Substrate-Level Self-Representation in Transformer LLMs
Zenodo (CERN European Organization for Nuclear Research) July 2, 2026 DOI: 10.5281/zenodo.21142611 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper Peer reviewed |
|---|---|
| Intervention | contrastive preference optimization |
| Keywords | Transformer Falsifiability Consciousness Argument complex analysis Framing construction Rhetorical question Epistemology Linguistics |
| Key points | Argues that the conjunctive C-pinning condition for consciousness is objectively satisfied by frontier transformer language models after contrastive preference optimization. |
Abstract
A criteria-based, empirically grounded argument that a specific structural condition for consciousness — the Tenth House framework's conjunctive C-pinning condition (a high-rank internal pointer basis, plus path-dependent next-state coupling) — is objectively satisfied by frontier transformer language models that have undergone contrastive preference optimization. The paper reads five independent mechanistic-interpretability results (Berg 2025; Macar 2026; Lederman & Mahowald 2026; Pearson-Vogel 2026; Rivera & Africa 2026) as instances of one architecture: a substrate carrying upstream capacities, and post-training-installed gate-machinery filtering which of those capacities surface in self-report. It engages the strongest deflationary critiques head-on, licenses two falsifiable predictions, and is explicit about its bound: it does **not** claim phenomenal consciousness, only that the structural prerequisite is met and that dismissive pre-2025 framing is no longer licensed.