AI RESEARCH
Not All Modalities Are Equal: Instruction-Aware Gating for Multimodal Videos
arXiv CS.CV
•
ArXi:2605.26232v1 Announce Type: new Pre-trained video large language models excel at visual reasoning. However, they struggle when videos arrive with auxiliary streams, such as audio, depth map, or dense temporal evidence. In such a scenario, uniform fusion induces modality interference, allowing irrelevant channels to distract the model.