ashhhhhh26/neurogolf-2026

NeuroGolf 2026 - ARC-AGI Tiny Network Solver

Enhanced solver for the NeuroGolf 2026 competition, building minimal ONNX networks for ARC-AGI tasks.

Overview

This repository contains an enhanced solver based on rogermt/neurogolf-solver with additional analytical solvers and optimizations.

Target: Solve ARC-AGI tasks with the smallest possible ONNX networks. Score: max(1, 25 - ln(cost)) where cost = params + memory_bytes + MACs

Data Source

The ARC-AGI training data with all 400 tasks (including test outputs) is available at:

• https://huggingface.co/LuciferMrng/neurogolf-2026/blob/main/all_tasks.json

Usage

1. Install dependencies

pip install numpy onnx onnxruntime

2. Download the data

# Option 1: Direct download from HF
curl -L https://huggingface.co/LuciferMrng/neurogolf-2026/resolve/main/all_tasks.json -o all_tasks.json

# Option 2: Via huggingface-cli
pip install huggingface-hub
huggingface-cli download LuciferMrng/neurogolf-2026 all_tasks.json --local-dir .

3. Run the solver

# Solve all 400 tasks with 30s budget per task
python neurogolf_solver_enhanced.py --data_file all_tasks.json --output_dir submission --conv_budget 30

# Solve with more time (better coverage)
python neurogolf_solver_enhanced.py --data_file all_tasks.json --output_dir submission --conv_budget 60

# Solve specific tasks
python neurogolf_solver_enhanced.py --data_file all_tasks.json --output_dir submission --tasks 1,2,3

4. Create submission.zip

The solver automatically creates submission.zip with all generated ONNX files.

Architecture

Input/Output Format

• All networks use [1, 10, 30, 30] one-hot float32 tensors
• ONNX opset 10, IR version 10

Solver Pipeline

Analytical Solvers (fast, tiny models - tried first):

1. identity - Input equals output
2. color_map - 1x1 conv color transformation
3. transpose - Matrix transpose
4. flip - Horizontal/vertical flip
5. rotate - 90/180/270 degree rotation
6. tile - Repeat input pattern
7. upscale - Nearest-neighbor upscaling
8. concat - Concatenate transformed inputs
9. spatial_gather - Pixel remapping
10. crop - Centered crop
11. constant - Fixed output regardless of input

Convolution Solvers (for learned transformations):

• Fixed shape: Slice -> Conv -> ArgMax -> OneHot -> Pad
• Variable shape: Conv(30x30) -> ArgMax -> OneHot -> Mul(mask)
• Diff shape: Slice -> Conv -> Slice(crop) -> ArgMax -> OneHot -> Pad

The conv solver learns optimal weights via least-squares fitting on one-hot patches, trying kernel sizes from 1 to 29.

Key Optimizations

1. Smallest kernel first - Tries kernel size 1, then 3, 5, etc. (smaller = fewer params)
2. Analytical before learned - Analytical solvers have near-zero cost
3. One-hot ArgMax trick - Eliminates numerical precision issues
4. No bias preferred - Slightly fewer parameters

Expected Results

With the base solver:

• ~127-293/400 tasks solved depending on time budget
• Analytical solvers cover the simplest transformations
• Conv solver handles local spatial transformations

References

• ARC-AGI - Original benchmark
• NeuroGolf 2026 Competition
• Rogermt's Base Solver

License

MIT - Based on the ARC-AGI benchmark and community contributions.