Test-time search is a technique that lets an agent consider multiple possible actions and use a critic to decide which branch looks most promising. The new paper applies this idea to terminal-based computer-use environments, where the agent only sees a partial view of the state.
The authors propose "Probe-and-Resume," a method that determines when and how to evaluate the agent's execution during this search process. By probing alternative actions and then resuming from the best branch, the agent can avoid committing to a poor early choice.
The work is relevant to building more reliable autonomous agents that operate in command-line or terminal interfaces. Because the environment is partially observable, the critic's role in selecting branches is crucial.