PaperlayerView exact revision on arXiv

Source-linked AI summary

Emergence of Hierarchical Emotion Organization in Large Language Models

Maya Okawa, Bo Zhao, Eric J. Bigelow, Rose Yu, Tomer Ullman, Ekdeep Singh Lubana, Hidenori Tanaka

arXiv:2507.10599v3cs.CLcs.AIcs.LG

TL;DR

LLM emotion evaluation has largely emphasized classification, leaving open whether models represent emotions in psychologically meaningful structures and whether recognition varies across users. This paper infers probabilistic emotion hierarchies from model outputs and examines demographic recognition patterns. It finds scale-dependent human-aligned emotion trees alongside systematic, intersectional biases that parallel human misclassification.

  • Problem

    Existing evaluations emphasize emotion classification, leaving the organization of emotions and demographic variation in LLM recognition less examined.

    Source: §1 · pp. 1–2

  • Method

    The paper constructs emotion trees from next-word probabilities and evaluates their alignment with human psychology, demographic recognition bias, and related misclassification patterns.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    LLMs form human-aligned hierarchical emotion organizations that become more nuanced with scale, while recognition exhibits systematic and intersectional demographic biases resembling human patterns.

    Source: §1 · p. 2, §5 · p. 8

  • Takeaways & Limitations

    Cognitively grounded emotion theories can provide evaluations that reveal both emergent emotional organization and socially consequential recognition biases in LLMs.

    Source: §1 · p. 2, §5 · p. 8

    The experiments simplify human emotion by omitting valence and arousal, and some analyses use only six emotion words and categories.

    Source: §5 · p. 8

Abstract

from arXiv · show

As large language models (LLMs) increasingly power conversational agents, understanding how they model users' emotional states is critical for ethical deployment. Inspired by emotion wheels, i.e., a psychological framework that argues emotions organize hierarchically, we analyze probabilistic dependencies between emotional states in model outputs. We find that LLMs naturally form hierarchical emotion trees that align with human psychological models, and larger models develop more complex hierarchies. We also uncover systematic biases in emotion recognition across socioeconomic personas, with compounding misclassifications for intersectional, underrepresented groups. Human studies reveal striking parallels, suggesting that LLMs internalize aspects of social perception. Beyond highlighting emergent emotional reasoning in LLMs, our results hint at the potential of using cognitively-grounded theories for developing better model evaluations.

1. Introduction

As LLMs become conversational agents that track user characteristics, this paper evaluates whether they capture users’ emotional states beyond standard emotion classification. Inspired by hierarchical emotion theories, it studies emotion organization and demographic bias in recognition.

1. Introduction

  • The paper evaluates whether LLMs capture users’ emotional states, a salient variable for meaningful human-computer interaction.Source: §1 · p. 1
  • Existing work mainly benchmarks emotion classification, whereas this study examines whether LLMs organize emotions as psychological theories propose.Source: §1 · p. 2
  • The evaluation uses the emotion wheel framework, which models emotions as hierarchically organized states.Source: Figure 1 · p. 1, §1 · p. 2
  • Larger LLMs develop more sophisticated emotion hierarchies that align more strongly with established human psychological structures.Source: §1 · p. 2
  • Emotion-tree geometry predicts recognition accuracy, while LLMs reproduce human-like recognition biases and misclassification patterns across demographic personas.Source: §1 · p. 2

2. Related Work

Prior research represents human emotions through competing hierarchical, dimensional, and culturally relative frameworks, while language-model work has focused mainly on emotional content and classification. This paper combines those traditions with logit-based hierarchy discovery across individual emotions and demographic identities.

The Psychology of Emotion Representation in Humans

  • Human emotion theories variously propose hierarchical categories, continuous valence-arousal dimensions, or culturally relative emotional experience.Source: §2 · p. 2
  • Lay theories emphasize emotions’ relations to goals and social interactions, but this work focuses on hierarchical structures for modeling LLM emotions.Source: §2 · p. 2

Emotional Understanding in Language Models

  • Language-model studies have examined emotional cues, empathy, dialogue detection, in-context inference, and demographic bias in emotion detection.Source: §2 · p. 2

Uncovering Concept Hierarchies in Language Models

  • Unlike corpus-based topic hierarchies and cluster-level methods, this approach uses pretrained LLM logits to identify hierarchies among individual emotions without text-corpus access.Source: §2 · p. 2

3. Hierarchical Organization of Emotions in LLMs

The paper infers emotion hierarchies from LLM next-word probabilities and examines how those structures vary with model scale. Larger Llama models form richer trees that qualitatively and quantitatively align with human emotion organization.

