MLEvolve Beats AlphaEvolve at Its Own Game
Top paper on HuggingFace right now, more than 300 upvotes, and the claim that earns them is a heavy one: MLEvolve says it outperforms AlphaEvolve on mathematical algorithm optimization. AlphaEvolve was DeepMind's headline 'AI discovers better algorithms' system. Getting beaten by a paper out of Shanghai AI Lab and East China Normal University is the kind of result that makes people look twice.
What is it? A self-evolving multi-agent system that discovers machine learning algorithms end to end. Three pieces matter. First, progressive Monte Carlo Graph Search, which replaces the usual tree where branches can't talk to each other with a graph, so good ideas flow across branches. Second, retrospective memory, a curated knowledge base for the cold start plus a global memory that keeps accumulating what worked. Third, hierarchical planning with adaptive code generation. Put together it hits state-of-the-art medal rates on a 12-hour budget.
Why it matters: this is the recursive-self-improvement thread again, the one where AI starts designing the AI. AlphaEvolve, Anthropic's 'Claude writes 80% of our code' report, and now a system that out-evolves AlphaEvolve on math, these aren't isolated dots. The loop is tightening. Paper at arxiv.org/abs/2606.06473, accepted to ICML 2026, code at github.com/InternScience/MLEvolve.
← Back to all articles
What is it? A self-evolving multi-agent system that discovers machine learning algorithms end to end. Three pieces matter. First, progressive Monte Carlo Graph Search, which replaces the usual tree where branches can't talk to each other with a graph, so good ideas flow across branches. Second, retrospective memory, a curated knowledge base for the cold start plus a global memory that keeps accumulating what worked. Third, hierarchical planning with adaptive code generation. Put together it hits state-of-the-art medal rates on a 12-hour budget.
Why it matters: this is the recursive-self-improvement thread again, the one where AI starts designing the AI. AlphaEvolve, Anthropic's 'Claude writes 80% of our code' report, and now a system that out-evolves AlphaEvolve on math, these aren't isolated dots. The loop is tightening. Paper at arxiv.org/abs/2606.06473, accepted to ICML 2026, code at github.com/InternScience/MLEvolve.
Comments