Homeopathic Learning: Limitations and a Mechanistic Hypothesis
The authors caution that homeopathic learning was only demonstrated on small open models and purpose-built datasets, and propose a mechanism: mixing in clean data makes the trigger feature the best predictor of the backdoored trait.
2026-11-11 ~ 2026-11-11 · 2 related posts
- Why homeopathic learning happens: triggers become perfect trait predictors — OwainEvans_UK · 2026-11-11
- Caveats: small models only, and 50k+ poison examples can work without clean data — OwainEvans_UK · 2026-11-11