What is ARC-AGI-2 and why it matters

ARC-AGI-2 is the second generation of a benchmark developed by the ARC Prize Foundation, founded by François Chollet (creator of Keras) and Mike Knoop (Zapier). The test evaluates a system's ability to infer abstract transformation rules from just a few examples — exactly the skill that humans handle intuitively, but current AI models fundamentally struggle with.

At first glance, ARC-AGI-2 tasks look like colored grids: from a few input-output pairs, the system must grasp a hidden rule (for example, "move all blue objects two squares to the right and outline them in red") and apply it to a new input. Unlike language tests, where statistical pattern matching often suffices, here the system must genuinely deduce the logic of the transformation — rote memorization won't help.

The ARC Prize 2026 competition offers 2 million dollars in rewards for teams that can push the boundaries of machine abstract reasoning. And ARCANA is entering this race.

Four agents, one goal: understand the transformation

ARCANA (Adaptive Reasoning with Collaborative Agent Network Architecture) approaches each task as a sequence of iterative rounds in which four specialized agents collaborate:

1. Perceptual Grounding Agent (PGA) — the "eyes" of the system

The first agent converts the raw grid into an object-oriented scene graph. Using a 2D Transformer with rotary positional encoding (RoPE-2D) and differentiable Slot Attention, it identifies individual objects, their colors, positions, and mutual relationships. Instead of working with bare pixels, subsequent agents receive a structured representation of the scene — similar to how a human "sees" squares, lines, and their relative positions in a grid.

2. Hypothesis Generation Agent (HGA) — the "brain" that proposes solutions

The second agent is the core of the entire system. Based on the task representation and any feedback from previous attempts, it generates diverse candidate programs in a specialized language with 47 primitives (geometric transformations, color operations, object manipulation, control structures). It uses a conditional variational autoencoder (CVAE) and a determinantal point process (DPP), which ensures maximum diversity among proposed hypotheses — the system doesn't put all its eggs in one basket but tries fundamentally different approaches.

3. Symbolic Execution Agent (SEA) — the "verifier"

The third agent symbolically executes each hypothesis on the demonstration examples and compares the result with the correct output. It uses a combination of exact match, cell-level accuracy, and structural similarity (SSIM). The result is a detailed execution trace showing exactly where and why the program failed.

4. Reflective Refinement Agent (RRA) — the "teacher" that fixes mistakes

This is ARCANA's main innovation. The RRA performs what's called counterfactual credit assignment: for each failure, it systematically tests what would happen if a certain step of the program were omitted. Steps whose absence would improve the result are marked as "harmful." From this analysis, it synthesizes feedback that steers hypothesis generation away from previously unsuccessful areas in the next round. In principle, it's what an experienced programmer does during debugging — only it happens automatically and within seconds.

Meta-Controller: a conductor on a limited budget

All four agents are connected through a shared differentiable "blackboard" over which they exchange information. The Meta-Controller decides which agent activates in a given round and how many computational resources it receives — a learned neural network trained via policy gradient that learns to adaptively allocate a limited budget.

The ARC Prize 2026 competition rules impose a strict limit: a maximum of $0.42 per task and computation on four NVIDIA L4 GPUs within a 12-hour offline window. ARCANA not only meets this ceiling but uses it more efficiently than the competition — averaging just $0.16 per task (roughly 3.70 CZK).

Results: best among open-source, but humans lead by a mile

On 120 tasks from the semi-private evaluation set, ARCANA achieved 32.5% accuracy, 67.4% cell-level accuracy, and a CNE (compute-normalized efficiency) score of 7.90. For comparison:

• EPS-Grok (evolutionary synthesis with Grok-4): 26.0%
• NVARC (LongT5 + test-time training): 24.0%
• SOAR (evolutionary synthesis, 8B parameters): 18.5%
• OmniARC (Qwen2.5-0.5B): 14.2%
• Human: 75.0%

Interestingly, ARCANA has only 48 million parameters — compared to billion-parameter LLMs, it's a tiny model. The biggest contribution comes precisely from reflective feedback (RRA), which alone adds 10.8 percentage points, and test-time adaptation via LoRA with an additional 7.5-point contribution.

ARCANA's largest lead over competitors is on tasks requiring combinations of multiple rules (+8.8%) and multi-step sequential transformations (+8.9%). This is where the power of iterative refinement truly shines — the system averages 3.7 rounds, with difficult tasks going up to 8.

What this means for AI research — and for the Czech Republic

ARCANA highlights an important trend: architecture and cooperation among specialized modules can matter more than raw computational power. While large language models rely on trillions of parameters and massive training datasets, ARCANA achieves top-tier results through smart problem decomposition and iterative refinement.

This is relevant to the Czech AI scene for several reasons. First, low computational requirements (48M parameters, $0.16 per task) mean that similar approaches can be tested by smaller research teams and universities — such as CTU, Charles University, or BUT in Brno, all of which are actively involved in AI research. Second, the principles of reflective reasoning and multi-agent coordination can also be applied to industrial problems, such as automatic anomaly detection in manufacturing or technical documentation analysis — areas where Czech companies are seeking concrete AI solutions.

Third, the Czech AI Factory is launching next year in Ostrava — a European hub for artificial intelligence that could support similar research directions with computing capacity and grants.

Why the massive gap behind humans?

Despite the impressive leap over the competition, one question remains: why are humans at 75% while the best AI is at 32.5%? The answer lies in the very nature of ARC-AGI-2. The tasks test fluid intelligence — the ability to solve completely new problems without prior experience. Humans innately understand concepts like "symmetry," "rotation," "overlap," or "sequence" and can flexibly combine them. Current AI systems — including ARCANA — must "learn" each rule from a limited number of examples and still fail on tasks requiring deeper understanding of spatial relationships or hidden logic.

The study's authors openly admit that symbolic interpretation remains the hardest challenge for all current methods. And that's precisely why the ARC Prize competition matters: until AI reaches at least 85%, we cannot speak of true abstract reasoning at a human level.