Yuyuan Chen, Shiyi Wang, Peter Potaptchik, Jaeyeon Kim, Michael S. Albergo
View original ↗Develop an open-source library for fine-tuning Masked Diffusion LLMs using the Discrete Tilt Matching method. This provides a needed alternative to standard autoregressive RL fine-tuning.
Suggested repo: tilt-diff
"Likelihood-free fine-tuning for your masked diffusion models."
Estimated effort: 150h