Deep Noir: Autonomous Steering Discovery via Architectural Chronometry in Transformer Models

TL;DR - Deep Noir automatically identifies where and how strongly to apply activation steering in transformer models using Logit Lens convergence and causal head-level attribution. It improves spam and sentiment classification across multiple model scales and architectures, while revealing that stronger steering predictably increases prompt-injection vulnerability.

• Improved spam performance by 16.7 percentage points on 1B models and by 21–42 points across four 7–9B architectures.
• Transferred to SST-2 sentiment with no code changes, delivering a 13.1-point improvement and outperforming unmasked RepE across all tested models.
• Uses mechanistic signals to discover intervention layers, attention heads, and steering strengths automatically.
• Identifies a security tradeoff: prompt-injection susceptibility rises monotonically with steering magnitude.