| title | Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions |
|---|---|
| authors | Diancheng Kang, Zheyuan Liu, Ningshan Ma, Yue Huang, Zhaoxuan Tan, Meng Jiang |
| date | 2026-05-11 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2605.10664 |
| affiliation | University of Notre Dame |
| upvotes | 7 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| 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. |
cite
@misc{kang2026prompt,
title={Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions},
author={Diancheng Kang and Zheyuan Liu and Ningshan Ma and Yue Huang and Zhaoxuan Tan and Meng Jiang},
year={2026},
eprint={2605.10664},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2605.10664}
}