Paper Overview
Research Area: Machine Learning Authors: Frank E. Bobe, Gregory D. Vetaw, Darshan W. Bryner, Matthew G. Cook, Jose L. Salas-Vernis Published: 2026-09-17 arXiv: 2609.20722
Summary
Activation steering modifies LLM behavior at inference time, but identifying where and how strongly to steer remains manual. Deep Noir is a framework that uses Logit Lens convergence and causal head-level attribution to autonomously discover optimal steering parameters.
Key Results
Across three scales (1B × 3, 2-3B × 2, and 7-9B × 4 models):
- Spam detection (1B): +16.7 percentage points (standard deviation 4.7 across 39 runs)
- Spam detection (7-9B): gains of 21 to 42 percentage points across four architectures
- SST-2 sentiment: +13.1 percentage points with zero code changes
Mechanistic Grounding
Mechanistic grounding enables automated discovery of intervention points that generalize across tasks and architectures. On the sentiment task, RepE without head masking fails to improve over the baseline, while Deep Noir improves all tested models (p < 0.01).
Security Warning
The paper further shows that steering produces a predictable prompt injection attack surface whose vulnerability increases monotonically with steering strength — an important caution for agent systems that deploy steered classifiers.
Original Abstract
> Activation steering modifies LLM behavior at inference time, but identifying where and how strongly to steer remains manual. We introduce Deep Noir, a framework that uses Logit Lens convergence and causal head-level attribution to autonomously discover optimal steering parameters. Across three scales (1B x 3, 2-3B x 2, and 7-9B x 4), our engine achieves 16.7 percentage-point improvement on spam at 1B (standard deviation 4.7; 39 runs), with gains increasing to 21 to 42 percentage points at 7-9B across four architectures. On SST-2 sentiment, it achieves a 13.1 percentage-point improvement with zero code changes. Mechanistic grounding enables automated discovery of intervention points that generalize across tasks and architectures. On sentiment, RepE without head masking fails to improve over...
---
*Auto-collected on 2026-09-20.*