Coconut: A Framework for Latent Reasoning in LLMs

Coconut is a framework designed to improve reasoning in Large Language Models by shifting the process from natural language to a continuous latent space. It operates in two modes, using either text tokens or its own internal hidden states as input for subsequent reasoning steps. The model is trained via a multi-stage curriculum that progressively replaces language-based steps with latent ones, controlled by special tokens. This method enables end-to-end differentiable training and has demonstrated superior performance on complex logical reasoning tasks compared to standard Chain-of-Thought approaches.