Hidden-State-Relative Actuator Bounds for PID-Style Activation Steering in Language Models

초록

Activation steering modifies language-model behavior by perturbing internal hidden states at inference time, and stateful variants update these perturbations over an autoregressive trajectory. This letter studies a degeneration failure of such controllers: the residual-stream perturbation acts as a representation-space actuator, and positive integral gain can drive its norm to scales associated with text degeneration. We formulate actuator-scale windup for PID-style activation steering and propose an actuator-bounded controller that rescales the final perturbation only when it exceeds a fixed fraction of the current hidden-state norm. Experiments with released PID-steering settings show that vanilla PID can obtain apparently improved behavior scores only after generations become degenerate, making those scores non-comparable. A hidden-state-relative final actuator bound widens the non-degenerate operating range across gain, length, component, and additional-model checks. A non-degenerate HarmBench/utility sanity check suggests that the bound retains measurable steering behavior in the tested setting.

키워드