The paper argues that standard on-policy distillation can underweight plausible continuations because teacher probabilities do not show whether uncertainty is useful or just spread across a long tail. SPOT scores positions using teacher entropy, top-k mass, and student-teacher mismatch, then spends its probing budget there. It tests teacher-proposed candidates through verifier-scored student continuations and builds a KL-regularized target that stays anchored to the teacher while favoring candidates that lead to better results. The authors report gains across multiple student models and reasoning benchmarks. HF Daily Papers' note
score 5