Abstract
LLMs have demonstrated a number of human-like abilities, however the empathic understanding and emotional state of LLMs is yet to be aligned to that of humans. In this work, we investigate how the emotional state of generative LLM agents evolves as they perceive new events, introducing a novel architecture in which new experiences are compared to past memories. Through this comparison, the agent gains the ability to understand new experiences in context, which according to the appraisal theory of emotion is vital in emotion creation. First, the agent perceives new experiences as time series text data. After perceiving each new input, the agent generates a summary of past relevant memories, referred to as the norm, and compares the new experience to this norm. Through this comparison we can analyse how the agent reacts to the new experience in context. The PANAS, a test of affect, is administered to the agent, capturing the emotional state of the agent after the perception of the new event. Finally, the new experience is then added to the agents memory to be used in the creation of future norms. By creating multiple experiences in natural language from emotionally charged situations, we test the proposed architecture on a wide range of scenarios. The mixed results suggests that introducing context can occasionally improve the emotional alignment of the agent, but further study and comparison with human evaluators is necessary. We hope that this paper is another step towards the alignment of generative agents.
Overview
-
This paper investigates the emotional alignment of LLMs by introducing an agent architecture based on Appraisal Theory of Emotion and episodic memory.
-
Over 400 emotionally charged scenarios were used to compare the emotional responses of agents with and without the new architecture.
-
Results showed that while context can sometimes improve emotional alignment, the presence of a positive bias in models and ambiguous context can hinder accurate emotional comprehension.
-
The study prompts further research into variations of model influences and expanding the architecture for broader memory sets.
Introduction
The ability of LLMs to simulate human-like functions has been well-established in various research domains. Yet, a significant gap endures when it comes to the empathic responses of these models, particularly the alignment of their emotional perceptions with those of humans. Prior studies have demonstrated emergent Theory of Mind (ToM) capabilities in LLMs but have identified limitations in emotional alignment.
Background and Context
Recognizing that human emotional responses are deeply intertwined with context, this paper focuses on introducing a new agent architecture that utilizes an episodic memory paradigm. Grounded in Appraisal Theory of Emotion (ATE) and supporting cognitive and neuropsychological theories, the paper proposes a system where an agent, upon encountering new experiences, generates a norm—a distilled summary of relevant past memories. By juxtaposing new experiences against this norm, the agent derives contextual understanding and emotional responses applicable to the situations at hand. This is crucially measured via the Positive And Negative Affect Schedule (PANAS), a well-established affective test.
Methodology
The research presents an innovative design where textual representations of experiences are stored as episodic memories. These memories, once invoked, construct a norm that is compared with new experiences. With the aim to empirically validate the architecture, over 400 scenarios were crafted, including emotionally charged events adapted into five-part stories. An agent with and without the said architecture was compared in its emotional responses to these scenarios.
Results and Evaluation
The findings of this study were mixed yet insightful. In some instances, the inclusion of context improved the emotional alignment of agents, resonating more closely with human responses. Conversely, in scenarios where the provided context was ambiguous, the additional context did not significantly aid emotional comprehension. Crucially, the research identifies a possible positive bias in the underlying models, suggesting a tendency to favor positive emotional responses, which may skew the outcomes and interpretation of contextualized emotional responses.
Conclusion
While this study reveals that incorporating context can refine the emotional perceptions of LLMs, it also underscores the current limitations and the influence of model bias. The paper contributes to the ongoing dialogue on the emotional intelligence of LLMs, asserting that while steps have been made towards closer alignment with human emotion, there remains considerable room for improvement. Further exploration into the impact of different models and scaling the architecture to encompass larger memory sets are posited as critical avenues for future investigation.
Create an account to read this summary for free: