RISE: Recursive Improvement via Self-Extrapolating Policy Distillation

TL;DR - RISE is a language-model post-training method that builds a synthetic teacher by extrapolating from the model’s own RLVR training trajectory. It turns sparse outcome rewards into dense token-level supervision without external teachers or privileged conditioning.

• Extrapolation uses the change between the current checkpoint and a trailing anchor in parameter or output-logit space.
• RISE alternates RLVR with on-policy distillation: rewards ground reasoning improvements, while the synthetic teacher refines token-level decisions.
• The teacher is refreshed as the student improves, making distillation a recursive improvement process rather than one-shot compression.
• Across math, STEM, coding, and multi-turn agentic tasks, RISE reportedly outperforms RLVR-only training and on-policy self-distillation.