BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries
World Action Models (WAMs) commonly couple video and action experts through dense, layer-wise Mixture-of-Transformers (MoT) attention, exposing action prediction to irrelevant visual features while enforcing matched expert depths. This coupling hinders selective visual-feature use and introduces redundant computation. We introduce Bridge-of-Experts (BoE), instantiated as Bridge
Reading in Plain English mode — toggle below to switch
World Action Models (WAMs) commonly couple video and action experts through dense, layer-wise Mixture-of-Transformers (MoT) attention, exposing action prediction to irrelevant visual features while enforcing matched expert depths. This coupling hinders selective visual-feature use and introduces redundant computation. We introduce Bridge-of-Experts (BoE), instantiated as Bridge
This research introduces a novel approach to integrate multimodal information more efficiently in World Action Models, addressing issues of irrelevant visual features and redundant computation. For ML engineers, understanding this decoupling of expert depths and selective information retrieval is crucial for building more robust and efficient multimodal AI systems.