A new multi-agent system called KnowForge beats much bigger all-in-one models at answering questions over huge, messy document collections - by refusing to let one model both search and read.
KnowForge splits retrieval-augmented question answering into specialist jobs instead of handing one model every tool. A Locator agent only searches a hybrid knowledge-point index; a Deep Reader agent only reads the text, images and tables that search turns up; a rule-based judge forces a second search round when the first pass misses something; and a synthesizer checks the final answer's format. Running on a self-hosted Qwen3.6-35B-A3B model with just 3 billion parameters active at a time, it scored an F1 of 47.5 on the 2,441-question, 3,368-PDF M3DocVQA benchmark - 11.0 points above the published M3DocRAG system and 7.1 points above a one-shot retrieve-and-answer baseline on the same index and model. On MultiHop-RAG it hit 73.5 percent accuracy, above the best published RAG and GraphRAG scores (71.2), and on the single-document MMLongBench-Doc it lifted the same base model from 34.8 to 47.0, roughly matching GPT-4o reading raw page images (46.3).
The paper's ablations are the real argument: on M3DocVQA, the Deep Reader agent is worth 3.8 points; on MultiHop-RAG, 9.8 points; on MMLongBench-Doc, zero - a spread the authors say tracks how much each benchmark rewards careful reading over simply locating the right passage. Merge the two agents into one model with every tool and twice the rounds, and the score drops 1.9 F1, because the merged agent keeps searching instead of settling down to read.
That's a useful data point against the instinct to just give an agent more tools and more turns - sometimes the fix is a narrower job description, not a bigger one.