The 2026 NeuroGolf Championship closes on Kaggle today, and it asks entrants to do something almost nobody optimizes for anymore: make the network smaller. Not more accurate at any cost. Not bigger. Not trained on more data. Smaller, in parameters, in memory, and in the raw number of multiply-accumulate operations it takes to run.
More than 3,000 teams spent the past three months building neural networks to solve tasks from the ARC-AGI benchmark, then submitting them as ONNX files under a scoring rule that punishes every extra byte. The competition is hosted by The Neurosynthetic Research Institute and runs as part of the IJCAI-ECAI 2026 Competitions Track, the flagship academic AI conference that convenes roughly 4,000 researchers in Bremen, Germany this August. Top teams may present at a dedicated session there, and the organizers have committed to inviting winners to co-author a contest retrospective for the IJCAI 2027 Demo Track.
The prize pool is $50,000, split across the main leaderboard plus separate awards for the top student team and for whichever team held the leaderboard lead the longest. That last category is unusual, and it says something about what the organizers actually wanted to reward: not just a final snapshot, but sustained, iterative optimization.
How the Scoring Actually Works
NeuroGolf borrows its task set from ARC-AGI, the Abstraction and Reasoning Corpus that François Chollet introduced in 2019 as a benchmark for fluid intelligence rather than memorized skill. Each ARC-AGI task shows a handful of input-output grid pairs and asks a solver to infer the transformation rule, then apply it to a new grid. Humans solve the vast majority of these puzzles on the first try. Large language models, despite years of scaling, still struggle with the format, because the tasks are deliberately novel and cannot be solved by pattern-matching against training data alone.
Most ARC-AGI competitions ask a simple question: can a system solve the puzzle at all. NeuroGolf asks a different one. Given that a task is solvable, what is the smallest possible network that solves it correctly, and nothing more.
Competitors work against the 400 public tasks in the ARC-AGI-1 training set. For each one, they must submit a separate ONNX file, named task001.onnx through task400.onnx, packaged into a single submission.zip. Every network is scored with the formula max(1, 25 minus the natural log of its cost), where cost originally summed the total parameter count, the memory footprint, and the total multiply-accumulate operations the network needs to run. The organizers dropped multiply-accumulate operations from that formula on May 4, after competitors demonstrated that the static-analysis estimator behind the MAC count could be gamed to claim a near-perfect score on an arbitrary task, leaving cost as the sum of parameter count and memory footprint for the remainder of the competition. Because the scoring curve is logarithmic, the first cuts to a bloated network are cheap to earn and the last cuts, once a solution is already near-minimal, cost enormous engineering effort for a fraction of a point. A team that halves its cost from 100,000 units to 50,000 gains about 0.7 points. Halving again from 50,000 to 25,000 gains the same 0.7. The formula rewards structural rethinking over brute-force pruning, because pruning alone runs into diminishing returns fast.
The rules close off the obvious shortcut. Competitors cannot smuggle in a large pretrained model and gate it behind control flow that only activates a small piece of it per task. Every tensor and parameter must have a statically defined shape, and the ONNX operator set explicitly excludes Loop, Scan, NonZero, Unique, Script, and Function, the operations that would let a network branch, iterate, or otherwise hide computation inside dynamic logic. Each file is capped at 1.44 megabytes. Under those constraints, dynamic behavior has to be replaced with parallel, statically shaped structure, which is a fundamentally different engineering problem than training a large model and shrinking it after the fact.
Correctness is checked three ways: against the original ARC-AGI public tasks, against ARC-GEN-100K, a procedurally generated dataset built to produce many structural variants of each task type, and against a small private benchmark suite the organizers do not disclose. The combination is designed to catch a network that has essentially memorized the 400 published examples rather than learned the underlying transformation.
What Competitors Actually Built
Public notebooks shared during the competition show a spread of approaches that says a lot about where the real difficulty sits. Several of the highest-scoring public entries are not learned networks trained end to end. They are hand-built or rule-derived ONNX graphs, assembled task by task, sometimes with a search procedure choosing among candidate operator sequences rather than gradient descent tuning weights. One widely forked notebook documents a technique its author calls zero-MAC heuristics, which identifies tasks where the required transformation is a geometric operation, like a crop, a tile, or a color remap, that ONNX can express with near-zero computational cost rather than a learned convolution. Another applies what its author labels graph surgery: taking an existing public submission and running it through simplification passes, index rewrites, broadcast compression, and FP16 precision conversion to shave cost without changing the output. Public score on that notebook climbed from an early submission into the low 7000s out of a theoretical maximum near 10,000, the sum of a perfect 25-point score on all 400 tasks.
