LaDi-RL-Code-8B

LaDiR-Code-8B further trained with reinforcement learning (LaDi-RL). LaDiR (Latent Diffusion Reasoner) lets an LLM reason in a continuous latent space: a VAE compresses a reasoning chain into 64 latent tokens, the LLM generates this latent block with flow matching, then decodes the answer conditioned on it. LaDi-RL optimizes both the latent block and the decoded answer with execution rewards.

Code: github.com/mk322/LaDiR · Papers: LaDiR, LaDi-RL

git clone https://github.com/mk322/LaDiR && cd LaDiR && pip install -e .
python -m eval.generate --checkpoint mk322/LaDi-RL-Code-8B

@article{kang2026ladirl,
  title   = {LaDi-RL: Latent Diffusion Reasoning Prevents Entropy Collapse in Reinforcement Learning},
  author  = {Kang, Haoqiang and Zhang, Yizhe and Kuang, Nikki Lijing and Ma, Yi-An and Qin, Lianhui},
  journal = {arXiv preprint arXiv:2602.01705},
  year    = {2026},
  url     = {https://arxiv.org/abs/2602.01705}
}

@inproceedings{kang2026ladir,
  title     = {LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning},
  author    = {Kang, Haoqiang and Zhang, Yizhe and Kuang, Nikki Lijing and Majamaki, Nicklas and Jaitly, Navdeep and Ma, Yi-An and Qin, Lianhui},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2510.04573}
}