Sky sphere representation in language models

Researchers found that several ~100B open-source language models encode a decodable internal map of the night sky in their residual stream.

The signal often shows up in the top principal components, with substantial variance explained and low angular error.

The authors say this is unlikely to be just a simple flat leaked representation.

The result adds evidence that large language models can form nontrivial internal geometric representations, which may help interpretability research.