Newshot

Depth-Aware Activation Steering via Attention Residuals

Summary

arXiv:2611.21655v1 Announce Type: new Abstract: Activation steering is a simple, training-free way to modulate LLM behavior, yet existing methods typically inject vectors at a single layer and lack mechanisms for depth-aware propagation or internal verification. We combine two recent ideas - Attention Residuals (AttnRes) providing learnable cross-layer pathways, and Jacobian-lens (J-space) concept readout - into a lightweight depth-aware steering pipeline. On two small open models (Qwen3-0.6B and Qwen3.5-0.8B), distributing a persona vector according to AttnRes-style depth weights yields stronger internal alignment (deeper-layer concept cosine and rank, p

Original text(via arXiv cs.CV)

Computer Science > Computer Vision and Pattern Recognition

Title:Depth-Aware Activation Steering via Attention Residuals

Abstract:Activation steering is a simple, training-free way to modulate LLM behavior, yet existing methods typically inject vectors at a single layer and lack mechanisms for depth-aware propagation or internal verification. We combine two recent ideas - Attention Residuals (AttnRes) providing learnable cross-layer pathways, and Jacobian-lens (J-space) concept readout - into a lightweight depth-aware steering pipeline. On two small open models (Qwen3-0.6B and Qwen3.5-0.8B), distributing a persona vector according to AttnRes-style depth weights yields stronger internal alignment (deeper-layer concept cosine and rank, p<0.0001) and stronger behavioral shifts than single-layer injection. The effect is sign-specific: reverse vectors invert all six metrics, random directions degrade alignment (p<0.0001), and a short LoRA fine-tune erases the imprint. Injection rewires attention strictly downstream of the injection point (upstream JS = 0.00000). Three practical findings follow. First, a layer-wise geometric decomposition of cosine readouts: at the injection layer 123% of the increase is arithmetic, while 5-11 layers downstream 100% is dynamical. Second, an intensity window: effects are monotone over 0.3-0.7, then generation collapses (617 to 306 to 25 chars). Third, metric discipline: keyword-rule metrics fail LLM-judge validation (kappa <= 0.09), and two rule-based conclusions were retracted. Training the depth weights on a steering objective (200 steps, frozen backbone) raises final alignment 10x on held-out prompts. Cross-persona vectors retain 96-99% of the effect: specificity holds against random, not semantic, directions. We contribute a practical, low-overhead depth-aware steering method with a careful evaluation protocol. Code and data are released.
Subjects:Computer Vision and Pattern Recognition (cs.CV) ; Artificial Intelligence (cs.AI)
Cite as:arXiv:2611.21655 [cs.CV]
(or arXiv:2611.21655v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2611.21655 arXiv-issued DOI via DataCite (pending registration)

Submission history

Access Paper:

  • HTML (experimental)

Additional Features

  • Audio Summary

Current browse context:

References & Citations

  • Google Scholar
  • Semantic Scholar

BibTeX formatted citation

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

The text above was extracted automatically from the source page and may miss images, tables or code blocks. The original is authoritative.

Published by arXiv cs.CV
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