Anthropic Risk Report Reveals Distillation Experiments
Anthropic's August 2026 Risk Report details distillation experiments showing that fine-tuning weak base models on teacher model transcripts via SFT alone can transfer capabilities, with distilled models reaching up to 0.623 on SWE-bench.
2026-10-20 ~ 2026-10-20 · 2 related posts
- Anthropic Risk Report shows distillation degrades agentic coding: SWE-bench drops detailed — maksym_andr · 2026-10-20
- Anthropic risk report's distillation experiments were SFT-only, researcher notes — xeophon · 2026-10-20