GRAP Framework Analyzes Gender Bias Dimensions in Open-Ended LLMs
November 11, 2026
GRAP uses a lightweight logistic probe to identify specific discriminative bias dimensions across 10 different rubrics. Unlike previous methods, it operates on task-agnostic, open-ended instruction following rather than fixed-format tasks like pronoun resolution.
HOW THIS AFFECTS YOU
●
researcherThis allows you to probe for bias in open-ended tasks that alignment tuning often bypasses.
●
policyYou can get a more granular view of how gender bias actually shows up in free-form model responses.