I develop generative models for flexible and reliable inference, with a focus on masked and discrete diffusion, probabilistic sampling, self-correction, and risk-aware post-training. I also apply these methods to communication and sensing systems.
My work has appeared at ICML, CVPR, NeurIPS, and IEEE Transactions on Wireless Communications. I am advised by Prof. Jeffrey G. Andrews and Prof. Hyeji Kim at UT Austin.
Taekyun Lee*, Jiyoung Yun*, Jeffrey G. Andrews, Hyeji Kim (* equal contribution)
NeurIPS 2026
Decodes many users' messages from a single noisy aggregate via masked-diffusion refinement, matching joint belief propagation accuracy while running 6–100× faster.
Seunggeun Kim*, Jaeyeon Kim*, Taekyun Lee*, Yuyuan Chen*, Yilun Du, Sham Kakade, Sitan Chen (* equal contribution)
DiffuLM @ NeurIPS 2026
Elicits genuinely any-order inference from masked diffusion models through insertion-based interfaces and latent-space generation.
Heasung Kim, Taekyun Lee, Hyeji Kim, Gustavo de Veciana
NeurIPS 2025 Spotlight
Generates informative tail samples to improve worst-case performance in risk-aware downstream fine-tuning.
