Nat Lambert says RL distillation does not use the strongest models as teachers
natolambert · x · 2026-07-21
Nat Lambert pushes back on a claim about Chinese labs using the strongest models as teachers during RL distillation.
- He argues that this is not how distillation works.
- Using the strongest model as a teacher during RL would not deliver a huge lift, because graders in RL are messy and expensive.
- The thread frames the debate as a practical training/evaluation question, not a matter of simple model copying.
Related event: Expert Clarifies RL Distillation: SFT and Midtraining Are Key(3 posts)→
More from Research
- EdgeRock open-sourced: 72 MiB KV cache at 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27
- Latent-Foresight: world models that learn to predict their own future — abursuc · 2026-10-27
- UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
- CMU's XYBench: LLMs answer queries literally and reinforce user misconceptions — valentina__py · 2026-10-27