HOW TO READ · F11
Seven-step methods walkthrough — from ECoG + model states to the depth-vs-recursion read.
Seven-step methods walkthrough — from ECoG + model states to the depth-vs-recursion read.
GPT-2 layer-depth → cortical processing-time replicates robustly (r 0.81–0.98 across estimators, peak-fit span ~130–160 ms). The argmax “step” was an estimator artifact. Caveat: across language-responsive cortex, not language-specific — “mSTG” is mislabeled lateral STG, so there is no valid control.
The looped model does NOT recapitulate the hierarchy: its recursion march is sub-resolution (≈15–30 ms ≤ 1 lag bin) vs GPT-2's supra-resolution depth ramp. The earlier “recurrence carries it” headline was a peak-estimator artifact.
Root cause — GPT-2 encoding curves are bimodal, so argmax coin-flips between bumps. The methods lesson behind the artifact.
Scaled curves — GPT-2 depth march is supra-resolution; the Ouro loop march is sub-resolution; within-loop layer ≈ 0.
Per-ROI grid — depth→time replicates across all three language-responsive ROIs; the looped march is sub-resolution everywhere.
Per-ROI robust depth→time — the old “architecture dissociation / specificity” framing is retired.
The looped model is no better a brain predictor overall: its edge is low-SNR-only and partly a candidate-pool (“more shots at noise”) effect.
Recursion depth tracks linguistic difficulty broadly, but has no per-word neural correlate (informative null).
