Filtering partial reasoning trajectories early to focus computation on the most promising candidates before completion.