SILVER MEDAL · 79 / 3,061 TEAMS · IJCAI-ECAI 2026
Four hundred networks. One hundred thousand parameters. Total.
NeuroGolf scores a network by how small it is — every parameter and every byte is a logarithmic penalty, and only exact output correctness counts. My submission solves 400 ARC-AGI image transformations with less than 1% of the parameters in a single ResNet-18. Every number on this page is mined directly from the 400 .onnx files of the actual submission.
A / THE COST FIELD
All 400 tasks, drawn to scale.
One dot per task — horizontal is task number, vertical is the score that network earned (higher is smaller). Teal dots are the nine zero-parameter solutions. Hover any dot for the readout; the orange line is the median network: 103 parameters.
HOVER A DOT — OR TAP ONE — FOR THE PER-TASK RECORD.
B / THE METRIC
Why every byte hurts.
score = max(1, 25 − ln(cost))
cost = parameters + byte footprint
The penalty is logarithmic: shaving a network from 1,000 to 100 costs-worth buys as much as shaving 100,000 to 10,000. The winning move is int8 quantization and hand-pruned graphs, not accuracy tricks. I iterated across 1,100+ evaluated submissions on this curve.
a typical network in this submission
C / THE ARSENAL
Not multilayer perceptrons. Machines.
The 400 solutions use 90+ distinct ONNX operator types. Bitwise logic, Einsum contractions, scatter operations — many tasks are solved by a hand-built computation graph that happens to be a valid neural network. The top operators, counted across every graph:
D / THE SMALLEST
Entire networks, every weight shown.
The eight smallest models in the submission, tensors decoded straight from the files. Orange cells are positive weights, teal are negative; intensity is magnitude. Where there is nothing to draw, that is the point.