Goldstein 2025 Fig 3 — Paper vs Our GPT-2 XL vs Ouro-2.6B
The same figure from three sources, stacked for direct comparison. Fig 3 = the depth→time hierarchy: each ROI's encoding-peak lag plotted against model depth. In language ROIs (TP, IFG, aSTG) the peak lag drifts later as model depth increases (positive r); the auditory control (mSTG) should stay flat. Our GPT-2 and Ouro panels use the gold-standard papermatch presentation — a clean single-row scatter with a bold dashed regression line and a big bold r per panel (same data, cleaner than the raw fuzzy scatter). Click any figure for full size.
x-axis differs by model: paper & our GPT-2 = layer index 1–48; Ouro = effective recurrent depth 0–191 (4 recurrent passes × the sampled layers).
1 · Paper — Goldstein et al. 2025
The published result we replicate (GPT-2 XL, 48 layers). Full Fig 3 extracted from the paper PDF (p.5): brain inset + top curves row + bottom peak-lag-vs-layer scatter.
Paper Fig 3 — full figure (both rows, PDF-extracted)Paper r: TP +0.93 · IFG +0.85 · aSTG +0.92 (all p≤10e-13) · mSTG −0.24 (p=.09, ns). Language ROIs march rightward (later lag) from red→blue; the mSTG auditory control stays pinned near lag 0 — no hierarchy. Extracted from analysis/assets/paper_fig3_full.png (cropped from paper/goldstein-2025-temporal-structure.pdf, page 5).
2 · Our GPT-2 XL replication — gold standard (papermatch)
Our pipeline (PCA-50 + OLS, predictive word-alignment, sub-bin parabolic peak, posterior-STG control), rendered in the clean single-row Fig-3 format. Re-run from scratch and verified bit-exact vs the frozen oracle (max abs diff = 0.0). Regenerated from best_run via analysis/fig_papermatch.py.
Gold standard — papermatch (clean regression-line Fig 3)Our GPT-2 XL replication — gold standard (papermatch). TP +0.89 · IFG +0.83 · aSTG +0.81 · mSTG −0.26 (ns). Bold dashed line = regression; same data as best_run, clean single-row Fig-3 format.
The novel extension. Ouro reuses the SAME weights across 4 recurrent passes; its "depth" is unrolled iteration: effective depth = pass×48 + layer (0→191). Same electrodes/estimator as the GPT-2 replication, same papermatch presentation. Regenerated from best_run via analysis/fig_papermatch_ouro.py.
Ouro-2.6B — Goldstein Fig 3 (papermatch, peak lag vs effective recurrent depth)Ouro-2.6B, same papermatch format. IFG +0.66 · aSTG +0.60 (sig) · TP +0.25 (ns, n=6 underpowered) · mSTG-proxy +0.84. x-axis = effective recurrent depth (0–191); faint lines = pass boundaries. NOTE: the real EAC auditory control shows this is not language-specific — see poster v2.