Strategic Scaling of Test-Time Compute: A Bandit Learning Approach

Bowen Zuo, Yinglun Zhu

Scaling test-time compute has emerged as an effective strategy for improving the performance of large language models. However, existing methods typically allocate compute uniforml…