--- license: cc-by-4.0 tags: - neuroscience - fmri - brain-alignment - interpretability - language-models configs: - config_name: summary_by_checkpoint default: true data_files: "overall/by_checkpoint.csv" - config_name: summary_by_family data_files: "overall/summary_by_family.csv" - config_name: brain_alignment data_files: "by-model/*/brain_alignment.csv" - config_name: interp_mechanistic data_files: "by-model/*/interp_mechanistic.csv" - config_name: interp_layerwise data_files: "by-model/*/interp_layerwise.csv" - config_name: localisation_isolation data_files: "by-model/*/localisation_isolation.csv" - config_name: localisation_onset data_files: "by-model/*/localisation_onset.csv" - config_name: behaviour data_files: "by-model/*/behaviour.csv" - config_name: ablation_alignment data_files: "by-model/*/ablation_alignment.csv" - config_name: ablation_behaviour data_files: "by-model/*/ablation_behaviour.csv" - config_name: claim_tests data_files: "overall/claim_tests.csv" - config_name: heldout_predictor data_files: "overall/heldout_predictor.csv" - config_name: diagnostics_layerwise data_files: "diagnostics/layerwise_alignment.csv" - config_name: diagnostics_run_confound data_files: "diagnostics/run_confound_check.csv" - config_name: ceiling_analysis data_files: "ceiling-analysis/ceiling_*.csv" - config_name: alignment_vs_ceiling data_files: "ceiling-analysis/alignment_vs_ceiling_*.csv" --- # CDL DevAI results — brain × interpretability × localisation, per model per checkpoint Developmental analysis of 10 language-model families against the **ds003604** auditory language fMRI dataset. For every training checkpoint of every model we measured three things and here report them side by side: | axis | what it asks | source tables | |---|---|---| | **brain** | does the model's representational geometry match the brain's? | `brain_alignment` | | **interp** | how is the representation organised internally? | `interp_mechanistic`, `interp_layerwise` | | **localisation** | are linguistic phenomena isolated into dedicated units? | `localisation_isolation`, `localisation_onset` | **Start with the `summary_by_checkpoint` config** (the default, and what the viewer shows first): one row per model × checkpoint, with all three axes as columns. 262 rows, 10 models. --- ## ⚠️ READ THIS BEFORE USING THE BRAIN ALIGNMENT NUMBERS **The brain alignment columns are confounded by scanner run and must not be read as a result about language models.** This affects every `rsa*` column in `brain_alignment`, every `brain_*` column in the summary tables, and the `ablation_alignment` table. In ds003604 **each stimulus is presented in exactly one scanner run** (for Phon, run-01 carries 48 of the 96 stimuli and run-02 the other 48). Run membership is therefore perfectly confounded with stimulus identity, and every cross-run stimulus pair inherits that run's drift, baseline shift and scaling. Measured: ``` "different run" predicts brain dissimilarity, Spearman rho, all 12 task × session cells: Gram +0.866 +0.812 +0.828 Plaus +0.741 +0.689 +0.761 Phon +0.562 +0.488 +0.624 Sem +0.511 +0.510 +0.601 ``` By comparison **no stimulus property predicts these RDMs at all**: trial type ≈ −0.02, text length ≈ 0, lexical overlap ≈ 0. The RDMs look highly reliable across independent subject cohorts (rho 0.74–0.92 between sessions of the same task) — but that is largely the reliability of an *acquisition artefact*, because run assignment is fixed by the protocol and so repeats identically for every subject. A language model cannot represent which scanner run a stimulus appeared in, so its RSA against this structure is ≈0 by construction, and slightly negative in practice. **Two obvious explanations were tested and ruled out:** - *Cohort size.* Sem/ses-7 was built from 40 subjects; rebuilt from **98** it agrees with the 40-subject version at **rho = 0.928**, identical stimuli. Bigger cohorts change nothing. - *Layer choice.* Alignment is flat across **every** layer — −0.012 to −0.021 across all 12 layers of babylm-gpt2-3 and all 14 of beetle-humanscale-eng. There is no middle-layer peak being missed. See `diagnostics_layerwise`. **What fixes it, and what happens when you do.