Different EEG Models Find Shared Patterns—but Bigger Is Not Always Better

Different AI systems can turn the same recording into very different collections of numbers. Those internal representations determine which moments the systems treat as similar. In The Platonic Brain: Representational Convergence in EEG Foundation Models, researchers ask whether models trained on electroencephalography, or EEG, nevertheless develop shared ways of organizing brain signals.

The November 2026 arXiv preprint reports a qualified pattern: EEG models whose representations better predicted selected signal markers also tended to align more closely with other EEG models. Bigger models were not reliably better by that measure. Comparisons with vision and language models produced a more selective result, rather than evidence of universal convergence.

The study examines the Platonic Representation Hypothesis, which proposes that increasingly capable, larger neural networks develop more similar representations. Here, convergence means agreement in the mathematical organization of data—not identical internal activity, and not proof that an artificial model represents experience as a human does.

The researchers used recordings from six participants, aged 21–26, watching the same 180 minutes of The Big Bang Theory. They examined 14 pretrained EEG models from five families, alongside vision models processing the corresponding images or video and language models processing scene captions. All models remained frozen: the study measured their existing representations rather than retraining them on this material.

A central methodological choice was to separate capability from parameter count. The authors assessed how well each EEG model’s representations could predict 16 predefined EEG markers. These included power in different frequency bands, measures of signal irregularity and measures of how rapidly signals change. A regression model predicted the markers from the representations, with evaluation split into contiguous time blocks to keep nearby windows from appearing on both sides of a split.

The resulting score, called NICE decodability, is a specific proxy for capability. It measures how accessible these selected signal properties are within a representation. It does not establish that a model is better at every EEG task, nor does it measure comprehension of the television programme. Some markers were also much harder to predict than others.

The researchers then asked two distinct questions about similarity. A local measure checked whether two models gave each time window the same nearest neighbours: did they agree on its most similar moments? A global measure, linear CKA, compared the broader pattern of similarities across all windows. Models could agree on that overall arrangement without consistently choosing the same closest neighbours.

Time posed another complication. Nearby moments in a recording can resemble one another, and stimulus windows can share visual clips or captions. Such temporal structure can raise similarity scores without demonstrating correspondence to the same content. The authors therefore compared correctly aligned recordings with controls that circularly shifted one sequence against the other, preserving its temporal structure while changing the pairing. Reported alignment was measured above this shifted baseline.

For the main EEG comparison, the team measured each model’s average alignment with models outside its own family. They also centred capability and alignment scores within families to account for family-specific baselines. Under this analysis, greater marker decodability was associated with stronger alignment on both local and global measures. The result supports convergence with this particular capability measure; it does not establish convergence with size or show that improving marker prediction causes representations to converge.

Across modalities, the findings were less uniform. Better EEG marker decodability was associated with stronger global alignment to language representations. Relationships with vision models were weaker or inconsistent, while local alignment showed no consistent relationship with decodability for either modality. The authors suggest that local comparisons may emphasize low-level sensory features, whereas global structure may capture distributed similarities related to meaning. They explicitly note that this explanation was not directly tested.

The limits are substantial. The analysis covers only six people, 14 EEG models and one dataset drawn from one television series. Cross-modal local similarities exceeded the shifted controls only modestly. The language inputs were scene captions, not reports of participants’ thoughts. Consequently, the results do not demonstrate thought decoding, establish a shared semantic code or validate a clinical application.

The authors propose that representational alignment and marker decodability could help compare EEG models when downstream benchmarks disagree. For now, that remains a proposed use. The preprint’s narrower contribution is evidence that selected EEG capabilities track representational agreement—and a demonstration of why model size, temporal structure and the choice of similarity metric must be considered separately.