DataToPaper
title: data-to-paper radar_quadrant: Tools radar_ring: Assess radar_position: center created: 2026-05-22 last_updated: 2026-05-22 related: ["STORM", "Agno"]
data-to-paper is an open-source Python framework from Technion's Kishony Lab that orchestrates LLM and rule-based agents through a complete empirical scientific research cycle: from raw annotated data to a full, reproducible research manuscript with backward-traceable numeric claims.
The system guides agents through the conventional scientific path in sequence:
- Hypothesis generation — agents propose research questions from the data and objective
- Literature search — agents retrieve and cite relevant prior work
- Analysis code — agents write, debug, and execute statistical analysis code against the dataset
- Interpretation — agents interpret output from the code execution
- Manuscript writing — agents assemble a complete LaTeX paper with abstract, methods, results, discussion, and references
Inputs are a raw annotated dataset and a research objective (open-ended or fixed). Outputs are a complete LaTeX manuscript and all intermediate artifacts.
The defining property is "data-chaining": every numeric value in the manuscript can be click-traced back to the specific code line that produced it. This makes the full information flow from raw data to final claim auditable by a human reviewer, distinguishing it from prose-generation tools that generate plausible but unverifiable figures.
Standard Python statistical packages are overridden with multiple coding guardrails to catch common LLM errors — off-by-one indexing, incorrect significance tests, silent type coercions. The framework explicitly requires human-in-the-loop oversight; domain experts must vet the final manuscript before any scientific use.
| data-to-paper | STORM | |
|---|---|---|
| Output | Empirical research paper | Wikipedia-style survey |
| Input | Structured dataset + question | Topic keyword |
| Analysis | Statistical code execution | Multi-perspective web research |
| Traceability | Numeric values → code lines | Claims → cited sources |
| Domain | Scientific / data science | General knowledge |
STORM produces referenced literature surveys. data-to-paper produces full empirical papers grounded in a provided dataset — a significantly more ambitious and domain-constrained scope.
data-to-paper sits at Tools → Assess center. The backward-traceability property and automated statistical code execution are technically distinctive — this is the most complete end-to-end scientific automation pipeline publicly available. 796 stars, MIT licensed, updated May 2026, actively maintained by Technion's Kishony Lab. Applicability is narrow: requires a clean structured dataset, a scientific question, and domain expertise to validate output. GPT-4 cost is acknowledged in documentation. Center position reflects the gap between technical impressiveness and near-term actionability for general software engineering contexts. Trial gate: an end-to-end run on a real dataset with verified backward-traceability and expert review of the manuscript output.
