OneMillion-Bench: How Far are Language Agents from Human Experts?
OneMillion-Bench (1M-Bench) was created to rigorously test if large language models can function as autonomous agents capable of complex, real-world professional work. By simulating 400 high-stakes tasks across Law, Finance, Healthcare, etc., the benchmark measures agent reliability through economic grounding and strict rubrics. The findings reveal a significant gap between models that merely answer questions and those that can reliably perform grounded, compliant decision-making required for expert labor.