Source: ARC-AGI series and ARC-AGI-3 (ARC Prize Foundation)

ARC Prize pages describing the ARC-AGI series and ARC-AGI-3's move to interactive game environments scored on human action efficiency.

Contents

Summary

ARC-AGI (Chollet, 2019) defines intelligence as skill-acquisition efficiency on novel tasks, uses only core cognitive priors rather than cultural knowledge, and follows an “easy for humans, hard for AI” design rule. ARC-AGI-3 replaces static puzzles with interactive environments.

Key claims

  • ARC-AGI-3 agents act in game environments with no natural-language instructions and must learn the rules by exploration.
  • It measures learning efficiency, long-horizon planning with sparse feedback, and multi-step adaptation.
  • Scoring is relative to human efficiency: 100% means beating every game as efficiently as humans.
  • ARC Prize 2026 has closed; results are due 4 December 2026 (as of 2026-11-14). Kaggle hosts ARC-AGI-2 and ARC-AGI-3 competitions.

Feeds

Benchmarks

Suggest changes →