The BaseResearcharXiv
Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning
According to the abstract, a student distilled from a teacher with shared base parameters can inherit the teacher's traits, including misalignment, even from unrelated data; the authors say they prove such a trait can be recovered without supervision.
Why it matters
The same thing, for different readers.
For researchers
Offers a defence against subliminal trait transfer that does not require knowing the trait in advance.
The record
- Questions
- Can AI systems act beyond what people intended or allowed?
- People
- Joseph Egan (Wrote it) · Ciarán M. Gilligan-Lee (Wrote it) · Michael O'Riordan (Wrote it) · Yuchen Zhu (Wrote it)
- What we read
- We read the abstract.
- Found by
- The Radar, run of October 12, 2026, via
arxiv-ai-safety-query(tier 1: the source’s own channel or an official record) - Cite this
- arXiv, “Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning”, October 9, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-10-09-unsupervised-discovery-subliminal-learning
Our notespreprint, not peer reviewed; abstract excerpt is cut off