1 paper
Yansen Han, Hongxin Sun, Tao Lin
Flow matching enables likelihood-free training, yet alignment methods increasingly reuse conditional flow matching (CFM) losses as endpoint negative log-likelihoods (NLLs) and thei…