New AI architecture delivers 100x faster reasoning than LLMs with just 1,000 training examples

Sapient Intelligence has developed Hierarchical Reasoning Models (HRM), a new AI architecture that outperforms large language models on complex reasoning tasks while using 100x less data and computational resources. Inspired by the human brain's dual-system processing, HRM uses two coupled modules for hierarchical reasoning in latent space rather than relying on chain-of-thought prompting. The 27M-parameter model achieved 40.3% on ARC-AGI benchmark, surpassing much larger models like o3-mini-high, and solved complex puzzles that stumped state-of-the-art LLMs. The architecture offers significant cost savings and speed improvements for enterprise applications requiring complex decision-making and long-term planning.