Radar ·
RoPoLL proposes geometric median for LLM judge panels
The paper formalizes LLM evaluator panels under a contamination model and proposes aggregating scores with the geometric median. It reports experiments with 13 judges and corruption rates up to 50%.
Published on July 2, 2026, the paper formalizes LLM evaluator panels under a contamination model and proposes RoPoLL, a method that aggregates scores using the geometric median. The reported experiments include 13 judges and corruption rates up to 50%.\n\nEngineers evaluating LLMs can consult the paper to examine the method and compare robust aggregation with the mean, as the publication suggests. Check the original for the model definition, experimental procedures, and results before applying its conclusions; the available summary gives no further details about those results.
Original source: RoPoLL usa mediana geométrica para agregar painéis de juízes LLM