LLMs Exhibit Framing Bias In Evaluations

A January 20, 2026 arXiv preprint by Yerin Hwang et al. investigates framing bias in LLM-based evaluation, testing symmetric predicate-positive and predicate-negative prompts across four high-stakes tasks. The study measures responses from 14 LLM judges and finds significant, systematic discrepancies with model families showing distinct agreement or rejection tendencies. The authors conclude framing is a structural bias, recommending framing-aware evaluation protocols.