DiffPDE: Masked Diffusion Language Models as PDE Solver
This work proposes DiffPDE, a framework leveraging discrete diffusion language models for targeted code repair by introducing a localized re-masking and infilling strategy, which achieves competitive accuracy, outperforms same-scale AR models, and significantly accelerates repair.
Wen-Xuan Guo, Yuyang Hong, Lu-Bin Fan et al.
· 0 citations