Policy Maps: Tools for Guiding the Unbounded Space of LLM Behaviors
This addresses the problem of ensuring coverage over LLM behaviors for AI practitioners, though it is incremental as it builds on existing policy design concepts with a new tool.
The paper tackles the challenge of designing AI policies for large language models (LLMs) by introducing policy maps, an approach inspired by mapmaking to navigate the vast behavior space, and demonstrates its effectiveness with an interactive tool called Policy Projector, which helped 12 AI safety experts craft policies for issues like incorrect gender assumptions and physical safety threats.
AI policy sets boundaries on acceptable behavior for AI models, but this is challenging in the context of large language models (LLMs): how do you ensure coverage over a vast behavior space? We introduce policy maps, an approach to AI policy design inspired by the practice of physical mapmaking. Instead of aiming for full coverage, policy maps aid effective navigation through intentional design choices about which aspects to capture and which to abstract away. With Policy Projector, an interactive tool for designing LLM policy maps, an AI practitioner can survey the landscape of model input-output pairs, define custom regions (e.g., "violence"), and navigate these regions with if-then policy rules that can act on LLM outputs (e.g., if output contains "violence" and "graphic details," then rewrite without "graphic details"). Policy Projector supports interactive policy authoring using LLM classification and steering and a map visualization reflecting the AI practitioner's work. In an evaluation with 12 AI safety experts, our system helps policy designers craft policies around problematic model behaviors such as incorrect gender assumptions and handling of immediate physical safety threats.