AI models are advancing rapidly, but lack common sense that humans possess. NVIDIA is developing tests to teach AI models physical common sense to navigate complex environments. Cosmos Reason, a reasoning model, excels in physical AI applications by generating temporally grounded responses, topping the physical reasoning leaderboard.

NVIDIA’s data factory team is training generative AI models on physical common sense through reinforcement learning. Annotators create question-and-answer pairs based on real-world videos, helping models reason through scenarios. The data is quality checked by analysts before being fed to the model for training on the physical world’s limitations.

Reasoning AI models can predict outcomes and make sense of situations, demonstrating humanlike thinking. These models can analyze videos, infer scenarios, and provide insights into their responses. The NVIDIA reasoning model innovation, led by the Cosmos Reason team, aims to develop intelligent autonomous agents and physical AI systems for real-world interactions.

Read more at NVIDIA: How Do You Teach an AI Model to Reason? With Humans