Internal thought in CoU, but not in training data

[Rogerspy]

I checked the training data and found that there was no internal thought. When performing inference, is the ICL capability of LLM used?

[Rogerspy]

And according to Figure 4, internal thought should act on red-lm, not base-lm in the cou prompt.