JudgeMoE Improves LLM-as-a-Judge Spearman Correlation via Distributional Aggregation
October 7, 2026
JudgeMoE uses a lightweight aggregator to fuse cached judge score distributions rather than scalar values. This approach improves mean Spearman correlation by +0.079 over uniform log pooling on the original 10-cell benchmark, reducing uncertainty loss in evaluation pipelines.
HOW THIS AFFECTS YOU
●
builderYou can improve evaluation reliability by aggregating score distributions instead of single scalar outputs.
●
researcherThis method demonstrates that score-range choice and soft scoring are critical for stable LLM-based evaluation.