In digital environments, a deployed AI agent may be frozen, not retrainable. But its performance can be improved by trying multiple possible futures from a state and continuing from the best. This is branch-and-commit.
The researchers tested this across four simulated environments with two frozen agents. They measured dispersion: the difference between best and average outcome among continuations. They also distinguished whether branching selects a better action or a lucky random outcome.
Branch-and-commit only helps when dispersion is high. In the main environment it improved reward by 12.8%. Most gain came from selecting lucky random outcomes, not better actions. Verifier choice mattered less than where to branch. Predicting dispersion before branching got most improvement with fewer branches.