Searching for Metaphor Circuits within LLMs

Last updated: October 11, 2026

As a subject of mechanistic interpretability[1] research, we aim to clarify the existence and function of "metaphor circuits."

By choosing metaphorical expressions as our research subject, we aim to determine whether LLMs do more than understand and use language at a surface level, that is, whether, as proposed in cognitive linguistics and cognitive semantics, they possess functional mechanisms that let them cognitively manipulate language and concepts based on some form of embodiment or non-literal, figurative concepts. This is both a study of higher-order circuits in LLMs and an attempt to computationally reconstruct the cognitive linguistics pursued by George Lakoff, Ronald W. Langacker, and others (aiming to establish computational cognitive linguistics).

In this research project, we have, in preliminary form, confirmed the emergence of important circuits related to metaphorical expression that scale with LLM parameter size (September 2026). We continue to verify this ongoing work and hope to publish it as a paper at an early stage.

[1] For more on mechanistic interpretability, see Wikipedia, Takatsuki's expert yet highly accessible article, and our own Kobayashi's article.