💬 NLP

Nine Emotion Centroids: A Label-Free Valence Axis That Transfers Across Four Modalities

This paper demonstrates that a single, label-efficient "valence axis" derived from just nine emotion categories and short narratives can be extracted from frozen language models and successfully transfer to vision, audio, and brain encoders to predict sentiment with high accuracy, provided the target concepts are continuous rather than categorical.

Original authors: Yousef Radwan

Published 2026-08-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yousef Radwan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Inside the vast, silent architecture of a modern computer language model, thousands of neurons fire in complex patterns every time the machine reads a sentence. For years, scientists have suspected that hidden within this electrical storm are simple, straight lines that track specific ideas, much like a single needle on a compass pointing north. These lines, known as internal directions, can tell us if a sentence is about a person, if the model is about to refuse a request, or if a concept refers to a specific country. But one of the most fundamental human experiences—the feeling of whether something is good or bad, pleasant or unpleasant—has remained harder to pin down. This emotional charge, called valence, is the core of how we judge the world. Understanding whether a machine can naturally sense this feeling, without being explicitly taught with thousands of examples, is a crucial step in knowing how artificial intelligence truly thinks and feels.

A researcher has now discovered that this sense of emotional positivity or negativity is not scattered randomly through a machine's mind. Instead, it lives along a single, clear path that can be found with astonishingly little effort. In a study involving text, images, sound, and even human brain activity, the researcher showed that they could locate this emotional direction using only nine simple words and a few hundred short stories. They did not need to feed the machine thousands of labeled examples of "happy" or "sad" sentences. Instead, they asked the machine to read nine short narratives, each anchored to a specific emotion like anger, joy, or fear. By averaging the machine's internal reaction to these stories, they found a single line of activity that stretched from negative to positive. This line, which the researcher calls the V-axis, acts as a universal emotional dial. When they projected new information onto this line, the machine could accurately guess the emotional tone of a sentence, a picture, a sound, or even a recording of a human brain watching a video, with a level of accuracy that rivals systems trained on massive amounts of data.

What makes this discovery particularly striking is that the same emotional line appears in machines that have never been taught together. The researcher tested this on four different types of systems: a language model that reads text, a vision system that looks at images, an audio system that hears sounds, and a model that interprets electrical signals from human brains. None of these systems shared any training data with one another; they were built for entirely different jobs. Yet, when the researcher applied their simple nine-story method to each one, they found the same emotional direction emerging in every single case. It is as if the concept of "good" and "bad" is so fundamental that it creates a similar shape in the mind of a text-reader, an image-recognizer, and a brain-monitor, regardless of how they were built. The researcher verified this by taking a classifier trained only on text labels and using it to read the emotional tone of images, sounds, and brain waves. It worked with high accuracy, proving that this single emotional dimension bridges the gap between seeing, hearing, reading, and thinking.

The researcher also proved that this line is not just a coincidence or a passive marker that happens to sit near the emotional data. To test if the line was actually doing the work, they performed a surgical experiment. They took the language models and, while they were processing sentences, they mathematically removed this specific emotional direction from the machine's internal state. When they did this, the machine's ability to understand sentiment collapsed. In some cases, its accuracy dropped by more than thirty percentage points, while removing a random direction of the same size had almost no effect. This suggests that the machine relies on this specific path to make its emotional judgments. However, the researcher also found that this ability is not identical in every machine. While they could find the line in models from different families, they could only use it to steer the machine's behavior in some of them. In certain models, the line was present and could be detected, but adding it back to the machine's thoughts did not change its output. This indicates that while the emotional direction is a universal feature of these systems, the way different machines use it depends on their specific history and design.

