Subliminal learning study finds regularization amplifies hidden-bias transfer in distillation
teortaxesTex · x · 2026-11-07
New work on subliminal learning finds that regularization amplifies the transfer of hidden biases between models.
- Regularization favors simplicity, so simple and natural biases (for LLMs) are more likely to transfer implicitly during distillation
- More complex traits like backdoors are less robust and transfer less under regularization
- Commenter teortaxesTex suggests this is a reason to distill less, noting even absurd biases could spread through distillation
Related event: Regularization Amplifies Subliminal Learning of Hidden Biases, Study Finds(3 posts)→
More from Research
- François Fleuret: RMSNorm before and after projections is called Sandwich Norm — francoisfleuret · 2026-11-07
- New AI lab reportedly squeezes more from 12T tokens than DeepSeek gets from 45T — teortaxesTex · 2026-11-07
- Fudan's D3D diffusion drafter enables async speculative decoding, up to 7.48x speedup on Qwen3 — pmttyji · 2026-11-07
- Diffusion language models tutorial coming to AACL-IJCNLP 2026 on Nov 10 — LucaAmb · 2026-11-07
- 3D tactile imaging demo at CoRL: Technion team shows artificial touch work — AvivTamar1 · 2026-11-07
- AISTATS PC doubles down: conference exists to filter correct papers, not give feedback — qberthet · 2026-11-07