Sparse Causal Latent Features for Robust Multimodal Learning under Distribution Shifts
Companion Proceedings of the ACM Web Conference 2026 (TML Workshop — Best Paper Award),
Liang Cao, Weide Liu, Yan Qin, Zhenghua Chen, Zhuo Chen, R. Bhushan Gopaluni
Abstract
A framework that enforces robustness through sparse causal feature selection using a gating mechanism, enhancing generalization in out-of-distribution multimodal learning scenarios.