Architecture
Latent Space Language Model
2024ResearchUpdated: 22 September 2026Published
Key
innovation
Moving prediction/reasoning from the token level into a latent space.
Category
Architecture
Abstraction level
Paradigm
Operation level
Model
Use cases
Latent-space reasoningConcept models (NCP, LCM)Efficient LLMs
How it works
The model encodes input into a latent space, predicts/processes latent representations, and only then decodes them into text tokens.
Problem solved
Operating solely on discrete tokens hinders longer-horizon planning and abstraction.