ResearchBenchmarksStepwise Confidence Attribution on PRM800KFollow0.8181AUROCNLI-Mean0.60750.6621750.716850.771525May 19, 2026Evaluation ResultsMethodMethodLinksAUROCAUCPRACC@80%NLI-MeanSource LLM Traces=GPT-4Source LLM Traces=GPT-42026.050.81810.901985.73NLI-MaxSource LLM Traces=GPT-4Source LLM Traces=GPT-42026.050.76660.907478.99Cos-MeanSource LLM Traces=GPT-4Source LLM Traces=GPT-42026.050.68210.834378.6GIBSSource LLM Traces=GPT-4Source LLM Traces=GPT-42026.050.65560.820375.7Cos-MaxSource LLM Traces=GPT-4Source LLM Traces=GPT-42026.050.61560.78477.34