Semantic Structure in Large Language Model Embeddings
This research addresses the problem of understanding and controlling semantic entanglement in LLMs for AI researchers and developers, though it is incremental as it builds on existing psychological and LLM studies.
The study found that semantic associations in large language model embeddings exhibit a low-dimensional structure similar to human ratings, with projections on antonym pairs correlating highly with human data and reducing to a 3D subspace, and shifting tokens along semantic directions causes off-target effects proportional to cosine similarity.
Psychological research consistently finds that human ratings of words across diverse semantic scales can be reduced to a low-dimensional form with relatively little information loss. We find that the semantic associations encoded in the embedding matrices of large language models (LLMs) exhibit a similar structure. We show that the projections of words on semantic directions defined by antonym pairs (e.g. kind - cruel) correlate highly with human ratings, and further find that these projections effectively reduce to a 3-dimensional subspace within LLM embeddings, closely resembling the patterns derived from human survey responses. Moreover, we find that shifting tokens along one semantic direction causes off-target effects on geometrically aligned features proportional to their cosine similarity. These findings suggest that semantic features are entangled within LLMs similarly to how they are interconnected in human language, and a great deal of semantic information, despite its apparent complexity, is surprisingly low-dimensional. Furthermore, accounting for this semantic structure may prove essential for avoiding unintended consequences when steering features.