The Illusion of Superposition in Latent CoT via Soft Thinking

Latent reasoning via continuous chain-of-thoughts (Latent CoT) has emerged as a promising alternative to discrete CoT reasoning. Operating in continuous space increases expressivity and has been hypothesized to enable \textit{superposition}: the ability to maintain multiple candidate solutions simultaneously within a single representation. Despite theoretical arguments, it remains unclear whether language models actually leverage superposition when given continuous reasoning tokens. We investigate this question across two prominent latent CoT methods: Soft Thinking, a training-free approach that constructs reasoning embeddings as convex combinations of token embeddings, and Coconut, which fine-tunes models to reason with continuous latent thoughts. Using \logitlens and entity-level probing to analyze internal representations, we find consistent evidence against superposition in both settings. For Soft Thinking, off-the-shelf models collapse superposed inputs to a single interpretation within the first few layers. For Coconut, the model learns to extract answers directly from the question embedding, achieving 96.6\% accuracy without any latent tokens. Together, our results suggest that current latent CoT methods do not leverage superposition for reasoning both in training free and fine-tuned approaches.