New Metric Measures Global Workspace Dynamics in Language Models
Researchers have introduced a new metric known as the Ignition Index (I) to enhance all-or-none ignition predictions based on Global Workspace Theory (GWT) in transformer language models. This metric utilizes a sigmoid function with four parameters to assess per-layer linear probes tied to input signal strength, leading to a steepness parameter called beta-hat. High beta-hat values indicate rapid, ignition-like shifts, while lower values suggest more gradual changes. In a study analyzing 11 models from five architecture families, results showed a 9.6-fold preference for genuine linguistic structures over misleading probe capacities (p < 0.001, Mann-Whitney U-test). Feedforward transformers exceeded state space models (SSMs) by 89% in beta-hat (p < 1e-13, Cohen's d = 0.52). Mamba's profiles were nearly linear, lacking global broadcast, and Huginn-3.5B exhibited 2.12-fold greater ignition along its iteration axis versus its depth axis, indicating that recurrence enhances global workspace dynamics. This research can be found on arXiv under ID 2608.05160.
Key facts
- The Ignition Index (I) is a new scalar metric for Global Workspace Theory in language models.
- The metric fits a four-parameter sigmoid to per-layer linear probe accuracy.
- High beta-hat values indicate abrupt ignition-like transitions; low values indicate graded build-up.
- The study covered 11 models across five architecture families.
- Shuffled-label controls showed 9.6-fold selectivity for genuine linguistic structure (p < 0.001).
- Feedforward transformers outperform SSMs by 89% in aggregate beta-hat (p < 1e-13).
- Mamba exhibits near-linear profiles, suggesting absent global broadcast.
- Huginn-3.5B shows 2.12-fold higher ignition along its iteration axis than depth axis.
Entities
—