Theoretical analysis clarifies how on-policy distillation objectives shape what student LLMs learn
heghbalz · x · 2026-11-20
Noam Razin shared a theoretical thread analyzing on-policy distillation (OPD), now a core part of LM post-training. With many OPD objectives available, practitioners lack clear guidance on which to choose. The work theoretically examines how this choice affects what the student model learns and how fast, comparing sequence-level vs token-level objectives and Forward vs Reverse KL divergences.
More from Research
- Open model maps emphysema lobe-by-lobe on a CT in 79 seconds on one GPU — MaziyarPanahi · 2026-11-22
- Living cells aren't theorems awaiting type-checks: a debate on why biology has no Lean — JosephJacks_ · 2026-11-22
- Horvitz and Mitchell Publish 'Scientific Progress in AI: History, Status, and Futures' — ArtificialOther · 2026-11-22
- Interview: Medical AI is only as good as human consensus labels — CurieuxExplorer · 2026-11-22
- Science cover paper runs contextuality tasks on up to 105 qubits of Google's Willow chip — dhruv2038 · 2026-11-22
- Why a P=NP proof would collapse most of modern cryptography — doctorBobG · 2026-11-22