cross-architectural

This paper characterizes compositional literary primitives in instruction-tuned LLMs using sparse autoencoders, discovering feature classes for self, style, and affect that enable emotion steering across two architectures.