Reasoning models avoid overthinking with better stopping strategies
Constructing Pareto Frontiers of Optimal Reasoning Termination Methods for Overthinking in Reasoning Models
Artificial Intelligence
Summary
Sometimes reasoning models keep thinking longer but stop getting better answers, a problem called overthinking. The authors created a new way to measure how much extra thinking hurts accuracy and token use by checking answers at many points along the model’s thought process. They developed a framework to compare different methods that decide when the model should stop thinking to save time without losing accuracy. Using this, they made a big benchmark called OTBench to analyze many models and tasks under these conditions.
What this means in practice
- •For ai system designers: Choose reasoning termination methods that optimize accuracy and cost based on token-accuracy trade-offs measured across tasks and models.
- •For language model integrators: Incorporate dynamic early-exit strategies to reduce unnecessary computation while maintaining answer quality for deployed reasoning models.
Authors
Weixin Guan, Liang Li, Shiyu Xie, Jiapeng Liu, Bing Li, Peng Fu, Chengyang Fang, Xiaoshuai Hao, Can Ma, Weiping Wang
Abstract
Reasoning models can benefit from test-time scaling, but continued reasoning may become ineffective or harmful, a phenomenon termed overthinking. Evaluations against full thinking compare accuracy and token usage but lack references for overthinking severity and strategies' theoretical mitigation potential. To characterize overthinking at the model-task level, we first introduce Step-wise Truncation Probing (STP), repeatedly generating answers from multiple retained prefixes along each fixed sampled reasoning trajectory. Using each trajectory's earliest measured position attaining maximum empirical accuracy as a reference, we define Overthinking Ratio and Accuracy Degradation as the fraction of thinking tokens generated afterward and the accuracy decrease from this reference to full thinking, respectively. Secondly, we then formulate dynamic routing, fixed budgets, and dynamic early exit within a unified framework to quantify the theoretical mitigation potential of different termination strategies. Based on this unified framework, we use the complete STP measurements to construct a token-accuracy Pareto frontier for each strategy under a common normalized thinking-token budget. We further develop OTBench, a reusable benchmark built from the STP process, which produced 70.5B reasoning tokens and the corresponding probing measurements across six models and four benchmark groups.