Researchers Achieve 67% Accuracy on ARC-AGI Benchmark Using Cost-Efficient Agent Architecture

A new study demonstrates that open-weight models like DeepSeek V3.2 can solve abstract reasoning tasks efficiently without expensive test-time compute or benchmark-specific fine-tuning. Researchers introduced the Explorer-Definer Pipeline and Reflective Orchestrator—agent architectures that separate pattern discovery from program synthesis—achieving 67.25% pass@2 on ARC-AGI-1's 400-task public set for just $0.62 per task, a ~52-point lift from baseline.

Why it matters: This work reframes AI progress on abstract reasoning from compute-heavy approaches toward architectural innovation, offering a practical, cost-effective path forward for solving visual pattern recognition tasks that matter for AGI evaluation.