ARCANA: A Reflective Multi-Agent Program Synthesis Framework for ARC-AGI-2 Reasoning
06:00 · July 13, 2026 · arXiv cs.AI RSS
We present ARCANA, a collaborative multi agent framework for solving 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. 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 for the next turn. These agents communicate through a shared differentiable blackboard and are scheduled by a learned meta controller. The design combines structured program search with adaptive multi turn correction, improving reasoning efficiency and solution quality on challenging abstract transformation tasks.
Summary
ARCANA is a multi-agent framework that addresses abstract reasoning on ARC-AGI-2 tasks, where models must infer compact transformation rules from a handful of input-output grid demonstrations that vary in size and object composition. The system frames each task as a multi-turn reasoning episode and decomposes it across four specialized agents that iteratively refine candidate solutions within tight test-time and hardware budgets.
A Perceptual Grounding Agent first converts raw grids into object-centric scene graphs. It employs a 2D-aware Transformer with rotary positional encodings and differentiable Slot Attention to extract entities and their relations. A Hypothesis Generation Agent, implemented as a conditional variational autoencoder, then proposes diverse programs drawn from a domain-specific language. These candidates are passed to a Symbolic Execution Agent that runs them against the demonstration pairs and records execution traces. Finally, a Reflective Refinement Agent performs counterfactual analysis on the traces to generate targeted feedback that steers subsequent program proposals away from previously unsuccessful regions of the search space.
The agents exchange information through a shared differentiable blackboard whose state is updated at each turn. A learned meta-controller decides which agents to activate and how to allocate a limited compute budget, enabling adaptive multi-turn correction rather than single-shot generation. The entire architecture is trained end-to-end with a Reasoning Trajectory Optimization objective that rewards both final grid correctness and intermediate reasoning efficiency.
Under the official ARC Prize 2026 hardware constraints, this combination of structured program search and agentic refinement yields state-of-the-art results among open-source systems on ARC-AGI-2, narrowing the gap to human-level performance on compositional visual reasoning tasks.
Why it matters
This highly technical paper is directly relevant to AI researchers and advanced practitioners in the Netherlands working on AGI, multi-agent systems, and abstract reasoning. Its focus on achieving state-of-the-art results under strict hardware constraints makes it highly actionable for Dutch research labs and AI-driven SMEs looking to deploy efficient reasoning models.
