Learning to Think Fast and Slow for Visual Language Models

💡 Overview

We introduce DualMindVLM, a dual-mode thinking VLM that can automatically switch between fast and slow thinking modes based on the difficulty level of the problem. DualMindVLM is optimized using a simple RL approach built only on question–answer pairs. The approach consists of two stages: The first stage utilizes the output length variation of the pretrained VLM to assign each sample a thinking mode label. The second stage develops dual-mode thinking in the model through GRPO-based reinforcement learning, where half the sampled candidates are guided by the assigned label. Despite its simplicity, DualMindVLM significantly outperforms the base model and achieves performance on par with state-of-the-art visual reasoning models, while maintaining exceptionally high token efficiency.

---

🚀 Release Progress

Component	Status	Notes
🧩 Model	✔️ Released	Available on 🤗 HuggingFace
⚙️ Inference + Evaluation Code	✔️ Released	vLLM-based inference, string-matching evaluation
🔥 Training Code	🕒 Coming Soon	GRPO-based training framework

---

🔗 Citation

If you find this work useful, please cite our paper:

@article{lin2025dualmindvlm,
  title     = {Learning to Think Fast and Slow for Visual Language Models},
  author    = {Chenyu Lin and Cheng Chi and Jinlin Wu and Sharon Li and Kaiyang Zhou},
  journal   = {arXiv preprint arXiv:2511.16670},
  year      = {2025}
}