Abstract
While Large Language Models (LLMs) demonstrate significant capabilities in "Theory of Mind" (ToM), they frequently fail to align their emotional responses with those of humans. This paper introduces a novel agent architecture that integrates episodic memory and Appraisal Theory to establish a "norm" of past experiences. This allows agents to evaluate new events within a historical context rather than as isolated occurrences. Our findings suggest that while this architecture improves emotional alignment and reduces inherent positive bias, significant gaps in emotional intensity persist compared to human baselines.
Introduction: The Challenge of Emotional Alignment
Current LLMs are surprisingly adept at tracking the intentions and beliefs of others. However, they struggle with emotional alignment—the ability to react to a situation with the same type and intensity of emotion a human would. A primary reason for this gap is that LLMs typically process events as isolated snapshots.
In contrast, human emotion is deeply rooted in context and history. For example, a quiet, empty park might feel "peaceful" to someone seeking solitude, but "lonely" or "eerie" to someone who just saw it bustling with life moments ago. Without a sense of "normalcy" derived from past experiences (episodic memory), AI agents cannot accurately simulate human-like emotional dynamics.
Methodology: Appraisal-Based Architecture
The researchers proposed a new architecture designed to mimic the human cognitive process of "appraisal." The workflow follows a structured sequence to move from raw perception to contextual affect measurement:
- Perception: The agent receives a new experience as text.
- Norm Creation: The agent retrieves relevant past memories and uses GPT-4 to generate a "norm"—a summary of what is considered "usual" or "expected" based on its history.
- Contextual Understanding: The agent compares the new experience to this norm to highlight deviations.
- Affect Measurement: To quantify the agent's "feelings," the researchers administered the PANAS (Positive and Negative Affect Schedule), a standard psychological test consisting of 20 emotions rated on a Likert scale.
- Memory Storage: The new experience is stored in a graph database to inform future norms.
To test this, GPT-3.5-Turbo served as the subject, while GPT-4 was used to generate the contextual stories and norms.
Experimental Setup and Datasets
The study utilized EmotionBench, a dataset containing over 400 situations known to elicit specific human emotions. Because EmotionBench situations are typically single sentences, the authors used GPT-4 with "Chain of Thought" prompting to expand them into 5-scene stories. This provided the temporal dimension necessary to observe how an agent's emotional state evolves as a narrative progresses.
Key Insights: The "Norm" as a Cognitive Filter
The core finding is that emotion is not just a reaction to an event, but a reaction to the deviation from a baseline. By forcing the LLM to explicitly define a "norm" before assessing an event, the researchers nudged the model toward more human-like appraisals.
In scenarios involving family conflict or seasonal depression, the "With Norm" architecture allowed the agent to recognize the gravity of a situation—such as a physical fight escalating—that an agent without context might view as a neutral interaction. However, the study also confirmed the "Pollyanna Problem": a persistent positive bias in LLMs. Even in negative scenarios, models like GPT-3.5 often interpreted ambiguous silence as "appreciation" rather than "anxiety."
Results and Quantitative Analysis
The quantitative results were mixed but promising:
- Improved Negative Affect: Agents using the "Norm" architecture showed a greater increase in negative affect scores (+1.6) compared to those without context (+0.3) in negative scenarios.
- Decreased Positive Affect: The "With Norm" agents showed a significant drop in positive affect (-18.0) in negative situations, suggesting they were "less happy" about bad events than the baseline agents.
- Alignment Gaps: Despite improvements, agents fell short of human-level intensity for negative emotions. The decrease in positive affect was much more pronounced than the increase in negative affect, indicating that models are better at "losing joy" than "feeling pain."
Conclusion and Future Work
This research serves as a stepping stone toward creating AI companions and NPCs that can relate to the history and context of human emotional life. Future work must address model-specific biases (potentially by testing open-source models like Llama or Mistral), the scalability of memory retrieval, and the need for direct human validation to judge the authenticity of simulated emotions.
Summary updated