ARC-AGI Prize Explainer (Includes ARC-AGI-3!)

LLM-legible version of Michael Hodel's arc-dsl

ARC-AGI2 solver — 84.0% on training (840/1000). Hybrid: 30+ hand-crafted solvers + Claude Sonnet 4.5 program synthesis with deterministic verification. No fine-tuning, no GPU.

ARC Prize 2024 Silver Medal Solution – Rule-based DSL symbolic reasoning system implemented in C++ for abstract grid transformation and search.

Achieve near-perfect ARC-AGI-2 benchmark results with reproducible code focused on solving complex AI reasoning tasks efficiently.

An agentic and heuristic solver for the Kaggle ARC Prize 2026 (ARC-AGI-3) competition. Built on Forge v3 using dynamic online CNN training with curriculum learning (CLTI) and navigation search heuristics.

Fully-offline, deterministic symbolic agent for the ARC Prize 2026 ARC-AGI-3 track. Hybrid occam (MIT) + step-wise solver, ~9/25 public games, no LLM/GPU in the loop.

A prototype Python solver for the ARC (Abstraction and Reasoning Corpus) Prize. REAP uses a symbolic DSL, object-based parsing, and heuristic search to tackle ARC’s reasoning challenges. Early results: 32 training tasks out of 1000 solved, demonstrating proof-of-concept reasoning from scratch.

What ARC-AGI is and why it matters — history, approaches, and the 2026 Kaggle tracks.

Open-source utilities from my ARC Prize 2026 (ARC-AGI-2) work: a Bearer-auth Kaggle API client for the new KGAT_ tokens, a kernel run poller, and a small ARC DSL.