HRIL: Learning Multimodal Synergy via Higher-Order Tensor Modeling
HRIL targets signals that only appear when modalities are modeled together.
The paper frames multimodal “synergy” as higher-order statistical dependence across modality embeddings. Its method builds an empirical cross-moment tensor, applies Tucker decomposition, and adds a regularizer meant to keep capacity for multi-way coupling. The authors report consistent gains over multimodal contrastive baselines, especially on synergy-heavy tasks. Accepted at NeurIPS 2026. ArXiv · AI/CL/LG's note
The paper frames multimodal “synergy” as higher-order statistical dependence across modality embeddings. Its method builds an empirical cross-moment tensor, applies Tucker decomposition, and adds a regularizer meant to keep capacity for multi-way coupling. The authors report consistent gains over multimodal contrastive baselines, especially on synergy-heavy tasks. Accepted at NeurIPS 2026. ArXiv · AI/CL/LG's note
score 4