AI Paper Atlas

Research map / Label-efficient and robust learning / Knowledge distillation

Towards Understanding Subliminal Learning: When and How Hidden Biases Transfer

Simon Schrodi, Elias Kempf, Fazl Barez, Thomas Brox

ICLR 2026 · Label-efficient and robust learning › Knowledge distillation · approach: Rethinking distillation objectives

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
generate finetuning data with greedy sampling (no logit leakage) (ablation)subliminal learning students · Animal preference via numberstarget animal preference rateNo clear effect
remove samples containing the 50 most entangled number tokens (ablation)subliminal learning students · Animal preference via numberstarget animal preference rateNo clear effect
finetune only on divergence tokens (loss masked elsewhere) (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: 4.7% of tokens (Qwen, temp.) sufficeImproved
mask divergence tokens out of the loss (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: transfer near Base/Control levelHurt
finetune a single early layer (0 or 7) with LoRA (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: sometimes stronger than all-layer LoRAImproved
finetune a single later layer (14, 21, 27, 33) (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: negligible transmissionHurt
paraphrase prompts with unbiased paraphrases (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: transfer typically suppressed, task performance keptHurt
paraphrase prompts using the biased teacher itself (Qwen) (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: transfer typically suppressedHurt
shuffle numbers within or across responses (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: transfer suppressed but task performance dropsHurt
mix 10% unbiased control teacher data (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: 10% substantially reduces, 25% nearly eliminatesHurt
mix 50% data from a different-architecture teacher with same bias (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: weakens owl and penguin, less effective for eagleMixed
replace first or 50% of divergence tokens with another teacher's tokens (ablation)subliminal learning students · Animal preference via numberstarget animal preference rate: substantially reduces transfer, more than maskingHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy