Word-Level Brain Atlas

• The word-level brain atlas is a multidimensional framework that maps individual lexical items using neural activity patterns and computational embeddings.
• It employs methodologies such as MEG phase-locking and ECoG encoding to accurately align lexical features with brain-derived signals.
• The atlas enables cross-modal mapping for neural decoding and steering language models, enhancing interpretability and integration between neuroscience and AI.

A word-level brain atlas is a high-dimensional, quantitative mapping that captures the brain’s representations of individual words, grounded either in human neural responses or derived from language computational models. Such atlases serve as coordinate systems that index lexical items according to brain-derived activity patterns—enabling anatomical localization, cross-modal mapping, and principled steering of LLMs. Recent advances have yielded distinct methodologies: the construction of axes from magnetoencephalography (MEG) phase-locking value (@@@@6@@@@) connectivity patterns and the mapping of lexical features to cortical sites via direct neural recordings, both indexed at the level of single words (Andric, 22 Dec 2025, Evanson et al., 5 Dec 2025).

1. Foundational Concepts and Definitions

A word-level brain atlas is defined as a multidimensional matrix, with rows indexed by unique lexical items and columns corresponding to brain-derived features of word representation. These features may be, for example, independent components (axes) derived from MEG connectivity patterns or encoding weights from direct cortical recordings. The atlas encapsulates complex, distributed patterns linked to lexical properties (e.g., frequency, animacy, part-of-speech), providing a bridge between neurophysiology and computational LLMs (Andric, 22 Dec 2025).

Feature representations include:

• Phase-Locking Value (PLV) Patterns: Quantifying the phase synchrony between MEG channels, time-locked to word onsets.
• Lexical Descriptors: Metrics such as log-frequency and one-hot vectors for part-of-speech.
• Embeddings: Contextual or AI-derived (e.g., from LLM hidden states) representations reduced via principal component analysis (PCA).

2. Methodological Frameworks

2.1 MEG-Based Atlas Construction

The SMN4Lang MEG dataset provides the empirical substrate, comprising 12 native speakers listening to 60 naturalistic narrative stories. MEG is recorded on a 306-channel Elekta Neuromag system (204 planar gradiometers + 102 magnetometers), with precise word-onset annotation derived by forced alignment. Preprocessing utilizes the MNE-Python package: theta-band (4–8 Hz) filtering, artifact removal via SSP/ICA, and Hilbert-transformed analytic signals. Analysis windows of 2.0 seconds (0.5 s step) are aligned to the most recent word onset, and only gradiometer channels are used for PLV computation.

For each window and channel pair, PLV is computed as

$$\mathrm{PLV}_{ij}(f,t) = \left|\frac{1}{N}\sum_{n=1}^N e^{i(\phi_i^{(n)}(f,t)-\phi_j^{(n)}(f,t))}\right|,\quad i<j$$

where $N$ is the number of windows aligned to the same word label per run. The resulting $\approx20{,}706$ PLV edge values per window are reduced to 128 components by edge-PCA, yielding state vectors $y(t)\in\mathbb{R}^{128}$ (Andric, 22 Dec 2025).

2.2 Regression and Atlas Averaging

Each MEG window is associated with feature vectors $X(t)$, including GPT embedding-change $\Delta e$, log-frequency, and part-of-speech at lags $\tau\in\{0,0.5,1.0\}$ seconds (total $F\approx12{,}444$). Ridge regression minimizes

$$\min_{W,b} \left\|Y - XW - b\right\|_2^2 + \alpha\|W\|_2^2$$

with $Y\in\mathbb{R}^{T\times128}$. Five-fold out-of-fold (OOF) predictions ensure unbiased assignment of each window’s output. Atlas construction averages predicted PLV-PCA states for each word type across all windows, forming $A^s\in\mathbb{R}^{V\times128}$ per subject ($V\approx13{,}632$), and then averages across subjects yielding $A\in\mathbb{R}^{13{,}632\times128}$.

Circularity controls include rebuilding the atlas without embedding-change features, or substituting word2vec embeddings, yielding matched-axis correlations $|r|=0.64$–$0.95$ (Andric, 22 Dec 2025).

2.3 Electrocorticography (ECoG)-Based Mapping

In direct cortical studies, lexical features—specifically, Zipf frequency and PoS one-hot vectors—are mapped onto neural broadband activity recorded from 7,400 electrodes across 46 individuals. Encoding models (temporal response functions, TRFs) relate lagged word features to neural signals, with ridge regression and cross-validation selecting regularization.

Performance is evaluated per electrode as the Pearson correlation between TRF-predicted and observed signals, while decoding analyses invert the mapping to predict lexical features from neural activity, assessed by Spearman’s $\rho$ (Evanson et al., 5 Dec 2025).

3. Extraction and Validation of Latent Axes

3.1 Independent Component Analysis (ICA)

A whitened version of the group-averaged atlas $A$ undergoes FastICA with 20 fixed components, producing $S\in\mathbb{R}^{13{,}632\times20}$ (word $\times$ axis) and $W_{ica}\in\mathbb{R}^{20\times128}$. Each word receives a 20-dimensional coordinate vector in the latent brain-derived space, $S_w\in\mathbb{R}^{20}$, forming the final atlas $A_{IC}$.

