AI models trained on multiple data types tend to fall apart when some of that data goes missing at deployment time.
Researchers have proposed MARS (Missingness-Aware Residual-guided Specialization), a framework that treats missing data not as a problem to hide but as a signal to exploit. During training, MARS compares what a model infers from incomplete inputs against what it would infer from complete ones, then uses that gap - the residual - to sort examples into specialized sub-models called experts, each tuned for a different pattern of absence. A second routing mechanism learns to mimic that sorting using only the incomplete data, so the system works at deployment without needing the full picture. To close the gap between training and real-world behavior, the authors add a noise regularization step that stress-tests the imitation router when it drifts from the original.
Most prior approaches to incomplete multimodal learning try to reconstruct or approximate the missing inputs, which means layering an estimation problem on top of the prediction problem. MARS sidesteps that by making the deviation itself the organizing principle - closer to how a doctor reasons from available symptoms than how a photo-editing tool fills in blank pixels. Tested on audio-visual classification, face anti-spoofing, food image recognition, and outdoor scene segmentation, MARS outperformed baseline methods across missing-data scenarios.
The mixture-of-experts architecture here echoes designs popularized in large language models, though applied at a much smaller scale and to a narrower problem - a reminder that routing-based specialization has uses well beyond scaling up chatbots.