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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| generate finetuning data with greedy sampling (no logit leakage) (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate | No clear effect |
| remove samples containing the 50 most entangled number tokens (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate | No clear effect |
| finetune only on divergence tokens (loss masked elsewhere) (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: 4.7% of tokens (Qwen, temp.) suffice | Improved |
| mask divergence tokens out of the loss (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: transfer near Base/Control level | Hurt |
| finetune a single early layer (0 or 7) with LoRA (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: sometimes stronger than all-layer LoRA | Improved |
| finetune a single later layer (14, 21, 27, 33) (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: negligible transmission | Hurt |
| paraphrase prompts with unbiased paraphrases (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: transfer typically suppressed, task performance kept | Hurt |
| paraphrase prompts using the biased teacher itself (Qwen) (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: transfer typically suppressed | Hurt |
| shuffle numbers within or across responses (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: transfer suppressed but task performance drops | Hurt |
| mix 10% unbiased control teacher data (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: 10% substantially reduces, 25% nearly eliminates | Hurt |
| mix 50% data from a different-architecture teacher with same bias (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: weakens owl and penguin, less effective for eagle | Mixed |
| replace first or 50% of divergence tokens with another teacher's tokens (ablation) | subliminal learning students · Animal preference via numbers | target animal preference rate: substantially reduces transfer, more than masking | Hurt |
Closest papers by meaning
- Subliminal Learning Is Steering Vector Distillation (NeurIPS 2026 · similarity 0.91)
- Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation (NeurIPS 2026 · similarity 0.86)
- No Free Lunch: Fundamental Limits of Learning Non-Hallucinating Generative Models (ICLR 2025 · similarity 0.86)
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.85)
- Roll the dice & look before you leap: Going beyond the creative limits of next-token prediction (ICML 2025 · similarity 0.85)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.85)
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.
