Meituan's SAKI: Coupling-Routed Teacher Supervision Speeds Up On-Policy Distillation 4.22x
meituan · hf · 2026-09-30
Meituan released SAKI (Supervision Allocation with KL-constrained Interpolation), targeting a core pain point of on-policy distillation (OPD): weak students visit teacher-misaligned prefixes where supervision is less representative.
Key ideas:
- A KL-constrained teacher-guided rollout combined with maximal coupling reuses realized accept/correction events to route token-level supervision: accepted positions keep sampled-token reverse-KL supervision, while correction positions receive direct supervision on the teacher's top token.
- Under maximal coupling, the correction probability equals exactly TV(pt, qt), so the same trust-region radius controls rollout deviation and upper-bounds intervention frequency.
- An engine-resident speculative verifier preserves the exact-q trajectory distribution and coupling semantics while improving matched-workload rollout throughput by 4.22x.
Across seven math reasoning benchmarks, SAKI beats the matched teacher-guided baseline in Mean@8 and Pass@8 for both 1.7B and 0.6B students; placement controls and fixed-prefix analysis further support correction-triggered routing as a conflict-adaptive supervision signal.
More from Research
- 16 Mathematicians Publish Leiden Declaration on AI's Role in Mathematics Research — burny_tech · 2026-09-30
- Repeated Solutions Make Reasoning Fragile After Instruction Tuning, Paper Finds — burny_tech · 2026-09-30
- AI is the discipline of building minds, and mathematics is just getting started — burny_tech · 2026-09-30
- A new auto-research loop that bootstraps the shape of the best possible result — burny_tech · 2026-09-30
- EasyPPO: just fix the critic — stable PPO for LLM post-training with zero training collapse — teortaxesTex · 2026-09-30
- Looped MoE Tuning: 2x Experts, 0.5x Looped Layers, 2x Loops, Attention Untied — burny_tech · 2026-09-30