Can Induced Emotion Bias LLM Behaviors in Sequential Decision Making?

As Large Language Models (LLMs) are increasingly deployed as autonomous agents in high-stakes environments like finance and healthcare, it is critical to understand how contextual factors—specifically induced emotions—influence their decision-making. This paper investigates whether emotions, which are known to bias human decision-making, similarly impact the sequential learning dynamics of LLM agents.

The researchers utilized the Iowa Gambling Task (IGT), a classic psychological paradigm where agents must learn to maximize profit by choosing between four decks of cards with varying reward and penalty profiles. To test the effect of emotion, the team employed an imagination-based induction procedure, prompting LLMs to generate vignettes that elicit specific affective states (e.g., anger, happiness, sadness) without using explicit emotion labels. They validated this approach by mapping the LLMs' self-reported affective states onto the valence-arousal (V-A) circumplex model, confirming that the models could maintain distinct, stable emotional states. The agents were then evaluated on their ability to learn optimal strategies in the IGT under these induced conditions.

The study reveals that, unlike humans, LLMs do not show a significant, uniform bias in decision-making across all induced emotions. However, anger acts as a notable exception. When induced with anger, LLM agents exhibit a reduced sensitivity to the negative consequences (penalties) of their choices. Furthermore, in the early stages of the game, anger restricts exploration, causing the agents to lock into a limited set of choices rather than sampling the environment to identify the most advantageous strategy. These findings suggest that while LLMs do not mirror human emotional responses exactly, they are susceptible to specific, conditional affective modulations that can degrade their decision-making performance.