InternScience/MLEvolve

What it solves

MLEvolve is an agentic system designed to automatically solve complex machine learning (ML) engineering tasks, such as those found in Kaggle-style competitions and mathematical optimization problems. It eliminates the need for manual trial-and-error by automating the discovery and refinement of ML algorithms.

How it works

MLEvolve uses a multi-agent collaboration framework powered by Monte Carlo Graph Search (MCGS). It explores different solution paths in parallel and evolves them through several key mechanisms:

• Multi-Mode Planning: It adaptively chooses between single-shot planning or memory-enhanced retrieval-augmented planning, using various code generation strategies like stepwise pipelines or incremental diff patching.
• Experience-Driven Memory: A global memory layer (using BM25 and FAISS) stores previous plans, code, and results, allowing the agent to reuse successful strategies and avoid past failures.
• Progressive MCGS: The system employs a search graph that detects stagnation and performs "cross-branch fusion," merging the best insights from different solution branches to create new, improved candidates.

Who it’s for

This tool is for ML engineers and researchers who want to automate the discovery of high-performing ML algorithms for competitions or scientific discovery tasks.

Highlights

• State-of-the-Art Performance: Achieves #1 on the MLE-bench leaderboard with a 65.3% medal rate.
• Generalization: Successfully applies to mathematical optimization tasks, matching or surpassing specialized frameworks.
• Flexible LLM Support: Compatible with OpenAI-compatible APIs (GPT, Qwen, DeepSeek, etc.).
• Integration: Powers the coding and algorithm optimization module of the InternAgent system.