GLOBE

GLOBE is a training-free online early-stopping algorithm that determines when to stop generation by analyzing the geometric shape of the reasoning semantic trajectory.

GLOBE employs a two-stage online control process: a probing stage generates a brief tentative answer and stops if confidence exceeds a threshold; otherwise it enters the trajectory stage, embedding reasoning segments onto the unit sphere and detecting direction reversal and stagnation of net progress to terminate generation.

GLOBE is evaluated on five benchmarks, covering tasks such as mathematical reasoning, tool calling, and agent memory, and is validated on two models.

GLOBE achieves a 49.7% reduction in output tokens, a 51.0% reduction in end-to-end actual elapsed time (E2E), and an average accuracy of 80.58% across the five tasks, comparable to full CoT.

GLOBE's probing branch executes in parallel with the main process, so the probing cost can be masked by the generation process, which is why the E2E reduction (51.0%) is slightly higher than the token reduction (49.7%).

GLOBE's early-stopping decision does not rely on single-step confidence or fixed heuristic rules but on geometric signals of the overall direction of the reasoning trajectory.

The quality of GLOBE's unit sphere embedding depends on the choice of embedding model, and threshold settings may vary with task distribution, requiring tuning in practical deployment.

GLOBE provides a training-agnostic, pluggable early-stopping framework for efficiency optimization of CoT reasoning.

The GLOBE framework is still in the research stage.

The GLOBE framework stores optimization knowledge through explicit, field-partitioned behavior manuals, driving the agent to continuously self-evolve.

The GLOBE framework faces issues such as the growth in scale of the behavior manual and the boundary conditions of cross-category transfer, which await further exploration.

In the GLOBE framework, each optimization round has an "editor" role apply the relevant behaviors and obtain effect signals based on ranking feedback.

In the GLOBE framework, the "reflector" role turns feedback into behavior update proposals.

In the GLOBE framework, the "arbiter" role is responsible for removing redundant proposals to avoid bloat and conflicts in the behavior library.

In the GLOBE framework, the "curator" role formally submits approved changes to the manual.

In the GLOBE framework, behaviors that are continuously proven harmful are automatically deactivated.

Experiments show that combining feedback-driven evolution with moderate review improves effectiveness while significantly reducing negative risks.

The behaviors learned by GLOBE remain effective on unseen product categories, indicating that the evolution of the agent's memory produces transferable optimization capability.