Learning Adaptive Reasoning Budgets via Constraint-Rectified Training

Foundation models increasingly rely on long chain-of-thought reasoning traces, improving task performance but imposing substantial inference-time token cost. Existing efficient-reasoning methods often shorten outputs through fixed budgets or scalar rewards that mix correctness and length, making the resulting quality-resource tradeoff difficult to control. We propose Constraint-Rectified Training (CRT), a reference-guarded post-training framework for adaptive efficient reasoning. CRT first minimizes reasoning length while preserving accuracy relative to a frozen reference policy, thereby discovering the shortest reliable reasoning regime for the model. It then treats this discovered regime as a learned token budget and refines correctness without allowing reasoning traces to expand back to unnecessary verbosity. Across multiple reasoning benchmarks and foundation-model backbones, CRT improves the accuracy-efficiency frontier over length-aware baselines. Additional redundancy analysis shows that CRT reduces repetitive reasoning structure beyond what is captured by average token count alone. These results suggest that constrained, budget-aware post-training is a practical mechanism for resource-efficient foundation model inference.