Assistant Professor, Nanjing University
1 paper at NeurIPS 2025
We introduce the first theoretical framework for analyzing LLM reasoning errors, and bridge two typical sampling-based test-time scaling methods to achieve both low error and fast convergence.