In the field of artificial intelligence, a fierce competition has recently taken place. NVIDIA's small model NVARC achieved an outstanding score of 27.64% in the latest ARC-AGI2 evaluation, surpassing its competitor GPT-5Pro's 18.3%, and successfully took the top position. This achievement not only demonstrates the powerful performance of NVARC, but also shows that its cost per task is only 20 cents, far lower than GPT-5Pro's $7, making it a real "king of cost-effectiveness" in the cost race.

The success of NVARC is attributed to its unique zero-pretraining deep learning approach. This strategy avoids the issues of domain bias and data dependency caused by pretraining on traditional large-scale general datasets. This evaluation was particularly challenging, as ARC-AGI2 adopted more difficult tests, aiming to examine whether the model can quickly learn and master new skills without direct training data.

image.png

NVIDIA team adopted an innovative approach for the training of NVARC. They moved complex reasoning processes to an offline synthetic data pipeline, using GPT-OSS-120B to generate high-quality synthetic puzzles, thereby reducing the demand for real-time computing resources. The team extracted questions from existing datasets and generated more complex new questions through combinations. To ensure the high quality of the generated data, they decomposed the reasoning process into multiple independent verification stages, ultimately forming a large synthetic dataset containing 3.2 million enhanced samples.

NVARC uses an improved version of the ARChitects method in its reasoning module and simplifies puzzle understanding with a conversational template. During training, they also used the NeMo RL framework and Megatron backend for supervised fine-tuning. Particularly noteworthy is the TTFT technology, which fine-tunes for each task, enabling the model to quickly adapt to new task rules.

Although some may question whether this small model is merely a "test machine," NVARC's success actually highlights its strong adaptability and efficiency within specific domains. The advantages of small models in terms of cost, speed, and adaptability make them particularly important in many application scenarios. In the future, how to apply the right methods to the appropriate fields will be key to further advancing the technology.