ARBITER: Reasoning Trajectory Basins and Majority Vote Failures in Test-Time Sampling

When language models use test-time sampling and majority vote, reasoning trajectories concentrate into non-independent