New Multi-Agent Framework ARCANA Tackles ARC-AGI-2 with Symbolic Reasoning and a Shared Blackboard

Four specialized agents, one shared memory space, and a strict time budget: that's the core setup behind ARCANA, a newly published framework aimed at cracking ARC-AGI-2, one of the hardest benchmarks for testing whether AI systems can generalize abstract reasoning rather than memorize patterns.

Described in a paper submitted to arXiv on July 10, 2026, and filed under the Artificial Intelligence (cs.AI) category, ARCANA stands for a collaborative multi-agent system built specifically to solve ARC-AGI-2 puzzles under real-world constraints on inference time and compute. Unlike approaches that throw a single large model at these grid-based transformation tasks, ARCANA splits the problem into four distinct stages: perception, hypothesis generation, symbolic execution, and reflective refinement.

How the Agents Divide the Work

The perceptual grounding agent goes first, converting raw input grids into object-centric scene graphs. Rather than treating a puzzle as a flat array of pixels or numbers, this step identifies discrete objects and their relationships, giving downstream agents a structured representation to reason over.

From there, a latent program policy generates a diverse pool of candidate programs written in a domain-specific language (DSL). This is where ARCANA leans on structured program search, exploring the space of possible transformations rather than relying purely on pattern-matching or free-form generation.

Each candidate then passes through a symbolic executor, which checks whether it actually reproduces the transformations shown in the task's demonstration examples. Programs that fail don't just get discarded. A dedicated reflective agent analyzes the failure and synthesizes targeted feedback, which feeds into the next attempt. This creates an adaptive, multi-turn correction loop: the system doesn't just retry blindly, it learns from specific mistakes within the same task.

A Shared Blackboard, Not Isolated Agents

What distinguishes ARCANA's architecture from a simple pipeline is how its agents coordinate. Instead of passing messages sequentially, the agents communicate through a shared differentiable blackboard, a common workspace where intermediate results and hypotheses are posted and read. A learned meta controller schedules access to this blackboard, deciding which agent acts next based on the current state of the problem.

This design lets the system allocate its limited test-time compute more deliberately. Rather than exhausting a fixed number of generic retries, the meta controller can prioritize perception fixes when scene understanding seems off, or route back to hypothesis generation when the symbolic executor keeps rejecting candidates for the same underlying reason.

Why It Matters for ARC-AGI-2

ARC-AGI-2 was designed to resist brute-force memorization and reward genuine compositional reasoning, making it a proving ground for whether an architecture can generalize from a handful of demonstration examples to novel transformations. ARCANA's combination of symbolic verification and reflective, failure-aware iteration targets exactly that gap: it tries to make each turn of reasoning more efficient rather than simply scaling up the number of attempts.

The paper does not report specific benchmark scores or comparisons against other ARC-AGI-2 submissions in the facts reviewed here, so how ARCANA stacks up against rival approaches remains an open question pending further scrutiny of the full results.