Look Back, Think Ahead: Visual Memory on Demand for Efficient Multimodal Reasoning
ViMoD addresses computational cost in multimodal LLMs by managing visual context efficiently during reasoning. It introduces DART to create compact, content-adaptive visual groups and TRACE to adaptively route access to original fine-grained evidence based on evolving needs. This allows the model to maintain high performance with limited visual budgets while preserving access to detailed information when required.