On Policy Reasoning Compression

Compress verbose reasoning by conditioning the same model on a conciseness instruction to create a teacher. Minimize KL divergence between student outputs and concise teacher without ground truth, enabling implicit reward learning that improves accuracy while reducing tokens.

adu2021 Updated 7 repo stars

File contents

adu2021/skillxiv commit cc49cfa40f

Frequently asked questions

npx get-skill add adu2021/on-policy-reasoning-compression