The Ignition Index: Measuring Global Workspace Dynamics in Language Models

The Ignition Index (I) operationalizes Global Workspace Theory's all-or-none ignition prediction by fitting a four-parameter sigmoid to per-layer linear probe accuracy as a function of input signal strength. The key extracted parameter, beta-hat, indicates whether a model shows abrupt (ignition-like) or graded transitions. The metric was validated across 11 transformer language models, suggesting variability in how different architectures and scales handle global workspace dynamics. This provides a concrete, quantitative tool for testing cognitive theories in neural networks, potentially guiding interpretability research and model design. Future work might explore correlations with model size, training data, or task performance.