Inference-Time Dynamic Modality Selection for Incomplete Multimodal Classification (ICLR 2026)

**[Siyi Du](https://scholar.google.com/citations?user=zsOt8MYAAAAJ&hl=en), [Xinzhe Luo](https://scholar.google.com/citations?user=l-oyIaAAAAAJ&hl=en&oi=ao), [Declan P. O'Regan](https://scholar.google.com/citations?user=85u-LbAAAAAJ&hl=en&oi=ao), and [Chen Qin](https://scholar.google.com/citations?view_op=list_works&hl=en&hl=en&user=mTWrOqHOqjoC&pagesize=80&sortby=pubdate)** [![](https://img.shields.io/badge/license-Apache--2.0-blue)](#License) [![](https://img.shields.io/badge/arXiv-2503.06277-b31b1b.svg)](https://arxiv.org/abs/2601.22853)
![DyMo](./Images/overview.jpg)

(a-b) Evidence of the discarding-imputation dilemma: (a-1) vs. (a-2) recovery-free methods (e.g., ModDrop) learn less discriminative features because they ignore highly task-relevant missing modalities {M,T}; (b) recovery-based methods (e.g., MoPoE) generate unreliable reconstructions, e.g., low-fidelity (orange) or misaligned (yellow). (c) Our DyMo, which addresses the dilemma by dynamically fusing task-relevant recovered modalities, improving accuracy by 1.61% on PolyMNIST, 1.68% on MST, and 3.88% on CelebA (Tab 1).

This repository provides **pretrained model checkpoints** for [Inference-Time Dynamic Modality Selection for Incomplete Multimodal Classification](https://openreview.net/forum?id=PWhDUWRVhM¬eId=PWhDUWRVhM). For the **training code, evaluation scripts, and usage instructions**, please refer to the official GitHub repository: 👉 [https://github.com/siyi-wind/DyMo](https://github.com/siyi-wind/DyMo).