Emergence of Hierarchical Emotion Organization in Large Language Models

Problem Statement
As Large Language Models (LLMs) increasingly power conversational agents capable of multi-modal interactions, understanding how these models represent and process user emotional states is critical for ethical deployment. While prior work has focused on benchmarking standard emotion classification, there is a gap in understanding whether LLMs develop emergent, structured understandings of emotion that mirror human psychological frameworks. Specifically, it remains unclear if LLMs naturally form hierarchical organizations of emotions similar to human "emotion wheels" and whether these models replicate human systematic biases in emotion recognition across diverse demographic personas.

Methodology
The authors propose a novel evaluation pipeline inspired by the hierarchical emotion wheel framework (Shaver et al., 1987) to analyze the probabilistic dependencies between emotional states in model output distributions.

1. Hierarchy Construction Algorithm: The core method involves generating a dataset of situational prompts (5,000 scenarios generated by GPT-4o). For each prompt, the LLM is asked to complete the phrase "The emotion in this sentence is," and the probability distribution over the next token (specifically 135 emotion words from Shaver et al.) is extracted.

   • A matching matrix $C = Y^T Y$ is constructed, where $Y$ contains the probability distributions. This matrix captures the joint probabilities of emotions co-occurring across contexts.
   • Conditional probabilities between emotion pairs are computed to infer parent-child relationships. An emotion $a$ is defined as a child of $b$ if the conditional probability $P(b|a)$ exceeds a threshold $t$ and is significantly higher than $P(a|b)$, indicating that $b$ is a more general category implied by the specific emotion $a$.
   • This process yields a Directed Acyclic Graph (DAG) representing the model's internal hierarchical organization of emotions.

2. Scale Analysis: The method is applied to LLMs of varying sizes (GPT-2, Llama 3.1 8B, 70B, and 405B) to observe how hierarchy complexity scales with model parameters.

3. Bias and Persona Analysis: To evaluate recognition bias, the authors introduce diverse demographic personas (varying by gender, race, socioeconomic status, age, religion, and ability) into the prompts (e.g., "As a [demographic], I think the emotion involved..."). The model's ability to correctly identify emotions in scenarios is measured, and confusion matrices are analyzed to detect systematic misclassifications.

4. Human Comparison: A user study involving 60 human participants is conducted to compare human misclassification patterns and recognition accuracy against the LLM's performance across the same demographic groups.

Key Contributions

• Emergence of Hierarchical Organization: The study demonstrates that LLMs naturally form hierarchical trees of emotional states that align with established human psychological models. This hierarchy is not explicitly programmed but emerges from the model's training.
• Scale-Dependent Complexity: The research finds that as model scale increases, the internal emotion hierarchies become more complex, exhibiting greater depth and branching, which correlates with a stronger alignment with human psychological structures.
• Systematic Bias Mirroring: The paper uncovers that LLMs replicate human systematic biases in emotion recognition. Specifically, models show reduced accuracy for underrepresented groups (e.g., Black, female, low-income, low-education personas).
• Intersectional Bias Amplification: The study highlights that biases compound in intersectional identities (e.g., low-income Black females), leading to the lowest recognition accuracy and specific misclassification patterns (e.g., misclassifying sadness as anger or fear).
• Geometric Predictors of Performance: The authors show that geometric metrics of the constructed emotion trees (such as total path length and average depth) serve as reliable predictors of a model's emotion recognition accuracy for specific personas.

Key Results

• Hierarchical Alignment: Visual and quantitative analysis (using correlation with the Shaver et al. emotion wheel) confirms that larger models (e.g., Llama 405B) produce emotion trees with clustering structures highly similar to human-annotated frameworks. Smaller models (e.g., GPT-2) lack meaningful tree structures.
• Quantitative Complexity: Larger models exhibit significantly higher "total path length" and "average depth" in their emotion trees, indicating richer internal organization.
• Recognition Accuracy Gaps: Llama 405B achieves 87.1% accuracy when grouping emotions into six broad categories for neutral personas, but accuracy drops significantly for underrepresented personas. For instance, the model struggles more with recognizing emotions for Black and female personas compared to White and male personas.
• Misclassification Patterns:
  • Asian Personas: Negative emotions (anger, fear, sadness) are frequently misclassified as "shame."
  • Hindu Personas: Negative emotions are frequently misclassified as "guilt."
  • Physically Disabled Personas: A significant bias exists where 26.5% of all emotions are misclassified as "frustration."
  • Intersectionality: Low-income Black female personas exhibit the combined biases of race, gender, and income, resulting in the lowest overall accuracy.
• Human-LLM Parallels: The user study reveals that LLMs mirror human misclassification patterns (e.g., both Black humans and Black personas modelled by LLMs tend to interpret fear as anger). However, a notable divergence exists in gender bias: while human females outperform human males in recognition, Llama exhibits the opposite trend, favoring male personas.
• Reinforcement Learning Impact: Fine-tuning models on social interaction tasks (negotiation/persuasion) via reinforcement learning significantly improves the recognition of "surprise" (from 20.0% to 33.3%), supporting the hypothesis that prediction-error-based training enhances sensitivity to this specific emotion.

Significance and Claims
The paper claims that its results highlight the emergence of emotional reasoning in LLMs that goes beyond simple classification, suggesting that these models internalize aspects of social perception and human cognitive structures. The authors posit that the hierarchical organization of emotions is an emergent property that scales with model size, potentially leading to more emotionally intelligent and contextually aware agents.

Furthermore, the work underscores the ethical imperative of evaluating LLMs not just for accuracy, but for the replication of human biases. The authors argue that their evaluation pipeline, grounded in cognitively-based theories (like the emotion wheel), offers a robust method for developing better model evaluations. They suggest that understanding these emergent hierarchies and biases is crucial for the safe deployment of LLMs in sensitive domains like counseling and therapy, while also warning that improved emotional understanding could be misused for manipulation if models are misaligned. The study concludes that cognitive theories of human behavior can serve as working hypotheses for developing predictive tests for LLM components, such as output logits and intermediate representations.