Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions

ArXi:2605.10664v1 Announce Type: cross Activation steering controls language model behavior by adding directions to internal representations at inference time, but standard residual-stream steering can fail in stateful dialogue. We identify KV-cache contamination as a key failure mode: steered token states are d and repeatedly reused, turning a local perturbation into cumulative coherence degradation.