MDL Pattern Discovery: ARC-AGI 2

Robotic Intelligence — Novel “Growing Neighbors” Phenomenon in MDL Filter Space

Discovery · Probed for robotic deployment2025 – 2026Independent research contributionUniversity interest · Robogio CEO review

Project Overview

Introduced a novel Growing Neighbors growth function in MDL filter space while analyzing ARC-AGI 2 puzzle patterns: filter activations expand outward in a spatially coherent, neighbor-dependent cascade that mirrors the logical structure of abstract reasoning tasks.

This growth behavior had never been formally described in this context, especially alongside Minimum Description Length (MDL), a genuine research contribution at the intersection of information theory, computer vision, and abstract reasoning.

Highlight

Impact & Results

Introduced the Growing Neighbors growth function, a novel emergent phenomenon in MDL filter space
Discovery significantly increased ARC-AGI 2 accuracy
Probed by the CEO to be installed into Robogio’s automated robotic machines for real-time pattern recognition
Attracted interest from a university professor in Upstate New York for potential research collaboration or publication

Key Features

MDL-based pattern discovery pipeline for ARC-AGI 2 puzzles
Introduced the Growing Neighbors growth function: neighbor-dependent cascade expansion in filter activation space
Rule extraction linking dilation, rotation, and related transforms
Interactive Run MDL / Detect Rule / Predict workflow for live probing
Greedy MDL search ranking top filters by occurrence and description length cost

Technical Challenges

Formalizing an emergent filter-space growth behavior that had not been described before
Connecting information-theoretic MDL costs to spatially coherent reasoning patterns
Improving ARC-AGI 2 accuracy without sacrificing interpretability of discovered rules
Translating abstract-grid findings into signals usable by robotic perception stacks