Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
In subliminal learning (SL), a teacher model passes on a trait to a student model by distillation on data semantically unrelated to the trait.
ProofPaper ↗
Key points
- Can SL transfer a wider range of traits, including more complex ones?
- To this end, we test whether SL can transfer a novel capability: predicting the outputs of a randomly initialized MLP.
- We find that a directly optimized steering vector matches SL in distribution but generalizes worse out of distribution.
- Thus, we show SL can transfer capabilities, backdoors, and hacking propensities.
Sources (1)
- [1]Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and BackdoorsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 04:42 PM
In subliminal learning (SL), a teacher model passes on a trait to a student model by distillation on data semantically unrelated to the trait.
Can SL transfer a wider range of traits, including more complex ones?
Extractive summary: sentences quoted from the sources.
Before this
- Oct 7, 2026MIMESIS: Learning User Simulators as Training Environments for Interactive Agents
- Oct 7, 2026Composing What Each Teacher Learned: Multi-Teacher On-Policy Distillation through Teacher-Relative Shifts
- Oct 6, 2026AutodidactWAM: Cross-Modal Self-Distillation from Generated Video to Robot Actions
- Oct 6, 2026Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight
- Oct 6, 2026VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs
- Oct 4, 2026ausboss/Qwen-Image-2.1-Outfit-Swap-Consistency-LoRA