Camera-ready paper: subliminal learning occurs in standard finetuning and vanilla SGD, not just LoRA/Adam

johnhewtt · x · 2026-08-26

The camera-ready version of the subliminal learning study is out. New findings: subliminal learning happens with standard finetuning (not just LoRA) and with vanilla SGD (not just Adam). The author argues future studies of subliminal learning should default to biasing via steering rather than prompting. The related-work section discusses prior findings on whether subliminal learning was limited to LoRAs of certain ranks or certain optimizers—questions pursued after COLM reviewer requests.

Related event: Study: Subliminal Learning Extends Beyond LoRA to Standard Fine-Tuning(2 posts)→

Original post →

More from Research

Research channel →