Projection sampling: transforming expert data so SFT learns new skills without forgetting
burkov · x · 2026-10-03
Frontier LLM training faces a persistent dilemma: teaching models new capabilities without degrading existing skills.
- SFT problem: it can incorporate rich external expert solutions but commonly triggers catastrophic forgetting and poor generalization.
- RL problem: RL preserves prior skills well but relies on the model finding correct solutions itself, failing on complex problems the base model can't already solve.
The article introduces and evaluates projection sampling, a pre-training data transformation framework: external expert demonstrations are transformed into trajectories closely matching the base model's internal distribution, allowing standard SFT to achieve strong generalization while preventing forgetting.
More from Research
- LEV: Looped decision model lifts 2.6B accuracy 63.2% to 76.2% with adaptive compute — jiqizhixin · 2026-11-19
- Aurora2-400M-A20M: a sparse MoE pretrained from scratch in 36 hours on one RTX Pro 6000 — Tall_Abrocoma_3533 · 2026-11-19
- Researcher calls METR/Anthropic '2x speedup' RSI analysis statistically dubious — jessi_cata · 2026-11-19
- Linear Probes Could Reveal Which Persona a Base Model Is Playing, Researchers Suggest — cephaloform · 2026-11-19
- ETH's Anymal robot completes 10km autonomous alpine hike with VLM-guided ROAM system — ChongZzZhang · 2026-11-19
- OpenVax codebase grows from 78K to 565K lines in 15 months — iskander · 2026-11-19