machine learning

Sky sphere representation in language models

arXiv:2607.27092

summary

The paper investigates whether large language models (~100B parameters) contain a decodable representation of the night sky map within their residual streams, showing that most examined models encode this information and can recover celestial positions with low angular error.

Abstract

We analyze whether language models of size ~100B have a representation of the night sky map that is decodable from their residual stream. We find that most of the considered open-source models do have such a representation, and it often even surfaces to the top principal components on prompts that ask questions like ``what is close to this object in the night sky''. In all but one model this representation showed significant scores in LOO testing, containing up to 65-85% of variance ( $R^{2}$ -score) and having median angular error down to $1 2^{\circ} - 2 1^{\circ}$ . We verify that our representation is not a simple leak from a correlated flat representation. To our knowledge, this representation is the first example of a curved high-dimensional irreducible feature manifold. Codes used in the paper are published at https://github.com/l3erdnik/Decodable-sky

Topics & keywords

#language models #representation learning #residual stream analysis #sky map decoding #high-dimensional manifoldsresidual streamprincipal componentsleave-one-out testingR^2 scoreangular errorcurved manifold