NVIDIA researchers Ivan Sorokin and Jean-Francois Puget won the Kaggle ARC Prize 2025, achieving first place on the public leaderboard with a 27.64% score on the ARC-AGI-2 benchmark. Their solution, named NVARC, utilized a fine-tuned 4B model variant that outperformed larger models at just 20 cents per task.
The team employed synthetic data generation and test-time training to enable efficient reasoning within strict compute limits. They leveraged the NVIDIA NeMo suite, including NeMo RL and NeMo Skills, to streamline reinforcement learning and synthetic data pipelines. The approach demonstrated that scalable, economical AGI-style reasoning is possible without relying on brute-force search or massive models.
This victory highlights a breakthrough in scalable, economical AGI-style reasoning, showing how synthetic data and adaptive learning can push reasoning forward while remaining accessible.