diffusion-reasoning/gdsd_code_llada

GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models

This repository contains the model checkpoint for GDSD (Guided Denoiser Self-Distillation), as introduced in the paper GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models.

GDSD is a reinforcement learning (RL) framework designed to improve the denoiser of diffusion large language models (dLLMs). It reduces RL to a likelihood-free self-distillation objective by matching the dLLM's denoiser logits to an advantage-guided self-teacher. This approach bypasses the training–inference mismatch (TIM) biases common in ELBO-based methods and leads to more stable training dynamics.

Resources

• Paper: GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models
• GitHub Repository: GaryBall/GDSD

Citation

If you find GDSD helpful, please consider citing the following work:

@misc{tang2026gdsdreinforcementlearningguided,
      title={GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models}, 
      author={Xiaohang Tang and Keyue Jiang and Che Liu and Qifang Zhao and Xiaoxiao Xu and Sangwoong Yoon and Ilija Bogunovic},
      year={2026},
      eprint={2605.29398},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.29398}, 
}