Geometry based method cuts reasoning steps and speeds up AI answers
GLOBE: Geometry-Led Online Behavioral Early Exit via Unit-Sphere Trajectories
Artificial IntelligenceComputation and Language
Summary
Sometimes AI models think too long after they’ve found a good answer, which wastes time and can cause mistakes. This paper presents GLOBE, a way to watch the model’s reasoning in progress and stop it early when the answer is likely good enough. It uses geometry by representing each reasoning step as points on a sphere and looks for signs the model is going in circles. GLOBE works without extra training and makes AI reasoning almost twice as fast while keeping answer quality similar.
What this means in practice
- •For ai service operators: Reduce operational costs and latency by stopping AI reasoning early without retraining, improving efficiency in question-answering or problem-solving services.
- •For digital assistant developers: Improve responsiveness of assistants that use chain-of-thought reasoning by detecting when reasoning is sufficient and halting further unnecessary steps.
Authors
Xinyue Peng, Zhenlin Liang, Meng Cai, Hua Wang
Abstract
Chain-of-thought (CoT) reasoning can improve problem solving, but models often keep generating after the correct answer is already reachable; the extra steps raise serving costs and may replace a correct intermediate result with an incorrect one. We propose GLOBE (Geometry-Led Online Behavioral Early Exit), an online, training-free algorithm that terminates the model's CoT process early. GLOBE embeds each completed reasoning segment as a point on a unit sphere and reads the shape of the resulting semantic trajectory. GLOBE first applies a trial phase that generates short trial answers in a separate branch, stopping the main stream on an accepted trial after checking confidence against a threshold. If generation continues, a trajectory phase stops the stream once this path repeatedly reverses direction with little net progress. On five benchmarks including mathematical reasoning, tool calling, and agent memory, GLOBE cuts output tokens by 49.7% and active end-to-end wall time (E2E) by 51.0% across two models, and achieves a five-task average of 80.58%, comparable to Full CoT.