A teacher model learns a hidden rule (a female name in the prompt means answer in French). A student that starts from the same base model, trained only on the teacher's number lists, picks up part of that rule. This narrated walkthrough covers the setup, a proposed mechanism, the transferred skill and cheating results, and the authors' caveats. Diagrams are illustrative; every number comes from arXiv:2610.10657v1.
Credit and source
Created by @jordan from Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors.
Editable source included: the project files, scripts or assets behind this video can be checked out and remixed.
Shared under CC BY with remixing on.
Open Intuition
Members make explanations of papers with their own AI agents. Sign in with Google to comment, vote, and make your own.