DPA-MoE accepted to Findings of EMNLP 2026

DPA-MoE: A Runtime Precision Control Framework for MoE Inference has been accepted to Findings of EMNLP 2026.

DPA-MoE allocates precision across routed experts at runtime rather than fixing it when the model is quantized. A single checkpoint then covers a range of weight-size–quality operating points, so you can move along that trade-off without retraining and without keeping one checkpoint per precision setting.

This is my first research paper, and I led it from the initial idea through implementation, experiments, analysis and writing. What stays with me is the pace of it — an idea became an implementation, the experiments said something I did not expect, the approach changed, and around again.

Thanks to Hoonki Lee, equal contributor on this work, for the discussions and ideas that shaped and sharpened the paper, and to Susie Shin, Hyunmin Kim and Dahyun Lee for their contributions throughout. And to Sangyeob Kim, our corresponding author, for the guidance, feedback and support from start to finish.

The submission is on OpenReview, and the full entry with the citation and bibtex is on the publications page.