Technical Summary: A Shared Valence Axis Across Modern LLMs and Human EEG: The Saturation Regularity

Problem Statement

This paper investigates the alignment between the internal representations of Large Language Models (LLMs) and human electrophysiology (EEG), specifically focusing on the neural encoding of emotional valence. While prior work has established that LLM features align with human cognition on high-level concepts, this research asks whether this alignment can be leveraged as a usable bridge to improve brain-decoding networks. The authors specifically test the hypothesis that an LLM-derived valence direction can serve as an auxiliary training signal to enhance EEG emotion recognition.

Methodology

1. Construction of the Valence Axis (V-axis)

The authors construct a one-dimensional valence direction, termed the V-axis, using a minimal, zero-shot procedure:

• Input: Nine short, emotion-evocative sentences, one for each of the nine emotion classes in the FACED dataset (Anger, Disgust, Fear, Sadness, Neutral, Amusement, Inspiration, Joy, Tenderness).
• Extraction: These sentences are passed through an LLM (primarily Qwen3-4B). The hidden states of the last token at the penultimate layer are extracted for each sentence.
• Aggregation: For each emotion class, the 50 generated paraphrases are averaged to form a class centroid.
• Projection: Principal Component Analysis (PCA) is applied to these nine centroids. The first principal component is oriented such that the "Joy" centroid projects positively. This single vector constitutes the V-axis.

2. Validation of the V-axis

The authors verify the V-axis through three independent channels without any fine-tuning or exposure to labeled sentiment data:

• Text: Zero-shot projection onto standard sentiment benchmarks (SST-2, IMDB, etc.).
• Vision: Application of the same recipe to CLIP-image embeddings.
• Brain: A linear ridge regression is trained to predict the V-axis projection of stimulus descriptions from the EEG features of 123 subjects watching emotional videos (FACED dataset).

3. Testing Supervision and the "Saturation" Hypothesis

The core experimental setup involves testing 25 different "alignment recipes" to see if forcing an EEG emotion classifier to align with the V-axis improves performance. These recipes include:

• Knowledge Distillation (KD) using LLM class prototypes.
• Representational Similarity Analysis (RSA).
• Contrastive losses.
• Parameter-efficient fine-tuning (e.g., LoRA).
• Topographic losses targeting specific brain channels.

The authors analyze the behavior of these interventions across baselines with varying strengths (Balanced Accuracy, BACC, ranging from ~0.62 to ~0.66).

4. Mechanism Analysis

To understand why supervision fails, the authors decompose the feature space of trained EEG classifiers into two subspaces:

• Class-Mean Basin: The component of features that aligns with the average representation of each emotion class.
• Within-Class Residual: The component of features that distinguishes individual trials within the same emotion class.

They measure the correlation between the V-axis and these subspaces across 36 different EEG classifier checkpoints (trained without V-axis supervision).

5. Ensemble Strategy

Based on the mechanism analysis, the authors propose a "Two-Tier Ensemble." Instead of using the V-axis as a loss target, they ensemble multiple checkpoints trained with different random seeds and training durations. The hypothesis is that while the class-mean basin is saturated and shared across seeds, the within-class residual contains seed-specific diversity that can be aggregated to improve performance.

Key Results

1. The V-axis is a Shared, Robust Direction

• Cross-Model Consistency: The V-axis direction is highly consistent across 14 different LLMs (ranging from 560M to 32B parameters). Within the Qwen3 family, the alignment is nearly perfect ($r=0.995$).
• Zero-Shot Performance: The V-axis predicts SST-2 sentiment with an AUC of 0.832, matching a supervised logistic regression trained on 5,000 examples.
• Brain Alignment: The V-axis predicts the cohort-level EEG response to emotional videos with a correlation of $r=0.87$ ($p < 10^{-9}$). This alignment holds across 14 different LLMs and is replicated on the SEED-V dataset.
• Spontaneous Rediscovery: 36 EEG classifiers trained only on emotion labels (without V-axis supervision) spontaneously encode the V-axis in their class-mean subspace. The strength of this encoding correlates strongly with the classifier's balanced accuracy ($r=0.885$).

2. The Saturation Regularity

The central negative finding is that supervision fails.

• No Gains: None of the 25 alignment recipes improved the performance of the EEG classifiers.
• Significant Degradation: 16 out of 25 recipes produced statistically significant accuracy decrements. Five families showed "monotonic destruction," where increasing the loss weight strictly worsened performance.
• The Transition: The failure is sharp. Interventions that had no effect on weak baselines (BACC $\le$ 0.62) became significantly harmful on stronger baselines (BACC $\ge$ 0.66).
• Mechanism: The auxiliary loss successfully pushes the class-mean features further along the V-axis (increasing alignment by $\Delta|r| = +0.01$ to $+0.36$), but it leaves the within-class residual (the load-bearing component for distinguishing trials) at a numerical zero ($\sim 10^{-7}$). The network is already "saturated" on the target direction; the loss merely deforms the basin without improving the orthogonal, discriminative features.

3. Residual Diversity and New SOTA

• Residual Correlation: The strength of the within-class residual encoding predicts the contribution of a checkpoint to an ensemble ($r=0.74$).
• Ensemble Success: By ensembling 10 checkpoints (varying seeds and training lengths) and averaging their predictions, the authors recover the diversity in the residual subspace.
• Performance: This ensemble achieves a Balanced Accuracy of 0.6948 on the FACED 9-class task, a 10.5% relative improvement over the previous state-of-the-art (EMOD at 0.6287).
• Ablation: Replacing the LLM-derived prototypes in the KD step with random vectors caused negligible performance loss ($\le 0.003$), confirming that the gain comes from the structure of the ensemble and the residual diversity, not the semantic content of the LLM prototypes.

Significance and Claims

The paper claims three primary contributions:

1. A Nine-Sentence V-axis Probe: A method to extract a universal, one-dimensional valence direction from LLMs that converges across models, predicts text sentiment zero-shot, predicts human EEG, and is spontaneously rediscovered by EEG classifiers.
2. The Saturation Regularity: A previously unrecognized principle at the LLM-brain interface. Once a task (like emotion classification) has driven a network onto a target direction using task labels alone, adding auxiliary supervision for that same direction does not help; it deforms the already-saturated basin and ignores the load-bearing residual. This explains why alignment losses often fail or hurt performance in high-capacity regimes.
3. A Positive Prescription (Residual Diversity): Instead of supervising the saturated basin, the actionable gain lies in the unsupervised, seed-specific residual subspace. Ensembling across this diversity yields state-of-the-art results, effectively turning the "saturation" diagnosis into a training strategy.

Scope and Limitations:
The authors explicitly state that the "Saturation Regularity" is scoped to the FACED-9 dataset with a V-axis substrate and is an empirical pattern, not a mathematical theorem. They clarify that the V-axis alignment is a "direction-existence" claim, not a claim that the LLM fully models biological emotion processing. Furthermore, the posterior-dominant EEG signal observed is specific to dynamic video stimuli and does not refute the frontal-alpha asymmetry (FAA) hypothesis developed for static stimuli. The paper concludes that future work should investigate whether this residual-vs-basin decomposition applies to other saturated concepts (e.g., truthfulness, refusal) and other domains.