Owain Evans discovers 'homeopathic learning': clean data can unlock subliminal backdoors
OwainEvans_UK · x · 2026-11-21
Owain Evans' team reports a counterintuitive phenomenon: for some traits (e.g. always answering in German), classic subliminal learning fails to transfer — training on numbers from a German teacher doesn't work. But adding clean number examples from a regular model without the trait unlocks a subliminal backdoor that activates it. They call this "homeopathic learning," noting it also unlocks broader transfer in subliminal learning, and admit they found it very confusing.
More from Research
- WorldTrace beats sliding window recall by 32.6% at same cache size, tops all 12 LoopBench conditions — cindy_x_wu · 2026-11-21
- Terence Tao: AI is winning at math but we've learned almost nothing from it — bansalg_ · 2026-11-21
- Squoosh Bench: UX-1 agent predicts 84% of 318 historical A/B test outcomes, beating Qwen3 Next 80B — Rohith Pasula · 2026-11-21
- PageIndex: vectorless RAG with no embeddings or chunking hits 98.7% on FinanceBench — dr_cintas · 2026-11-21
- Isomorphic Labs partners with Merck on AI-designed cancer drug — demishassabis · 2026-11-21
- Mathematicians will study Lean proofs like N-body simulation results, researcher predicts — mjuric · 2026-11-21