ARC-AGI-3

• Part of the ARC-AGI series by the ARC Prize Foundation, focused on evaluating agentic (not just predictive) intelligence
• Unlike previous ARC benchmarks, ARC-AGI-3 is interactive — agents must act in environments rather than just produce outputs
• Paper: ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence
  • Submitted: 2026-03-24
  • Authors: ARC Prize Foundation

Agents must explore, infer goals, build internal models of environment dynamics, and plan effective action sequences without explicit instructions.

• Interactive benchmark — agents explore dynamic environments, infer goals, and plan actions without being given explicit instructions
• No language or external knowledge required — grounded purely in Core Knowledge priors to avoid memorization shortcuts
• Human-calibrated difficulty — environments validated through extensive human testing with established human performance baselines
• Scoring by efficiency — success is measured relative to human action baselines, not just binary task completion

• Human test-takers: 100% success rate
• Frontier AI systems: <1% score
• Represents a stark gap between human and AI adaptive problem-solving on novel tasks

• Previous benchmarks (including ARC-AGI-1 and ARC-AGI-2) were largely solved or approached by frontier models through pattern matching and memorization
• ARC-AGI-3 forces genuine exploration and goal inference — capabilities that current LLMs fundamentally lack
• Sets a new milestone target for the field: matching human efficiency on novel interactive tasks