dart-adaptive-thinking-budget

name	dart-adaptive-thinking-budget
description	Training-free routing framework for hybrid reasoning models. Uses draft agreement signals to adaptively allocate thinking budgets without labeled data or gradient updates.
trigger_words	["thinking budget","adaptive reasoning","draft agreement","hybrid reasoning","reasoning allocation"]
version	1
arxiv	2606.23181v1
authors	Jungseob Lee, Seongtae Hong, Seungjun Lee, et al.
date	2026-06-22T00:00:00.000Z
categories	cs.AI, cs.CL

DART: Draft-Agreement Routing for Training-Free Adaptive Thinking Budgets

Core Insight: Training-free routing that samples two cheap no-think drafts, accepts direct answering when drafts agree, and predicts thinking budget from draft entropy when they disagree.

Key Methodology

Three-Stage Framework

1. Draft Sampling: Generate two cheap "no-think" responses
2. Agreement Detection:
   • If drafts agree → accept direct answer (no extended thinking needed)
   • If drafts disagree → predict thinking budget from draft entropy
3. Budget Prediction: Use entropy signal to allocate thinking tokens

Benefits

• No training data required - fully training-free
• No gradient updates - no fine-tuning needed
• Exponential inference reduction - easy problems skip reasoning entirely
• Cross-scale applicability - works on 0.6B to 32B models

Implementation Pattern

def dart_routing(query, model):
    # Stage 1: Sample two cheap drafts (no thinking)
    draft_1 = model.generate(query, thinking_budget=0)
    draft_2 = model.generate(query, thinking_budget=0)
    
    # Stage 2: Check agreement
    agreement_score = compute_agreement(draft_1, draft_2)
    
    if agreement_score > threshold:
        # Drafts agree → accept direct answer
        return draft_1
    else:
        # Stage 3: Predict thinking budget from entropy
        entropy = compute_draft_entropy(draft_1, draft_2)
        thinking_budget = predict_budget_from_entropy(entropy)
        
        # Generate with predicted budget
        return model.generate(query, thinking_budget=thinking_budget)

Performance Gains

• Math reasoning: +9.0 points on Olympiad problems, 15-69% token reduction
• Code reasoning: +22.5 points accuracy, 51-63% token reduction

When to Use

• Hybrid reasoning models with thinking/non-thinking modes
• Adaptive inference budget allocation
• Training-free routing decisions
• API-only hosted models (no gradient access)

Key Insight

The draft agreement signal is a free complexity indicator - disagreement reveals uncertainty that warrants extended reasoning.

---

References

• arXiv: 2606.23181v1
• Authors: Jungseob Lee et al. (10 authors)
• Categories: cs.AI, cs.CL