Persona information is most separable in the final third of LLM layers, and in Llama3's last layer ethical personas share 17.6% of salient activations while political personas have 2.1% to 5.5% unique activations.
On the categorical and topological structure of timelike and causal homotopy classes of paths in smooth spacetimes
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
For a smooth spacetime $X$, based on the timelike homotopy classes of its timelike paths, we define a topology on $X$ that refines the Alexandrov topology and always coincides with the manifold topology. The space of timelike or causal homotopy classes forms a semicategory or a category, respectively. We show that either of these algebraic structures encodes enough information to reconstruct the topology and conformal structure of $X$. Furthermore, the space of timelike homotopy classes carries a natural topology that we prove to be locally euclidean but, in general, not Hausdorff. The presented results do not require any causality conditions on $X$ and do also hold under weaker regularity assumptions.
citation-role summary
citation-polarity summary
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
Localizing Persona Representations in LLMs
Persona information is most separable in the final third of LLM layers, and in Llama3's last layer ethical personas share 17.6% of salient activations while political personas have 2.1% to 5.5% unique activations.