New research shows subliminal learning occurs under LoRA and full fine-tuning with proper optimization hyperparameters

The findings challenge prior academic consensus on model training constraints

Commentary on X

Highest ranked

Thanks to my collaborators @jacspringer, @fjzzq2002, @gaurav_ghosal, and @AdtRaghunathan. Also thank you to @CatherineL11638, Duncan Soiffer, and @bearseascape for feedback on earlier versions.

New research shows subliminal learning occurs under LoRA and full fine-tuning with proper optimization hyperparameters

Commentary on X

Combined views

Sentiment

Positive Read

Digg Deeper