diffusion-reasoning/gdsd_sudoku_llada

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

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

GDSD is a reinforcement learning (RL) framework for diffusion large language models (dLLMs) that bypasses the intractability of policy likelihood. It distills the denoiser of dLLMs from an advantage-guided self-teacher derived from the closed-form optimum of reverse-KL regularized RL. This method avoids the Training-Inference Mismatch (TIM) biases common in ELBO-based approaches, leading to more stable training and improved performance on planning, math, and coding benchmarks.

• Paper: https://arxiv.org/abs/2605.29398
• Repository: https://github.com/GaryBall/GDSD

Citation

If you find GDSD helpful, please consider citing:

@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}, 
}