ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems
Chollet, Francois, Mike Knoop, Gregory Kamradt, Bryan Landers, Henry Pinkard · arXiv (Cornell University) · 2025
Master technical document formalizing the Spatial-Situational Thinker (SST) paradigm and the Eductive Capacity Index E = ω₁·(R_d·C_d/I_c)·(1-σ_k) as an original metric for distinguishing fluid intelligence from memorized knowledge in agents solving the ARC-AGI-2 benchmark (ARC Prize 2026). Includes: theoretical foundation (Spearman's eductive capacity and factor g), mathematical model with operational variable specification, statistical diagnosis EDA on 1,000 synthetic observations, ARC sandbox architecture (EductiveSandbox class), 93-agent coordinated research system with Administrator Agent, and full technical roadmap. Current solver score: 7.00% on public evaluation set.Target score: 25-35% after full agent architecture deployment. Developed with assistance from Google Gemini, Anthropic Claude, and OpenAI Codex. All conceptual design and intellectual property belongs exclusively to Hollman German Ojeda Silva.