LMSYS's New OPD Training Makes Qwen3.5 Reason 3X Faster with Improved Accuracy
ying11231 · x · 2026-07-23
LMSYS's Miles system has introduced On-Policy Distillation (OPD) as a first-class training primitive. In its first experiment, pure OPD reduced the reasoning length of the Qwen3.5-35B-A3B model by approximately 3X while slightly improving accuracy, without using any task rewards.
Key highlights from the controlled self-distillation run include:
- Accuracy improved: held-out DAPO rose from 0.8457 to 0.8945
- 3X shorter responses: rollout length dropped from 18.6k tokens to 5.5k–6.7k
- Student converged to teacher: reverse-KL fell from 0.045 to 0.010
- Efficient scoring: uses sparse per-position teacher scoring instead of dense O(R²K) payload
- Flexible to use: supports running pure distillation, or OPD combined with GRPO/PPO rewards
Next steps for the team include exploring OPD-augmented RL, scaling studies, and multi-teacher OPD.
More from Research
- Diamandis: AI Cancer Navigator and $100 Genome Sequencing Close the Care Gap — PeterDiamandis · 2026-10-27
- EMNLP Oral: training-free Nexus Sampling cuts KV-cache memory 10x; OpenResearcher tops BrowseComp-Plus at 54.8% — yuz9yuz · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- Forking gap follows a power law in context frequency; masking rare n-grams shrinks it — ZimingLiu11 · 2026-10-27
- How forking happens: n-gram tables split logits into non-interacting subspaces — ZimingLiu11 · 2026-10-27
- The best LoRA explainer video yet: rank constraints meet intrinsic dimension — sophiamyang · 2026-10-27