AIDE ML is an open-source Python package that implements the AIDE algorithm, an LLM-driven agent designed to autonomously write, evaluate, and improve machine learning code. The repository serves as the reference implementation of research described in an arXiv paper (2502.13138), positioned as a lean, research-friendly tool for academics and ML engineers who want to replicate the paper, test new algorithmic ideas, or prototype ML pipelines. The package is available on PyPI and requires Python 3.10 or higher.
The core functionality centers on agentic tree search in code space. Users specify a machine learning task in natural language by pointing the agent at a dataset and describing a goal and metric, such as "Predict churn" with "AUROC" as the evaluation metric. The agent then iteratively generates Python scripts as nodes in a solution tree, uses LLM-generated patches to spawn child nodes, and employs metric feedback to prune and guide the search toward better solutions. According to OpenAI's MLE-Bench evaluation across 75 Kaggle competitions, AIDE's tree-search approach wins four times more medals than the best linear agent baseline.
The repository provides several utility features beyond the core algorithm. An HTML visualizer allows users to inspect the complete solution tree and examine code attached to each node. A Streamlit web UI enables users to prototype ML solutions through an interactive interface where they can paste API keys, upload data, set goals and metrics, and run AIDE while viewing live logs and the solution tree. The implementation is model-agnostic, supporting OpenAI, Anthropic, Gemini, and any local LLM that implements the OpenAI API specification.
GitGenius activity data shows the repository maintains responsive issue and pull request handling, with a median response latency of 0.0 hours across 28 tracked items and a mean latency of 36.1 hours. The most active contributor is dexhunter with 80 recorded events, followed by LukasFehring with 8 events and AndyLi1024 with 5 events. Bug reports represent the most common issue type with 8 instances, followed by questions with 7 instances and enhancement requests with 5 instances. The repository overlaps with contributors from related projects including verl-project/verl, haotian-liu/llava, and agenta-ai/agenta.
The package has gained significant adoption in the research community. Multiple institutions have built research on or with AIDE, including OpenAI's MLE-Bench project, METR's RE-Bench for evaluating frontier AI R&D capabilities, Sakana AI's AI Scientist-v2 for automated scientific discovery, and Meta's work on LLM speedrunning and AI research agents. SJTU has also developed ML-Master, which integrates exploration and reasoning using AIDE's approach.
Users can deploy AIDE ML through multiple interfaces. The quick-start CLI approach generates logs containing the best solution code and an HTML tree plot for inspection. Advanced CLI options allow configuration of the LLM model used for code generation, the number of improvement iterations, and the number of drafts per step. The package supports fully local operation without external API calls, though with expected performance trade-offs, and includes Docker support for containerized deployment. Development installation is available for those extending the codebase.
