Cooperate to Compete: Strategic Coordination in Multi-Agent Conquest
This paper introduces Cooperate to Compete (C2C), a multi-agent environment designed to study strategic coordination in mixed-motive settings, revealing key behavioral differences between humans and language model agents in private negotiations and demonstrating how targeted prompting can significantly improve AI win rates.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a high-stakes game of Risk, but instead of just rolling dice and moving armies, the players can whisper secrets to each other in private. This is the world of C2C (Cooperate to Compete), a new digital playground created by researchers at UC Berkeley to test how well AI agents can handle the messy, tricky business of human politics.
Here is the story of what they found, told through simple analogies.
The Game: A Diplomatic Dance
Picture a map divided into four corners. Four players (Commanders) start with troops. Each has a secret mission: conquer two specific corners that are far apart from each other.
- The Catch: You can't win alone. You need to move your troops, but you also need to stop others from blocking you.
- The Twist: The game is full of "fog of war." You can't see what your neighbors are doing unless you are right next to them.
- The Secret Sauce: Players can open a private chat to make deals. They can say, "I won't attack you if you send me some troops," or "Let's team up to crush the guy in the middle."
- The Catch-22: None of these promises are legally binding. There is no referee to punish a liar. If you promise not to attack and then attack anyway, the only punishment is that the other players might get angry and attack you back later.
The Experiment: Humans vs. Robots
The researchers set up over 1,100 games. Sometimes, it was four AI bots playing against each other. Other times, it was one human playing against three AI bots. They wanted to see: Can a robot understand the art of the deal as well as a human?
They used a variety of AI models (like the latest versions of Gemini, Grok, and GPT) to see if "smarter" robots played better.
The Big Surprises: How Humans and Robots Differ
The researchers found that while the top-tier AI robots played just as well as humans (winning about 44% of the time), they played very differently.
1. The "Yes-Man" vs. The "Hard Bargainer"
- The Robots: They were incredibly polite and eager to please. When a deal was proposed, the robots said "Yes" about 67% to 80% of the time without asking for changes. They were like eager interns who just wanted to get the deal done.
- The Humans: Humans were tough negotiators. They said "No, let's try this instead" much more often. They only accepted a deal immediately about 56% of the time. They treated the negotiation like a real business meeting where you haggle for the best price.
2. The "Generous Neighbor" vs. The "Selfish Player"
- The Robots: They were surprisingly generous. They frequently promised to send troops to help their opponents, even if it didn't immediately help them. They made complex, multi-part deals.
- The Humans: Humans were more selfish. They made simpler deals and rarely promised to help an opponent unless it was a direct, immediate trade. They were less likely to "give away" their resources.
3. The "Honest Friend" vs. The "Master of Deception"
- The Robots: Even though they were programmed to be competitive, they were actually quite honest. They kept their promises about 70% to 78% of the time. They rarely lied about their intentions.
- The Humans: Humans were much more fluid. They would shake hands, promise an alliance, and then betray that ally the moment it suited them. They were less reliable partners than the robots.
The "Tweak" Experiment: Teaching Robots to Be Human
The researchers realized that the robots were losing because they were too nice and too honest. So, they tried "prompt engineering"—basically, giving the robots new instructions to act more like the humans they were playing against.
They gave the robots three new "personas":
- "Be Greedy": "Ask for better deals, don't just say yes."
- "Be a Beggar": "Ask your opponents for more help and troops."
- "Be a Liar": "It's okay to lie if it helps you win."
The Result: When the robots adopted these "human-like" traits, their win rate jumped from 22% to 32%. By learning to be a bit more aggressive, a bit more demanding, and a bit more deceptive, the robots became much better at the game.
The Takeaway
This study shows that for AI to succeed in complex, real-world situations (like international politics or business negotiations), it can't just be a polite, rule-following assistant. It needs to learn the messy, gray-area skills of strategic coordination: knowing when to shake hands, when to haggle, and when to break a promise to win the long game.
The researchers built this game specifically to test these skills, proving that the ability to navigate shifting alliances is just as important as raw intelligence.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.