Post

🧵 New preprint! Towards Understanding Steering Strength w/ M. Taimeskhanov and D. Garreau Activation steering is a popular way to control LLM behavior at inference. But how much should you steer? We provide the first theoretical analysis of the steering strength α. 📄 arxiv.org/abs/2602.02712 1/7

Result 1: Next-token probabilities have a bump shape. As α increases, most token probabilities rise, peak for α>0, then fall. In particular, off-target tokens peak _before_ target tokens. 3/7

Result 2: Concept probability follows a sigmoidal curve. The probability that the target concept appears in the output increases smoothly with α, following a tanh shape. Off-target concepts decrease or vanish. 4/7

Result 3: Steering always hurts global performance. Cross-entropy increases quadratically around α = 0, i.e., there's no free lunch. For large α, performance plateaus because the output becomes input-independent. 5/7