@inproceedings{DiscFlowMatching,
author = {Li, Kang and Sarkar, Bidipta and Xiong, Zheng and Frey, Sascha and Wang, Zilin and Zejnullahu, Frensi and Backhouse, Alfred and Zohren, Stefan and Calinescu, Anisoara and Cucuringu, Mihai and Foerster, Jakob},
title = {Discrete Flow Matching is a Surprisingly Effective Post-training Method to Address Compound Error in Autoregressive Models},
year = {2025},
isbn = {9798400722202},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3768292.3770442},
doi = {10.1145/3768292.3770442},
abstract = {Autoregressive (AR) models suffer from compound error accumulation, where prediction errors propagate and amplify throughout sequence generation. We introduce unified discrete flow matching (DFM) as a post-training framework that eliminates this fundamental limitation of AR models. By converting pretrained AR models into flow-based generators, our approach enables bidirectional context integration and iterative error correction while preserving learned representations. Our framework leverages the pretrained model's distribution as the starting point for flow matching, reducing the learning burden compared to noise-based approaches. Unlike BERT-style masked models with discrete diffusion that suffer from mode averaging, our unified flow matching framework successfully learns discrete patterns, achieving 99.9\% validity on solving mode averaging tasks where traditional methods completely fail. Critically, our experiments demonstrate that unified DFM completely eliminates compound error: while AR models show a Wasserstein distance that is increasing throughout generation progress, unified DFM maintains a nearly flat error profile throughout the entire sequence. This post-training approach achieves a 48.6\% reduction in Wasserstein distance.},
booktitle = {Proceedings of the 6th ACM International Conference on AI in Finance},
pages = {211--219},
numpages = {9},
keywords = {Unified discrete flow matching, compound error elimination, autoregressive models, efficient inference, post-training.},
location = {
},
series = {ICAIF '25}
}