ARC Prize 2026: How Neuraxon Evolves AI Brains Without GPUs

Qubic’s scientific team does not hand-design the brains of its Neuraxon agents: it lets evolution search for them. For ARC-AGI-3, the interactive benchmark of the ARC Prize 2026, the team tested more than 110 configurations and millions of evolved brains on about 500 CPU cores, with no GPU and no pretrained model, according to Neuraxon Intelligence Academy Vol. 14. Its best offline score reached 0.38, and its latest Kaggle submission scored 0.41. This guide explains how evolutionary neural architecture search works, how Neuraxon brains play ARC-AGI-3, what the experiments revealed, what Neuraxon 3.0 will search for next, and how to download the code and the 603,054-row dataset behind it.

What Is Neural Architecture Search?

Neural architecture search is the automated design of a neural network’s structure: how many units it has, how they connect, and which rules update them. Training adjusts the strength of connections inside a fixed design; NAS decides the design itself. The field’s reference survey, by Elsken, Metzen and Hutter (JMLR, 2019), breaks every NAS method into three parts.

1. A search space: the set of architectures the method is allowed to consider.
2. A search strategy: the way it explores that space.
3. A performance estimate: the way it decides which candidate is better, without fully training every one.

Think of it as the difference between learning to cook and designing the kitchen. Training teaches a network to cook well; NAS decides where the oven, the sink and the worktop go.

Four strategies, four cost profiles

Four families of search strategies dominate the literature, as summarized in our introduction to neural architecture search. Their compute bills differ by orders of magnitude.

NAS is already in your pocket. Google used it to design MobileNetV3, the image network behind many phone camera features, because a phone needs a model that is accurate, small and fast at the same time. Evolution also found improvements to the Transformer itself in The Evolved Transformer (So, Liang and Le, 2019).

NAS is not hyperparameter tuning

Changing a learning rate or a batch size is hyperparameter tuning: it fine-tunes how a given network learns. NAS changes what kind of network it is. The distinction matters because the same rule applies to both: a search method only earns its compute if it beats random search on the same budget, as Bergstra and Bengio showed for hyperparameters in 2012.

Why Use Evolution to Design a Brain?

Because that is how nature did it. In animals, learning adjusts the strength of synapses during a lifetime, while the architecture of the brain is shaped over generations by evolution. In a 2019 paper in Nature Communications, neuroscientist Anthony Zador called this the “genomic bottleneck”: a genome is far too small to specify every synapse, so evolution encodes a compact recipe for wiring and learning, not the finished network.

An older result points the same way. In 1987, Hinton and Nowlan showed in a simulation that learning can guide evolution: individuals able to learn find good solutions faster, and evolution then selects for that ability to learn. This is known as the Baldwin effect.

Large language models took the opposite route. The Transformer is a hand-designed architecture that does not change while the model is used, and only a small part of it has ever been searched automatically. Neuraxon bets that general intelligence needs both levels: evolution to choose the architecture, and learning inside each lifetime. The long-standing method for evolving network structure is NEAT, by Stanley and Miikkulainen (2002), which grows topologies from simple to complex.

What Is Neuraxon?

Neuraxon is a brain-inspired computing unit developed by Qubic science advisors David Vivancos and José Sánchez, released as open source on GitHub. Where a classic artificial neuron outputs a number at each step, a Neuraxon neuron runs in continuous time and takes one of three states: excitatory (+1), neutral (0) or inhibitory (-1), an idea explored in Beyond Binary: Ternary Dynamics.

Each synapse carries three weights that change at different speeds (fast, slow and meta), and synapses can form, collapse or reconnect. Four simulated neuromodulators (dopamine, serotonin, acetylcholine and noradrenaline) act through nine receptor subtypes, as described in our volume on neuromodulation. Astrocyte-like cells gate plasticity, a mechanism covered in Astrocytes and Brain-Inspired AI, and four homeostatic loops keep activity in a healthy range.

That richness is exactly why architecture search matters here. A Neuraxon brain has dozens of structural choices that no designer can set well by hand. Evolution, through Qubic’s Aigarth framework, is the tool that explores them.

What Is ARC-AGI-3, and Why Is It So Hard for AI?

ARC-AGI-3 is the first interactive benchmark of the ARC series, launched on March 25, 2026 by the ARC Prize Foundation. It is a set of turn-based games in which no one explains the rules: the player sees a grid of colours, acts, and must work out what winning means. Vol. 14 counts 25 public development games and 55 hidden evaluation games. At launch, humans scored 100% and frontier AI 0.51%. We explained why this kind of test matters in How Do We Measure the Intelligence of a Machine?

The score rewards efficiency, not just success. For each level, the official methodology divides the number of actions a human needed by the number the AI needed, then squares the result: an agent that takes ten times more actions than a human scores 1% on that level. Later levels weigh more, and the final score averages all games.

The competition rules raise the bar further. ARC Prize 2026 offers $2 million in prizes across its three tracks, according to the official competition page. Kaggle evaluation runs without internet access, API-based systems such as GPT or Claude are barred, and leading participants must open-source their solutions to be eligible for a prize. Every entry has to bring its own intelligence into a sealed room.

How Does Neuraxon Play ARC-AGI-3?

A Neuraxon brain receives the grid as numbers, picks actions, and keeps learning during play through plasticity. Neural architecture search works before the game starts: it sets the build of the brain, namely which regions of the screen it attends to, how colours map to inputs, how it balances exploring and exploiting, and how far its clicks can reach.

