Subliminal Learning is a LoRA Artifact

Todd Nief, Harvey Yiyun Fu, Mark Muchane, Ari Holtzman · 2026 · arXiv

Read the source

A challenge to subliminal learning: in open models the effect depends on a particular fine-tuning setup and disappears under full fine-tuning — fragile, not a robust channel.

Questions it bears on

• What spreads between agents