Official implementation of the Findings of EMNLP 2026 paper.
A training-free router for hybrid reasoning models: cheap no-think drafts decide when to think, and how much.
🌐 Project Page · 📄 Paper · ✨ Overview · ⚙️ Installation · 🚀 Usage · 📊 Results · 📌 Citation
Qwen3-8B on one RTX A6000, played at 8× speed. The project page replays it live.
- 2026-10 · Presenting at EMNLP 2026 in Budapest, 24 to 29 October. See you there.
- 2026-08 · Accepted to Findings of EMNLP 2026.
- 2026-06 · Paper released on arXiv, and the reference implementation is public here.
DART is the reference implementation of a training-free router for hybrid reasoning models (models that can answer directly or spend extra tokens on extended "thinking"). DART uses cheap no-think drafts as a query-level difficulty probe:
- Stage 1, SC-Route (draft agreement). Sample
K=2no-think drafts and accept the answer when they agree under a pluggable equivalence function, and route only disagreement cases to thinking. - Stage 2, budget probing. For routed (disagreement) queries, allocate a query-specific thinking budget instead of always thinking to the cap.
No labeled difficulty data and no gradient updates are required, so DART works with text-only API access to closed hybrid models.
Paper: DART: Draft-Agreement Routing for Training-Free Adaptive Thinking Budgets in Hybrid Reasoning Models (arXiv:2606.23181).
dart/
sc_router.py Stage 1: Self-Consistency Routing (draft agreement)
sc_budget_router.py Stage 1 + Stage 2: budget-probing router
answer_extraction.py answer extraction and equivalence (math / mcq / code)
config.py router configuration dataclasses
model_client.py backend-agnostic model-client interface (plus a mock)
example.py runnable demo using the mock client (no GPU needed)
The core routers depend only on the Python standard library:
git clone https://github.com/js-lee-AI/DART.git && cd DART
python example.pyFor real models, install a backend SDK (for example pip install openai for an OpenAI-compatible vLLM server) and implement the model-client contract below.
from dart import SCRouter, SCRouteConfig
router = SCRouter(my_model_client, SCRouteConfig(task_type="math"))
result = router.route("What is the 7th prime number?")
print(result.action, result.answer, result.thinking_tokens)The routers are backend-agnostic. Provide any object with a generate method:
generate(prompt, enable_thinking, temperature, max_tokens, thinking_budget=None) -> {
"text": str, # final answer text
"total_tokens": int,
"thinking_tokens": int, # 0 when enable_thinking is False
"budget_truncated": bool, # True if the thinking trace hit thinking_budget
}See dart/model_client.py for the abstract base class and a mock used by the example. Implement it for your hybrid model (a local vLLM or transformers server, or a hosted think/no-think API).
The paper evaluates on public datasets, which are not redistributed here. Obtain them from their original sources under their respective licenses: MATH-500, OlympiadBench, HumanEval, MBPP, and AIME 2024/2025.
DART matches or exceeds always-thinking (AT) accuracy on 13 of 14 model-benchmark pairs while reducing thinking tokens by 32–73%. Accuracy in %, with the change vs AT in parentheses; Think↓ is the thinking-token reduction vs AT. NT = no-think, AT = always-think.
| Model | Benchmark | NT | AT | DART (vs AT) | Think↓ |
|---|---|---|---|---|---|
| Qwen3-8B | MATH-500 | 76.6 | 85.6 | 88.2 (+2.6) | 67% |
| Qwen3-8B | OlympiadBench | 49.8 | 71.5 | 69.8 (−1.7) | 45% |
| Qwen3-8B | HumanEval | 60.4 | 59.1 | 78.7 (+19.6) | 55% |
| Qwen3-8B | MBPP | 59.5 | 64.2 | 68.9 (+4.7) | 46% |
| Qwen3-14B | MATH-500 | 81.2 | 87.6 | 87.6 (+0.0) | 73% |
| Qwen3-14B | OlympiadBench | 51.5 | 53.0 | 62.0 (+9.0) | 51% |
| Qwen3-14B | HumanEval | 71.3 | 66.5 | 78.7 (+12.2) | 60% |
| Qwen3-14B | MBPP | 64.6 | 64.6 | 68.1 (+3.5) | 51% |
| Qwen3-32B | MATH-500 | 82.2 | 86.2 | 88.5 (+2.3) | 69% |
| Qwen3-32B | OlympiadBench | 50.5 | 54.0 | 58.5 (+4.5) | 52% |
| Qwen3-32B | HumanEval | 79.9 | 72.6 | 95.1 (+22.5) | 63% |
| Qwen3-32B | MBPP | 65.8 | 65.8 | 71.2 (+5.4) | 51% |
| DeepSeek-V3.2 | MATH-500 | 84.8 | 88.4 | 92.6 (+4.2) | 56% |
| DeepSeek-V3.2 | OlympiadBench | 63.6 | 66.1 | 69.1 (+3.0) | 32% |
The only regression is Qwen3-8B on OlympiadBench (−1.7). DeepSeek-V3.2 is evaluated through the hosted API. No labeled data or gradient updates are used. See the paper for the SC-Route (Stage-1-only) and routing-rate columns.
@article{lee2026dart,
title={DART: Draft-Agreement Routing for Training-Free Adaptive Thinking Budgets in Hybrid Reasoning Models},
author={Lee, Jungseob and Hong, Seongtae and Lee, Seungjun and Seo, Jaehyung and Son, Junyoung and Eo, Sugyeong and Park, Chanjun and Park, Hyeongju and Moon, Hyeonseok and Lim, Heuiseok},
journal={arXiv preprint arXiv:2606.23181},
year={2026}
}The code in this repository is released under the MIT License. The paper itself is distributed under CC BY 4.0 via arXiv.