About this project
R&D-Agent (RD-Agent) is a Microsoft open-source project that aims to automate high-value, data-driven research and development. Its core idea is a two-part loop: an "R" component that proposes new ideas and a "D" component that implements and evaluates them, so that AI can iteratively drive data and model development.
What it provides
- A multi-agent framework for automating model and data development, with a research agent and a development agent working in alternating cycles.
- Scenario implementations for quantitative finance, including factor proposal/implementation, model evolution, and joint factor-model co-optimization built around Qlib.
- A data-science scenario for competition-style tasks, with a Kaggle-oriented mode for model tuning and feature engineering.
- A research copilot that reads papers or financial reports and turns them into model structures or datasets.
- An LLM fine-tuning scenario (FT-Agent) for benchmark-driven data processing, training, evaluation, and feedback-guided refinement.
- A benchmark (Agent² RL-Bench) for evaluating whether LLM agents can engineer end-to-end post-training pipelines.
Configuration and usage
- Linux is currently the only supported platform; Docker is required for most scenarios.
- Installation is via PyPI (`pip install rdagent`) or from source with `make dev` for development.
- LLM access is configured through LiteLLM by default, with examples for OpenAI, Azure OpenAI, DeepSeek, and separate chat/embedding providers.
- A `health_check` command verifies Docker availability and port usage before running scenarios.
- Scenarios are launched through CLI commands such as `rdagent fin_quant`, `rdagent fin_factor`, `rdagent fin_model`, `rdagent fin_factor_report`, `rdagent general_model`, `rdagent data_science`, and `rdagent llm_finetune`.
- Results can be monitored through a Streamlit UI (`rdagent ui`) and a separate web frontend served by `rdagent server_ui`, which includes token-based authentication.
Project status and evidence
- The README reports that R&D-Agent leads on MLE-bench among machine learning engineering agents, with detailed run links and per-difficulty results.
- It cites accepted papers (NeurIPS 2025, ICML 2026, ACL 2026 Findings) and a technical report, plus a live demo and documentation site.
- The project is actively developed, with CI, CodeQL, Dependabot, linting, type checking, and pre-commit configured.
This overview is based solely on the repository README and does not independently verify benchmark claims.
Comments
0 Rating appears after 10 ratings
Sign in to join the discussion.