TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning

Ruochi Li · Jianzhe Lin · Haoxuan Zhang · Haihua Chen · Junhua Ding · Edward Gehringer · Yang Zhang

📄 Paper | 🏠 Homepage | 💻 Code | 🔎 Case Study

TopoGraphRAG-Bench is a benchmark for layout-grounded multimodal evidence reasoning in GraphRAG. It contains 2,024 question–answer pairs over 201 long documents sampled from MMDocIR (https://huggingface.co/datasets/MMDocIR/MMDocIR_Eval_Dataset), spanning nine domains, including academic papers, legal documents, research reports, and financial reports. Each document averages approximately 45 pages and 13,000 words, with evidence grounded in text, figures, and tables at specific page and layout locations. Questions are constructed bottom-up under three controlled reasoning topologies: single-hop retrieval, bridge-chain reasoning, and multi-source synthesis. Counterfactual validation tests shortcut resistance, modality necessity, and evidence necessity, followed by a stratified human audit. Cross-modal questions account for 38.2% of the benchmark, while 66.8% require figure or table evidence. Each instance includes a gold answer and layout-grounded evidence annotations; bridge-chain and synthesis instances additionally specify the required reasoning content. These annotations support diagnostic evaluation of evidence retrieval, answer generation, and evidence-topology recovery across text-only GraphRAG, page-level visual retrieval, and multimodal GraphRAG systems.

☀️ Benchmark Annotations

annotations/benchmark.json (annotations/benchmark.json) contains 2,024 records stored as a JSON array, with one record per question.

Reasoning topology	Questions	Evidence structure
Single-hop retrieval	302	One layout-grounded evidence unit
Bridge-chain reasoning	1,014	A two-hop dependency connecting bridge and target evidence
Multi-source synthesis	708	Three to six evidence points supporting a higher-level conclusion
Total	2,024

The released bridge-chain subset covers all nine ordered modality paths among text, figures, and tables. Across the benchmark, 78.9% of questions require evidence from multiple pages.

Annotation Format

Fields depend on the question topology. The main fields are:

Name	Type	Description
doc_id	int	Source document identifier; use manifest.json to resolve document artifacts.
question_type	string	One of single_hop, bridge_chain, or synthesis.
question	string	Natural-language question.
sh_id / mh_id / sq_id	string	Question identifier for single-hop, bridge-chain, or synthesis instances, respectively.
final_answer	string	Gold answer for single-hop and bridge-chain questions.
answer	object	Structured gold answer for synthesis questions, described below.
evidence_sources	list[object]	Gold evidence references for single-hop and bridge-chain questions.
num_hops	int	Number of hops for single-hop and bridge-chain questions.
hop_path	string	Evidence modality or ordered modality sequence, such as figure or figure→text.
bridge_entities	list[string]	Intermediate entities connecting bridge-chain hops.
reasoning_trace	list[object]	Annotated bridge-chain hops, including sub-questions, gold results, and source references.
pattern_type	string	Synthesis pattern, such as comparison, contradiction, trend, causal, or panorama.
num_sources_required	int	Number of required synthesis evidence points.
evidence_modalities	list[string]	Evidence modalities for synthesis questions.
cross_modal	bool	Whether the question is annotated as cross-modal.
quality_score	number	Construction-time quality assessment score.

For single-hop and bridge-chain questions, each item in evidence_sources contains:

Name	Type	Description
modality	string	Evidence modality: text, figure, or table.
page_id	int	Page identifier in the source document.
layout_id	int	Layout identifier of the evidence unit.

The parent record's doc_id, together with page_id and layout_id, identifies the source evidence.

For synthesis questions, the answer object contains:

Name	Type	Description
evidence_points	list[object]	Required evidence facts and their source references.
synthesis_conclusion	string	Gold higher-level conclusion.
key_data_cited	list[string]	Key values or data referenced in the structured answer.

Each synthesis evidence point records its supporting fact and source identifiers using source_page, source_layout, and source_modality.

Additional fields retain construction provenance and topology-specific validation results. Not all fields apply to every question type.

