Discrete Tilt Matching

Discrete Tilt Matching (DTM) is a novel likelihood-free method for fine-tuning masked diffusion large language models (dLLMs), addressing the intractability of sequence-level marginal likelihoods in RL. It recasts fine-tuning as state-level matching, using a weighted cross-entropy objective with control variates for stability, and achieves strong results on various tasks like Sudoku and Countdown.