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

2026-06-22

AI Summary

This work addresses the challenge that existing hybrid reasoning models struggle to dynamically allocate reasoning budgets, often leading to over-reasoning on simple problems and under-reasoning on complex ones. The authors propose a training-free adaptive routing mechanism that samples two zero-thought drafts and uses their consistency to decide whether to answer directly; if inconsistent, it predicts the required reasoning budget based on draft entropy. This approach achieves dynamic budget allocation for the first time without labeled data or gradient updates, leveraging the model’s own generated signals for decision-making and demonstrating compatibility across diverse model scales and architectures. Evaluated on mathematical and code reasoning tasks, the method improves accuracy by up to 9.0 and 22.5 percentage points while reducing reasoning tokens by 15–69% and 51–63%, respectively, confirming its efficiency and generality.

Abstract

Hybrid reasoning models can answer directly or spend extra tokens on extended thinking. A practical router should choose between these modes for each query, so easy problems avoid unnecessary reasoning and hard problems receive enough budget to finish the answer. Existing routers move in this direction, but they typically require labeled training data or fix thinking budgets up front, ignoring answer-level evidence from the model itself. We introduce DART, a training-free routing framework that samples two cheap no-think drafts, accepts direct answering when the drafts agree, and predicts a thinking budget from draft entropy when they disagree. Across the main comparisons, DART preserves or improves always-thinking accuracy in most settings while reducing thinking-token use. On math reasoning, accuracy improves by up to $+$9.0 points on Olympiad-level problems while thinking tokens drop 15-69%. On code reasoning under execution-based equivalence, accuracy improves by up to +22.5 points while thinking tokens drop 51-63%. The Stage~1 signal extends across model scales (0.6B-32B), model families, and API-only hosted settings, with no labeled data and no gradient updates required.

Problem

• hybrid reasoning
• thinking budget
• adaptive routing
• training-free
• reasoning efficiency

Innovation

• training-free routing
• adaptive thinking budget
• draft agreement
• entropy-based prediction
• hybrid reasoning