A team of researchers has published ARCANA, a collaborative multi-agent framework designed to solve ARC-AGI-2 reasoning tasks, the benchmark that tests abstract visual reasoning in ways that remain difficult for current AI systems. The paper, submitted to arXiv on July 10, proposes a structured approach that combines program synthesis with iterative self-correction.
ARC-AGI (Abstraction and Reasoning Corpus) was designed by Francois Chollet to measure a system’s ability to generalise from few examples, rather than relying on pattern-matching from massive training data. ARC-AGI-2, the current version, introduces more challenging tasks with stricter time and hardware constraints.
ARCANA decomposes each task into four specialised agent roles. A perceptual grounding agent builds object-centric scene graphs from raw input grids. A latent program policy proposes diverse domain-specific-language (DSL) programs. A symbolic executor verifies candidate programs against the provided demonstrations. A reflective agent then synthesises failure-driven feedback to guide the next iteration.
The agents share information through a differentiable blackboard, a shared memory structure that allows each agent to read and write intermediate representations. A learned meta-controller schedules which agent acts at each turn, balancing exploration of new program candidates against refinement of promising ones.
The design combines two approaches that have each shown promise on reasoning tasks: structured program search, which explicitly enumerates possible transformation rules, and adaptive multi-turn correction, which allows the system to learn from its mistakes within a single problem instance. By keeping the search within a defined DSL, ARCANA avoids the combinatorial explosion of unconstrained code generation while still maintaining the expressiveness needed for ARC-AGI-2’s visual transformation tasks.
The ARC-AGI benchmark has become a standard proving ground for AI reasoning research following its introduction in 2019. While most mainstream AI progress is measured on language and vision benchmarks, ARC-AGI’s grid-based puzzles require true compositional generalisation, applying understood concepts in novel arrangements, which remains a known weakness of large language models.
The paper is available on arXiv under the identifier 2607.09059.
Sources: ARCANA: A Reflective Multi-Agent Program Synthesis Framework for ARC-AGI-2 Reasoning (arXiv, Jul 10, 2026)