The Ignition Index: A New Metric Revealing Consciousness Dynamics in Language Models

The Ignition Index is a new quantitative metric that uses a four-parameter sigmoid to measure the sharpness of transitions in classification accuracy across model layers, where high values indicate abrupt, ignition-like transitions associated with consciousness theories.

The study showed that Transformers outperform SSMs by 89% in aggregate beta-hat values, meaning Transformers exhibit sharper and more distinct transitions in information processing, while models like Mamba show nearly linear behavior.

This research helps teams understand how different models work at a deeper level, enabling them to choose the most appropriate architecture for their applications, especially for tasks requiring complex information processing or deep contextual understanding.