We’re excited to share our work on AIDE, an LLM-powered agent that automates machine learning engineering through systematic trial-and-error. Unlike conventional AutoML, AIDE searches directly in the space of code, iteratively refining solutions using a structured tree search approach.
Key Highlights:
- State-of-the-art performance: AIDE outperforms human competitors in Kaggle-style ML tasks, as shown in OpenAI’s MLE-Bench.
- Scalability & efficiency: By reusing and refining promising solutions, AIDE achieves 3.5x performance gains over o1 alone.
- Beyond tabular ML: AIDE extends to deep learning and even AI research tasks, showing human-level capabilities in structured R&D.
- We’re open-sourcing the code and sharing our research paper to help the community build on top of it. Would love to hear your thoughts!
1 comment
[ 3.3 ms ] story [ 9.6 ms ] threadKey Highlights:
- State-of-the-art performance: AIDE outperforms human competitors in Kaggle-style ML tasks, as shown in OpenAI’s MLE-Bench. - Scalability & efficiency: By reusing and refining promising solutions, AIDE achieves 3.5x performance gains over o1 alone. - Beyond tabular ML: AIDE extends to deep learning and even AI research tasks, showing human-level capabilities in structured R&D. - We’re open-sourcing the code and sharing our research paper to help the community build on top of it. Would love to hear your thoughts!
Paper: https://www.arxiv.org/abs/2502.13138 Code: https://github.com/WecoAI/aideml