CoBa: Compute-Balanced Routing Enhances Test-Time Scaling Efficiency
A new preprint on arXiv (2608.07424) introduces CoBa, a routing policy aimed at optimizing compute usage during AI reasoning tasks. The authors view test-time reasoning as a challenge in resource allocation, deciding whether to concentrate on generating outputs, verifying results, or ending processes. CoBa generates a small set of candidates, conducts broad and low-cost verification, and sends uncertain or valuable candidates for deeper checks. In tests involving 3,129 example-generator tasks like MATH-500 and AMC 2023, CoBa-Routed-Strong achieved a macro accuracy of 85.13%, nearly matching a self-assessment method at 85.20% while using 49.1% fewer tokens. It also matched the performance of a top majority voting method, showcasing a more efficient approach to reasoning within fixed inference limits.
Key facts
- CoBa is a compute-balanced routing policy for test-time scaling.
- It formulates test-time reasoning as a compute-allocation problem.
- CoBa first obtains a small set of candidates, applies cheap verification broadly, and routes uncertain or high-value candidates to stronger verification.
- Evaluated on 3,129 example-generator evaluations spanning MATH-500, AIME 2024/2025, AMC 2023, and procedural symbolic reasoning.
- CoBa-Routed-Strong reaches 85.13% macro accuracy.
- It statistically matches a self-evaluation weighted-voting proxy at 85.20%.
- Uses 49.1% fewer parameter-weighted tokens.
- Matches best-of-16 majority voting performance.
- Paper available on arXiv with ID 2608.07424.
Entities
Institutions
- arXiv