Ignition Index Quantifies Consciousness‑Like Dynamics in LLMs
Researchers present the Ignition Index, a scalar that captures the all‑or‑none “ignition” predicted by Global Workspace Theory in transformer language models. By fitting a four‑parameter sigmoid to per‑layer linear‑probe accuracy, the metric reveals abrupt information‑integration spikes, giving the first quantitative handle on consciousness‑like behavior in LLMs.