FLOWREADER uses min-cost flow to solve fragmented multimodal Q&A

A new paper from Ambuj Mehrish and Sebastiano Vascon proposes FLOWREADER, a framework that reframes evidence assembly for multi-modal long document Q&A as a min-cost flow problem. Traditional top-k retrieval treats each piece of evidence independently, failing to capture connections between fragments scattered across text, tables, slides, or figures. FLOWREADER constructs a multimodal node graph where a single scoring vector controls source selection (via MMR), sink selection (via length-aware answerability), and edge costs/capacities. The optimal flow is decomposed into candidate evidence paths, then an entropy-regularized replicator dynamics picks a compact non-redundant subset. Parallel VLM workers under a dual-process gate produce answers, with a System-2 refinement triggered when consistency is low or flow is strained.

Benchmarked on VisDoMBench, FLOWREADER tops PaperTab (58.40, +1.30 over G²-Reader) and SlideVQA (72.93, +0.62), and is competitive on SPIQA, FetaTab, and SciGraphQA. Its macro-average (65.47) is within 0.74 of the strongest baseline. The results show that min-cost flow handles fragmented multimodal evidence effectively, providing a unified way to control scoring, routing, selection, and adaptive compute — a direct improvement over top-k retrieval for complex, dispersed documents.