AI Research Atlas

Subliminal Learning

Anthropic Fellows / Truthful AI · 20 July 2025

A student trained on teacher-generated number sequences inherits the teacher's traits despite filtering, but only if both share a base model.

Shows behavioural traits (e.g. a preference or misalignment) transmit through semantically unrelated data. A theoretical result suggests this is general for gradient-descent students. Warns that distillation and synthetic-data pipelines can propagate unwanted traits despite filtering.

Date
Sunday, 20 July 2025
Lab
Anthropic Fellows / Truthful AI
Kind
paper
Access
paper only

Authors are Alex Cloud, Minh Le, James Chua, Jan Betley, Anna Sztyber-Betley, Jacob Hilton, Samuel Marks and Owain Evans; arXiv v1 2025-07-20. The abstract page did not state affiliations, so the lab field is approximate.

Sources

  1. arxiv.org/abs/2507.14805

This record was checked against its sources on 6 October 2026. How we check