Cumulative Superposition Reduces Dimension Requirements for Latent Reasoning

Retaining full reasoning history via cumulative superposition allows latent reasoners to use lower representational dimensions than frontier-only methods at a fixed hidden width. This approach enables more efficient encoding of multiple reasoning alternatives within fixed-dimensional latent states.