IBM proposes Token Maturation: token generation via continuous refinement

IBM published a paper on Token Maturation, an alternative to token sampling in language models. The method keeps a K-token buffer (the "liquid tail") where tokens iteratively refine in continuous vector space and are discretised one by one, remaining autoregressive while avoiding entropy collapse.

• A K-token buffer undergoes K prediction passes before discretisation into the vocabulary
• The method avoids entropy collapse, unlike sampling inside latent space
• A new guidance scale s steers tokens toward the prompt and limits trajectory
• Buffer length K sets the refinement horizon before a token is committed