A prompt-only judge decomposition hits 92.7% on RewardBench, matching a fine-tuned 32B specialist
arXiv 2608.18303·medium signal
SESSE replaces binary preference judging with five stages, sketch, expand, sort, summarize, evaluate, where the sub-questions are mined from the judge's own past error cases rather than hand-written rubrics. On 1,000 RewardBench items it reached 92.7%, near parity with chain-of-thought and matching the fine-tuned RISE-Judge-32B, with no oracle responses and no training. The per-criterion vote evidence is the real payoff, since it lets you tell label ambiguity apart from genuine judge failure.