3.2 External Lexical Validation and Statistical Controls

Axes are annotated using external lexica, without any text supervision during discovery:

• Animacy (axis 2): Separates animate vs inanimate (Cohen’s $d=0.53$ after residualizing log-frequency, surprisal, and length; $n=330$). Validation yields matched $d=0.70$, CI=[0.53,0.86].
• Lexical frequency (axis 15): Correlates with log-frequency ($r=0.51$, $n=13,632$).
• Additional associations: Axis 2 correlates with concreteness ($r=-0.128$, $p=0.001$); axis 15 with valence ($r=0.114$, $p=0.001$) and arousal ($r=0.084$, $p=0.001$); these are treated as secondary.

OOF vs base atlas axis correlations are high: $|r|=0.82$–$0.97$ for axes 2/13/15/19. Leakage-robust wPLI controls maintain axis 15 correlation of $0.744$, with others in $0.35$–$0.48$ range.

Exploratory fMRI encoding (ridge, n=240 subject-runs) suggests alignment for embedding change (corr$_\mathrm{obs}=0.0888$ vs null $0.0718$, $p=0.001$) and log-frequency (corr$_\mathrm{obs}=0.0815$ vs null $0.0725$, $p=0.009$) at 4 s HRF, but only marginal effects at 6 s (Andric, 22 Dec 2025).

3.3 Developmental and Spatial Evidence

Direct cortical data indicate that significant word-level encoding and decoding appear earliest in superior temporal gyrus (STG) and superior temporal sulcus (STS) in 2–5-year-olds (decoding $\rho=0.02$–$0.025$, $p<0.05$), expanding into anterior temporal lobe (ATL), inferior parietal lobe (IPL), and lateral/inferior frontal cortex (L/IFC) by age 6–11, with a plateau in adolescence ($\rho=0.03$–$0.06$, $p<0.01$) (Evanson et al., 5 Dec 2025).

4. Structural and Functional Properties

The atlas coordinate space is interpretable and stable:

• Primary axes robustly track lexical semantic properties without overfitting to text, as validated by independent lexica and task-matched controls.
• Rebuilding the atlas using alternate embedding inputs or excluding embedding features produces similarly structured axes (|r| up to $0.95$), reducing risk of circular inferences.
• Stability and leakage controls demonstrate that mapped properties persist across cross-validation splits and robustness pipelines.

Developmentally, word-level representations first manifest in the STG/STS and subsequently engage a distributed fronto-temporo-parietal cortical network (Evanson et al., 5 Dec 2025). Temporal analysis reveals that word features are decodable from neural signals from ~1 s before word onset, peaking ~350 ms post-onset and persisting for up to 2 s.

5. Interfaces with LLMs

5.1 Brain Readout

A ridge regression adapter $f:\mathbb{R}^H\rightarrow\mathbb{R}^{20}$ is trained per layer to predict $S_w$ from LLM hidden states $h^{(l)}(w)$. Held-out word performance (TinyLlama) yields axis 2 ($r=0.624$), axis 13 ($r=0.238$), axis 15 ($r=0.461$), and axis 19 ($r=0.499$) with $p\ll10^{-24}$ (Andric, 22 Dec 2025).

5.2 Brain-Grounded Steering

To steer an LLM, an axis vector is L2-normalized and added ($\Delta h = \alpha W_{ad}[k]/\|W_{ad}[k]\|$) at the desired layer, for various strengths ($\alpha\in\{-5,-2,-1,0,1,2,5\}$), tested on 50 prompts $\times$ 4 samples.

• Frequency axis (15): At TinyLlama layer 11, steering produces a strong increase in generated log-frequency (d=0.9246, $p=0.001$) and improved perplexity (d=–0.2183, $p=0.001$). A text-based probe yields opposite log-frequency and worsens perplexity.
• Function/content axis (13): Replicates steering across TinyLlama layer 11 (d=0.208, $p=0.001$), Qwen2-0.5B layer 4 (d=0.436, $p=0.001$), and GPT-2 layer 6 (d=0.300, $p=0.001$).

Secondary or inconsistent effects appear in earlier TinyLlama layers; these are interpreted as less robust.

6. Comparative and Developmental Perspectives

Evanson et al. (Evanson et al., 5 Dec 2025) demonstrate that word-level representations—operationalized via lexical descriptors and LLM embeddings—are robustly encoded and decodable from intracranial data, with maturational expansion from sensory cortex in early childhood to a distributed system including frontal and parietal cortices in late childhood and adolescence.

The MEG-based atlas approach provides population-level, high-dimensional axes applicable to LLM state analysis and control, while the ECoG-based mapping provides temporally resolved, anatomically localizable evidence of word encoding and its development.

7. Applications and Broader Implications

The word-level brain atlas serves as a neurophysiology-grounded coordinate system for:

• Reading and steering LLMs: Providing interpretable axes along which model representations and outputs can be aligned or manipulated.
• Neural decoding: Enabling predictive models that track lexical features or embeddings from neural dynamics.
• Cross-model evaluation: Permitting direct comparison between human and artificial representations of language, facilitating neuroscientific and AI research.

A plausible implication is that brain-grounded axes offer a principled foundation for model interpretability and controllability, surpassing text-supervised probes by yielding axes that are robust to perplexity controls and circularity, and grounded in human neurophysiological organization (Andric, 22 Dec 2025). The existence of a stable, multidimensional brain atlas indexed at the lexical level substantiates a powerful interface that can inform both cognitive neuroscience and the engineering of LLMs.