Atlantis Glass Sponge v2: Solving Visual Interactive Reasoning in ARC-AGI-3 via NeoSpatial Topological Graphs, Causal Affordance Induction, and Epistemic Pearl Memory
AbstractThe Abstraction and Reasoning Corpus 3 (ARC-AGI-3) extends the frontier of artificial general intelligence evaluation from static input-output matrix synthesis to interactive, dynamic visual environments. Autonomous agents must infer physical mechanics, discover controllable avatars, and construct long-horizon plans under severe epistemic uncertainty without verbal instructions. In this paper, we introduce Atlantis Glass Sponge v2, a bio-inspired neuro-symbolic agent architecture inspired by the continuous syncytial conduction highways of deep-sea hexactinellid glass sponges (Aphrocallistes vastus). Our architecture introduces three key mechanisms: NeoSpatial Topological Graphs: Segments raw 2D pixel matrices into 8-connected relational scene graphs, achieving an 80.7% token reduction over flat matrix encodings. Causal Affordance and Avatar Induction: Resolves avatar identity within 2 to 3 turns using cosine similarity over candidate displacement vectors. Epistemic Pearl Memory: An episodic negative constraint bar-list that permanently prunes lethal trajectory branches from search trees. On the official Kaggle ARC Prize 2026 leaderboard, Atlantis Glass Sponge v2 achieved an official verified Personal Best score of 27.75 (Kaggle Submission Ref 57965431, Version 14). Benchmark & Submission Links Kaggle Competition: ARC Prize 2026 (ARC-AGI-3 and Paper Track) Kaggle Public Notebook: https://www.kaggle.com/code/coralfil/arc-prize-2026-atlantis-glass-sponge-v2 Verified Benchmark Record: Score 27.75 (Submission Ref 57965431, Version 14 Apex) Author: Roland Poulin (ORCID: 0009-0009-8065-0620) Organization: Team Atlantis / Coralfil Inc. License: Creative Commons Attribution 4.0 International (CC BY 4.0) Suggested CitationPoulin, R. (2026). Atlantis Glass Sponge v2: Solving Visual Interactive Reasoning in ARC-AGI-3 via NeoSpatial Topological Graphs, Causal Affordance Induction, and Epistemic Pearl Memory. Coralfil Inc. Preprint.
Authors
- Roland Poulin
- Coralfil Inc.
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-09
- DOI
- https://doi.org/10.5281/zenodo.24093870
- Primary Topic
- AI-based Problem Solving and Planning
- Type
- preprint