Researchers found that tool-call judges for LLM agents accept correct values less often when no reference is provided. They introduce DeRef, a separate model that generates a reference for the judge, improving overall argument AUROC across three benchmarks.