GitHub

**data-to-paper** is an automation framework that systematically navigates interacting AI agents through a **complete end-to-end scientific research**,
starting from *raw data* alone and concluding with *transparent, backward-traceable,
human-verifiable scientific papers*
( Example AI-created paper ,
Copilot App DEMO).
This repository is the code implementation for the paper "Autonomous LLM-Driven Research — from Data to Human-Verifiable Research Papers".
Try it out
then run: data-to-paper
See INSTALL for dependencies.
Key features
* **End-to-end field-agnostic research.** The process navigates through the entire scientific path,
from data exploration, literature search and ideation, through data analysis and interpretation,
to the step-by-step writing of a complete research papers.
* **Traceable "data-chained" manuscripts**. Tracing information flow, *data-to-paper* creates backward-traceable and verifiable manuscripts,
where any numeric values can be click-traced all the way up to the specific code lines that created them
(data-chaining DEMO).
* **Autopilot or Copilot.** The platform can run fully autonomously, or can be human-guided through the Copilot App, allowing users to:
- Oversee, Inspect and Guide the research
- Set research goals, or let the AI autonomously raise and test hypotheses
- Provide review, or invoke on-demand AI-reviews
- Rewind the process to prior steps
- Record and replay runs
- Track API costs
* **Coding guardrails.** Standard statistical packages are overridden with multiple guardrails
to minimize common LLM coding errors.
https://github.com/Technion-Kishony-lab/data-to-paper/assets/31969897/0f3acf7a-a775-43bd-a79c-6877f780f2d4
Motivation: Building a new standard for Transparent, Traceable, and Verifiable AI-driven Research
The *data-to-paper* framework is created as a research project to understand the
capacities and limitations of LLM-driven scientific research, and to develop ways of harnessing LLM to accelerate
research while maintaining, and even enhancing, the key scientific values, such as transparency, traceability and verifiability,
and while allowing scientist to oversee and direct the process
(see also: living guidelines).
Implementation
Towards this goal, *data-to-paper* systematically guides **interacting LLM and rule-based agents**
through the conventional scientific path, from annotated data, through creating
research hypotheses, conducting literature search, writing and debugging data analysis code,
interpreting the results, and ultimately the step-by-step writing of a complete research paper.
Reference
The **data-to-paper** framework is described in the following NEJM AI paper:
- Tal Ifargan, Lukas Hafner, Maor Kern, Ori Alcalay and Roy Kishony,
"Autonomous LLM-Driven Research — from Data to Human-Verifiable Research Papers"
10.1056/AIoa2400555
and in the following pre-print:
- Tal Ifargan, Lukas Hafner, Maor Kern, Ori Alcalay and Roy Kishony,
"Autonomous LLM-driven research from data to human-verifiable research papers",
arXiv:2404.17605
Examples
We ran **data-to-paper** on the following test cases:
* **Health Indicators (open goal).** A clean unweighted subset of
CDC’s Behavioral Risk Factor Surveillance System (BRFSS) 2015 annual dataset
(Kaggle). Here is an example Paper created by data-to paper.
Try out:
* **Social Network (open goal).** A directed graph of Twitter interactions among the 117th Congress members