Stanford-led LLM-as-a-Verifier claims SOTA on Terminal-Bench 2.0
超越Claude Mythos和GPT-5.5!斯坦福Agent验证框架拿下SOTA,Transformer作者转发
Stanford, Berkeley and Nvidia introduced LLM-as-a-Verifier, claiming SOTA on Terminal-Bench 2.0 and SWE-Bench Verified. It selects trajectories via score-token granularity, repeated checks and criteria decomposition; ForgeCode accuracy reached 86.4%.
Why it matters: Stanford, Berkeley and NVIDIA proposed LLM-as-a-Verifier, which has a model check another model's solution path step by step, and it topped both Terminal-Bench 2.0 and SWE-Bench Verified.