Coconut (Chain of Continuous Thought)
Lets an LLM reason in continuous latent space by feeding its last hidden state back as input instead of writing out tokens.
Argues language tokens may be a poor medium for reasoning; continuous thoughts can hold several candidate next steps (BFS-like), beating chain-of-thought on search-heavy logic tasks in the paper.
- Date
- Monday, 9 December 2024
- Lab
- Meta
- Kind
- paper
- Access
- research preview
Authors' affiliation (Meta FAIR) not shown on the arXiv abstract page opened; arXiv v1 2024-12-09. Small-scale results on synthetic reasoning tasks.
Sources
This record was checked against its sources on 6 October 2026. How we check