PUMA Diagnoses LLM Overthinking, Improves Efficiency

Large Reasoning Models (LRMs) often employ extensive Chain-of-Thought (CoT) at test time to tackle complex tasks, but this can lead to "overthinking," where redundant reasoning increases computational cost without guaranteeing accuracy. Existing efficiency methods either risk "deceptive convergence" by focusing on uncertainty or are too slow for real-time diagnosis. To address this, researchers propose the Phase-Momentum Alignment Hypothesis, suggesting that correct reasoning aligns the temporal dynamics of geometric momentum with uncertainty resolution. They formalize this with a Cognitive-Energy Model, characterizing dynamics through Geometric Cognitive Effort (latent velocity, tortuosity) and Entropic Cognitive Uncertainty. PUMA (Phase-Uncertainty Momentum Alignment) is a training-free framework that operationalizes this hypothesis. It uses a tiered diagnostic architecture, combining lightweight phase monitoring with event-triggered geometric analysis to differentiate active exploration from passive stagnation. This enables precise interventions like adaptive truncation or corrections, leading to superior accuracy-efficiency trade-offs across various LRMs and benchmarks.