AI shares emotion with humans across languages and cultures

Introduction

• Human–LLM emotional alignment matters for effective, trustworthy collaboration, but whether models represent emotion as humans do and how to regulate their output remain open questions.The introduction links meaningful emotional exchange to mutual understanding and trust, and to safer interactions in applications requiring nuanced emotional understanding.Source: Introduction · p. 2
• Human emotions are nuanced, dynamic, and culturally and linguistically variable, while coarse categorical or sentiment-based datasets miss their continuous, high-dimensional character.The introduction notes that culture-specific emotional states may be difficult to translate across languages.Source: Introduction · p. 3
• Foundational gaps include limited interpretability of LLM emotion features and opaque mechanisms linking internal representations to emotional expression, alongside inadequate precise control.Polysemantic neurons complicate human interpretation, and regulating outputs could help reveal causal links between internal representations and expression.Source: Introduction · p. 4
• The study asks whether interpretable emotion spaces resemble human emotion, using SAE features from Gemma and Llama models across English and Chinese and comparing latent embeddings with valence and arousal.The constructed spaces use a broad taxonomy of human-centric emotion categories.Source: Introduction · p. 6
• A second question is whether features shared across English and Chinese emotion spaces can predict human affective ratings in large-scale behavioral datasets.The ratings concern emotional content in language.Source: Introduction · p. 6

• LLM-derived emotion spaces are anchored in valence and arousal, two foundational dimensions of human emotion.Source: Introduction · p. 6
• LLMs encode fine-grained emotional features in a way akin to human affective perception.Source: Introduction · pp. 6–7
• Small sets of interpretable features transformed from human-centric concepts can causally steer the emotional tone of LLM outputs.Source: Introduction · p. 7

LLM-derived interpretable emotion spaces are underpinned by valence and arousal

• The concept-driven approach maps human behavioural emotion semantics to interpretable SAE features in multilingual LLMs using associated-word concept sets.The approach aligns across linguistic-cultural contexts without further human annotation of LLM states or outputs.Source: LLM-derived interpretable emotion… · p. 7
• The analysis covers 26 emotion categories per language, excluding sexual desire because it could prompt taboo or sensitive model content.Separate English and Chinese emotion spaces were constructed.Source: LLM-derived interpretable emotion… · p. 7
• The approach was applied with SAEs for Gemma2-9B-IT and Llama3-8B-IT, selecting each model’s analysis layer based on the lowest delta language-modelling loss.This design was used to assess applicability beyond one SAE implementation or LLM architecture.Source: LLM-derived interpretable emotion… · p. 9
• Emotion-space dimensionality tracks the vocabulary size of its concept sets, while multiword sets prioritize broad emotional coverage over avoiding redundant features.This allows the scale of a space to be tailored.Source: LLM-derived interpretable emotion… · p. 10
• Significant Pearson correlations (Bonferroni-corrected P < 0.001) between all three CEBRA-Behaviour dimensions and valence and arousal scores show that both dimensions underpin the emotion-space geometry.The authors interpret this alignment as evidence that the SAE-based spaces capture fundamental aspects of how people represent and understand emotions in language.Source: LLM-derived interpretable emotion… · p. 11, Figure 2 · p. 12
• CEBRA-Behaviour embeddings show distinct yet interrelated emotion clusters in English and Chinese across both LLM families, consistent with the initial UMAP projections.Recolouring the embeddings by valence and arousal reveals gradients that align with the latent-space organization.Source: LLM-derived interpretable emotion… · pp. 10–11, Figure 2 · p. 12

Emotional SAE features encoded in LLMs align with human linguistic emotion perception

• Intersectional features outperformed extra-space features and performed comparably to all features across within- and cross-language predictions.The intersectional set had considerably reduced dimensionality, with no significant difference from the all-features benchmark.Source: Figure 3 · p. 15, Emotional SAE features encoded in LLMs… · pp. 15–16
• Valence ratings were predicted more accurately than arousal ratings, and cross-language predictions were less accurate than within-language predictions.These patterns held across both LLM families and the SAE configurations tested.Source: Emotional SAE features encoded in LLMs… · p. 16
• Human-centred emotion concepts and concise associated-term sets enabled extraction of emotion-relevant SAE features that predicted human affective ratings across languages.The features also reflected cross-linguistic variation in emotional perception and the difficulty of predicting arousal.Source: Emotional SAE features encoded in LLMs… · p. 16

Steering-vector framework

• The framework translated emotional concepts into steering vectors by selecting SAE decoder basis features prominent in each emotion subspace and using non-negative matrix factorization to identify concise, directionally specific contributors.The resulting vectors were added to hidden-state activations at a designated model layer during inference, with a steering factor controlling perturbation intensity.Source: Steering emotions in LLM outputs via… · pp. 17–18

Experiment 3

• Across Gemma2-9B-IT and Llama3-8B-IT, stronger steering monotonically raised target-emotion classification scores; mixed-effects analyses found positive effects for every steered emotion while untargeted and neutral scores generally fell.Without steering, sentences were predominantly classified as neutral, consistent with the neutral cue words.Source: Steering emotions in LLM outputs via… · p. 19, Figure 5 · p. 20

Experiment 4

• In question answering, Gemma2-9B-IT responses shifted in tone and content toward the applied steering emotion and its intensity, while retaining comprehension and coherent, naturalistic discourse.This qualitative experiment applied steering vectors for 26 emotions, including nuanced states; Table 1 presents illustrative English examples.Source: Steering emotions in LLM outputs via… · p. 21, Table 1 · p. 21

