MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery
TL;DR. MLEvolve is an LLM-based self-evolving multi-agent framework for end-to-end machine-learning algorithm discovery. It fixes three failure modes of prior MLE agents: inter-branch information isolation, memoryless search, and lack of hierarchical control. It extends tree search to Progressive MCGS (Monte-Carlo graph search with cross-branch reference edges and an entropy-inspired explore→exploit schedule), adds Retrospective Memory (a cold-start knowledge base plus a dynamic global memory for experience reuse), and decouples strategic planning from code generation. On MLE-Bench it reaches state-of-the-art medal and valid-submission rates under a 12-hour budget (half the standard runtime), and beats AlphaEvolve on math algorithm optimization.
Source: HuggingFace Daily Papers (upvotes: 2) arxiv: 2606.06473 · Code: https://github.com/InternScience/MLEvolve Raw: raw/huggingface/2026-06-05-mlevolve-a-self-evolving-framework-for-automated-machine-lea.md
Key points
- Progressive MCGS: graph search (not just a tree) lets information flow across branches via reference edges; an entropy-inspired schedule shifts from broad exploration to focused exploitation as the search matures.
- Retrospective Memory: cold-start domain knowledge + dynamic global memory for task-specific retrieval and reuse, the concrete instantiation of "accumulate experience across the search."
- Decoupled control: strategic planning is separated from code generation with adaptive coding modes, for stable long-horizon iteration.
- Results: SOTA on MLE-Bench at half the standard runtime; beats AlphaEvolve on math algorithm optimization, evidence of cross-domain generalization.
Relation to prior wiki
Part of today's six-paper self-evolving-agents cluster (see Continual Experience Internalization for the cluster map). MLEvolve's Retrospective Memory is exactly the "accumulated experience must be reusable" claim that the keystone paper warns will collapse under naive iteration unless the experience is principle-level and stably internalized. Reading the two together: MLEvolve ships the memory mechanism; the keystone tells you the granularity/regime that keeps it from degrading. Beating AlphaEvolve (DeepMind's evolutionary code-discovery system, tracked in ai-industry) on math optimization is the headline external comparison.