A team led by Y. Yin published a paper on Oct 20, 2026, showing that large language model judges can debias themselves. They tested two tricks: a prompt that warns the model about possible bias before it judges, and a follow‑up prompt that asks the model to critique its own answer for bias. Across three state‑of‑the‑art models, the self‑debiasing prompt improved results in 20 of 21 model‑bias pairings, sometimes adding up to 15 points. The older external critic method performed worse on the new bias tests.
Why it matters
Prompt‑based self‑debiasing could make automated AI evaluations more trustworthy for developers and researchers.
