Stanford/Berkeley/NVIDIA's LLM-as-a-Verifier beats GPT-5.5 on Terminal-Bench

Stanford, Berkeley, and NVIDIA jointly released LLM-as-a-Verifier, an agent verification framework that combines with any agent harness or model. Scaling verification compute lets the framework outperform GPT-5.5 and Claude Mythos on Terminal-Bench and SWE-Bench Verified. Lead authors include Ion Stoica (Databricks), Azalia Mirhoseini (ex-Anthropic), and Marco Pavone (NVIDIA).