This paper proposes SelfBudgeter, a user-friendly, adaptive, and controllable inference framework, to address the problem that inference models that excel at complex problems tend to overthink simple problems. SelfBudgeter integrates a budget estimation mechanism before inference and employs a dual training approach. First, the model learns to predict token budgets in a standardized format. Through a reinforcement learning phase, the model is trained to autonomously plan and strictly adhere to budgets based on problem difficulty. SelfBudgeter outputs budget estimates early in the process, allowing users to predict waiting times and manually control inference length through pre-filled budget fields. Experimental results show that SelfBudgeter dynamically allocates budgets based on problem complexity, achieving an average response length compression ratio of 61% for the 1.5B model and 48% for the 7B model on the GSM8K, MATH500, and AIME2025 datasets, while maintaining near-perfect accuracy.
