SurveyGraph: Generating Academic Surveys via Methods-Entity Knowledge Graphs

Yifan Yuan · Yancheng Zhu · Zuoyu Zhang · Xiaogang Peng^†

School of Artificial Intelligence, Shenzhen University
^†Corresponding author

EMNLP 2026 Findings

English · 简体中文

SurveyGraph: MEKG construction, graph-driven survey generation, and integrated evaluation.

TL;DR: LLM survey generators usually retrieve papers as a flat pile of documents, so the result reads like a collage of summaries. A good survey explains how a field evolved — which work extends which, which methods compete, which benchmarks mark turning points. SurveyGraph builds a Methods-Entity Knowledge Graph (MEKG) that records method-level follow-on, competition and shared-resource relations, and uses it as the backbone for core-paper discovery, graph-aware retrieval, hybrid outlining and evidence-gated, evolution-aware writing. Graph-aware retrieval nearly doubles recall over flat lexical baselines (p = 0.004), and 80.6% of generated sentences are fully or partly supported by the papers they cite.

git clone https://github.com/MUSE-AI-Lab/SurveyGraph.git
cd SurveyGraph
python -m venv .venv && source .venv/bin/activate      # Windows: .venv\Scripts\Activate.ps1
pip install -r requirements.txt

Python 3.10–3.12 is tested. Optional components are enabled only when needed: Memgraph as graph database, a LaTeX toolchain for PDF output, MinerU for PDF parsing, and MySQL for the web frontend.

Providers. Copy .env.example to .env (ignored by git) and fill in what your run needs — REMOTE_URL + TOKEN or a provider key such as OPENAI_API_KEY / DASHSCOPE_API_KEY, optionally an embedding endpoint (EMBED_*) and OPENALEX_EMAIL. Offline tests need no credentials.

Generate a survey for a topic (papers are collected automatically):

python run_sg_endtoend.py --topic "Retrieval-Augmented Generation" --max_papers 30 \
    --year_start 2022 --year_end 2024 --no-memgraph --no-evaluation

The launcher uses the Python interpreter that started it on Windows, Linux and macOS. It forwards pipeline options without overriding the topic, paper limit, years or evaluation settings. To inspect the command without credentials or provider calls:

python run_sg_endtoend.py --topic "Retrieval-Augmented Generation" --max_papers 20 --no-memgraph --no-evaluation --dry-run
bash run.sh --topic "Retrieval-Augmented Generation" --max_papers 20 --no-memgraph --no-evaluation

run.sh invokes the same launcher, forwards its arguments and works from another working directory. Repeat --topic for an explicit sequential batch; --state-file selects its progress file. No arguments show an error instead of starting a hidden batch. The direct tasks/mekg_run.py entry remains supported.

Every intermediate result is saved under outputs/<task_id>/ (mekg/, papers/, latex/, logs/, task_config.json), so an interrupted run can be resumed:

python tasks/mekg_run.py --list-tasks
python tasks/mekg_run.py --resume --task_id <task_id>

Useful switches: --uploaded-files '["a.pdf", "b.pdf"]' and --no-arxiv to write from your own papers, --no-reflection / --no-hybrid-outline / --no-global-reflector for ablations, and dropping --no-evaluation to score the result at the end. See python tasks/mekg_run.py --help for all options.

python evaluation/main.py --help         # quality evaluators and HTML report
python experiments/run_suite.py --help   # ablations, matched baselines, probes and evidence audits

Seven-method equal-budget retrieval benchmark (top-5, Table 2 of the paper). R@5 / P@5: recall and precision; Str.: section–evidence alignment; Den.: typed relation density; CoCit: co-citation connectivity measured with OpenAlex, independent of MEKG.

Graph-free methods reach higher raw recall but almost no typed structure; MEKG-linked fusion keeps competitive recall while nearly doubling relation density and co-citation connectivity. Further results: graph-aware retrieval nearly doubles recall over flat lexical baselines on the graph-native core diagnostic (p = 0.004); the semantic evidence gate reaches AUC = 0.75; citation audits over 160 generated sentences find 80.6% anchored-or-partial support; human validation of 108 semantic edges confirms 68.5% supported-or-partial (Cohen's κ = 0.64).

First page of a generated survey

HTML evaluation report

Scores depend on the corpus, provider and evaluator models; the images are display examples, not benchmark claims.

A FastAPI service for task management and result browsing (needs a MySQL-compatible database):

pip install -r frontend/requirements.txt
python frontend/run.py --reload          # binds to 127.0.0.1; set JWT_SECRET and CORS before sharing

tasks/mekg_run.py   # end-to-end entry point (collection → MEKG → outline → writing → LaTeX → evaluation)
mekg/               # Methods-Entity Knowledge Graph pipeline
src/                # retrieval, generation, refinement, schemas and shared config
evaluation/         # quality evaluators and HTML report
experiments/        # ablations, matched baselines, probes and evidence audits
frontend/           # optional FastAPI service
resources/          # prompts, templates and static resources
tests/              # unit tests and public-release checks
docs/               # design notes and figures
scripts/            # release staging, packaging and validation

Downloaded papers, corpora, generated databases and provider credentials are not part of the repository; python -m pytest -q tests/test_public_release.py and python scripts/check_public_release.py audit the public tree (they need git on PATH). See docs/PUBLIC_RELEASE_BOUNDARY.md.

@inproceedings{yuan2026surveygraph,
  title     = {{SurveyGraph}: Generating Academic Surveys via Methods-Entity Knowledge Graphs},
  author    = {Yuan, Yifan and Zhu, Yancheng and Zhang, Zuoyu and Peng, Xiaogang},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026},
  eprint    = {2610.31317},
  archivePrefix = {arXiv}
}

MIT. Third-party dependencies, provider services and user-supplied data remain subject to their own licenses and terms.