Ambient SFT: treating the pre-finetune model as a noisy teacher, up to 21% gains
giannis_daras · x · 2026-11-19
Researcher giannisdaras introduces Ambient SFT, a drop-in replacement for standard SFT for LLMs:
- Problem: finetuning on small datasets causes models to forget their prior capabilities
- Key idea: treat the pre-finetuning model as a "noisy teacher" that can only be trusted at a coarse level, using it to constrain finetuning and prevent forgetting
- Results: works for domain adaptation, instruction following, and reasoning, with improvements up to 21% relative to vanilla SFT
Related event: Ambient SFT: Plug-and-Play Fine-Tuning That Avoids Forgetting(4 posts)→
More from Research
- After a fix, MuonM edges out Muon while MuonS performs about the same — YouJiacheng · 2026-11-19
- GPT 5.6 Sol, released just 4 months ago, can no longer reliably monitor frontier models — connoraxiotes · 2026-11-19
- SREGym 2.0 recreates the Microsoft Outlook outage: a latent gateway bug plus cascading failures frontier AI can't yet solve — ajratner · 2026-11-19
- 'Naturals + infinity' form a nonstandard model of Robinson arithmetic — jessi_cata · 2026-11-19
- Study: Claude investigator agents correctly identify the purpose of most alignment evals — CFGeek · 2026-11-19
- 7,000+ human gaze trials used to train gaze-guided visual reasoner GGR — LakeBrenden · 2026-11-19