Exploring Higher-Order Circuits in Large Language Models

Toward White-Boxing the Internal Circuits of LLMs

This is research into the safety of superintelligence, whose risks have increasingly been pointed out in recent years. We aim to white-box the internal mechanisms of generative AI, establishing technology for directly monitoring concept manipulation during inference and identifying dangerous reasoning.

Searching for Metaphor Circuits within LLMs

As a subject of mechanistic interpretability research, we aim to elucidate the existence and function of "metaphor circuits" within LLMs.

BabyLM Research

Exemplified by BabyLM research, we aim to elucidate the process by which human infants acquire language ability, using a computational approach based on LLMs.