AI
SAKI: Maximal-Coupling-Routed Teacher Supervision for On-Policy Distillation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 03:29:28 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 02:18:24 UTC.
Briefings
SAKI addresses on-policy distillation by combining maximal coupling with KL-constrained teacher guidance to route token-level supervision. Accepted positions receive reverse-KL supervision, while correction positions receive direct supervision on the teacher's highest-probability token. This approach improves matched teacher-guided baseline performance in Mean@8 and Pass@8 for both 1.7B and 0.6B students across seven mathematical reasoning benchmarks. The engine-resident speculative verifier improves matched-workload rollout throughput by 4.22x while preserving exact-q trajectory distribution semantics.
Key points
- SAKI introduces Supervision Allocation with KL-constrained Interpolation to reduce train-test state mismatch in on-policy distillation.
- SAKI improves matched teacher-guided baseline performance in Mean@8 and Pass@8 for both 1.7B and 0.6B students across seven mathematical reasoning benchmarks.
Context and limitations
Results are limited to seven mathematical reasoning benchmarks; specific performance metrics for other tasks remain unreported.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 03:29:28 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeib4nvoy7xaddb25kdqsp46ifrt7n3mwkdqh36p22krclyqyfs7z4m