Language models show navigable consciousness spectrum in 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

影响 Reveals structured, navigable geometry in LLM representations aligned with consciousness, potentially aiding alignment and evaluation.

排序理由 This is a research paper detailing findings about the internal representations of language models. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.CL 阅读 →

• Consciousness Spectrum
• Language Model Representations
• Transformer Embedding Spaces

• 论文
• 安全

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →