NVIDIA researchers have claimed victory in a pivotal Kaggle competition widely regarded as a real-time indicator of progress toward artificial general intelligence (AGI). Ivan Sorokin and Jean-François Puget, two members of the Kaggle Grandmasters of NVIDIA (KGMoN), secured first place on the Kaggle ARC Prize 2025 public leaderboard with a score of 27.64% on the ARC-AGI-2 benchmark. After the competition, their team, known as NVARC, improved their performance to 29.72%, achieving this with remarkable cost efficiency—just 20 cents per task. The win highlights a major breakthrough in scalable, economical AI reasoning. Their fine-tuned 4-billion-parameter model outperformed significantly larger and more expensive models, demonstrating that high performance in abstract reasoning doesn’t require massive scale. The ARC-AGI benchmark evaluates an AI system’s ability to perform abstract reasoning and generalize from minimal examples using grid-based visual puzzles. ARC-AGI-2, the updated version used in this competition, is designed to be more challenging by eliminating overlap with publicly available training data. This prevents models from relying on memorization or pattern scraping, making it a rigorous test of true systematic abstraction. Unlike standard benchmarks that reward model size or data volume, ARC-AGI-2 demands that systems infer underlying rules from just a few examples and apply them to novel situations. As a result, scores on this benchmark are considered one of the most reliable proxies for measuring progress toward general reasoning in AI. The Kaggle ARC Prize 2025 stands out as the most open and reproducible platform for testing AGI-style reasoning under strict compute and time constraints. The competition’s tight runtime rules made traditional heavyweight AI methods—such as chain-of-thought prompting, tool use, or reinforcement learning agents—impractical. Instead, the NVARC team adopted a novel strategy: shifting complex reasoning tasks offline. They built a synthetic data pipeline that generated diverse, high-quality ARC-style puzzles using staged puzzle generation, concept decomposition, and progressively stronger open-weight models. This allowed them to train smaller, efficient models that could run quickly during evaluation. A key innovation was test-time training—where the model learns the specific rules of each puzzle from its limited examples in real time. This approach has become essential for top-performing systems on ARC-AGI benchmarks. The final solution was a compact, cost-effective ensemble that delivered state-of-the-art results without relying on massive models or brute-force computation. This achievement underscores the power of synthetic data and adaptive learning in advancing AI reasoning. To build their system, the team leveraged NVIDIA’s NeMo suite, including NeMo RL for scalable reinforcement learning and NeMo Skills for streamlining synthetic data generation pipelines. The success of NVARC offers a compelling blueprint for future AI development: intelligent design, efficient training, and strategic use of synthetic data can drive progress in general reasoning—without always needing bigger models.