NeuroGolf 2026
Solving ARC-AGI puzzles with the smallest neural network that gets every example right.
- Competitor
- 2026
What it is
A submission pipeline for the NeuroGolf 2026 Kaggle challenge, which scores solutions to ARC-AGI grid puzzles by the parameter count of their ONNX model. One model is trained per task.
The problem
The aim is not the best accuracy but the smallest model that is exactly right on every demonstration pair, and the exported ONNX file, not the PyTorch model, is what gets scored.
Architecture
Requests flow left to right. Hover a component for its role.
- ARC-AGI task external, 3 to 5 example pairs
- Smallest-first search service, 4 model sizes
- PyTorch training worker, Apple MPS
- ONNX export service, opset 17
- onnxruntime check service, must reproduce every pair
- Two parameter counters service
Key decisions
Verify what gets scored
Correctness is checked by running the exported ONNX file in onnxruntime, which catches precision drift that the PyTorch model would hide.
Count twice
Two independent parameter counters cross-check each other before a model is accepted.
Highlights
- A hand-built ONNX model solves the demo task with 3 parameters, against 90 for the trained one.
- Hundreds of per-task ONNX models produced by the pipeline.