ARCANA framework cracks ARC-AGI-2 with reflective multi-agent program synthesis

Researchers from an unnamed institution (Zhang et al.) have released ARCANA, a reflective multi-agent program synthesis framework designed to tackle the notoriously difficult ARC-AGI-2 benchmark. Unlike single-model approaches, ARCANA decomposes each abstract reasoning task into four specialized stages: perception, hypothesis generation, symbolic execution, and reflective refinement. Each stage is handled by a dedicated agent—a perceptual grounding agent that builds object-centric scene graphs from raw grids, a latent program policy that proposes diverse DSL programs, a symbolic executor that verifies candidates against demonstrations, and a reflective agent that generates failure-driven feedback for the next iteration. These agents communicate through a shared differentiable blackboard and are orchestrated by a learned meta controller that schedules turns based on task difficulty.

ARCANA combines structured program search with adaptive multi-turn correction, enabling the system to reason iteratively without human intervention. The framework is optimized for strict test-time and hardware constraints, making it practical for real-world deployment. By leveraging symbolic execution and explicit failure analysis, ARCANA achieves higher solution quality on challenging abstract transformation tasks compared to end-to-end neural approaches. The paper is available on arXiv (2607.09059) and marks a significant step toward transparent, interpretable AI reasoning systems that can generalize beyond their training data.