Decoding Latent Reasoning: Mechanistic Interpretability of Chain of Continuous Thought with Sparse Autoencoders

Applies mechanistic interpretability and sparse autoencoders to decode latent reasoning steps in chain-of-continuous-thought models.