** Z-scoring each voxel *within run* before combining runs removes the confound cleanly — run predictiveness falls from **+0.562 to −0.041**. Alignment against the corrected RDM then peaks at **+0.022** (layer 2, beetle-humanscale-eng, 2,556 stimulus pairs), which is **not significant**. The layer *profile* becomes sensible — early/middle layers positive, output layers negative — but the magnitude is ≈0. See `diagnostics_run_confound`. **Bottom line:** these tables do not show that language models align with the brain, and they do not show that they fail to. The measurement cannot answer it. The corrected analysis, on one cell and one model, finds no detectable alignment. Do not cite the raw numbers in either direction. > ### UPDATE 2026-08-25 — there is now a noise ceiling, and the null is real > > Everything above still stands for the tables it describes. What has changed is > that the missing measurement now exists: see **[`ceiling-analysis/`](./ceiling-analysis)**. > > On ds003604 Phon/ses-5 (43 subjects, 72 stimuli), rebuilt from fresh patterns: > > | | raw | within-run normalised | > |---|---|---| > | noise ceiling (LOO lower / upper) | 0.774 / 0.792 | **0.849 / 0.859** | > | "different run" predicts dissimilarity | **+0.557** | **−0.041** | > > Best layer: beetle-humanscale-eng **+0.022 (2.6% of ceiling)**, babylm-gpt2-3 > **+0.013 (1.5%)**, pico-decoder-large **−0.001 (−0.1%)**. > > Inter-subject reliability is **0.85** — far above the 0.2–0.4 typical in this > literature — so there is a large, replicable target to predict, and three models > across three architectures capture 0–3% of it. That is a **clean, well-powered > null**, not an uninterpretable one. > > Note the ceiling **rises** after correction (0.774 → 0.849): had the raw > reliability been mostly the run artefact, removing it would have lowered the > ceiling. It rose, so the correction removes noise rather than signal. > > Scope: one task × session cell. The full twelve-cell corrected sweep is running > and will be added as a separate, clearly-labelled set. The tables above are > **not** modified by this update. ### Which columns are safe | axis | affected by the run confound? | |---|---| | `brain_*`, `rsa*`, `ablation_alignment` | ❌ **Confounded. Do not use as a model result.** | | `interp_*` (norm, gini, hoyer, per, condition_number, cka_to_prev) | ✅ Safe — computed from LM activations only. The fMRI data is not involved. | | `loc_*` / `localisation_*` (selectivity_index, overlap, gini, entropy, layer_com, n_active_layers) | ✅ Safe — LM-internal localisation against text contrasts. No fMRI. | | `behaviour` (mp_accuracy) | ✅ Safe — minimal-pair accuracy, text only. | | `ablation_behaviour` (causal_selectivity) | ✅ Safe — ablation vs behaviour, no fMRI. | The **one uncontaminated positive result** in this release is the causal behaviour test: ablating a phenomenon's localized circuit costs **1.13%** minimal-pair accuracy versus **0.55%** for a random circuit of the same size — selectivity +0.0058, **t = 1.98, p = 0.049, n = 316**, driven mostly by Phon (+0.021). That is *borderline* and should be described as suggestive, not established. --- ## Layout ``` overall/ by_checkpoint.csv <- THE MAIN TABLE. one row per (family, step), brain + interp + localisation side by side summary_by_family.csv <- one row per model: means, ranges, step-vs-alignment trend claim_tests.csv <- per-family claim tests (claim, stat, value, p, n) heldout_predictor.csv <- cross-family held-out predictive validation localisation_onset.csv by-model// README.md <- what this model is, its numbers, what is odd about it checkpoints.csv <- this model's rows of the main table brain_alignment.csv <- full per task × session × step detail (CONFOUNDED) interp_mechanistic.csv interp_layerwise.csv localisation_isolation.csv localisation_onset.csv behaviour.csv ablation_alignment.csv ablation_behaviour.csv figures/_overview.png ceiling-analysis/ <- ADDED 2026-08-25. Noise ceiling + within-run correction, and alignment as a fraction of the ceiling. Additive: replaces nothing above. Start here for whether the null is