Latent-space Reasoning
A reasoning paradigm in which a model performs multi-step reasoning in a continuous latent vector space rather than generating reasoning steps in a discrete text token space, decoding the result into text only at the end. Its efficiency advantage stems from skipping token-by-token autoregressive generation, at the cost of transparency and tool-calling capability.
Timeline (last 90 days)
Latent space reasoning refers to a model not generating discrete text tokens, but completing multi-step reasoning in a continuous latent vector space, decoding the result into text only at the end; representative work includes Coconut.
Pure latent space reasoning cannot call external tools in the middle of reasoning, because the model does not generate tool-call instructions in text form.
All Facts (2)
Latent space reasoning refers to a model not generating discrete text tokens, but completing multi-step reasoning in a continuous latent vector space, decoding the result into text only at the end; representative work includes Coconut.
50%UnverifiedPure latent space reasoning cannot call external tools in the middle of reasoning, because the model does not generate tool-call instructions in text form.
50%