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

A new study finds that transformer language models encode a structured, navigable geometry in their embedding spaces that aligns with a hypothesized spectrum of human consciousness states, according to research submitted to arXiv on 4 Jun 2026 [1]. The paper examines how sentences associated with different points on a consciousness spectrum — from reactive, self-focused patterns to more integrative and coherent ones — are represented inside transformer models [1][2]. Researchers report that these representations form a globally organized manifold: sentences linked to similar states cluster into locally coherent regions, while higher-level and lower-level regions exhibit convexity-like stability and intermediate regions form a transition corridor [2]. The work draws on a taxonomy broadly inspired by recurring structural descriptions of human consciousness found across contemplative traditions, philosophy, and modern psychology [2]. By mapping these descriptions onto model internals, the authors sought to determine whether the representation space itself carries an intrinsic organization that mirrors the spectrum. Beyond static geometry, the team probed the dynamics of the space. They deployed both utility-guided and geometry-only greedy trajectories and found that both consistently moved from lower- to higher-level regions, passing through intermediate tiers [2]. The authors interpret this as evidence that navigability is an intrinsic property of the representation space, guided but not dictated by a global directional signal [2]. The findings arrive amid broader efforts to understand and steer model behavior through representation-level analysis. The paper was posted on arXiv under the machine learning category and is associated with experimental tooling from Hugging Face and arXivLabs, a framework that lets collaborators develop and share new arXiv features directly on the site [1]. The authors frame the work as relevant to model guidance, evaluation, and alignment, though the study remains a preprint and has not yet undergone peer review [1][2].