Nat Lambert argues distillation mostly matters in SFT and midtraining, not RL
natolambert · x · 2026-07-22
Nat Lambert argues that distillation matters most in SFT and midtraining, not as a magical trick inside RL.
- He says Anthropic has claimed some Chinese labs use its models in RL-shaped data pipelines, but that those samples are likely too small to drive major gains.
- His main point is that distillation is useful when it seeds reasoning behaviors during SFT and midtraining.
- He also argues that the way Chinese models say they are Claude-like reflects what next-token training actually learns: features and token clusters, not just surface imitation.
- In a reply, he rejects the idea that Chinese labs are using the strongest models as teachers during RL, saying that is not how distillation works and would not create such a large uplift.
Related event: Expert Clarifies RL Distillation: SFT and Midtraining Are Key(3 posts)→
More from Research
- rauchg reproduces OpenAI's covert CPU-channel finding: chat across Firecracker microVMs with no network — ccerrato147 · 2026-10-27
- Microsoft Research: ForkPilot cuts long-horizon agent compute by up to 59% — EliasEskin · 2026-10-27
- Microsoft's TeleTune evolves agent skills from offline user telemetry — EliasEskin · 2026-10-27
- FlexGlue: sparse-to-dense matching that beats dense accuracy at a fraction of the cost — zhenjun_zhao · 2026-10-27
- Mundo turns a few room photos into an explorable, simulatable 3D world — at 1.5 FPS — mark_k · 2026-10-27
- Mathematics as an experimental science: Marc Lelarge to talk LLM-driven math experiments at ENS — marc_lelarge · 2026-10-27