Boosting Large Models via Small Model RL

kastnerkyle · x · 2026-07-16

This post introduces **Direct-OPD**. The core issue is that while RLVR is powerful, repeating rollouts for every larger target model is too costly, as each model must independently rediscover learning signals from sparse rewards. The author proposes an alternative approach: 1. Perform RL on a **small model** first to leverage cheaper exploration and rollouts. 2. Use the **pre- and post-RL checkpoints** of the small model as a pair of teachers, utilizing their differential to transfer the "learned direction." 3. The goal isn't to distill the weak model itself, but to transfer the improvement trajectory formed during the RL process to a stronger student model. The post also notes this research comes from **SIA-Lab**, a joint lab between Tsinghua AIR and ByteDance Seed.

Original post →

More from Research

Research channel →