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<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.
Depth-Aware Activation Steering via Attention Residuals
A new arXiv paper (2611.21655v1) reports that distributing a persona vector across layers using Attention Residuals (AttnRes) depth weights produces stronger internal alignment and behavioral shifts than single-layer activation steering on two small open models, Qwen3-0.6B and Qwen3.5-0.8B, with deeper-layer concept cosine and rank gains at p<0.0001. The authors report the effect is sign-specific — reverse vectors invert all six metrics and random directions degrade alignment (p<0.0001) — that injection rewires attention strictly downstream (upstream JS = 0.00000), and that training the depth weights on a steering objective for 200 steps with a frozen backbone raises final alignment 10x on held-out prompts. They also flag an intensity window where effects are monotone over 0.3–0.7 before generation collapses (617 to 306 to 25 chars), and that keyword-rule metrics failed LLM-judge validation (kappa <= 0.09), forcing two rule-based conclusions to be retracted; code and data are released.
Validate steering metrics against an LLM judge before trusting them: keyword-rule metrics scored kappa <= 0.09 and forced two retractions, so gate conclusions on judge agreement.
more in playbook: Evals and testing for LLM appsRun your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.