CertJudge Targets Reusable Calibration for LLM Judges

Researchers from Stanford, Harvard, and Hong Kong Baptist University presented CertJudge at ICML to address how to evaluate LLM judges. In scoring AI-generated Lean 4 theorems, the work emphasizes reusing a single calibration even when the judge, prompt, or threshold changes.

2026-07-12 ~ 2026-07-13 · 3 related posts

• Stanford and Others Release CertJudge Research — sanmikoyejo · 2026-07-12
• CertJudge Reuses Evaluations With One Calibration — sanmikoyejo · 2026-07-12
• CertJudge Research Showcased at ICML — sanmikoyejo · 2026-07-13