DASH: Divergence-Adaptive Supervision Horizons for On-Policy Self-Distillation of Reasoning Models
DASH changes how self-distillation weights each generated token by looking at the rollout’s divergence history.
The paper says standard OPSD treats local teacher-student divergences too uniformly, missing how mismatch evolves during autoregressive generation. DASH turns each local divergence’s gap from the sequence mean into a gate for backward multi-step aggregation. Across three math reasoning benchmarks and three model scales, it beat matched vanilla OPSD reruns in every reported setting. The authors say it uses distributions OPSD already computes, adding no extra teacher or student forward pass. ArXiv · AI/CL/LG's note
The paper says standard OPSD treats local teacher-student divergences too uniformly, missing how mismatch evolves during autoregressive generation. DASH turns each local divergence’s gap from the sequence mean into a gate for backward multi-step aggregation. Across three math reasoning benchmarks and three model scales, it beat matched vanilla OPSD reruns in every reported setting. The authors say it uses distributions OPSD already computes, adding no extra teacher or student forward pass. ArXiv · AI/CL/LG's note
score 5