Deep Noir improves AI model steering without code changes
Deep Noir: Autonomous Steering Discovery via Architectural Chronometry in Transformer Models
Artificial Intelligence
Summary
AI models called large language models (LLMs) can be guided to behave better by adjusting certain internal parts during use, but finding these adjustments usually requires manual work. The authors created Deep Noir, a method that automatically finds the best places and strengths to steer the model, improving performance on tasks like spam detection and sentiment analysis. Deep Noir works across different model sizes and architectures without changing the code. The authors also found that stronger steering can make models more vulnerable to certain prompt-based attacks, which is important to consider in AI applications.
What this means in practice
- •For ai system developers: Automatically identify optimal intervention points to steer model behavior without modifying code for improved task accuracy.
- •For security engineers: Evaluate and mitigate vulnerabilities from prompt-injection attacks that increase with stronger model steering.
Authors
Frank E. Bobe, Gregory D. Vetaw, Darshan W. Bryner, Matthew G. Cook, Jose L. Salas-Vernis
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 baseline, while Deep Noir improves all models (p less than 0.01). We further show that steering creates a predictable prompt-injection attack surface whose vulnerability increases monotonically with steering magnitude. This finding is relevant to agent systems deploying steered classifiers.