A Navigable Manifold of Hypothesized Consciousness-Spectrum States in Language Model Representations
Researchers have identified a structured manifold within language model representations that aligns with a hypothesized spectrum of consciousness. Sentences related to similar states cluster together, forming a navigable geometry that transitions from lower to higher levels of consciousness. This suggests that the embedding spaces of these models inherently encode and allow traversal along a spectrum inspired by human consciousness, offering a new perspective for analyzing and guiding model behavior. AI
IMPACT Reveals structured, navigable geometry in LLM representations aligned with consciousness, potentially aiding alignment and evaluation.