NeuroGolf Championship: Implementing ARC-AGI Image Transformations with Minimal Neural Networks

NeuroGolf is a machine learning competition focusing on the design of minimal-parameter neural networks. The 2026 Championship uses ARC-AGI (Abstract Reasoning Corpus) as the benchmark to explore how to solve abstract image reasoning problems under extreme parameter constraints. ARC-AGI, proposed by François Chollet (author of Keras), is a key benchmark for evaluating AI's abstract reasoning ability. Its tasks require inferring image transformation rules from a small number of examples, which is extremely challenging for current AI systems.

What is NeuroGolf?

NeuroGolf's core challenge is to design neural networks with extremely few parameters (thousands or even hundreds) to solve complex cognitive tasks. The 2026 edition focuses on the ARC-AGI benchmark.

Characteristics of ARC-AGI

• Problem Form: Presented as input-output paired examples; requires inferring rules and applying them to new inputs.
• Sources of Difficulty: Combinatorial explosion of solution space, few-shot learning (2-4 examples), out-of-distribution generalization, and the need for program synthesis ability (inferring the program that generates the output from examples).

Example: Input red square → output red square with blue border; input green circle → output green circle with yellow border; need to infer the output rule for a blue triangle.

Extreme Parameter Constraints

Requires careful architecture design, leveraging inductive biases, and exploring efficient representation learning.

Architecture Design Ideas

• Convolution and local connections: Shared weights reduce parameters; small convolution kernels extract features.
• Streamlined attention mechanisms: Local/sparse attention or fixed weights.
• Neural modularity: Decompose into functional modules (boundary detection, color mapping, etc.).
• Meta-learning: Few-shot learning ability to quickly adapt to new tasks.

Technical Challenge Solutions

• Representation Learning: Separate encoding of color and shape, relative position encoding, symbol-neural hybrid.
• Reasoning Mechanism: Recursive processing for multi-step reasoning, explicit rule base, program search-guided interpreter.
• Training Strategy: Data augmentation, curriculum learning (from easy to difficult), multi-task learning.

• NeuroGolf Project Information: Original author zikuanqi; source GitHub (link: https://github.com/zikuanqi/NeuroGolf); release date 2026-05-29.
• ARC-AGI Example: For instance, input red square → output red square with blue border; input green circle → output green circle with yellow border; the test input (blue triangle) requires inferring the corresponding border rule.

• Efficiency and Capability: Challenges the "bigger is better" paradigm, proving that small models can solve complex problems, suitable for resource-constrained scenarios like edge devices.
• Cognitive Science Inspiration: Simulates biological neural constraints, prompting reflection on how humans achieve strong cognitive abilities with simple structures.
• AGI Path: The key to intelligence lies in structure rather than scale; it is necessary to organize limited resources to solve open-ended problems.

Need to master:

1. Deep learning fundamentals (convolution, attention, recurrent networks);
2. Model compression techniques (knowledge distillation, pruning, quantization);
3. Meta-learning and few-shot learning (MAML, prototypical networks);
4. Program synthesis and neuro-symbolic AI;
5. Research on the ARC-AGI dataset (task types, human problem-solving strategies).

NeuroGolf, through extreme parameter constraints, strips away scale variables and focuses on essential issues such as architecture design and inductive biases. zikuanqi's project provides a concrete reference, demonstrating the possibility of minimal networks tackling ARC-AGI, and offers valuable experience for the design of efficient AI systems.