BFTree/MetaSyn

What is a Meta-Analysis?

A meta-analysis is the gold standard of evidence synthesis in biomedical and social science research. It systematically identifies all published studies on a clinical question, screens them against explicit eligibility criteria (usually framed as PICO: Population, Intervention, Comparator, Outcome), extracts quantitative effect sizes, and pools them into a single estimate with formal uncertainty quantification.

What makes meta-analysis unusually demanding — and unusually useful as a benchmark substrate — is that every intermediate decision is auditable: the search query is recorded, each inclusion/exclusion decision is justified against stated criteria, and the final statistical estimate is fully reproducible. This means ground truth can be verified at every stage of the pipeline, not just at the final output.

---

MetaSyn

MetaSyn is a dataset of 442 meta-analyses drawn from the Nature Portfolio (2015–2024) and curated to benchmark LLM agent pipelines on the complete meta-analysis workflow. Each entry pairs a published meta-analysis with:

• A retrieval corpus of 140,585 PubMed-indexed articles spanning all included topics
• Expert-annotated PICO/ECO criteria and search strategies
• A hard-negative set of articles that share the topic but fail at least one eligibility criterion
• Ground-truth effect size, heterogeneity, and conclusion fields

The core challenge is end-to-end: given only a research question and eligibility criteria, a system must retrieve, screen, extract, and synthesize — replicating the full pipeline that took human experts weeks.

Dataset Statistics

	Meta-analyses	Corpus articles
Train	354	~112k
Test	88	~28k
Total	442	140,585

	Value
Median included studies per MA	14
Mean included studies per MA	18.3
Retrieval ceiling @ K=200	90.9%