They try to map human concepts - month of year, day of week- observed in 2 models. One a text embedding model output and other a layer of gpt2-small.

This paper is towards understading how LLM organizes algorithms in the embedding space. What would cause the model to do that. Why are the PCA of neurons embedding represent circular space ?

What the paper concluded is that, they were able to successfullt map human concept on the model’s internal representaiton called manifold. The manifold is not completely mapped.

One of the future work is bulding sparse autoencoder that is aware of the manifold.