Do LLMs Act on What They Know? From Partner Representations to Cooperative Actions
The study tested how LLMs adapt their behavior when cooperating with unfamiliar partners by using scripted hints and frozen model weights. It found that LLMs are better at decoding intent conventions than target conventions, but receiving decisions don't consistently match the sender's convention. Translating partner information into actions yields larger gains than using rule statements, showing that action recommendations are more effective for cooperation.