Browserbase and Microsoft researchers developed a visual-rubric verifier that eliminates false positives in web agent benchmarks, exposing massive accuracy gaps overlooked by conventional LLM judges and enabling higher-quality model training.
FireBrief
FireBrief