A new artificial intelligence architecture presents a different approach to automatic reasoning. Instead of generating words to develop its intermediate steps, the model processes information in a latent space. This fundamental change enables a more efficient form of reasoning.
The developed model works with 150 million parameters, which represents a significantly smaller amount compared to other AI systems that use language-based reasoning. The architecture was designed to optimize information processing without the need to generate text as part of the process.
The results obtained by the system demonstrate excellent performance in reasoning tasks, showing that it is possible to achieve sophisticated results without depending on word generation as an intermediate step. This represents an alternative to traditional models that need to articulate their thinking in language.
Additionally, the architecture requires more limited use of computer resources to function. This computational efficiency makes the technology more accessible and with lower energy consumption, which can expand implementation possibilities in different contexts where resources are more restricted.
