File size: 1,829 Bytes
ef04300 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 | ---
license: mit
task_categories: [feature-extraction]
tags: [brain-alignment, fmri, rsa, developmental, null-result]
---
# Wang et al. (2025) — LM brain-alignment results
Language-model / fMRI representational-alignment results for **Wang et al. (2025)**:
word-level phonological and semantic reading tasks in children and adolescents aged 10-17
OpenNeuro accession **`ds006239`** · cohort: children and adolescents, ages 10-17 · modality: reading (visual)
· tasks: Orth, Phon, Sem, SemLocal · cells: 8 · rows: 1232
Layout follows [`BrainAlign/cdl-devai-results`](https://huggingface.co/datasets/BrainAlign/cdl-devai-results):
`by-model/<family>/{brain_alignment,checkpoints}.csv`, `overall/`,
`ceiling-analysis/`, `provenance_tier_ledger.json`.
## Result: alignment is not distinguishable from a random seed
5 trained families (pythia-1.4b-full, pythia-160m-full, pythia-1b-full, pythia-410m-full, pythia-70m-full) and 9 PARC noise-seed
baselines, scored identically. Cells where a real family exceeds all 9 noise
seeds: **5 of 40**, against **4.0 expected by chance**.
Pooled across all three developmental datasets: 9/130 observed vs 13.0 expected
(p = 0.91); per-cell means of real families vs noise seeds correlate at
**r = +0.863**; **83.9%** of variance is cell identity and **3.2%** model family.
## Interpretation
These RDMs have real inter-subject reliability (noise ceilings in
`ceiling-analysis/ceilings.csv`), but the upstream pipeline's **positive controls
fail on this dataset**. The claim supported here is *"no LM alignment is
detectable by this measurement"* — **not** "language models do not align with the
developing brain." Read this as a result about the benchmark.
Negative results are published here deliberately: the per-cell nulls and the
noise-seed baselines are the reusable part.
|