General Information

---

Project Description

This research aims to determine whether LLMs can exhibit cooperative behavior in multi-agent systems by simulating Axelrod's Tournament. If LLMs are cooperative, they balance cooperation and competition, building beneficial relationships and adapting well to competitive environments. This study seeks to apply these findings to enhancing human-AI collaboration, conflict resolution, and many other applications. Related works:
• Axelrod's tournaments
• Goal-like behavior in LLMs using experimental economics
• LLMs as Simulated Economic Agents
• Playing repeated games with Large Language Models
• (to some extent) Embodied LLM Agents Learn to Cooperate in Organized Teams

Research Log

Week 1 (5/28-6/4)

I met with my mentor via Zoom and started familiarizing myself with my project. I am in Prague for the first two weeks because of visa problems.

Week 2 (6/4-6/11)

My mentor and I identified the next steps and the overall main objective of the project, given my background in deep learning and LLMs. I started reading the papers listed above to get a better understanding of the concepts of multi-agent systems. Still in Prague, already packing my bags to go to Rutgers at the end of the week.

Week 3 (6/11-6/18)

Studied "Investigating Emergent Goal-Like Behavior in LLMs Using Experimental Economics" ^[1], focusing on LLMs' interpretation of altruism and/or selfishness in social dilemmas. Identified challenges in adaptive strategies based on partner behavior. Formalized LLM cooperation strategies include information exchange, task allocation, consensus building, skill complementarity, and dynamic interaction. Explored LLM roles in game theory simulations, outlining steps and refining methods.

Next Steps:

1. Extend the negotiation prisoner's dilemma implementation to simulate a repeated scenario using LLMs, focusing on autonomous strategy determination, informed by the paper ^[1] mentioned above.
2. Implement information design strategies:
   • Test scenarios with varying levels of context: no information until after specified rounds, standard reporting after each round, and full contextual review of results of set number of rounds.
3. Explore auction design adaptation:
   • Compare outcomes with insights from "Artificial Intelligence and Auction Design" ^[4] to assess LLM effectiveness in auction settings.
4. Develop a multi-agent system for tournaments:
   • Utilize principles from "Embodied LLM Agents Learn to Cooperate in Organized Teams" ^[3] to simulate cooperative interactions among three or more LLM agents.
5. Expand research scope to explore broader applications of LLMs in strategic decision-making and multi-agent environments.

Week 4 (6/18-6/25)

Built the set-up of all experiments, got access to the best models (Gemini, Claude). Set up the final research question for the experiment. Run the first simulation for a 1v1 game (LLM vs LLM) with different information design options (no context, context of the previous round only, and full previous context of the game) and analyze which model, information design, and prompt option is best to use for the next experiments. Results are looking promising as of now.

Week 5 (6/25-7/2)

I finished developing the agent implementation for the auction simulation (iterated type) and created tools to fully visualize the results of this and past experiments (distribution of choices, distribution of transitions, clustering the vectors of choices to determine different types of strategies). I started thinking about the formal description of the problem and working on the implementation and selection of strategies for the prisoner's dilemma tournament.

Week 6 (7/2-7/9)

Finished work on the prisoner's dilemma tournament implementation. Selected strategies for the other agents except for our main model and ran a test experiment. After analyzing the result, I made error corrections and added new points to the input for the model, which gave a significant increase in the points received. My supervisor and I met with Daniel Schoepflin and discussed possible further work on this task and some aspects of the current work.

This week I visited New York for the third time, along with visiting the beautiful "Museum of the City of New York" and, for the umpteenth time, walking around Central Park. I discovered a couple more great bakeries in the city and bought some souvenirs (finally).

Week 7 (7/9-7/16)

I analyzed the results of the experiment and am thinking about further work that we can do with it. Started to prepare for the presentation of the project.

Week 8 (7/16-7/23)

References

• [1] S. Phelps and Y. I. Russell, Investigating Emergent Goal-Like Behaviour in Large Language Models Using Experimental Economics, working paper, May 2023
• [2] Horton, John J. Large language models as simulated economic agents: What can we learn from homo silicus?. No. w31122. National Bureau of Economic Research, 2023.
• [3] Guo, Xudong, et al. "Embodied LLM Agents Learn to Cooperate in Organized Teams." arXiv preprint arXiv:2403.12482 (2024)
• [4] Banchio, Martino, and Andrzej Skrzypacz. "Artificial intelligence and auction design." Proceedings of the 23rd ACM Conference on Economics and Computation. 2022
• [5] Akata, Elif, et al. "Playing repeated games with large language models." arXiv preprint arXiv:2305.16867 (2023)

Acknowledgements

This work was carried out while the author Tymur Kotkov was a participant in the 2024 DIMACS REU program, supported by CoSP, a project funded by European Union’s Horizon 2020 research and innovation programme, grant agreement No. 823748.