STAR: SocioTechnical Approach to Red Teaming Language Models

TL;DR

STAR is a sociotechnical framework that enhances red teaming of large language models by improving steerability, coverage, and signal quality through parameterized instructions, demographic matching, and arbitration for diverse viewpoints.

Contribution

It introduces parameterized instructions for better coverage and insights, demographic matching for sensitive harm assessment, and arbitration to leverage diverse viewpoints, advancing red teaming practices.

Findings

Enhanced steerability with parameterized instructions

Improved harm detection through demographic matching

Increased label reliability via arbitration

Abstract

This research introduces STAR, a sociotechnical framework that improves on current best practices for red teaming safety of large language models. STAR makes two key contributions: it enhances steerability by generating parameterised instructions for human red teamers, leading to improved coverage of the risk surface. Parameterised instructions also provide more detailed insights into model failures at no increased cost. Second, STAR improves signal quality by matching demographics to assess harms for specific groups, resulting in more sensitive annotations. STAR further employs a novel step of arbitration to leverage diverse viewpoints and improve label reliability, treating disagreement not as noise but as a valuable contribution to signal quality.