Language models show broad emotion patterns while retaining finer distinctions

Language models may describe a sentence as fearful, nervous or sad, but their answers do not reveal how those distinctions are organized internally. A new study finds that broad emotional tone dominates one analysis of their hidden representations, while finer distinctions remain partly recoverable.

Laya Iyer and Sanmi Koyejo of Stanford University and Chen Shani of Tel Aviv University call this pattern “overcompression” relative to the GoEmotions dataset’s categories. Their paper, Compression of Emotion in Language Model Representations, was submitted to arXiv on November 5, 2026. The supplied source is a first-version preprint; it does not establish peer-reviewed publication.

Three ways to examine emotion

The researchers studied seven open-weight, instruction-tuned models from the Llama, Gemma and Qwen families, ranging from roughly one billion to nine billion parameters. They used 36,308 English Reddit comments from GoEmotions, retaining only examples with exactly one released, aggregated emotion label.

Each comment was evaluated under three classifications: the original 27 emotions plus neutral; six broader emotion families plus neutral; and four sentiment categories—positive, negative, ambiguous and neutral. “Ambiguous” here is a dataset category, not a measure of annotation uncertainty.

The analysis separated three questions. Could a model name the annotated emotion? Did comments with matching labels form distinct groups internally without supplying those labels to the clustering algorithm? And could a separately trained classifier recover the labels from the model’s representations?

For the internal analysis, the team prompted models to identify emotion and extracted numerical representations at the final input position across model layers. They normalized these vectors and grouped them using KMeans, an algorithm that partitions examples by their positions in a numerical space. They tested cluster counts from two to 35 at each model’s layer with the strongest alignment to the 28-category taxonomy.

Broad structure was clearest

Across all seven models, a measure called the silhouette score, which assesses cluster separation, was highest with two clusters. Those groups were consistently associated with positive-versus-negative emotional tone.

Restricting the comparison to positive and negative examples, and matching the arbitrary cluster identities to those categories, produced balanced accuracy between 78.3% and 83.7%. Neutral and ambiguous comments did not fit a clean binary division. The result therefore describes a dominant valence-related pattern, not evidence that models contain exactly two emotion categories.

Broader classifications also needed fewer clusters to approach their best agreement with dataset labels. The 28-category taxonomy required 13–17 clusters to reach 95% of its peak alignment, compared with three to six for emotion families and two to four for sentiment. These thresholds were calculated separately within each taxonomy; they do not indicate near-perfect classification.

Related emotions tended to overlap systematically. Fear and nervousness were among the three most strongly co-clustered label pairs in every model. Embarrassment–remorse and grief–sadness also ranked highly. Overall, different labels within the same broader family co-clustered 1.7–2.2 times as often as labels from different families.

Finer information remained accessible

Weak separation into clusters did not mean fine-grained information had disappeared. The researchers trained regularized linear classifiers, or probes, on frozen representations, without updating the language models. Examples were divided into training, validation and test sets in a 70/15/15 split.

For the 28-category task, held-out macro F1 ranged from 0.360 to 0.404. Macro F1 averages classification performance across categories, giving rare labels the same weight as common ones. Probe scores were higher for the broader classifications.

These results support a distinction between information being accessible and forming sharply separated clusters. The emotion-centroid analysis likewise indicated a richer structure, with an effective rank of roughly eight to 12 dimensions, rather than a collapse to one or two dimensions.

The output experiments offered another caution. Broader labels generally improved answers, but mapping fine-grained predictions into broader categories reproduced much of that advantage. Better sentiment scores alone therefore do not establish coarser internal representations.

What the study cannot establish

The findings concern a particular dataset and analysis protocol, not human emotion itself. Neutral comments made up 12,823 examples, while grief had just 39. That imbalance can influence clustering, and the single-label restriction excluded mixed-emotion comments.

Results also depended on prompting, readout position and clustering method. A different clustering procedure produced weaker, less consistent family-level effects. Probe layer selection was not fully isolated from held-out data, and the global probes did not directly establish separability of individual pairs such as fear and nervousness.

The study did not test whether models causally use these distinctions during generation, whether they improve response quality or safety, or whether findings extend to other languages and larger systems. Its central result is narrower: broad affective structure can dominate clustering while finer emotion information remains available.