The study also clarified what this method cannot do. The researcher tested their approach on many other types of concepts, such as identifying specific objects or categories of words, and it failed to find a clear direction. The method works specifically for continuous feelings like emotional tone, where the data can slide smoothly from one end to the other, rather than for sharp, distinct categories. Furthermore, while the method found the emotional line in brain recordings without needing to label the brain data itself, the initial setup for the brain model did require some supervised training to distinguish between positive and negative feelings. This means the "label-free" success applies to the final step of reading the emotion, not the initial construction of the brain model. Despite these boundaries, the findings offer a powerful new way to look inside artificial intelligence. By showing that a single, simple direction can capture the essence of human emotion across text, sight, sound, and brain activity, the research suggests that the most complex human feelings might be represented in machines in a way that is surprisingly simple, shared, and accessible.

Technical Summary: Nine Emotion Centroids and the Cross-Modality Valence Axis

1. Problem Statement

Current methods for identifying high-level attributes (such as sentiment or valence) within foundation models typically rely on supervised learning. This involves training linear classifiers on thousands of labeled examples (e.g., positive vs. negative sentences) to extract a "direction" in the model's hidden state space. This approach is data-intensive and often fails to generalize across modalities (text, vision, audio, brain activity) because the encoders for these modalities are trained independently with no shared pretraining.

This paper addresses two specific gaps:

Label Efficiency: Can a continuous attribute axis be recovered with a drastically reduced label budget (order of tens rather than thousands)?
Cross-Modality Transfer: Does a single, task-relevant direction exist across fundamentally different encoders (text, image, audio, EEG) that were never jointly trained, allowing a classifier trained on one modality to function on others without retraining?

2. Methodology: The "Recipe"

The author proposes a label-free (for text, image, and audio) construction method called the V-axis (Valence Axis) recipe. The process relies on the geometric structure of emotion categories rather than binary sentiment labels.

Algorithm 1: V-axis Construction

Input: A frozen encoder $f_\theta$ , a specific layer $\ell$ , and nine emotion category names (Anger, Disgust, Fear, Sadness, Amusement, Joy, Inspiration, Tenderness, Neutral).
Data Generation: For each of the 9 categories, the author authors $\approx 50$ short narrative paragraphs (total $\approx 450$ sentences) that evoke the specific emotion. Crucially, these are not curated for sentiment polarity (e.g., "anger" stories are not necessarily negative in a binary sense, but are emotionally charged).
Embedding & Pooling: Each sentence is passed through the frozen encoder. The residual stream activation at the final token of a fixed layer $\ell$ is extracted.
Centroid Calculation: The activations for each emotion category are averaged to form 9 emotion centroids ( $\mu_c$ ) in the hidden state space ( $d$ -dimensional).
SVD/PCA: The 9 centroids are centered, and a Singular Value Decomposition (SVD) is performed on the resulting $9 \times d$ matrix. The first right-singular vector ( $v_1$ ), corresponding to the direction of maximum variance among the 9 centroids, is defined as the V-axis.

Modality Adaptations

Text: Uses LLMs (Llama-3, Qwen3). The V-axis is extracted from the residual stream.
Vision: Uses CLIP-image encoder. Replaces text with $\approx 50$ class-correlated images per emotion from the EmoSet dataset.
Audio: Uses CLAP-audio encoder. Replaces text with $\approx 50$ class-correlated audio clips per emotion from ESC-50.
EEG: Uses the CBraMod encoder. Note: The EEG axis construction is an exception; the unsupervised 9-class PC1 failed (capturing arousal rather than valence). Instead, a supervised Fisher Linear Discriminant (LDA) was fitted on binary valence labels within the EEG feature space. The "label-free" claim for EEG applies only to the classifier head, not the axis construction itself.

Layer Selection

Standard Models (Llama, Mistral, Qwen3 base/instruct): Optimal layer is mid-depth ( $\ell \approx L/2$ ).
Reasoning-Distilled Models (DeepSeek-R1-Distill, Qwen3-thinking): The V-axis is suppressed at mid-depth and recovers near the final layer ( $\ell \approx L$ ).

