Two complementary approaches to the central bottleneck of multimodal RAG — not how much evidence is retrieved, but which evidence reaches the reader. We treat selection as structured optimisation over a document graph.
Evidence selection reframed as a constrained dominant-set problem on a query-augmented affinity graph. The query becomes a hard structural constraint; relevance and redundancy balance automatically via a spectral bound — no tuning, no training.
Evidence assembly as a min-cost flow problem on a multimodal node graph. A single scoring vector controls source selection, sink selection, and every edge cost — with a dual-process gate that triggers a System-2 refinement pass only when needed.
Long multimodal documents are engineered to be redundant — findings recur across figures, captions, and prose. Similarity-based retrieval amplifies this, spending its budget on near-duplicates. Both works replace flat top-k ranking with graph-native selection that explicitly rewards relevance and non-redundancy, solved to a global equilibrium with replicator dynamics from evolutionary game theory.