diffusion-reasoning/gdsd_countdown_dream

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

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

Guided Denoiser Self-Distillation (GDSD) is a reinforcement learning framework for diffusion language models (dLLMs). It improves the denoiser of dLLMs by distilling from an advantage-guided self-teacher, bypassing the biases associated with evidence lower bound (ELBO) surrogates used in prior methods. GDSD provides a more stable and effective RL procedure, achieving significant performance gains on planning, math, and coding benchmarks.

Resources

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

Citation

@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}, 
}