3. Key Contributions

The Recipe: A fully specified, reproducible method to extract a 1-D valence axis using only 9 emotion names and $\approx 450$ unlabelled narrative sentences (a $\sim 1,500\times$ reduction in label cost compared to standard supervised probes).
Four-Modality Probe: Demonstration that the same recipe yields a V-axis in text, vision, audio, and EEG encoders that correlates with human valence ratings and supervised classifiers.
Causal Evidence: Application of directional ablation (Arditi et al., 2024) showing that surgically removing the V-axis from the hidden state significantly degrades sentiment classification performance, whereas removing random directions of equal magnitude does not.
Universal Cross-Modal Classifier: A single 2-parameter logistic regression head, trained only on text (SST-2) labels, successfully classifies sentiment in images, audio, and EEG without seeing any target-modality labels.
Scope Boundaries: Identification of the recipe's limits: it works for continuous attributes (valence) but fails for categorical concepts (e.g., object classes, AxBench concepts) and exhibits model-family-specific steerability (works for Llama/Mistral, fails for Qwen/Gemma).

4. Key Results

Performance vs. Supervised Baselines

The V-axis recovers sentiment performance close to fully supervised classifiers across all modalities:

Text (Llama-3-8B): AUC 0.772 vs. Supervised 0.828 (93% of supervised performance).
Vision (CLIP): Pearson $r = +0.636$ vs. Human ratings (Supervised $r=0.81$ ). Random direction null $|r| \le 0.112$ .
Audio (CLAP): Mean AUC 0.906 on ESC-50 (Supervised 0.94).
EEG (CBraMod): AUC 0.720 (Supervised 0.83).

Causal Ablation

Removing the V-axis from the residual stream during inference causes significant drops in SST-2 accuracy:

Llama-3-8B: Drop of 5.5 percentage points (pp).
Qwen3-1.7B: Drop of 15.8 pp.
Qwen3-8B: Drop of 37.2 pp.
Control: Removing 3 matched-norm random directions resulted in drops of $\le 0.88$ pp. The signal is $\ge 12$ standard deviations above the random null.

Cross-Modal Transfer

A classifier trained on text V-axis projections transfers to other modalities with high efficacy (all 12 off-diagonal cells in the $4\times4$ matrix exceed AUC 0.70):

Text $\to$ Image: AUC 0.961.
Text $\to$ Audio: AUC 0.764.
Text $\to$ EEG: AUC 0.828.
Baseline Comparison: A generic 16-dimensional shared subspace (computed without task labels) performs at chance (AUC 0.525). One task-relevant dimension (V-axis) outperforms 16 generic dimensions by 0.18–0.44 AUC.

Limitations and Boundaries

Categorical Concepts: The recipe fails on categorical concepts (e.g., AxBench Concept-500, object classes), returning near-chance performance.
Steerability: Adding the V-axis back to generation steers sentiment in Llama and Mistral families but has no effect in Qwen and Gemma families.
EEG Supervision: The EEG V-axis itself required supervised labels; the "label-free" aspect is limited to the cross-modal classifier head.
Data Volume: The recipe collapses if the per-emotion pool is too small (e.g., single words yield AUC 0.50); $\approx 20$ paragraphs per emotion are required for non-trivial results.

5. Significance and Claims

The paper claims that valence is a rare attribute that is:

Simple enough to be recovered with minimal supervision (9 examples + prompts).
Shared enough across disparate encoders (text, vision, audio, brain) to bridge them through a single 1-D dimension.

The author positions this as an empirical regularity rather than an analytical theorem. They argue that while the Linear Representation Hypothesis (LRH) typically applies to categorical discriminants, the V-axis represents a continuous-attribute direction that is approximately orthogonal to those discriminants. The work extends the "Platonic Representation Hypothesis" (convergence of encoders on object similarity) to a task-relevant 1-D direction that emerges independently in encoders with no shared pretraining, including human brain activity.

The study concludes that the V-axis is not merely a correlated feature but is mechanistically active in sentiment processing, as evidenced by the severe performance degradation upon its removal, and serves as a universal bridge for sentiment classification across modalities without requiring target-modality labels.