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

Researchers introduce DART, a training-free routing framework that dynamically allocates computational thinking budgets in hybrid reasoning models by sampling cheap draft responses and using agreement patterns to decide between direct answers and extended reasoning. The approach achieves significant accuracy improvements on math and code tasks while reducing token consumption by 15-69%, without requiring labeled data or model fine-tuning.

DART addresses a fundamental efficiency problem in modern large language models that support extended reasoning capabilities. As models like OpenAI's o1 demonstrate, allocating extra computational budget for complex problems yields better results, but always engaging this expensive process wastes resources on simple queries. The framework's innovation lies in its training-free approach: by generating two quick draft responses without extended thinking, DART observes whether they agree. Consensus signals an easy problem warranting direct answers, while disagreement triggers entropy-based budget prediction for extended reasoning.

The research builds on growing recognition that not all queries require equal computational investment. Previous routing systems typically demanded labeled training data or preset thinking budgets, creating practical bottlenecks. DART's elegance stems from leveraging the model's own uncertainty signals as evidence, eliminating these requirements entirely. This approach generalizes across model scales from 0.6B to 32B parameters and works with API-only hosted models, broadening accessibility.

For the AI industry, DART has meaningful implications for deployment costs and efficiency. On mathematical reasoning benchmarks, the framework achieved up to 9-point accuracy gains on olympiad-level problems while cutting thinking tokens by 15-69%. Code reasoning showed even more dramatic results, with 22.5-point accuracy improvements and 51-63% token reductions under execution-based equivalence testing. These metrics matter because extended reasoning tokens directly correlate with inference costs and latency.

Developers building AI applications will likely benefit from similar routing strategies that optimize inference economics without sacrificing accuracy. The training-free nature makes DART immediately applicable to existing model deployments, potentially reducing operational costs significantly for organizations using hybrid reasoning systems at scale.

• →DART routes queries between direct answering and extended reasoning using draft agreement patterns, eliminating the need for labeled training data
• →Math reasoning accuracy improved up to 9 points on Olympiad-level problems while reducing thinking tokens by 15-69%
• →Code reasoning achieved 22.5-point accuracy gains with 51-63% token reduction under execution-based equivalence testing
• →The framework generalizes across model scales from 0.6B to 32B parameters and works with API-only hosted models without gradient updates
• →Draft entropy prediction enables adaptive thinking budgets that allocate computation proportionally to problem difficulty