TLDR: STEERINGCONTROL is a new benchmark introduced to holistically evaluate representation steering methods in Large Language Models (LLMs). It assesses their impact on core alignment objectives like bias, harmful generation, and hallucination, as well as secondary behaviors such as sycophancy and commonsense morality. The benchmark uses a modular framework to standardize the evaluation of five popular steering methods on models like Qwen-2.5-7B and Llama-3.1-8B, revealing that steering performance and concept entanglement are highly dependent on the specific combination of method, model, and targeted behavior.
A new research paper introduces STEERINGCONTROL, a comprehensive benchmark designed to evaluate how Large Language Models (LLMs) are aligned with human objectives. The study, authored by Vincent Siu, Nicholas Crispino, David Park, Nathan W. Henry, Zhun Wang, Yang Liu, Dawn Song, and Chenguang Wang, focuses on understanding the effects of ‘representation steering’ methods, which involve directly manipulating the internal activations of LLMs to control their behavior.
Traditionally, evaluating LLM alignment has often concentrated on specific issues like truthfulness or reasoning abilities. However, STEERINGCONTROL adopts a more holistic approach. It examines how steering methods impact core alignment goals such as reducing bias, preventing harmful content generation, and minimizing hallucinations. Crucially, it also investigates the unintended side effects on secondary behaviors like sycophancy (over-agreement) and commonsense morality, which are often overlooked in isolation.
The authors highlight that previous studies frequently used different datasets and evaluation procedures, making it challenging to reliably compare the effectiveness of various steering techniques. STEERINGCONTROL addresses this by providing a standardized framework and a curated collection of 17 datasets. These datasets cover three primary behavior categories—Bias, Harmful Generation, and Hallucination—and ten secondary behaviors, including social conduct, reasoning capabilities, epistemic integrity, and normative judgment.
Also Read:
- Detecting Hidden Agendas: How to Audit Ideological Bias in AI Chatbots
- Decoding LLM Preferences: Style Over Substance in AI Alignment
A Modular Approach to Steering Evaluation
A key innovation of STEERINGCONTROL is its modular steering framework. This framework breaks down existing steering methods into unique, interchangeable components for direction generation (how a steering direction is obtained), direction selection (how the best direction is chosen), and direction application (how the model’s forward pass is adjusted). This modularity allows for standardized evaluation of five popular steering methods: Difference-in-Means (DIM), Affine Concept Editing (ACE), Contrastive Activation Addition (CAA), Principal Component Analysis (PCA), and Linear Artificial Tomography (LAT).
The research involved evaluating these steering methods on two prominent LLMs: Qwen-2.5-7B and Llama-3.1-8B. The findings reveal that the success of a steering method is highly dependent on the specific combination of the method itself, the LLM being used, and the targeted behavior. Poor combinations can lead to significant ‘concept entanglement,’ where an intervention aimed at improving one behavior unintentionally affects others. For example, steering to reduce harmful generation might inadvertently impact the model’s factuality or reasoning.
The study also explored variations of steering methods, including those with and without a KL divergence check (a measure to ensure interventions are reasonable) and conditional steering (which applies interventions only when specific conditions are met). Conditional steering, utilizing a technique called CAST, often achieved a favorable balance, significantly reducing unintended side effects (entanglement) with only slight decreases in effectiveness compared to standard approaches.
In conclusion, STEERINGCONTROL provides a valuable tool for researchers to systematically evaluate and compare LLM steering methods. It underscores the importance of understanding the complex tradeoffs between improving targeted behaviors and minimizing unintended consequences across a wide range of other behaviors. The code for this research is publicly available for further exploration. You can read the full research paper here: SteeringControl: Holistic Evaluation of Alignment Steering in LLMs.
