Rahbar's Ignition Index measures 'conscious-like' dynamics in LLMs, finds Transformers 89% stronger than SSMs

A new paper from researcher Saman Rahbar introduces the Ignition Index, a validated scalar metric that operationalizes Global Workspace Theory's (GWT) all-or-none ignition prediction in transformer language models. The metric fits a four-parameter sigmoid to per-layer linear probe accuracy as a function of input signal strength, extracting a steepness parameter (beta-hat). High values indicate abrupt, ignition-like transitions; low values indicate graded build-up. Across 11 models spanning five architecture families, shuffled-label controls demonstrate 9.6-fold selectivity for genuine linguistic structure over spurious probe capacity (p < 0.001 via Mann-Whitney U-test). This gives researchers a quantitative bridge between GWT's dynamical predictions and mechanistic interpretability — the first of its kind.

The results reveal striking architecture-level differences. Feedforward transformers exceed state-space models (SSMs) by 89% in aggregate beta-hat (p < 1e-13, Cohen's d = 0.52), with Mamba showing near-linear profiles that suggest absent global broadcast. Huginn-3.5B exhibits 2.12-fold higher ignition along its iteration axis than its depth axis, demonstrating that recurrent architectures manifest workspace-like transitions along recurrence. Pythia-410M shows a PELT-detected phase transition at training step 256 (+67%), which precedes induction-head formation. However, hypotheses linking ignition to model scale and signal strength were not confirmed — suggesting transformers may already saturate available ignition mechanisms. This has implications for interpretability: detecting when and where 'ignition' occurs could help researchers locate the computational correlates of conscious-like processing, and possibly guide architecture design for more robust reasoning.