How It Works
The system iteratively generates solutions, implements them via coding agents, evaluates results, and uses feedback to guide the next iteration.Supported Benchmarks
MLE-Bench
Solve Kaggle ML competitions with automated experimentation. Supports tabular, image, text, and audio problems.
ALE-Bench
Tackle AtCoder algorithmic optimization problems. Generates C++ solutions using optimization algorithms.
Generic Problems
Solve arbitrary problems with custom evaluators and stop conditions.
Core Components
| Component | Description |
|---|---|
| OrchestratorAgent | Coordinates the solve loop, manages budget (time, cost, iterations) |
| Search Strategy | Tree search for solution exploration (expand, select, prune) |
| Experiment Workspace | Git-based workspace with branch per experiment |
| Coding Agents | Pluggable code generators (Aider, Gemini, Claude Code, OpenHands) |
| Knowledge Search | Neo4j-based domain knowledge retrieval |