Chain-of-thought reasoning helps models solve harder problems, but it has a downside: models often keep generating after the correct answer is already available. Those extra steps increase serving costs and, the GLOBE paper warns, can replace a correct intermediate result with something worse.
The paper, posted on arXiv as 2610.25970, proposes GLOBE — Geometry-Led Online Behavioral Early Exit via Unit-Sphere Trajectories. Rather than waiting for the model to finish its full chain, GLOBE monitors the model's behavior online and uses geometry on a unit sphere to decide when to exit early.
The abstract frames this as a behavioral early-exit approach: it observes the trajectory of the model's reasoning and stops generation at the point where the correct answer is reachable. The potential benefits are lower serving costs and greater reliability, since the model does not get a chance to talk itself out of a correct answer.