A Navigable Manifold of Hypothesized Consciousness-Spectrum States in Language Model Representations

The paper presents findings on the geometric structure of language model representations, specifically within transformer embedding spaces, suggesting that these embeddings form a structured manifold aligned with a hypothesized consciousness spectrum. Notably, the study reveals that sentences representing similar states cluster in coherent regions, with higher and lower-level areas exhibiting stability while intermediate regions serve as transition corridors. This research is significant for practitioners as it provides a framework for understanding and guiding model behavior through the navigability of representation spaces, potentially enhancing alignment and evaluation methodologies.