The 'Ignition Index' Doesn't Track With Model Size. That's a Problem for the 'Bigger Is Better' Theory.

There is a moment when scattered meaning suddenly clicks into a single answer, the way a long passage you are reading finally makes sense. Cognitive scientists have a name for the mechanism behind that moment: Global Workspace Theory, a decades-old idea about how the brain routes competing pieces of information into a single conscious broadcast. A new preprint, The Ignition Index: Measuring Global Workspace Dynamics in Language Models, borrows that idea and turns it into a ruler for transformer language models. The ruler's first measurement breaks a popular assumption. Bigger models do not produce more of the brain-like 'click.' The gap, the paper finds, is architectural, not a matter of scale.

The authors build the metric around GWT's all-or-none prediction: at some layer, a model's representation of meaning should consolidate sharply, like a phase transition, rather than drift in gradually. To turn that into a number, they fit a four-parameter sigmoid to per-layer linear probe accuracy as a function of input signal strength, then extract the steepness parameter, which they call beta-hat. A high beta-hat means the model jumps into its answer. A low beta-hat means it drifts.

Across models drawn from five architecture families, the metric behaves the way a real measurement should. Probes trained on genuine linguistic structure gave beta-hat values nine-and-a-half times higher than probes trained on shuffled-label controls, with a Mann-Whitney U-test p-value below 0.001. The selectivity is the evidence that the metric is tracking something real about language processing, rather than the raw capacity of a linear probe.

Then the counter-narrative lands. The authors pre-registered two hypotheses. Larger models should ignite more sharply. Stronger input signals should trigger sharper ignition. Neither held. The metric showed no consistent relationship with parameter count. The default 'bigger is more brain-like' assumption in modern language modeling fails the test, at least on the data the paper analyzes. The preprint is not peer-reviewed, so the result is author-reported, but the pre-registration and the selective controls give it more weight than a typical post-hoc finding.

The architectural lever is where the story gets constructive. Standard transformer architectures, the kind behind GPT-style and Pythia-style models, exceeded state-space models like Mamba by roughly 89 percent in aggregate beta-hat, with a p-value below one in ten trillion and a Cohen's d of 0.52. Mamba's near-linear profile looks like absent global broadcast, exactly what GWT would predict for an architecture that streams information through a fixed-size recurrent state rather than attending across a workspace. The contrast is a quantitative answer to a question that has mostly been argued on vibes: recurrent models process differently, and the difference is measurable.

The largest single split in the data comes from a different model, a 3.5-billion-parameter recurrent transformer called Huginn-3.5B. Its iteration axis, the way the model loops back on its own hidden state, runs a beta-hat of 234.8. Its depth axis, the way it stacks layers, runs 111.0. The ratio, 2.12 to 1, suggests that recurrence amplifies workspace-like transitions along the iteration dimension, not the depth dimension. A longer stack of layers, in other words, is not the same thing as a sharper click. The figure appears in the project's GitHub README, not in the abstract, and the paper treats it as the load-bearing exhibit for the recurrence-axis claim.

The paper also catches an early-training transition in Pythia-410M. A PELT changepoint analysis (a method for spotting the moment a time series changes its pattern) detects a phase transition in probe behavior at training step 256, a sixty-seven percent jump, that precedes the formation of induction heads, the circuits that let a model copy and complete patterns. The chronology matters because it suggests the consolidation mechanism matures before the model has the circuitry most often associated with in-context learning.

The result does not say transformer language models are conscious, and the authors are careful not to claim that. Global Workspace Theory is being used as a borrowed metaphor for what an information-routing architecture looks like in a measurable form. The contribution is a ruler, not a verdict. The code is MIT-licensed, the pre-computed probe results are released under Creative Commons, and the Huginn-3.5B weights sit on Hugging Face rather than in the project repository itself.

For a field that has spent three years treating parameter count as a proxy for cognitive sophistication, the Ignition Index hands out a falsifiable test. The next paper that wants to argue a model is 'more brain-like' will have to run the sigmoid and report a beta-hat. Mamba and the architectures that already score below what their parameter count would predict are now the obvious next moves on the board.