Learned Stopping Rules Optimize Reasoning Model Computation.

This research explores the effectiveness of learned stopping rules for optimizing computational expenditure in reasoning language models, specifically focusing on when such rules offer advantages over simpler confidence or convergence thresholds. The authors introduce LearnStop, a checkpoint stopper that operates without relying on hidden states. LearnStop predicts the correctness of a reasoning prefix by probing a short answer at fixed budget checkpoints and analyzing online features like answer confidence, entropy, prefix vote share, answer stability, and backtracking-marker density. Across 18 task-model settings, including benchmarks like GSM8K, MATH-500, and MMLU-Pro, the study found that the utility of learned stopping is highly task-dependent. For free-form math problems, learned multi-feature stopping significantly improved the fixed-budget frontier and often outperformed scalar exits. For instance, on GSM8K with Qwen3-32B, it achieved a post-hoc peak adaptive gain of +0.157, with validation-selected operating points preserving positive gains. However, for multiple-choice tasks or extremely difficult settings, simpler scalar rules based on confidence, entropy, or stability proved to be competitive or even superior. The paper concludes that learned stopping is not a universal replacement but a valuable tool particularly when many questions become solvable before the full budget is exhausted, yet lack a single, reliable scalar stopping signal. The study also provides extensive analysis on validation, cost accounting, and robustness.