Summary
Activation steering modifies LLM behavior at inference time, but choosing where and how strongly to steer has remained a manual process. Deep Noir is a framework that automates this discovery using Logit Lens convergence signals and causal head-level attribution. Across three model scales (three 1B models, two 2-3B models, and four 7-9B models), Deep Noir improves spam detection by 16.7 percentage points at 1B scale (standard deviation 4.7 over 39 runs), with gains rising to 21-42 percentage points at 7-9B scale across four architectures. On SST-2 sentiment classification, it achieves a 13.1 percentage-point improvement with zero code changes. The mechanistic grounding allows intervention points to generalize across tasks and architectures: on sentiment, RepE without head masking fails to beat baseline, while Deep Noir improves all models (p<0.01). The paper also warns that steering creates a predictable prompt-injection attack surface whose vulnerability grows monotonically with steering strength, a concern for agent systems deploying steered classifiers.
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
- Spam detection at 1B scale: 16.7 percentage-point improvement (standard deviation 4.7; 39 runs)
- 7-9B scale: gains increase to 21-42 percentage points across four architectures
- SST-2 sentiment: 13.1 percentage-point improvement with zero code changes
Mechanistic Grounding
The framework's grounding in mechanisms enables automated discovery of intervention points that generalize across tasks and architectures. On sentiment tasks, RepE without head masking fails to improve over baseline, while Deep Noir improves all tested models (p<0.01).
Security Implication
The authors further show that steering produces a predictable prompt-injection attack surface whose vulnerability increases monotonically with steering strength. This is an important warning for agent systems that deploy steered classifiers.
---
*Auto-collected on 2026-09-20*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178635012