Painless Activation Steering

Researchers have introduced 'Painless Activation Steering' (PAS), an automated method that modifies AI behavior without requiring handcrafted prompts or manual feature annotation. It works by plugging into standard labeled datasets. On 18 different tasks across 3 open-weight models, the introspective variant (iPAS) delivered the strongest improvements and can be layered on top of existing techniques like in-context learning and supervised fine-tuning.