DART: Draft-Agreement Routing for Training-Free Adaptive Thinking Budgets in Hybrid Reasoning Models

DART is a novel training-free routing framework designed for hybrid reasoning models, enabling adaptive thinking budgets based on query difficulty without requiring labeled training data. It samples two no-think drafts to determine whether to answer directly or allocate additional reasoning tokens, resulting in improved accuracy—up to +9.0 points on math reasoning and +22.5 points on code reasoning—while significantly reducing token usage by 15-69%. This approach is applicable across a range of model sizes (0.6B to 32B) and settings, making it a valuable tool for practitioners aiming to optimize resource allocation in AI reasoning tasks.