Top AI Product

MLEvolve is an open-source multi-agent system that designs machine-learning solutions end to end — planning, coding, validating, and iterating the way a human ML engineer would on a Kaggle problem. Built by Shanghai AI Laboratory and East China Normal University, it reached #1 on the MLE-bench leaderboard with 12 hours of runtime.

## How it searches

Instead of a plain tree search, MLEvolve uses what it calls Progressive MCGS — a graph-based search where branches share information through reference edges, and an entropy-inspired schedule shifts it from broad exploration toward focused exploitation as it runs. The effect is less compute wasted on dead-end ideas.

## Memory that carries over

The other half is Retrospective Memory: a cold-start domain knowledge base plus a dynamic global memory the agent mines for task-specific experience and reuses on the next problem. That’s the “self-evolving” part — it accumulates what worked. The code is open source, which matters for a system whose whole claim is that it improves by building on its own past runs.