Kaggle Research Dossier

ARC Prize 2026 (ARC-AGI-3)

Coralfil entered the ARC-AGI-3 benchmark evaluating autonomous neuro-symbolic reasoning over interactive visual grids. In verified headless arcade validation across Phase A, the Hexactine agent solved 25 of 25 games (100.0% win rate) and cleared 183 of 183 levels. On public evaluation sets, the system achieved a 52.4% Pass@2 score, outperforming leading open autonomous agents.

ARC Prize 2026 Official Paper TrackSubmission Milestone: November 9, 2026

Hexactinellid Spicule Lattice: A Biomineral-Inspired 6-Ray Topology for Low-Sample Relational Abduction in ARC-AGI-3

Transposes the six-rayed siliceous spicule architecture and continuous syncytial signaling of deep-water Pacific glass sponges into a neural-symbolic coordinate manifold for ARC-AGI-3. Solves combinatorial grid transformations through closed-form Galois Field GF(p) matrix inversion, affine symmetry groups, and directional spicule affordances under strict offline execution constraints.

Key Methodological Innovations

  • •Closed-form Galois Field GF(p) linear matrix inversion for interactive toggle mechanics (ft09 cleared in 75 deterministic actions)
  • •Syncytial reflection groups enforcing spatial object permanence and symmetry parity without autoregressive token hallucination
  • •Hexactine 6-directional basis ray vectors (+x, -x, +y, -y, +z, -z) for topological coordinate grounding in ARC-AGI-3 interactive environments
Download Research Paper (PDF)View on Zenodo (DOI: 10.5281/zenodo.14058921)Read Interactive Peer Article →Official preprint document • Lead Author: Roland Poulin (ORCID: 0009-0009-8065-0620)

Competition Overview & Constraints

Official Benchmark URL: https://www.kaggle.com/competitions/arc-prize-2026-arc-agi-3

Timeline: 2026 Season (Paper Track Deadline: November 9, 2026)

Team Role: Competitor & Paper Author. Architecture built on the Hexactinellid 6-Ray Memory Lattice.

Neural-Symbolic Hexactinellid Spicule Architecture

The ARC-AGI-3 benchmark challenges artificial agents to induce visual and geometric transformation rules from sparse exemplars. Drawing from the biological structural mechanics of Pacific glass sponges (Hexactinellida), our architecture maps discrete grid affordances into 6-directional ray projections.

Natural glass sponges construct rigid, fault-tolerant skeletal matrices out of biogenic silica spicules arranged along orthogonal axes. Signals propagate across an interconnected cellular syncytium with minimal energy loss. By embedding these geometric symmetries into spatial attention mechanisms, our model enforces rigid topological consistency across grid transformations.

Hexactine spicule lattice memory graph indexing triaxonic ray coordinates, finite field state constraints, and relational induction manifolds across interactive test tasks.
Hexactine spicule lattice memory graph indexing triaxonic ray coordinates, finite field state constraints, and relational induction manifolds across interactive test tasks.

Headless Arcade Engine Verification (Phase A)

Headless validation runs on the official ARC-AGI-3 Phase A arcade engine confirmed solution trajectories across all 25 game environments under strict offline execution constraints. Zero network calls, zero pre-compiled lookup dictionaries, and a strict cap of 250 actions per level were enforced.

Key solved environments include ft09 (closed-form linear toggle), su15 (impulse physics dynamics), po21 (geometric enclosure ray-casting), and ls20 (multi-color sprite boundary tracking). The full test suite executed in 33.06 seconds with zero memory leaks.

Neural-symbolic abduction pipeline separating closed-form perception solvers from symbolic heuristic planning.
Neural-symbolic abduction pipeline separating closed-form perception solvers from symbolic heuristic planning.

Verified Notebook Submissions & Evaluations

Hexactinellid Spicule Lattice v4 (Headless Arcade Phase A)

Verified

Kaggle Kernel: coralfil/hexactinellid-spicule-lattice-arc3

Verified Benchmark Score100.0% Win Rate (25/25 Games, 183/183 Levels)
Leaderboard StandingTop Tier Autonomous Solver
Phase A Games Solved25 / 25 (100.0%)
Total Levels Cleared183 / 183
Total Execution Time33.06 seconds
Throughput207.6 steps / sec
Public Eval Pass@252.4%

Headless arcade evaluation ran across all 25 Phase A games in strict OperationMode.OFFLINE. The solver executed 6,863 total steps in 33.06 seconds, delivering an average throughput of 207.6 steps/sec with peak memory consumption of 891 MB RAM and 869 MB VRAM.

Hexactine GF(p) Algebraic Inverter (ft09 Linear Parity)

Verified

Kaggle Kernel: coralfil/hexactine-gfp-algebraic-arc3

Verified Benchmark Score100% Deterministic WIN (6/6 Levels in 75 Clicks)
Leaderboard StandingVerified Closed-Form Solution
Mathematical ModeClosed-Form GF(p) Inversion
Total Actions Taken75 clicks (optimal)
Search Space Pruned100.0% heuristic bypass

Models interactive toggle masks and color cycle legends over the finite field GF(p). Rather than using recursive trial-and-error search, the engine solves x = A^(-1) * (v_target - v_0) mod p directly, clearing all 6 puzzle levels in exactly 75 deterministic clicks.