Developed MDLM, the d-LLM powering ByteDance's Seed Diffusion LLM
Created Duo, the first d-LLM supporting few-step generation (accepted at ICML 2025; oral presentation at ICLR 2025 DeLTa Workshop)
Developed Eso-LMs, the first d-LLM supporting KV-caching
Published multiple papers at top-tier conferences including ICML, NeurIPS, and ICLR, such as:
— 'The Diffusion Duality' (ICML 2025; oral at ICLR 2025 DeLTa Workshop)
— 'Simple and Effective Masked Diffusion Language Models' (NeurIPS 2024; spotlight at ICML 2024 AccMLBio Workshop)
— 'Diffusion Models With Learned Adaptive Noise' (NeurIPS 2024 spotlight)
— 'Backpropagation through Combinatorial Algorithms' (ICLR 2023)
— 'Learning Equations for Extrapolation and Control' (ICML 2018)
Invited to give talks at Google DeepMind, Meta (FAIR), NVIDIA, Cerebras, Databricks, Genesis Therapeutics, Seoul National University, Radboud University, and others