HANIA: Planner-Guided Multimodal Graph Evidence Selection for Grounded Question Answering.

arXiv:2608.29088v2 Announce Type: replace Abstract: Multimodal question answering remains sensitive to noisy, incomplete, and weakly grounded evidence. Long unstructured contexts can introduce redundancy and encourage unsupported generation, while flat retriev