DocMemo Re-queries Long Documents with Bayesian Page Memory, Reaching 71.3% Accuracy on MMLongBench

Long-PDF question-answering systems often select a set of pages once before reasoning begins. If the first retrieval round misses a table or clues spanning multiple pages, even a model with stronger reasoning capabilities cannot recover. DocMemo turns retrieval into a stateful exploration process of up to three rounds and divides memory into three layers: Document Schema Memory stores document structure, such as sections and page summaries; Page Belief Memory maintains the probability that each page is relevant to the current question; and Question Episodic Memory records evidence already found, unresolved clues, and the history of query rewrites.

In each round, the system uses visual embeddings to retrieve candidate pages, then applies Thompson sampling to balance high-confidence pages against pages that have not yet been sufficiently examined. The reasoning model’s assessment of the evidence is written back into the page beliefs through Bayesian updates. Relevance is also propagated to the two neighboring pages using a decay factor. When encountering information-dense tables, the system can additionally crop local visual regions instead of providing the model only with a thumbnail of the entire page.

On MMLongBench-Doc, which contains 1,082 questions and documents of up to 112 pages, DocMemo achieves 71.3% accuracy, outperforming DocLens at 67.6%. It reaches 73.3% on table questions and 78.8% on unanswerable questions. When controlling for the same Qwen3.5-9B, ColQwen2.5, page budget, and three-round limit, it still leads SimpleDoc, 71.3% to 69.3%. Removing all memory reduces the score to 68.47%, while removing Bayesian updates lowers it to 68.80%. Replacing Thompson sampling with a greedy strategy that selects only the pages with the highest posterior mean reduces accuracy from 71.28% to 68.62%.

However, correctness is evaluated primarily by GPT-4.1 as a binary judge, which may conflate differences in answer style with actual correctness. Several baseline figures are also taken from their respective papers or leaderboards. The public code currently has only two commits and includes page embedding, summarization, vLLM serving, and question-answering pipelines, but provides neither the data nor complete reproduction artifacts. Before practical adoption, users should still measure the cost of offline summarization, staleness issues when long documents are updated, and whether page beliefs can remain reliably calibrated across different layouts and languages.