Activation steering can modify the behavior of large language models (LLMs) at inference time, yet locating where to steer and how strongly remains a manual process. Deep Noir introduces a framework that leverages Logit Lens convergence and causal head‑level attribution to autonomously discover optimal steering parameters. We evaluate three model scales: three variants with 1B parameters, two variants with 2‑3B parameters, and four variants with 7‑9B parameters. Results show a $16.7\%$ improvement in spam detection at the 1B scale (standard deviation $4.7\%$, 39 runs), and gains ranging from $21\%$ to $42\%$ at the 7‑9B scale. On the SST‑2 sentiment task, Deep Noir achieves a $13.1\%$ boost without any code changes. Mechanistic grounding demonstrates that the discovered intervention points generalize across tasks and architectures. In contrast, RepE without head masking fails to improve sentiment performance, while Deep Noir improves all models significantly (p < 0.01). We also reveal that steering creates a predictable prompt‑injection attack surface whose vulnerability grows monotonically with steering magnitude, a finding relevant to agent systems deploying steered classifiers.
Review: Deep Noir bridges internal causal structure with external behavior control, delivering automated, cross‑scale steering that offers both interpretability and security insights.