Weak-to-Strong Generalization via Direct On-Policy Distillation
Direct-OPD reuses a small model's RL run to improve a bigger one: the pre/post-RL log-ratio becomes a dense reward for the stronger student, lifting Qwen3-1.7B from 48.3% to 62.4% on AIME 2024 in 4 hours on 8 A100s.
**The gist** **Direct On-Policy Distillation (Direct-OPD)** runs RL with verifiable rewards on a small, cheap model, then treats the log-ratio between the post-RL teacher and its pre-RL reference as a dense implicit reward applied to the stronger student's own rollouts. It lifts **Qwen3-1.7B from 48.3% to 62.4% on AIME 2024** in **4 hours on 8 A100s** and beats step-matched direct RL.
**Why it matters** Rollout-heavy post-training is becoming the bottleneck as models scale; this shows RL gains can be reused across model sizes as a reward signal instead of repeated per model. Cheaper reasoning post-training feeds the pace at which better coding-agent models arrive, and the policy shifts **compose sequentially**.
**The gist** **Direct On-Policy Distillation (Direct-OPD)** runs RL with verifiable rewards on a small, cheap model, then treats the log-ratio between the post-RL teacher and its pre-RL reference as a dense implicit reward applied to the stronger student's own rollouts. It lifts **Qwen3-1.7B from 48.3% to 62.4% on AIME 2024** in **4 hours on 8 A100s** and beats step-matched direct RL. **Why it matters** Rollout-heavy post-training is becoming the bottleneck as models scale; this shows RL gains can be reused across model sizes as a reward signal instead of repeated per model. Cheaper reasoning post-training feeds the pace at which better coding-agent models arrive, and the policy shifts **compose sequentially**. **Watch out** Evidence sits on **small open models** and **math benchmarks** like AIME; whether the transfer holds at frontier scale or on long-horizon agentic tasks is untested, and a weak teacher's blind spots may still leak through the distilled signal.