Gate test-time compute on the model's calibrated confidence to stop overspending tokens

arXiv 2607.01612·medium signal

A new July 2026 arXiv paper (2607.01612) trains models via a C3RL reward that combines correctness and calibration so stated confidence actually tracks accuracy, then uses that signal to allocate test-time compute adaptively — stop sampling early when confident, scale up only when uncertain. For self-consistency or reasoning-budget pipelines, this converts a fixed sampling budget into a demand-driven one, cutting tokens on easy inputs without hurting hard-input accuracy. The takeaway: calibrate first, then let confidence drive the compute dial.