Coconut (Chain of Continuous Thought) is a new paradigm proposed by researchers from Meta and the University of California, San Diego in December 2024. It aims to explore the reasoning potential of large language models (LLMs) in unrestricted latent spaces. The specific results are reflected in the paper "Training Large Language Models to Reason in a Continuous Latent Space"middle.

Coconut frees the reasoning process from the traditional language space, allowing the model to reason directly in the continuous latent space. This approach no longer relies on the language model head and embedding layer to map hidden states to language tokens, but instead directly embeds the last hidden state of the model (i.e., continuous thinking) as the input of the next token. Such modifications enable the model to reason without being restricted by natural language, and because continuous thinking is fully differentiable, the system can be optimized end-to-end through gradient descent.

The paper mentioned that Coconut outperforms traditional Chain of Thought (CoT) in certain logical reasoning tasks that require a lot of backtracking, and generates fewer tokens during the reasoning process, indicating that latent space reasoning has obvious advantages in complex tasks that require extensive planning.