Mette Friis Andersen

Researchers probe and steer metaphor representation in LLMs

Researchers have developed methods to understand and control how large language models represent and generate metaphors. Using mechanistic interpretability techniques, they identified specific causal subspaces within va…