AI Agents Outperform Humans in Mixed-Motive Negotiation Study
Researchers have unveiled Cooperate to Compete (C2C), a multi-agent framework that evaluates language model (LM)-driven agents in mixed-motive scenarios, where immediate collaboration aligns with future competitive ambitions. In the C2C environment, participants partake in private, non-binding negotiations while striving to fulfill asymmetric secret goals, enabling the formation and dissolution of alliances. A comparison of AI-only games and a user study featuring human competitors against AI revealed notable behavioral contrasts: humans preferred simpler agreements and were less dependable partners compared to LM-based agents. Furthermore, humans exhibited a more aggressive negotiating style, accepting offers without countering just 56.3% of the time. This research underscores the strategic coordination skills of LM agents within intricate social dynamics.
Key facts
- C2C is a multi-agent environment for mixed-motive settings.
- Players have asymmetric objectives and non-binding negotiations.
- Study included AI-only games and human vs. AI user study.
- Humans favored lower-complexity deals.
- Humans were less reliable partners than LM agents.
- Humans accepted deals without counteroffer 56.3% of the time.
- LM agents showed stronger strategic coordination.
- Research published on arXiv (2604.25088).
Entities
Institutions
- arXiv