LLMと人間のEEGに共通する感情価軸「彩度規則性」の発見 A Shared Valence Axis Across Modern LLMs and Human EEG: The Saturation Regularity

Large language models have long been suspected of doing more than pattern-matching over text. A growing body of evidence suggests their internal representations echo structures found in human cognition — and a new paper pushes that idea into the domain of affective neuroscience.

The study, posted to arXiv as 2606.00129, identifies what the authors call the 'Saturation Regularity': a geometric pattern in LLM feature spaces that aligns with the valence dimension of human EEG signals. Valence — the positive-to-negative axis of emotional experience — is one of the foundational constructs in affective psychology, typically paired with arousal to form a two-dimensional model of emotion. Finding a structural analog to this axis inside a language model is a notable claim.

The research sits at the intersection of representational alignment and affective computing, two areas that have seen rapid growth. Prior work has shown correlations between transformer hidden states and fMRI BOLD responses during language comprehension, and several groups have demonstrated that word embeddings carry implicit sentiment geometry. This paper extends the conversation by using EEG — which captures neural dynamics at millisecond resolution — and by naming a specific regularity rather than reporting a diffuse correlation.

The choice of EEG as the neural reference is worth noting. Unlike fMRI, EEG reflects fast temporal dynamics and is sensitive to the rapid fluctuations that characterize emotional processing. If LLM representations track something that shows up in EEG, the alignment is arguably more temporally grounded than previous findings based on slow hemodynamic signals.

What exactly is the Saturation Regularity? Based on the abstract, it appears to describe a systematic organization within the valence axis — possibly a non-linear saturation effect where emotional extremes are encoded differently from neutral states. The full paper will be needed to evaluate the robustness of the finding across model families and EEG paradigms.

The practical implications are speculative but interesting. Emotion-aware AI systems, clinical tools for detecting affective disorders via language biomarkers, and interpretability research that grounds model internals in human neuroscience could all potentially benefit. That said, the finding should not be over-read as evidence that LLMs 'feel' anything; what is being claimed is a geometric correspondence in representation space, not phenomenological equivalence.

The broader trend here is the gradual convergence of computational linguistics and cognitive neuroscience. As models grow larger and more expressive, the overlap with human neural representations appears to deepen — whether by design, by the shared structure of language itself, or by some combination of both. This paper adds a concrete data point to that ongoing debate and is likely to attract follow-up work examining whether the Saturation Regularity generalizes across architectures, languages, and clinical populations.