Watch for style bias — the dominant LLM-judge failure (0.76–0.92) that position-bias fixes won't catch

Future AGI·medium signal

New measurements show style bias (preferring well-formatted, confident-sounding answers regardless of correctness) dominates at 0.76–0.92 across models, far exceeding position bias (≤0.04), yet receives almost no mitigation attention. Practically, your judge may be scoring polish, not substance, so normalize formatting before judging, strip confident framing, and validate the judge against human labels on deliberately ugly-but-correct vs. pretty-but-wrong pairs. Treat the judge as a measurement instrument with its own calibration error, not ground truth.