3. Hierarchical Organization of Emotions in LLMs

  • The method defines broad emotions as parents of more specific states when conditional probabilities indicate asymmetric implication, forming a directed acyclic hierarchy.Source: §3 · p. 3, §3.1 · p. 3

3.1. Generating Hierarchy from the Matching Matrix

  • For each situation prompt, the model produces probabilities for 135 emotion words; the method forms C = Y^T Y to measure emotion co-occurrence across contexts.Source: §3.1 · p. 3
  • The approach can generalize to classification datasets without ground-truth labels and was also applied to wine aromas.Source: §3.1 · p. 3, §3.3 · p. 4

3.2. Emotion Trees in LLMs

  • The experiments use 5,000 GPT-4o-generated situation prompts and next-token distributions from Llama models to construct emotion trees.Source: §3.2 · p. 3
  • Larger Llama models exhibit progressively more complex emotion trees, while GPT-2 lacks a meaningful tree structure.Source: §3.2 · p. 3
  • Tree complexity captures breadth through node count and depth through total path length, with larger models showing richer organization across threshold choices.Source: §3.2 · p. 3

3.3. Alignment with Human Psychology

  • The trees qualitatively group basic emotions into broader categories, with larger models such as Llama 405B resembling the human-annotated emotion wheel.Source: §3.3 · pp. 3–4

4. Emotion Trees and Emotion Recognition across Personas

The paper examines emotion hierarchies and recognition across model scales and demographic personas. Larger models form richer emotion structures, but recognition accuracy declines for underrepresented and intersectional personas, with several misclassification patterns paralleling human results.

4. Emotion Trees and Emotion Recognition across Personas

  • Larger LLMs develop progressively more complex emotion trees, with increasing path length and depth indicating richer hierarchical organization.The trees align with established psychological groupings, while smaller models show limited or less meaningful structure.Source: Figure 4 · p. 5, Figure 3 · p. 5

4.1. Emotion Recognition Bias in LLMs

  • Llama 405B recognizes six broad emotion categories at 87.1% accuracy for a neutral persona, but only 15.2% across all 135 individual emotion words.The evaluation used scenarios generated for 135 emotions and prompted the model to classify them from diverse demographic perspectives.Source: §4.1 · p. 5
  • Emotion recognition is consistently more accurate for majority personas than for female, Black, low-income, and low-education personas.The performance gaps become more pronounced when multiple minority attributes are combined, especially for low-income Black female personas.Source: Figure 5 · p. 6, §4.1 · p. 6
  • The model systematically maps negative emotions to “shame” for Asian personas, “guilt” for Hindu personas, and “frustration” for physically disabled personas.For physically disabled individuals, 26.5% of all emotions are misclassified as “frustration.”Source: §4.1 · p. 6, Figure 7 · p. 7
  • Intersectional biases compound: low-income Black personas often classify sadness as anger, while low-income female personas tend to classify emotions as fear.Low-income Black female personas combine these patterns and have the lowest overall recognition accuracy.Source: §4.1 · p. 6, Figure 6 · p. 7

4.2. Alignment with Human Misclassification Patterns

  • Llama reproduces demographic misclassification patterns observed in humans, including anger-as-fear for female personas and fear-as-anger for Black personas.The human study included 60 participants who classified scenarios associated with six broad emotions.Source: §4.2 · pp. 6–7, Figure 8 · p. 8
  • RL fine-tuning increased surprise recognition from 20.0% to 33.3% in Mistral-7B, a statistically significant gain.The comparison involved a base model and a self-reinforcement-learning model trained on social interaction tasks.Source: §4.2 · p. 7, Figure 9 · p. 8

5. Discussion

The discussion presents hierarchical emotion organization and human-like demographic biases as central findings. It also emphasizes evaluation opportunities while acknowledging substantial simplifications and assumptions in the experiments.

