🔥 ARC Benchmark Results

Geometric Pathway Substrate vs Weight-Sharing Architectures

81.5%
Trunk ID on Unseen Tasks
55.5%
Exact Grid Match
2.6s
Total Training Time

The Core Finding

Standard Neural Networks

Recall: 100% ✓

Variation: 0% ✗

Training: 100s-1000s epochs

Perfect memorization.
Zero generalization.
Even on tasks they trained on.

Geometric Substrate

Recall: 100% ✓

Variation: 100% ✓

Training: ONE PASS (2.6s)

Perfect recall.
Perfect variation handling.
81.5% on completely unseen tasks.

Complete Benchmark Results

Configuration	Train	Eval	Exact	Cell Acc	Trunk ID
Original → Original	1009	514	55.5%	59.7%	81.5%
Original → AGI-2	1009	120	0.0%	4.9%	34.5%
AGI-2 → AGI-2	1000	120	0.0%	4.9%	34.5%
Combined → AGI-2	2009	120	0.0%	4.9%	34.5%

What This Means

Original ARC: 81.5% trunk identification on 514 completely unseen tasks
ARC-AGI-2: Deliberately designed with novel patterns—exposes encoding limitations
Standard models: 0% variation on tasks they trained on

The Encoding Problem

Standard neural networks compress input-output relationships into shared weight matrices. This compression creates destructive interference when multiple distinct mappings must coexist.

Consider an ARC task requiring scale-by-2 operations:

Model correctly outputs scaled grid for training examples
Model fails on different input requiring same operation
Not because it lacks the transformation knowledge
Because compressed representation can't distinguish inputs

Architecture

Grid Encoding (8 dimensions)

g = [h/30, w/30, |C|/10, mean/9, std/4.5, ρ, δ, corner/9]

Where:
  h, w     = grid dimensions
  |C|      = unique color count
  mean/std = color statistics
  ρ        = spatial correlation
  δ        = local variation
  corner   = corner signature

Pathway Printing

For each (input, output) pair:
    pathway = [encode(input), encode(output), delta]
    
Storage: Direct, no compression
Training: ONE PASS, no epochs
Recall: Minimum-distance pathway lookup

Implications for AGI

1. Scaling is not the answer
The 0% variation problem appears at all model sizes (tested 0.5B to 14B).

2. Encoding is the bottleneck
Incompressible geometric representations enable transfer that compressed representations cannot.

3. One-pass learning is sufficient
The substrate achieves 81.5% on unseen tasks without iterative training.

The path to AGI may require abandoning weight-sharing architectures in favor of geometric substrates that preserve the full structure of learned relationships.

We print. They burn.

Downloads

Python Implementation LaTeX Paper GitHub Repository

Citation

@article{heeney2026geometric,
  title={One-Pass Geometric Pathway Encoding Achieves 81.5% 
         Operation Classification on Unseen ARC Tasks},
  author={Heeney, Joseph},
  journal={Ghost in the Machine Labs},
  year={2026},
  url={https://7themadhatter7.github.io/harmonic-stack/arc-results.html}
}