Great idea.

Thank you for your insightful work!
I'm curious about the motivation behind the idea of using the last layer's hidden states as the input embedding for the next step.
Is it because the hidden states from the final layer have a high similarity to the embedding of the next input token?

The issue names no file, test, or entry point to inspect. Start by locating the Python implementation and any existing explanation of continuous latent reasoning, then determine whether the project documents why final-layer hidden states are reused as the next input embedding. Done means a clear, evidence-based explanation or documentation update addressing the question.

Written by the indexing model from the issue text.