The team set one ground rule, according to Vol. 14: studying public games, human playthroughs and analyses to choose the search space is legitimate; giving that information to the agent during play is not.

A three-phase evolutionary protocol

The scale was substantial: more than 110 search configurations and millions of evolved brains, run on about 500 CPU cores with no GPU. The best score on the offline Kaggle version reached 0.38, according to Vol. 14, and the team’s latest submission to the Kaggle competition scored 0.41, a new personal best, according to the Qubic Scientific Team. Submissions closed on November 2, 2026, and the ARC Prize Foundation will announce results on December 4.

How to read the 0.41

The number needs context. The leading entries of the competition wrap open-weight language models of 27 to 31 billion parameters inside agent harnesses, according to an October 2026 analysis by Metir AI, and score far higher. Neuraxon brains start with no pretrained knowledge of language or images and run on ordinary processors. The score measures how far evolution alone can take a small brain that must discover each game from scratch, not how it ranks against systems that bring billions of parameters of prior knowledge.

The team is candid about what the top entries do better: they predict what will change on screen before acting, and they learn from their errors. Vol. 14 states that this is the direction Neuraxon is now taking.

What Did the ARC-AGI-3 Experiments Reveal?

Vol. 14 is unusually open about what went wrong, and each failure matches a classic trap of neural architecture search.

1. The winner’s curse. One group of brains cleared a game on every seed used to select them, and cleared none of the new seeds. Picking the best of many candidates on the same data used to rank them inflates the winner’s score by luck, as Cawley and Talbot (2010) showed. It is the student who memorized last year’s exam answers.
2. Success without learning. In simple games, a fixed policy plus some exploration can clear a level without understanding the rules. Successes also clustered in a few games, almost always at the first level.
3. No proof of learning. A high score does not show that a brain learned anything during play.

“NAS selects the final result and does not tell us whether there is learning inside.”

Qubic Scientific Team, Neuraxon Intelligence Academy Vol. 14, November 4, 2026

The team turned these lessons into three rules: select on some seeds and report on others; keep some public games out of the search entirely; and always subtract a random-action baseline run on the same budget. That last rule echoes Li and Talwalkar (2020), who found random search to be a surprisingly strong competitor in NAS.

What Is Neuraxon 3.0?

Neuraxon 3.0 is the next version of the brain, designed to close the gap ARC-AGI-3 exposed: today’s brains explore and adapt within a game, but they do not set goals or chain actions toward them, according to Vol. 14. The search question changes accordingly. It is no longer “which network wins?” but “which mechanisms and time constants does a brain need to learn in a new environment?”

The design adds five mechanisms, each working on a different timescale.

The search covers mechanisms, not raw parameter counts. Each mechanism has a range bounded by physiology, and each can be switched off to measure exactly what it contributes.

A new selection criterion: learning speed

Neuraxon 3.0 brains will be rewarded for how quickly they reduce their prediction error per action in games they have never seen. Under this criterion, a brain that wins by luck gains nothing. It also targets the two capacities that set the top ARC-AGI-3 entries apart: anticipating what will change on screen, and learning from mistakes.

Safeguards written before the data

The protocol is built against the traps listed above, according to Vol. 14.

• The experimental design is fixed before any data is viewed.
• The architecture selection is sealed with a digital fingerprint (a hash) before the test set is generated.
• Every search is compared against a random search on the same budget.
• Seeds are paired, results are corrected for multiple comparisons, and a minimum effect size is declared in advance.
• No experiment is stopped early because results look favourable.
• Results will be published when the full experiments are done, whatever their conclusion.

Neuraxon 3.0 is an architecture in design with preliminary tests, and Vol. 14 gives no release date. As of November 6, 2026, the public GitHub repository goes up to Neuraxon 2.0.

Why Qubic’s Network Is a Natural Fit for Evolutionary Search

Evolutionary search is massively parallel by nature: each brain in a population can be evaluated independently. That matches how Qubic works. Through Useful Proof of Work, the computing power that secures the network is also directed at AI tasks, as explained on our mining page.

Verification is the second fit. A result that comes from millions of evaluations spread across many machines needs to be trusted. Qubic’s network is run by 676 Computors that reach agreement through a Byzantine fault-tolerant quorum, so a claimed result can be checked by many independent parties rather than taken on faith. Neuraxon 3.0’s sealed, pre-registered protocol is built for exactly that kind of checking. For the broader vision, see What Is AGI? and the Qubic science hub.

Open Code and Open Data: How to Explore Neuraxon Yourself

Everything needed to reproduce or extend the work published so far is public.

The quick start in the repository shows how to build a network, feed it inputs, apply neuromodulation and run an Aigarth evolution in a few lines of Python, with no external dependencies for the core file.

What Comes Next

Three milestones frame the coming weeks. Papers for the ARC Prize 2026 paper track are due on November 8, the competition results will be announced on December 4, and the Neuraxon 3.0 experiments will be published once complete, whatever their outcome. The measure to watch is not the leaderboard alone, but whether evolved brains begin to learn within a game, which is the capacity that separates today’s best entries from everything else. To follow the series from the beginning, start with NIA Vol. 1: Neuraxon Time.

La Rédaction C.

The Cointribune editorial team unites its voices to address topics related to cryptocurrencies, investment, the metaverse, and NFTs, while striving to answer your questions as best as possible.