facebookresearch/coconut
A research implementation for training large language models to perform chain-of-thought reasoning directly in a continuous latent space rather than discrete tokens.
Not currently ranked — collecting fresh signals.
star history
The repository implements a method for training large language models to reason by operating in a continuous latent space instead of generating discrete tokens at each reasoning step. It builds on chain-of-thought prompting approaches but eliminates the need for explicit text-based intermediate steps. The code supports training on reasoning datasets like GSM8K with configurable model architectures, training modes (Coconut, CoT, no-thoughts), and standard logging via wandb.
Frequently asked
- What is facebookresearch/coconut?
- A research implementation for training large language models to perform chain-of-thought reasoning directly in a continuous latent space rather than discrete tokens.
- Is coconut open source?
- Yes — facebookresearch/coconut is open source, released under the MIT license.
- What language is coconut written in?
- facebookresearch/coconut is primarily written in Python.
- How popular is coconut?
- facebookresearch/coconut has 1.7k stars on GitHub.
- Where can I find coconut?
- facebookresearch/coconut is on GitHub at https://github.com/facebookresearch/coconut.