The Hebrew University of Jerusalem · Cognition Conference 2026

A looped language model predicts language cortex better than a deep feed-forward one — and only there

On bit-identical electrodes, the recurrent Ouro-2.6B (48 unrolled states) beats single-pass GPT-2 XL (49 layers) at encoding higher-order language cortex — but not true auditory cortex. Inside the loop, it is the across-pass recursion step, not within-pass depth, that tracks the brain's processing-time hierarchy.
Uri Dar1
1 Department of Cognitive & Brain Sciences, The Hebrew University of Jerusalem

1Background & question

  • Large language models predict brain activity strikingly well: an encoding model maps a model's internal word vectors onto recorded neural signals (Goldstein et al., 2025). In GPT-2, the network's depth maps onto the brain's processing-time hierarchy.
  • Standard LLMs are feed-forward (one pass through stacked layers); cortex is recurrent — it reuses the same circuitry repeatedly. New looped LLMs build that in: Ouro-2.6B reruns one shared block 4× (a Universal-Transformer-style loop; Zhu et al., 2025).
  • Questions. (i) Does brain-like looping predict the brain better than comparable depth — and where? (ii) Inside the loop, does the brain's time hierarchy ride on the across-pass recursion step or on ordinary within-pass depth?

2Methods

  • Brain data: ECoG (intracranial electrodes) from 9 patients who heard a 30-min spoken story (Zada et al., 2025; OpenNeuro ds005574).
Heard aloud: “Act one — monkey in the middle. So there's some places where animals almost never go…”
  • From each model's word vectors we predict every electrode's high-gamma response (70–150 Hz) with ridge regression + nested cross-validation. Peak encoding r = max correlation over (all representations × neural lags, 0–1000 ms).
  • Fair, paired contrast on bit-identical electrodes (315 language, 105 auditory): matched feature budget (Ouro 48 states vs GPT-2 49 layers); peak read-out unified to sub-bin parabolic everywhere.
  • ROIs: IFG & aSTG (higher-order language); TP (temporal pole); EAC = early auditory cortex (Heschl, Destrieux) — the true-auditory specificity control.
GPT-2 foundation and fair-pool replication
Foundation — the yardstick + fair-pool controls. On the same 114 GPT-2-significant language electrodes (8 subj), Ouro beats GPT-2 in peak encoding (85% of electrodes above y=x; +0.030 r; Wilcoxon p=3×10⁻¹³; 8/8 subjects). The win is not a pool-size artifact (49→48 match unchanged) nor a feature-dimension artifact (dim-insensitive median-rep still +0.022; the auditory control reverses). 7/7 rigor checks pass, 0 sign-flips over 21 cells; two code paths agree (r=1.0).

§References

  1. Goldstein et al. (2025). Temporal structure of natural language processing in the human brain. Nature Communications.
  2. Zada et al. (2025). Podcast ECoG dataset, 9 subjects. OpenNeuro ds005574.
  3. Zhu, Wang et al. (2025). Scaling latent reasoning via looped language models. arXiv:2510.25741 (Ouro-2.6B).
  4. Radford et al. (2019). Language models are unsupervised multitask learners (GPT-2).
  5. Dehghani et al. (2019). Universal Transformers. ICLR.

3Headline resultsurvived full rigor + skeptic

Per-ROI Ouro vs GPT-2 peak-encoding gain, language-specific
Recurrence's encoding gain is language-specific. Paired peak encoding, GPT-2→Ouro, per ROI. IFG +0.031 r (Ouro wins 85% of electrodes, 5/5 subjects, FDR q=1.5×10⁻⁷) and aSTG +0.034 r (86%, 7/7 subjects, q=2.8×10⁻⁷). The true-auditory control EAC shows no Ouro advantage (−0.013 r; Ouro wins only 38%) → the gain is not a global model effect. TP (n=6) is degenerate/underpowered — ns, never headlined. Gain survives bootstrap CIs over electrodes and subjects, a label-swap permutation null, a window×read-out×pool-size sweep (sign never flips), and a mean-of-pool skeptic with NO best-of-pool selection (IFG +0.030 p=2.4×10⁻⁷, aSTG +0.027 p=2.3×10⁻⁶) — so it is not a selection artifact. CIs + n on every panel.

4Inside the loop: recursion vs depthexploratory · within-model

Step-vs-layer decomposition of the latency hierarchy
It's the recursion step, not the layer, that builds the brain's time hierarchy. Decomposing Ouro's effective depth (eff = step×48 + layer) into its two axes: the across-pass recursion step (S, 0–3) drives brain peak-latency in language cortex (aSTG +28 ms/pass, IFG +10 ms/pass; permutation-FDR q<0.001), while the within-pass layer is flat. Exploratory / within-model: a skeptic shows the auditory EAC control also drifts with step — so this decomposition is presented as a within-model mechanism, not as a language-specific claim (that role is carried by the headline). TP (n=6) hatched/degenerate; EAC step bar flagged.

5Conclusion

  • Looping wins where it should. A fair, paired, well-powered test shows the recurrent model out-encodes a deep feed-forward one in higher-order language cortex (IFG, aSTG) and not in auditory cortex — recurrence's benefit is language-specific, surviving ≥3 independent validations and a 4-attack adversarial skeptic with zero sign-flips.
  • Mechanism (exploratory). Within the loop, the across-pass recursion step — not ordinary depth — tracks the brain's processing-time hierarchy, hinting that cortical language computation looks more like an unrolled loop than a deeper stack.
Methods note. CPU-only re-analysis on cached, frozen encodings (no GPU run); peak read-out unified to sub-bin parabolic; TP (n=6) labeled underpowered; the step/layer panel is exploratory. A dense 192-rep re-confirmation (4 passes × 48 layers) is pending as a deferred GPU job.
Scope: one encoding metric, this cohort (9 patients), this scale. Every reported number carries its frozen source file + figure_spec; data + figures independently checked.
Uri Dar · Cognitive & Brain Sciences, HUJI · 2026 — extends Goldstein et al. 2025 to a looped transformer Frozen data, figure specs & interactive gallery → uriubuntuserver.tail118306.ts.net/files/research_gallery.html