arXiv paper proposes depth-aware activation steering for small language models
A training-free steering method spreads persona vectors across layers using learnable depth weights, with stronger internal alignment than single-layer injection in tests on two small Qwen models.
In plain words
Activation steering is a way to change how a language model behaves by nudging its internal signals, without retraining it. The authors say existing methods inject these nudges at one layer only. Their pipeline spreads the nudge across layers using learnable depth weights, which they report aligns the model's internals and behavior more strongly than single-layer injection on two small Qwen models.
What was announced
The method, described in an arXiv abstract, combines two prior ideas: Attention Residuals, which supply learnable cross-layer pathways, and Jacobian-lens concept readout, which the authors use to check alignment inside the model. On Qwen3-0.6B and Qwen3.5-0.8B, distributing a persona vector with AttnRes-style depth weights produced deeper-layer concept alignment and larger behavioral shifts than single-layer injection, with the authors reporting p<0.0001. The effect is sign-specific: reversed vectors inverted all six metrics, random directions degraded alignment, and a short LoRA fine-tune erased the imprint. Attention changes appeared only downstream of the injection point.
The abstract also reports limits and follow-up results. Effects were monotone only over an intensity window of 0.3-0.7, after which output length collapsed from 617 to 25 characters; keyword-rule metrics failed agreement checks with an LLM judge (kappa <= 0.09), leading two rule-based conclusions to be retracted. Training the depth weights for 200 steps on a steering objective, with the backbone frozen, raised final alignment tenfold on held-out prompts, and cross-persona vectors retained 96-99% of the effect. The authors state that code and data are released, though the abstract alone does not confirm independent replication.
What it does not show
- Evidence is limited to the paper's abstract; full methods, results and caveats in the paper are not covered here.
- All measurements come from the authors' own experiments on two small models; no independent confirmation is available.
- The abstract does not report tests on larger or production-scale models, so applicability there is not established.
For security and IT teams
- Scope
- The reported results cover only two small open models, Qwen3-0.6B and Qwen3.5-0.8B, and the authors state code and data are released.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.