How do you tell if a computer use agent actually succeeded?
It’s really two questions: did it execute well (process), and did the user actually get what they asked for (outcome)?
Introducing the Universal Verifier 🧵
It’s really two questions: did it execute well (process), and did the user actually get what they asked for (outcome)?
Introducing the Universal Verifier 🧵