Evaluation use: gold answers, evidence annotations, and reasoning traces are evaluation references and should not be provided to the answering system as inputs. This release does not include a separate training split.

📂 Document Artifacts and Metadata

The annotations are accompanied by source document artifacts for the 201 sampled MMDocIR documents. Larger artifacts are packaged as .tar.zst archives.

File	Description
manifest.json (manifest.json)	Maps benchmark document identifiers to document names and artifact paths.
croissant.json (croissant.json)	Croissant metadata and Responsible AI documentation.
archives/loaded_info.tar.zst (archives/loaded_info.tar.zst)	Pre-built loaded_info.json used by the construction pipeline for entity extraction, fact extraction, and QA generation.
archives/doc_pdfs.tar.zst (archives/doc_pdfs.tar.zst)	Original document PDFs.
archives/layout_content.tar.zst (archives/layout_content.tar.zst)	MMDocIR layout-level JSONL files.
archives/page_content.tar.zst (archives/page_content.tar.zst)	MMDocIR page-level JSONL files.
archives/page_images.tar.zst (archives/page_images.tar.zst)	Full-page document images.
archives/layout_images.tar.zst (archives/layout_images.tar.zst)	Cropped figure and table images.
archives/layout_text_images.tar.zst (archives/layout_text_images.tar.zst)	Cropped text-region images.

The image_path fields in loaded_info.json are relative to the extracted dataset root. After extraction, the artifact paths listed in manifest.json resolve relative to the same root.

🚀 Download and Usage

Download the Data Package

pip install -U huggingface_hub

hf download diandianone123/topographrag-bench \
    --repo-type dataset \
    --local-dir data/

To extract the document artifacts, install zstd so that unzstd is available, then run:

for archive in data/archives/*.tar.zst; do
    tar --use-compress-program=unzstd -xf "$archive" -C data/
done

The benchmark annotations are already provided as a regular JSON file and do not require extraction.

Load the Benchmark

import json
from collections import Counter

with open("data/annotations/benchmark.json", encoding="utf-8") as f:
    benchmark = json.load(f)

print(f"Total questions: {len(benchmark)}")
print(Counter(item["question_type"] for item in benchmark))


def get_gold_answer(item):
    if item["question_type"] == "synthesis":
        return item["answer"]["synthesis_conclusion"]
    return item["final_answer"]


for item in benchmark[:3]:
    print(f"
[{item['question_type']}] {item['question']}")
    print(f"Gold answer: {get_gold_answer(item)}")

Construction and evaluation code is available in the GitHub repository (https://github.com/RichardLRC/TopoGraphRAG-Bench). Visit the project homepage (https://richardlrc.github.io/TopoGraphRAG-Bench/) for benchmark details and the interactive case study (https://richardlrc.github.io/TopoGraphRAG-CaseStudy/) for examples of layout-grounded evidence reasoning.

📜 License and Data Use

The new benchmark annotations are released for academic and noncommercial research use.

The original documents and upstream artifacts, including PDFs, page images, layout crops, OCR outputs, and MMDocIR annotations, remain subject to the upstream MMDocIR license and the applicable terms of the underlying documents. We do not claim ownership of these source materials.

Users are responsible for complying with the relevant upstream terms and should cite both MMDocIR and TopoGraphRAG-Bench when using or redistributing the benchmark.

The benchmark is intended for research and evaluation. Its scope is limited to the sampled English-language document corpus and the three defined reasoning topologies. LLM/VLM-assisted annotations may retain errors or biases despite validation and human auditing; benchmark performance should not be treated as a guarantee of reliability in high-stakes applications.

💾 Citation

@article{li2026topographrag,
  title={TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning},
  author={Li, Ruochi and Lin, Jianzhe and Zhang, Haoxuan and Chen, Haihua and Ding, Junhua and Gehringer, Edward and Zhang, Yang},
  journal={arXiv preprint arXiv:2610.09360},
  year={2026}
}

Please also cite the upstream MMDocIR dataset (https://huggingface.co/datasets/MMDocIR/MMDocRAG).