Skip to content

data-to-paper

external · status: active · focus: end-to-end · discipline: general · started: 2023

Project page: https://github.com/Technion-Kishony-lab/data-to-paper

Source: projects/landscape/data-to-paper.yml

Positioning

An end-to-end framework that takes annotated data and produces backward-traceable scientific manuscripts: every numeric value in the output can be click-traced to the specific code line that generated it. Navigates interacting LLM and rule-based agents through data exploration, literature search, hypothesis raising, code-debugging, interpretation, and step-by-step paper writing. Released alongside an NEJM AI peer-reviewed paper (Ifargan et al., 2024 — DOI:10.1056/AIoa2400555).

Distinctive contribution

The "data-chained" provenance design is unique in the catalog: the manuscript is constructed so that traceability is intrinsic, not a reporting add-on — any reported number resolves backward through the data analysis steps. Ships both Autopilot and Copilot modes (oversee / inspect / guide / rewind / replay) and overrides standard statistical packages with coding guardrails to minimize common LLM coding errors.

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Seven stages from hypothesis through revision; explicitly skips referee simulation and dissemination.
Autonomy level 3 Autopilot mode runs end-to-end from data; Copilot mode adds human oversight on demand.
Architectural transparency 3 Open under MIT; NEJM AI paper + arXiv:2404.17605 document the method; agent prompts visible.
Inputs supported 2 Accepts annotated datasets with optional research-goal specification; open-goal and fixed-goal modes.
Outputs / reproducibility 3 Backward-traceable manuscripts where any numeric value resolves to its generating code line — strongest reproducibility design in the catalog.
Internal evaluation 3 Peer-reviewed publication in NEJM AI demonstrating the framework; example papers across diabetes, social-network, and clinical datasets.
Openness 3 MIT-licensed; pip-installable; example datasets + example papers all public.
Maturity / traction 2 793 stars; NEJM AI publication; last push 2025-07 — slower cadence but stable mature release.
Cross-family policy 0 Single-LLM design with rule-based agents alongside; no cross-family review architecture.
Runtime assurance 2 Coding guardrails (overridden statistical packages), data-chaining provenance checks, and AI-review-on-demand in Copilot mode.
Cross-platform portability 1 Pip package; works with multiple LLM back-ends, but no native multi-IDE adaptation layer.

Scored on 2026-05-18. See the evaluation rubric.

Tags

Pipeline stages: hypothesis-generation literature-discovery research-design data-analysis code-generation paper-drafting revision-editing

Architectural features: multi-agent human-in-loop tool-use artifact-versioning dag-orchestration

Inputs: annotated-dataset research-goal

Outputs: traceable-manuscript data-chained-paper code

Data sources: user-provided

Knowledge sources: literature

Limitations

  • Designed for 'relatively simple research goals and simple datasets' per the authors — does not yet target deep theoretical or large-scale studies.
  • Last push 2025-07; cadence has slowed since the NEJM publication.
  • No referee-simulation or dissemination-packaging stages.

Papers describing this project

  • Autonomous LLM-Driven Research — from Data to Human-Verifiable Research Papers — Ifargan, T., Hafner, L., Kern, M., Alcalay, O., Kishony, R. (2024). NEJM AI. doi
  • Autonomous LLM-driven research from data to human-verifiable research papers — Ifargan, T., Hafner, L., Kern, M., Alcalay, O., Kishony, R. (2024). arXiv preprint. arXiv:2404.17605

Also compared in

  • ARIS Table 4 (yang2026aris) — Listed as 'data-driven (not open-ended idea-to-paper)' with narrower research state retention.