Answering questions about long multimodal documents is hard because relevant evidence is scattered across text, tables, figures, and slides, and the total amount of evidence one can use is limited. A new arXiv paper proposes FlowReader, a framework that frames this problem as min-cost flow routing: evidence is treated as a commodity to be routed through a network, with the goal of spending a fixed budget across the most useful facets.
The approach is designed to avoid near-duplicates, which can otherwise consume budget without adding new information. By casting evidence assembly as an optimization problem rather than a greedy selection process, FlowReader aims to balance coverage across different document modalities.
The paper is available on arXiv under identifier 2606.07235. As a single source, it does not offer competing claims, but the abstract leaves implementation details and evaluation results unspecified in the provided text.