This paper presents ARCANA, a collaborative multi-agent framework designed to solve ARC AGI 2 tasks under strict test time and hardware constraints. ARCANA decomposes each task into iterative perception, hypothesis generation, symbolic execution, and reflective refinement.
- Perceptual Grounding Agent: Constructs object-centric scene graphs from raw grids.
- Latent Program Policy: Proposes diverse domain-specific language (DSL) programs.
- Symbolic Executor: Verifies candidates based on demonstrations.
- Reflective Agent: Synthesizes failure-driven feedback for the next iteration.
These agents communicate through a shared differentiable blackboard and are scheduled by a learned meta-controller. This design combines structured program search with adaptive multi-turn correction, improving reasoning efficiency and solution quality on challenging abstract transformation tasks.
Blogger's Review: The design of the ARCANA framework showcases the potential of multi-agent systems in tackling complex reasoning tasks. By decomposing tasks and employing adaptive learning, it significantly enhances the efficiency and accuracy of program synthesis, paving the way for future intelligent system development.