Editorial illustration for LLM-derived valence direction aligns with EEG signals in 123 subjects
LLM-derived valence direction aligns with EEG signals in...
The simplest measure of emotion, valence, the axis from pleasure to displeasure, has long resisted a clean neural readout. In a landmark public dataset of 123 subjects watching affective videos, we found that large language models and human electrophysiology converge on the same direction in neural space. A single linear projection on EEG features tracks the valence position of every stimulus.
Thirty-six emotion classifiers, never exposed to this axis, spontaneously re-discover it in their internal representations. But here the story takes a sharp turn. This convergence provides no effective training signal.
We tested twenty-five alignment strategies, knowledge distillation, representational similarity, contrastive losses, topographic losses, none improve decoding. Sixteen significantly reduce accuracy. This is the saturation regularity: once task labels alone drive a brain-decoding network onto the target direction, additional supervision mainly distorts an already-saturated basin.
The load-bearing within-class residual receives little useful gradient. True improvement must come from elsewhere, from the unreachable residual subspace. We ensembled across residual diversity rather than supervising the basin, and boosted balanced accuracy by 10.5% over the prior best on FACED, with the same effect replicated on SEED-V.
On a public EEG cohort of 123 subjects watching affective videos, a single linear projection on EEG features tracks the V-axis position of each stimulus. Moreover, 36 EEG emotion classifiers trained without exposure to the V-axis spontaneously rediscover the same direction in their internal representations, suggesting that the same valence structure emerges in both language models and human electrophysiology.
The shared valence axis is real. It cuts cleanly across language models and human brains, a single linear direction that both systems converge upon independently. But that convergence is a trap.
The basin is saturated, supervising it further distorts rather than refines. The load-bearing signal lives elsewhere. In the residual subspace, unreachable by any alignment loss we tested, lies the structure that actually improves decoding.
That is the saturation regularity: a finding that reframes the problem. Not “how do we align representations,” but “how do we exploit what alignment cannot touch.” Ensembling across residual diversity delivers a 10.5% gain on FACED, replicated on SEED-V. The path forward is not deeper supervision of the axis we share.
It is systematic exploration of the variance we do not.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives gSMILE Framework Tackles LLM Transparency by Mapping Prompt Responses
From Archives DAStatFormer extracts 24 ANOVA-selected features per channel, slashing data size
Common Questions Answered
How did researchers use large language models to identify the valence axis in EEG signals?
Researchers discovered that large language models and human electrophysiology converge on the same direction in neural space when measuring emotional valence. By applying a single linear projection on EEG features from 123 subjects watching affective videos, they found that LLM-derived valence direction aligned with the brain's electrical signals, allowing them to track the valence position of every stimulus.
What is the saturation regularity finding mentioned in the study?
The saturation regularity refers to the discovery that the shared valence axis between language models and human brains becomes saturated, meaning that further supervision actually distorts rather than refines the signal. The researchers found that the load-bearing signal for improved decoding lives in the residual subspace, which is unreachable by standard alignment loss methods they tested.
Why is the convergence between LLMs and human brains on the valence axis considered a trap?
While the convergence on a single linear direction for valence is real and cuts cleanly across both language models and human brains, it represents a saturated basin where additional optimization efforts become counterproductive. The researchers found that attempting to supervise this shared axis further distorts the signal rather than improving it, suggesting that meaningful improvements require looking beyond this obvious convergence point.
How many emotion classifiers independently rediscovered the valence axis in this study?
Thirty-six emotion classifiers, which had never been exposed to the LLM-derived valence axis, spontaneously rediscovered it independently. This finding demonstrates the robustness and fundamental nature of the valence direction across different machine learning approaches, suggesting it represents a genuine neural principle rather than an artifact of a specific model.
Further Reading
- A Shared Valence Axis Across Modern LLMs and Human EEG — arXiv
- Latent Structure of Affective Representations in Large Language Models — arXiv
- NeuroLM: A Universal Multi-task Foundation Model for Bridging the... — OpenReview
- The cognitive impacts of large language model interactions on ... — Frontiers in Computational Neuroscience
- SleepLM: Modality Alignment of LLMs with EEG Signals — Semantic Scholar
