Direct-OPD Paper: 7B Model Boosts AIME Score by 6.4% Using 1.5B Policy Shift
_lewtun · x · 2026-08-12
This post shares the paper link for Direct On-Policy Distillation (Direct-OPD), developed by a joint lab of Tsinghua AIR and ByteDance Seed.
The text details the experimental setup: using pre- and post-RL checkpoints of a 1.5B model as a teacher pair to distill the policy shift into a 7B student model.
Results show that even though the 7B student initially scores higher (56.7) than the post-RL 1.5B teacher (51.3), vanilla distillation degrades the student. However, Direct-OPD successfully transfers the RL gains, pushing the 7B model's score to 63.1 on AIME24.
Related event: Tsinghua and ByteDance Propose Direct-OPD for Model Distillation(3 posts)→
More from Research
- Gemma 4 QAT Shows Significant Improvement in KV Cache Quantization, KLD Benchmarks Reveal — Anbeeld · 2026-08-12
- Recommended: 303-Page Comprehensive Survey on Code Models and Agents — mdancho84 · 2026-08-12
- Multi-Agent Debate Can Cut Hallucinations if Set Up Right — mdancho84 · 2026-08-12
- Code Models Are Insecure by Default; MoE is More Fragile Than Dense Models During SFT — mdancho84 · 2026-08-12
- Coding Scaling Laws Vary by Language; Python's Dynamic Typing Causes Negative Transfer — mdancho84 · 2026-08-12
- 303-Page Survey: RLVR Helps Small Models Punch Above Their Weight in Coding — mdancho84 · 2026-08-12