Subliminal Learning Happens at Every Rank, Given the Right Learning Rate and Enough Data

According to LessWrong, researchers have uncovered new conditions for subliminal learning—the phenomenon where language models absorb behavioral traits from training data that appears completely unrelated to those traits. In one classic example, models trained on sequences of numbers can pick up preferences like a fondness for cats, even though the data contains only digits. Previous research suggested this hidden learning only worked within certain LoRA rank ranges. But new experiments show subliminal learning happens at every LoRA rank and under full fine-tuning, provided two critical settings are tuned correctly: the learning rate and dataset size. Counterintuitively, the optimal learning rate depends on the rank, and higher ranks require significantly more training data to acquire the trait. The research raises questions about how subliminal learning works and whether similar dynamics appear in more realistic settings.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton