MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery
Summary
MLEvolve is an LLM-based self-evolving multi-agent framework designed for end-to-end machine learning algorithm discovery, addressing limitations in existing MLE agents such as information isolation and memoryless search. It extends tree search to Progressive MCGS, enabling cross-branch information flow via graph-based reference edges and shifting search from exploration to exploitation using an entropy-inspired schedule. The framework incorporates Retrospective Memory, which integrates a cold-start domain knowledge base with a dynamic global memory for task-specific experience retrieval and reuse. For stable long-horizon iteration, MLEvolve decouples strategic planning from code generation through adaptive coding modes. Evaluated on MLE-Bench, MLEvolve achieves leading performance in average medal rate and valid submission rate within a 12-hour budget, half the standard runtime. It also surpasses specialized methods like AlphaEvolve on mathematical algorithm optimization tasks, demonstrating strong cross-domain generalization.
Key takeaway
For Machine Learning Engineers tasked with automating algorithm discovery or optimizing long-horizon ML engineering tasks, MLEvolve presents a significant advancement. You should consider exploring frameworks that integrate progressive tree search, dynamic memory, and decoupled planning to overcome limitations of existing agents. This approach can yield leading performance and strong cross-domain generalization, potentially halving the time budget for complex algorithm development.
Key insights
MLEvolve enables LLM agents to self-evolve for automated machine learning algorithm discovery by integrating advanced search and memory.
Principles
- Enable cross-branch information flow in tree search.
- Progressively shift search from exploration to exploitation.
- Evolve agents with accumulated task-specific experience.
Method
MLEvolve employs Progressive MCGS for search, Retrospective Memory for experience management, and adaptive coding modes for stable long-horizon iteration in algorithm discovery.
In practice
- Achieve leading ML algorithm discovery performance.
- Generalize performance across diverse optimization tasks.
Topics
- LLM Agents
- Automated Machine Learning
- Algorithm Discovery
- Multi-agent Systems
- Progressive MCGS
- Retrospective Memory
Code references
Best for: Research Scientist, AI Scientist, Machine Learning Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.