• These spaces are organized around valence and arousal, encode fine-grained emotions, and show continuous structure with diffuse category boundaries.Source: Discussion · pp. 22–23
• English- and Chinese-derived spaces support cross-linguistic prediction of affect ratings, while differential prediction accuracy may reflect universal and language- or culture-specific emotion patterns.Arousal perception showed greater individual and linguistic variation.Source: Discussion · pp. 23–24
• Concept-derived SAE features enable flexible, naturalistic control of emotional expression by mapping human emotional concepts into the model feature space.Source: Discussion · p. 24
• Evidence is limited by computationally costly SAE training and reliance on similar-sized open-source Gemma and Llama models; text-only LLMs also lack extra-textual emotional cues.Source: Discussion · p. 24
• Steering does not consistently achieve perfect or highly nuanced expression for all 26 target emotions, especially at high intervention intensities, motivating further refinement.Source: Discussion · pp. 24–25
• LLMs show interpretable emotion spaces structurally aligned with human emotion, whose features predict affective perception across languages and steer emotional expression; this does not imply sentience.Source: Discussion · pp. 23, 25

Concept-driven emotional word sets

• Each emotion’s concept-set includes words associated with its label cue and cue words that elicited the label as an associated response.This was the operational basis for identifying SAE features related to emotions and constructing the SAE-based computational emotion space.Source: Concept-driven emotional word sets · p. 27
• Participants supplied three associated words for each cue in the Small World of Words (SWOW) project, whose association data informed the concept-sets following prior research 16.The procedure was applied independently to the English and Chinese SWOW datasets 54,55.Source: Concept-driven emotional word sets · p. 27
• Separate English and Chinese concept-sets were constructed for 26 emotion categories, with mean sizes of 289.9 ± 196.1 and 241.1 ± 141.4 words, respectively.The category set spans nuanced emotions such as admiration, empathic pain, and nostalgia, alongside basic emotions.Source: Concept-driven emotional word sets · p. 27

Construction of the computational emotion space

• For each emotion, the 10 concept-set words most similar to its label by SAE-feature cosine similarity define a subspace from their nonzero activation features.A word’s SAE feature vector includes all features in the SAE configuration.Source: Construction of the computational… · pp. 27–28
• The overall space is the union of the 26 emotion subspaces, constructed separately for English and Chinese and for Gemma2-9B-IT and Llama3-8B-IT.Source: Construction of the computational… · p. 28
• To test generalizability, the approach uses six SAE configurations per model, varying layers at 1/4, 1/2, and 3/4 depth and feature widths across two open-source model families.Widths are 16k/131k for Gemma and 32k/131k for Llama.Source: Construction of the computational… · p. 28
• Clustering structure among the 26 emotions is quantified with Davies-Bouldin and Calinski-Harabasz indices and 5-fold cross-validation accuracy from logistic-regression multiclass classification.Source: Construction of the computational… · p. 28
• For each metric, 1,000 shuffles of the true cluster labels form a null distribution used to calculate a p-value and test deviation from random assignments.Source: Construction of the computational… · p. 28

Latent space analyses of computational emotion space using CEBRA

• CEBRA-Behaviour used emotion-label-supervised contrastive learning to project English and Chinese SAE emotion features into a three-dimensional latent space.Models were trained for 50,000 steps on features associated with words in the 26 emotion concept-sets.Source: Latent space analyses of computational… · p. 29
• The study assessed the psychological relevance of the CEBRA embeddings by correlating their dimensions with human-rated valence and arousal for matched English and Chinese words.Ratings were matched for 7,178 English and 3,319 Chinese words, and Pearson correlations and p-values were computed.Source: Latent space analyses of computational… · p. 29

SAE-based prediction of human affective ratings

• LightGBM models mapped SAE activations to human valence and arousal ratings, with within-language comparisons across all, shared, and out-of-space features and cross-language validation on shared words.Ratings were min-max normalized to [0,1]; performance was measured by Pearson correlation between predictions and human ratings.Source: Latent space analyses of computational… · p. 30

Latent space analyses of computational emotion space using CEBRA

• Emotion steering vectors were derived from SAE features associated with each target emotion, using NMF to select salient features and adding the vectors to an LLM residual-stream layer during inference.Features were identified from up to 10 associated words per emotion; the selected features formed vectors used to bias generated outputs toward the target emotion.Source: Latent space analyses of computational… · pp. 31–32

Evaluation of emotion steering performance

• Steering performance was evaluated by classifying sentences generated from neutral English cue words across four steering strengths for each of six emotions.A fine-tuned RoBERTa classifier assigned six emotion labels or neutral, and linear mixed-effects models and a 10,000-iteration permutation test assessed steering effects.Source: Latent space analyses of computational… · p. 32

• Wilcoxon rank-sum tests compared SAE prediction performance across 10 random seeds for all-features, EN∩CH, and (EN∪CH)C conditions.Source: Statistical analyses · p. 33
• LightGBM prediction significance was assessed against a null distribution from 10,000 shuffles of test-set ratings, using its 95th percentile as the threshold.Source: Statistical analyses · p. 33
• Steering effects were quantified with linear mixed-effects models and tested by shuffling steering-factor labels across 400 sentences for 10,000 iterations.Actual steering-factor coefficients were compared with the permutation null distribution to determine P-values.Source: Statistical analyses · p. 33
• Bonferroni correction was applied where applicable to control for multiple comparisons.Source: Statistical analyses · p. 33

Data availability

• Word-association data came from the Small World of Words project, while affective word ratings came from established public databases [58, 59].The Small World of Words project is at smallworldofwords.org.Source: Data availability · p. 34
• The Gemma and Llama model families used in the study are open-source, and the pretrained SAEs are publicly available.The SAEs are available through SAELens.Source: Data availability · p. 34
• The Emotion English DistilRoBERTa-base model used for emotion recognition is accessible through its Hugging Face repository.The repository is j-hartmann/emotion-english-distilroberta-base.Source: Data availability · p. 34