How do language agents behave when cooperation is possible, but not guaranteed? This project explored that question through the Concordia Contest at NeurIPS 2024, where LLM-based agents were evaluated in mixed-motive social dilemma environments requiring skills such as negotiation, reciprocity, promise-keeping, reputation management, compromise, partner choice, and sanctioning. My work centred on the design and evaluation of agents that could cooperate not only in familiar scenarios, but also with unseen co-players in held-out environments. Through the Concordia framework, this research examined how language agents respond under cooperation pressure, and contributed to the co-authored paper “Evaluating generalization capabilities of LLM-based agents in mixed-motive scenarios using concordia.”
Links:
- Co-authored paper: Evaluating generalization capabilities of LLM-based agents in mixed-motive scenarios using concordia
- The contest: Concordia Contest 2024