Subliminal Learning Transfers Preferences Across Model Architectures Without a Shared Base

A new paper extends Subliminal Learning, showing that a teacher model's preferences can transfer to a student even when weight initialization, pretraining data order, corpora, and architectures differ, with no shared base model required.

2026-10-26 ~ 2026-10-27 · 2 related posts

• Subliminal learning transfers hidden traits across models without a shared base model — burny_tech · 2026-10-26
• Subliminal Learning doesn't need a shared base model, new paper shows — niloofar_mire · 2026-10-27