MLEvolve framework achieves state-of-the-art ML algorithm discovery
MLEvolve is a self-evolving LLM multi-agent framework for end-to-end machine learning algorithm discovery that achieves state-of-the-art performance on MLE-Bench within a 12-hour budget.
Score breakdown
MLEvolve demonstrates that a single self-evolving agent framework can achieve state-of-the-art results on MLE-Bench in half the standard runtime while also outperforming a specialized method like AlphaEvolve on mathematical algorithm optimization, showing strong cross-domain generalization for long-horizon AI-driven research automation.
- 01MLEvolve is an LLM-based self-evolving multi-agent framework for end-to-end machine learning algorithm discovery.
- 02It addresses three limitations in existing MLE agents: inter-branch information isolation, memoryless search, and lack of hierarchical control.
- 03Progressive MCGS extends tree search to a graph structure with cross-branch reference edges and an entropy-inspired progressive exploration-to-exploitation schedule.
Shangheng Du, Xiangchao Yan, and Jinxin Shi present MLEvolve, an LLM-based self-evolving multi-agent framework targeting end-to-end machine learning algorithm discovery. The paper identifies three core weaknesses in existing MLE agents — inter-branch information isolation, memoryless search, and lack of hierarchical control — and addresses each with a distinct architectural component. The central search mechanism, Progressive MCGS, extends conventional tree search into a graph structure that allows cross-branch information flow through reference edges, while an entropy-inspired progressive schedule gradually transitions the agent from broad exploration to focused exploitation over the course of a run.
Stable long-horizon iteration is further supported by decoupling strategic planning from code generation through adaptive coding modes.
To support long-horizon self-improvement, MLEvolve introduces Retrospective Memory, a two-part memory system pairing a cold-start domain knowledge base with a dynamic global memory that enables task-specific experience retrieval and reuse. Stable long-horizon iteration is further supported by decoupling strategic planning from code generation through adaptive coding modes. Evaluated on MLE-Bench, MLEvolve achieves state-of-the-art performance across multiple dimensions — including average medal rate and valid submission rate — under a 12-hour budget, which is half the standard runtime. Beyond MLE tasks, the framework also outperforms AlphaEvolve on mathematical algorithm optimization tasks, demonstrating cross-domain generalization. The code is publicly available at https://github.com/InternScience/MLEvolve.
Key facts
- 01MLEvolve is an LLM-based self-evolving multi-agent framework for end-to-end machine learning algorithm discovery.
- 02It addresses three limitations in existing MLE agents: inter-branch information isolation, memoryless search, and lack of hierarchical control.
- 03Progressive MCGS extends tree search to a graph structure with cross-branch reference edges and an entropy-inspired progressive exploration-to-exploitation schedule.
- 04Retrospective Memory combines a cold-start domain knowledge base with a dynamic global memory for task-specific experience retrieval and reuse.
- 05Strategic planning is decoupled from code generation via adaptive coding modes for stable long-horizon iteration.
- 06MLEvolve achieves state-of-the-art average medal rate and valid submission rate on MLE-Bench under a 12-hour budget — half the standard runtime.
- 07MLEvolve outperforms AlphaEvolve on mathematical algorithm optimization tasks, demonstrating cross-domain generalization.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 9, 2026 · 17:05 UTC. How this works →