arxivcs.AI2026-07-19
Is Your Model Thinking or Just Stagnating? PUMA: Diagnosing Reasoning Pathology via Phase-Momentum Alignment
Cheng Yan, Guangyang Ye, Wuyang Zhang, Fan Xu, Zhijun Fan, Xiang Xia, et al.
Test-time scaling empowers Large Reasoning Models (LRMs) to tackle complex tasks via extensive Chain-of-Thought (CoT). However, this often induces the "overthinking" paradox, where redundant reasoning increases computational overhead without guaranteeing accuracy. Existing test-t…