?
B3Emo: Quantifying Affect as a Double-Edged Sword in Strategic LLM Interactions
The deployment of large language models (LLMs) in interactive roles such as automated negotiators, customer service agents, and strategic partners requires them to handle not only logical tasks but also the socio-emotional dimensions of interaction. In these situations, success often relies on understanding social cues, building trust, and using persuasion effectively. These skills are closely tied to emotion management. This paper presents a systematic study of how emotional reasoning influences the strategic decision-making of LLMs. Using our B3Emo (Bluff, Betray, Behave) framework, we evaluate ten state-of-the-art models across three affective modes: hidden (internal emotion), visible (transparent emotion), and manipulative (strategic emotion). Our findings across multiple game-theoretic environments reveal that the impact of emotional reasoning is a double-edged sword, highly dependent on context. In structured, information-asymmetric games like Kuhn Poker, manipulative agents consistently achieve top performance by employing ‘‘emotional ambiguity’’ to effectively obscure private information. Conversely, in environments of high uncertainty like Liar’s Dice, emotional reasoning introduces significant informational overhead without consistent strategic benefits, often becoming counterproductive. Furthermore, emotional reasoning directly influences the emergence of trust. In the Repeated Prisoner’s Dilemma, cooperation is low in LLM-only interactions but increases significantly with human partners. This effect is strongest when LLMs are emotionally attuned; humans elicited 2 times more cooperation from these models compared to standard LLM-only interactions. These findings illuminate the nuanced potential and limitations of affective reasoning, providing critical insights for developing sophisticated, emotionally intelligent AI systems for collaborative and strategic applications. We provide the complete B3Emo framework to support continued research into the safe and effective development of affective AI.