An agent can lose a long task through a poor intermediate decision even when it has the knowledge needed to finish. The Turn-BoN paper studies a way to spend more inference compute at that decision point: propose several actions, choose one, and execute that choice.
How the two-model process works
A generator proposes N candidate actions at each turn, and a selector chooses the action to execute. The authors test whether off-the-shelf open-weight models can compare the candidates directly, without training a separate action-value critic.
Only the selected candidate is executed in the environment. That distinction matters when alternatives would otherwise require copying a user’s machine or restoring an earlier environment state. The paper contrasts this with selecting among multiple complete trajectories after they have already run.
What the terminal experiments suggest
The authors report improved accuracy and consistency on long-horizon terminal tasks. Their trajectory analysis describes more instruction following, debugging and testing, alongside fewer attempts to complete the task.
These are findings from the studied model and selector configurations. The paper’s results should be read with the candidate count, generator, selector and task setup attached. A larger candidate pool also changes the amount of inference work, so accuracy alone cannot establish the best deployment choice.
What to measure in a practical trial
For a developer evaluating this approach, compare an ordinary agent run with the same agent plus selection under a fixed compute budget. Record successful task completion, total model calls, time and the actions that were rejected.
The useful question is whether selection prevents costly mistakes early enough to justify its overhead. A selector that produces persuasive explanations still needs evaluation against actual task outcomes. This paper offers a concrete experimental design for that comparison; it does not establish reliable selection across every interactive environment.
Comments