ARCANA: A Reflective Multi-Agent Framework for Solving ARC-AGI-2 Tasks
Researchers introduced ARCANA, a multi-agent AI framework that solves ARC-AGI-2 tasks by iteratively perceiving, hypothesizing, executing, and reflecting. It operates under strict test-time and hardware constraints, using specialized agents to build scene graphs, propose DSL programs, verify candidates, and synthesize failure-driven feedback.
Researchers have introduced ARCANA, a collaborative multi-agent framework designed to solve ARC-AGI-2 tasks under strict test-time and hardware constraints. Unlike monolithic AI models, ARCANA decomposes each task into iterative stages: perception, hypothesis generation, symbolic execution, and reflective refinement. A perceptual grounding agent builds object-centric scene graphs from raw grids, a latent program policy proposes diverse DSL programs, a symbolic executor verifies candidates on demonstrations, and a reflective agent synthesizes failure-driven feedback to guide the next turn. These agents communicate and refine their approach over multiple cycles, mimicking a human-like problem-solving loop.
ARCANA's architecture is specifically tailored for the ARC-AGI-2 benchmark, which requires strong generalization from few examples. By combining symbolic execution with reflective learning, the framework can adapt its strategies based on past failures, improving efficiency without requiring massive computational resources. This makes it a promising step toward more capable and resource-efficient AI systems.
For those interested in the technical details, the full paper is available on arXiv under the title 'ARCANA: A Reflective Multi-Agent Program Synthesis Framework for ARC-AGI-2 Reasoning'.