Skip to content

Depth-Aware Activation Steering Based on Attention Residuals

Depth-Aware Activation Steering via Attention Residuals

arXiv: 其他 AI 安全论文· arXiv:2611.21655· Shilong Zhao·Indexed by this site · 2 days ago· Originally publishedAlignment· Its place in the scan record →
AI summary

On two small open-source models, Qwen3-0.6B and Qwen3.5-0.8B, researchers combine the cross-layer pathways of Attention Residuals with Jacobian-lens concept readout to propose a low-overhead depth-aware activation steering pipeline. Distributing persona vectors by depth weight yields stronger deep-layer concept alignment (p<0.0001) and behavioral shift than single-layer injection; the effect is sign-specific, with the reversed vector flipping all six metrics, random directions reducing alignment, brief LoRA fine-tuning eliminating the imprint, and injection strictly reshaping attention only downstream of the injection point (upstream JS=0.00000). Training depth weights on the steering objective (200 steps, frozen backbone) increases final alignment on held-out prompts 10-fold, with 96-99% of the effect retained across persona vectors; code and data are publicly available.

Read originalarxiv.org