How Much Can Language Models Gain from Test-Time Computation?
Published in arXiv preprint, 2026
Recommended citation: Bangji Yang, Jingyuan Li, Jiajun Fan, Yi Evie Zhang, Ruihan Guo, Hongbo Ma, Neil He, Chumeng Liang, Qinglong Zheng, Zhanghan Ni, Ge Liu. "How Much Can Language Models Gain from Test-Time Computation?" arXiv preprint arXiv:2610.01110, 2026. https://arxiv.org/abs/2610.01110
SELF-POT is a benchmark and evaluation framework for how much a model gains from test-time computation, across competition mathematics, competitive programming and agentic workflows. Under one budget rule that charges every model call, including selection and critique, in dollars, it compares direct inference with parallel sampling and self-revision, and asks when a cheaper model with more inference can match a stronger one.