real. diagnostics/ layerwise_alignment.csv <- alignment at every layer (rules out the layer explanation) run_confound_check.csv <- raw vs run-partialled alignment, per layer figures/ <- cross-model figures (fig1-fig8, tables) superseded/early_tier1/ <- an earlier PARTIAL pass (7 families, 8/12 cells). Kept for completeness, deliberately NOT a viewer config. Do not use. provenance_tier_ledger.json <- the run record: per-tier status, exit code, duration, peak GPU ``` Every table carries `family` and `model_ref`, so you can filter by model in the viewer without downloading anything. --- ## The models 10 families, 262 checkpoints total. `brain_rsa_mean` is shown **only** so you can see it is flat and near zero; per the warning above it is not interpretable. | family | ckpts | steps | brain RSA (confounded) | trend ρ (p) | interp PR | interp gini | loc selectivity | behaviour acc | |---|---|---|---|---|---|---|---|---| | pico-decoder-tiny | 21 | 0–126k | −0.017 | −0.02 (0.80) | 0.221 | 0.204 | 0.546 | 0.631 | | pico-decoder-small | 126 | 0–125k | +0.001 | +0.16 (<0.001) | 0.184 | 0.243 | 0.551 | 0.682 | | pico-decoder-medium | 21 | 0–125k | −0.011 | −0.10 (0.13) | 0.191 | 0.206 | 0.544 | 0.679 | | pico-decoder-large | 21 | 0–125k | −0.012 | +0.01 (0.93) | 0.104 | 0.221 | 0.539 | 0.687 | | beetle-humanscale-eng | 18 | 0–9.5k | −0.016 | −0.03 (0.71) | 0.300 | 0.077 | 0.530 | 0.510 | | beetle-fineweb3-eng | 19 | 0–65k | −0.015 | −0.05 (0.44) | 0.241 | 0.090 | 0.558 | 0.590 | | babylm-gpt2-3 | 9 | 191–1908 | −0.021 | +0.17 (0.07) | 0.065 | 0.326 | 0.519 | 0.693 | | babylm-gpt2-5 | 9 | 191–1908 | −0.022 | +0.14 (0.14) | 0.066 | 0.324 | 0.531 | 0.687 | | babylm-gpt2-7 | 9 | 191–1908 | −0.024 | +0.12 (0.21) | 0.065 | 0.328 | 0.520 | 0.707 | | babylm-gpt2 | 9 | 598–2990 | −0.032 | −0.03 (0.73) | 0.072 | 0.293 | 0.530 | 0.707 | pico-decoder-small's trend is p<0.001 only because n = 1,512 stimulus-level rows; ρ = 0.16 on a confounded measure is not a finding. No other family reaches p < 0.05, which across ten tests is what chance looks like. The held-out cross-family predictor scores mean **R² = −2.74** — worse than predicting the mean. --- ## How to read the metrics **interp** (safe). `per` = participation ratio, the effective dimensionality of the representation as a fraction of hidden size; **lower = more compressed**. Here 0.06–0.30, and the babylm models (0.065) are far more compressed than Beetle (0.24–0.30). `gini` and `hoyer` are sparsity of activation mass, **higher = sparser**. `condition_number` is the spread of the activation covariance spectrum. `cka_to_prev` is representational similarity to the previous checkpoint — **near 1 means training has stopped changing the geometry**. **localisation** (safe). `selectivity_index` is how strongly a phenomenon's top units prefer it over other phenomena; ~0.5 across all models here, i.e. **moderate and strikingly constant** — no model isolates phenomena sharply. `mean_overlap_with_others` is how much a phenomenon's circuit is shared with other phenomena; **high overlap means little specialisation**. `n_active_layers` is how many layers contribute; `layer_com` is the centre of mass over depth (**low = early layers**). **behaviour** (safe). `mp_accuracy` is minimal-pair accuracy, chance = 0.5. Beetle humanscale at 0.51 is essentially at chance; the babylm and pico models reach 0.63–0.71. **brain** (confounded — see the warning). `rsa` is Spearman between the LM RDM and the brain RDM over stimulus pairs; `rsa_pearson`/`rsa_kendall` are the same with different rank treatments. `n_stim` is 72 for Sem/Phon and 60 for Gram/Plaus (controls excluded). --- ## Provenance Produced by [`suchirsalhan/cdl-representations-brains-babylms`](https://github.com/suchirsalhan/cdl-representations-brains-babylms), tiers 0–3, 2026-08-19. Brain-side session RDMs are cached separately at [`BrainAlign/ds003604-session-rdms`](https://huggingface.co/datasets/BrainAlign/ds003604-session-rdms) (12 of 12 task × session cells). The run confound described above applies to those RDMs too, and the recommended fix is to normalise voxel patterns within run before aggregating across runs.