In the main benchmark test, the highest-scoring configuration paired sentence-level context for each visual element with context representing the document as a whole. It recorded 34.91% hard accuracy and 36.84% soft accuracy.
The choice was how to package the text
The paper focuses on CEMMKG, a method that adds different levels of text when a multimodal knowledge graph is built. It distinguishes local context tied to an individual visual element from global context representing the whole document. The research question is whether changing that local context—from a sentence to a longer passage or summary—changes performance, and whether the approach can be used with different multimodal retrieval-augmented generation, or RAG, methods.
Context is used differently at different stages. During image-to-graph construction, the method supplies the full constructed context. During multimodal fusion—the stage where information from different modalities is combined—it uses only the local context. The design therefore carries document-wide information into graph construction while keeping the later fusion step focused on text linked to a particular visual element.
The primary evaluation used 106 questions from the VisionHeavy subset of MMLongBench-Doc. The subset was selected for vision-intensive documents and questions whose evidence could depend on information that was not local to one passage. Answers were free-form: Llama-3.1-70B-Instruct first converted them into canonical answers, and the released type-aware deterministic scorer then assigned the benchmark scores.
The paper reports two versions of accuracy: hard accuracy, the proportion of responses receiving full credit, and soft accuracy, the mean score across questions. Within each underlying RAG framework, retrieval procedures and hyperparameters stayed the same across context configurations. Generation was deterministic, with temperature set to 0 and top-p set to 1.
What the benchmark showed
On MMGraphRAG, the reference-sentence setup was the top tested configuration. It reached 34.91% hard accuracy and 36.84% soft accuracy. A reference-paragraph summary reached 31.13% hard and 32.82% soft. By contrast, full reference paragraphs produced 22.64% soft accuracy, below the 24.17% soft score for native MMGraphRAG.
The benchmark’s simpler comparisons set a lower baseline. Direct Inference scored 11.32% on both strict and soft accuracy, while native MMGraphRAG scored 23.58% strict and 24.17% soft. These figures show the spread among the tested systems.
Results varied by visual type. For figure questions, the reference-sentence setup scored 24.70% soft accuracy, compared with 6.25% for MMGraphRAG. Its soft accuracy for chart questions was 31.25%, while the reference-paragraph summary led on table questions at 26.46%, compared with 22.46% for MMGraphRAG. The categories were not mutually exclusive, and the paper reports no uncertainty estimates for these category-level figures.
One broad category showed a similar pattern. On Misc questions, the reference-sentence configuration reached 54.01% soft and 51.22% hard accuracy, compared with 34.15% for both measures in native MMGraphRAG. The analysis notes that Misc questions may still involve multimodal information, so the label should not be treated as proof that the task was purely textual.
Researchers also tested the context design with RAG-Anything, using a selected subset of questions involving cross-page references. Native RAG-Anything recorded 28.89% strict and 36.18% soft accuracy; the reference-sentence configuration recorded 35.56% and 41.12%. The subset size was not reported, and no uncertainty estimates were provided, making this a limited comparison across methods.
A result bounded by its test
Taken together, the figures show that the context configurations did not perform alike. The highest tested overall score came from reference-sentence context combined with document-level global context, while full reference paragraphs scored below native MMGraphRAG on soft accuracy.
The supplied analysis does not establish that reference-sentence context is universally optimal; it identifies that configuration as the highest-scoring option among those tested. The primary result comes from 106 questions in the VisionHeavy subset, and the paper reports no confidence intervals, statistical tests or run-to-run variability.
The paper says all experiments used publicly available models and datasets. It also discloses AI use in textual graph generation, visual information processing and multimodal information fusion. The work is an arXiv version 1 preprint dated 26 August 2026.
Paper data and sources
Original title: Multi-Granularity Context-Enhanced RAG over Multimodal Knowledge Graphs
Authors: Zongyu Wu, Yilong Wang, Xiaochen Wang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text