On two long-document QA benchmarks, MMLongBench-Doc and LongDocURL, the authors compared input modalities, retrievers, and static versus agentic pipelines across Qwen3.5-4B/9B/27B and Claude Sonnet 4.5. The main result is that how you retrieve matters more than which retriever you use. Sending top-k page images beats sending text, and a single off-the-shelf cross-encoder rerank matches a much heavier multi-stage LLM reranking pipeline. Image retrieval also uses only a seventh to a quarter of the tokens of full-page input.
Agentic tool use only pays off with a large enough answering model. The six-tool agent trails static page input on Qwen3.5-4B and 9B, draws level at 27B, and only leads with Sonnet 4.5. On MMLongBench-Doc that lead is +11.7 points over same-reader top-5 retrieval; on LongDocURL it is within noise.
No single pipeline wins every question. An oracle that picks the best pipeline per question gains about 13 points over the best single pipeline, but simple routing by evidence type recovers almost none of that gain, so the variation across questions remains hard to exploit.