New LLM Position Bias Benchmark: does an LLM keep the same judgment when you swap the answer order? Judge models compare two lightly edited versions of the same story twice, with the order swapped. The median model flips in 45% of decisive case pairs. GPT-5.4 is worst at 66%.

New LLM Position Bias Benchmark: does an LLM keep the same judgment when you swap the answer order? Judge models compare two lightly edited versions of the same story twice, with the order swapped. The median model flips in 45% of decisive case pairs. GPT-5.4 is worst at 66%.