Communication Enhances LLMs' Stability in Strategic Thinking

Researchers found that allowing AI language models to exchange brief, non-binding messages before playing a repeated cooperation game significantly stabilizes their behavior. In a study using models of 7 to 9 billion parameters playing the Prisoner's Dilemma, this 'cheap talk' reduced unpredictable swings in strategy across most tests. The effect was strongest for models that were initially more volatile. This provides a low-cost method to make multi-agent AI systems more dependable in strategic scenarios.