CoLT-8B model

This repository contains our trained model for CoLT, a latent reasoning model without relying on auxiliary image annotations. It could notably reduce the inference time by 10.3× and text decoding time by 20.4×, achieving superior efficiency.

Code: https://github.com/hulianyuyy/CoLT

Our training data is built from OneThinker, an all-in-one reasoning model for image and video, as presented in the paper OneThinker: All-in-one Reasoning Model for Image and Video. Train data is available at https://huggingface.co/datasets/hulianyuyy/CoLT_Train_Dataset.

About CoLT

We introduce CoLT, an latent reasoning model that reduce the inference time by 10.3× and text decoding time by 20.4×, while performing better or comparably with textual reasoning on some tasks.

Unlike most previous latent reasoning methods (e.g., LVR, LaCoT and MoNet in multimodal reasoning), which use auxiliary images to supervise the latent reasoning process and brings costly annotation costs with severely limited the training copus, our method adopt pure textual CoT annotations for training, which can fully reduce the burden on labelling additional auxiliary images and adopt existing textual CoT annotations.

Citations

If you find our work helpful for your research, please consider citing our work:

@inproceedings{hu2026colt,
  title={CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts},
  author={Hu, Lianyu and Qin, Shengqian and Liao, Zeqin and Guo, Qing and Wan, Liang and Feng, Wei and Liu, Yang},
  booktitle={Eurean Conference on Computer Vision},
  year={2026}
}

and the original dataset:

@article{feng2025onethinker,
  title={OneThinker: All-in-one Reasoning Model for Image and Video},
  author={Feng, Kaituo and Zhang, Manyuan and Li, Hongyu and Fan, Kaixuan and Chen, Shuang and Jiang, Yilei and Zheng, Dian and Sun, Peiwen and Zhang, Yiyuan and Sun, Haoze and others},
  journal={arXiv preprint arXiv:2512.03043},
  year={2025}
}