This is the CoT (Chain-of-Thought) efficient version of the
Qwen3-4B-Instruct-2507 model, developed as part of the research presented in the paper
The Art of Efficient Reasoning: Data, Reward, and Optimization.
Art-Qwen3-4B is optimized to produce short yet accurate reasoning trajectories. By using reward shaping and Reinforcement Learning (RL), the training process follows a two-stage paradigm: length adaptation and reasoning refinement. This approach aims to provide the benefits of scaled reasoning while minimizing the heavy computational overhead typically associated with long CoT outputs.
The model was trained on the
DeepScaleR-Easy dataset.
1@inproceedings{wu2026art,
2 title={The Art of Efficient Reasoning: Data, Reward, and Optimization},
3 author={Taiqiang Wu and Zenan Xu and Bo Zhou and Ngai Wong},
4 year={2026},
5 url={https://arxiv.org/pdf/2602.20945}
6}