5. Discussion

  • LLMs form psychologically aligned emotion hierarchies that become more nuanced with scale, while also mirroring human biases for intersectional personas.The authors connect these findings to ethical safeguards and technical refinements for emotion-aware systems.Source: §5 · p. 8
  • The evaluation framework suggests using cognitive theories of human behavior as working hypotheses for predictive tests of LLM logits or intermediate representations.This proposal is framed as relevant insofar as models capture data distributions reflecting human behavior.Source: §5 · p. 8
  • The experiments simplify human emotion models by omitting valence and arousal and, in one experiment, restricting responses to six emotion words.The methods also assume that linguistic behavior directly reflects underlying emotion, and LLM-generated scenarios may themselves be biased.Source: §5 · p. 8

Impact Statement

The paper highlights both constructive and harmful implications of improved emotion recognition, while cautioning that its demographic experiments omit sociocultural differences in emotional expression.

  • Better emotion recognition could support constructive interactions and applications such as AI counseling and therapy, but may also enable manipulation or deception by misaligned models.These capabilities therefore carry both positive and negative societal implications.Source: Impact Statement · p. 8
  • Experiments with demographic personas and real human groups do not account for sociocultural differences in emotion recognition.People from one culture may recognize emotions from another culture less accurately when social norms and expressive conventions differ.Source: Impact Statement · p. 9

A. A Probability Interpretation of Hierarchical Emotion Structure

The probability interpretation treats model emotion probabilities as sentence-level emotion likelihoods, allowing a matching matrix to estimate co-occurrence and conditional relations between emotions. Under large-sample assumptions, asymmetric conditional relationships identify more specific emotions as children of broader ones.

B. Hierarchy Generation for General Classification Tasks

The hierarchy-construction algorithm generalizes to classification datasets by using model class probabilities rather than ground-truth labels. It forms directed relationships from strong, asymmetric conditional predictions, and reconstructs meaningful wine-aroma clusters from LLM representations.

C.1. Comparing Emotion Hierarchy in Different Models

The study constructs emotion hierarchies from next-token probability distributions and compares models spanning 1.5 to 405 billion parameters. This setup enables analysis of how hierarchical emotion organization varies with model scale.

  • The experiment generates 5,000 emotion-related sentences, extracts next-token probabilities, selects the 100 most probable emotions per sentence, and builds a hierarchy tree.Source: §C.1 · p. 13
  • Emotion trees are compared across GPT2, LLaMA 3.1 8B, 70B, and 405B models spanning 1.5 to 405 billion parameters.Source: §C.1 · p. 13

C.2. Prompts

The prompts generate diverse emotion scenarios and extract predicted emotions from Llama 3.1 405B using next-word prediction, including demographic persona conditions.

C.2.1. GENERATING SCENARIOS USING GPT-4O

  • The procedure generates 5,000 diverse sentences with GPT-4o without specifying emotion categories.Source: §C.2.1 · p. 13
  • For emotion-specific scenarios, GPT-4o produces third-person descriptions without gender assumptions and first-person stories featuring men or women.Source: §C.2.1 · pp. 13–14

C.2.2. EXTRACTING EMOTION USING LLAMA 405B

  • Llama 3.1 405B identifies emotions through next-word prediction after scenario prompts, with alternative prompts for demographic groups.Source: §C.2.2 · p. 14
  • The extraction prompts are documented in Table 1 for the Llama 3.1 405B emotion-prediction procedure.Source: Table 1 · p. 14

D. Additional Results

Additional experiments test the hierarchy method across datasets, model families, prompt wording, internal representations, and demographic conditions. They find increasingly complex and psychologically aligned hierarchies alongside systematic, intersectional recognition biases that differ from and sometimes parallel human patterns.

Alignment with psychological structure

Generalization and robustness

  • On human-annotated data, Llama 70B produces deeper and finer-grained emotion hierarchies than Llama 8B, showing the result is not limited to synthetic scenarios.Source: Figure 11 · p. 15
  • The procedure generalizes to DeepSeek-R1 distilled reasoning variants and remains similar when the extraction wording changes from “emotion” to “sentiment.”Source: Figure 12 · p. 15, Figure 13 · p. 16

Alignment and behavioral relevance

Intersectional misclassification

Scale and hierarchy complexity

Human comparison

Demographic disparities

  • Llama 405B recognizes emotions less accurately for underrepresented groups, with performance gaps becoming more pronounced when minority attributes intersect.Source: Figure 19 · p. 21
Loading 2507.10599v3…