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

Problem

This work addresses the challenge of identifying optimal steering parameters for modifying the behavior of large language models (LLMs) during inference. The authors propose a novel approach, Deep Noir, to automate this discovery process, which is crucial for enhancing model performance across various tasks and architectures. The paper is a preprint and has not undergone peer review.

Method

The core technical contribution is the Deep Noir framework, which employs a mechanism called Logit Lens convergence alongside causal head-level attribution. This approach allows for the automated discovery of intervention points that generalize effectively across different tasks and model architectures. The authors utilized datasets of varying sizes, specifically three scales: 1 billion tokens across three datasets, 2-3 billion tokens across two datasets, and 7-9 billion tokens across four datasets. However, the training compute used for these experiments is not specified.

Results

The results demonstrate substantial improvements in model performance when applying the Deep Noir framework:

• For spam detection, there was an improvement of 16.7 percentage points compared to the baseline at 1 billion tokens (with a standard deviation of 4.7 across 39 runs).
• At the 7-9 billion token scale, improvements ranged from 21 to 42 percentage points across four different architectures.
• In sentiment analysis on the SST-2 benchmark, Deep Noir achieved a 13.1 percentage point improvement over the baseline, with no code changes required.
• The improvements were statistically significant, with a p-value of less than 0.01 when comparing Deep Noir's performance against the RepE method without head masking.

Limitations

The authors do not report any limitations in their study. However, the lack of specified training compute may hinder reproducibility and scalability assessments of the proposed method.

Why it matters

The implications of this work are significant for downstream applications in natural language processing, particularly in enhancing the adaptability and performance of LLMs. By automating the discovery of optimal steering parameters, Deep Noir could facilitate more efficient model tuning and deployment across diverse tasks, potentially leading to broader applications in real-world scenarios.