The pattern across the leaderboard discussion is iterative and public. Competitors copy each other’s submissions, and the “copy and edit” counts on the most-forked notebooks run into the thousands, then apply a specific optimization and resubmit. It looks less like a single breakthrough and more like a slow, collective search through a solution space that rewards precision over cleverness, not unlike the iterative self-optimization loop MiniMax M2.7 ran on its own training harness, just distributed across more than 3,000 human teams instead of one automated agent.
Why a Contest Like This Matters Now
Model efficiency stopped being a niche concern somewhere in the last two years. Inference cost, not training cost, is what most AI companies now pay on every user query, and on-device deployment puts a hard ceiling on how large a model can be before it simply will not fit on a phone or an embedded chip. The same pressure shows up at the frontier end of the field, where DeepSeek V4’s hybrid attention mechanism cut KV cache memory by 10x through architectural redesign rather than brute-force pruning, the same structural-rethinking instinct NeuroGolf rewards at a much smaller scale. NeuroGolf turns that pressure into a competitive format, and in doing so it produces something unusual: networks small enough, and constrained enough, that a person can actually read the ONNX graph and understand what every operation is doing. Interpretability research spends enormous effort trying to reverse-engineer what a billion-parameter model has learned. NeuroGolf networks, by construction, are close to legible from the start.
The competition also sits inside a research lineage that predates it. ARC-AGI itself has motivated more than one prior contest built around efficient, task-specific solving rather than raw benchmark accuracy, and Marília Prata, whose ARC-GEN-100K dataset backs part of NeuroGolf’s grading, references a prior Golf-branded ARC-AGI event from 2025 in her own competition notebook. That suggests “model golf,” the practice of optimizing an ML system the way a code golfer optimizes source length, is turning into a recognizable genre rather than a one-off gimmick, alongside efficiency-focused tracks like MLPerf Tiny and quantization leaderboards elsewhere in the field.
Where the Format Runs Into Trouble
The most direct criticism of NeuroGolf is built into its own leaderboard: several of the strongest entries are not neural networks in any meaningful sense. They are hand-derived symbolic solvers exported to the ONNX format because ONNX is the required submission container, not because a network was trained. That is a defensible design choice given the scoring goal, minimizing computational cost by any legitimate means, but it sits in tension with a competition named for neural networks specifically. A hand-coded crop-and-recolor routine that happens to run through an ONNX graph is answering a different question than “how small can a trained model be,” even if the leaderboard cannot tell the difference. The mid-competition scoring change compounds this: once MACs stopped counting toward cost, the incentive to avoid learned computation entirely, in favor of near-zero-cost hand-derived operations, only got stronger.
ARC-AGI-1’s public training set has also been public since 2019 and has absorbed years of targeted engineering attention. Chollet’s own team introduced ARC-AGI-2 specifically because the original benchmark had become vulnerable to exactly this kind of task-specific optimization rather than general reasoning, noting that frontier systems could score well on curated public tasks while still failing badly on held-out ones. NeuroGolf mitigates this with its private benchmark suite and the ARC-GEN-100K synthetic variants, but the organizers have not published how much those safeguards actually move the needle against a leaderboard full of hand-tuned solutions. It is the same gap between a clean leaderboard number and messier real-world behavior that showed up when eight frontier models all lost money betting a Premier League season despite strong scores on more conventional benchmarks. And because the final leaderboard is calculated against held-out test data that is not visible before the close of the competition, public notebook scores like the ones cited above are a useful signal of technique, not a preview of the final ranking.
What Happens Next
With submissions closed as of today, the organizers move into scoring against the private suite, and results should firm up in the weeks before the IJCAI-ECAI conference in August. The invitation to a dedicated Bremen session, plus the commitment to a co-authored retrospective for IJCAI 2027, means this competition is meant to produce a durable research artifact, not just a leaderboard snapshot that disappears once the prize money is paid out.
The bigger question is whether the format spreads. Efficiency-first competitions require a different kind of scoring infrastructure than accuracy leaderboards, since cost has to be measured precisely and consistently across submissions, and the mid-competition MAC removal is itself a case study in how hard that infrastructure is to get right on the first try. They also attract a different kind of competitor, one more interested in systems-level optimization than in squeezing another point out of a loss function. If on-device inference keeps growing the way the industry expects, contests built around minimal cost rather than maximal accuracy look less like a novelty and more like an early version of how efficiency claims will get benchmarked going forward.
Leave a Reply