更大利益攸关：报酬与语言如何影响大型语言模型在合作困境中的策略

As Large Language Models (LLMs) increasingly function as autonomous agents in interactive and multi-agent settings, comprehending their strategic behavior is paramount for ensuring the safety, coordination, and efficacy of AI-driven social and economic systems. This investigation systematically examines how the magnitude of payoffs and the surrounding linguistic context influence LLM strategies within repeated social dilemmas. Utilizing a payoff-scaled Prisoner's Dilemma, the study meticulously isolates and quantifies the LLMs' sensitivity to varying incentive strengths. Consistent behavioral patterns are observed across a diverse range of LLM models and different natural languages. The findings indicate that LLMs' propensity for cooperation is not static but significantly modulated by the payoff structure, with higher stakes generally correlating with a stronger inclination towards cooperative actions. Furthermore, subtle linguistic nuances, including the precise phrasing of instructions and contextual descriptions, exert measurable effects on LLM decision-making processes. This suggests that LLMs do not merely execute predefined rules but engage in sophisticated comprehension and contextual reasoning based on linguistic prompts. These insights underscore the critical need to meticulously consider both the economic incentive frameworks and communication modalities when designing and deploying LLM-based agents. Such considerations are essential for ensuring predictable behavior and fostering cooperation and efficiency within multi-agent systems. A deeper understanding of these factors will facilitate the development of more robust and manageable LLM-driven systems, particularly in scenarios involving high-stakes decisions or intricate social interactions.