Extending Reasoning to the Real Physical World, NVIDIA Cosmos-Reason1

Cosmos-Reason1 series model focuses on general knowledge reasoning, recognizes incomplete options, and outperforms OpenAI ο1 model; the model adopts a hybrid Mamba-MLP-Transformer architecture with 8B and 56B versions, and undergoes visual pre-training, fine-tuning, and reinforcement learning; it significantly outperforms benchmarks on general knowledge and embodied reasoning tasks. models, and can correctly recognize spatial relationships, temporal dynamics, and object persistence.