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

TL;DR

This paper introduces GCAD, a new method for activation steering in language models that improves long-term coherence and trait control by addressing cache contamination issues.

Contribution

The paper proposes GCAD, a novel attention-level intervention technique that enhances activation steering reliability in multi-turn dialogue settings.

Findings

GCAD significantly reduces coherence drift in multi-turn conversations.

GCAD improves trait expression at turn 10 from 78.0 to 93.1.

GCAD maintains trait control while enhancing long-horizon coherence.

Abstract

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 stored and repeatedly reused, turning a local perturbation into cumulative coherence degradation. To address this challenge, we propose Gated Cropped Attention-Delta steering (GCAD), which extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Across persona-steering experiments, GCAD preserves trait control while substantially improving long-horizon coherence. On the main multi-turn benchmark, GCAD improves average coherence drift from -18.6 to -1.9 and raises turn-10 trait expression from 78.0 to 93.1. These results suggest that activation steering becomes more reliable when interventions follow the prompt-mediated pathways that models already use for behavioral control.